<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Izayohi on Medium]]></title>
        <description><![CDATA[Stories by Izayohi on Medium]]></description>
        <link>https://medium.com/@izayohi?source=rss-74e79e92759b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*3Lb9yIJFrTkxlOwzwHd7kQ.gif</url>
            <title>Stories by Izayohi on Medium</title>
            <link>https://medium.com/@izayohi?source=rss-74e79e92759b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 04:18:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@izayohi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Revised Language Process Theory: A Recording Framework for Production, Reception, and Norms]]></title>
            <link>https://medium.com/@izayohi/revised-language-process-theory-a-recording-framework-for-production-reception-and-norms-a8cb9addbff2?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/a8cb9addbff2</guid>
            <category><![CDATA[linguistics]]></category>
            <category><![CDATA[communication]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 06:16:52 GMT</pubDate>
            <atom:updated>2026-10-06T06:23:59.190Z</atom:updated>
            <content:encoded><![CDATA[<h3>An introduction to a theory that includes AI-generated text, language-like animal vocalisations, and human inner speech</h3><p><em>Motohisa Ishibe</em><br><strong>October 6, 2026 — revised English introduction</strong></p><p><strong>Starting proposal.</strong> Revised Language Process Theory was proposed as an account that treats AI-generated texts, language-like animal vocalisations, and human inner speech as language, alongside human interpersonal expression.</p><h3>Abstract</h3><p>Revised Language Process Theory begins with an inclusive proposal about the scope of language and develops a framework for recording the processes that this proposal brings into view. Its central task is to distinguish four claims: understanding, compliance, adoption of an evaluative criterion, and change in a collective norm. Production and reception are recorded as processes; expressions as material signals when externalisation occurs; and learned correspondences, evaluations, requirements, and implementation arrangements as states or conditions with specified bearers. A recording protocol distinguishes observations, inferences, and the stipulations of constructed examples. The article applies it to workplace terminology and a reported human–AI experiment, and maps it to language management, dialogue alignment, and convention research. These applications demonstrate analytical consequences, rather than empirical superiority. The framework preserves questions from Japanese language-process theory without assuming identical mechanisms or conscious experiences across humans, animals, and artificial systems.</p><p><strong>Keywords:</strong> language process theory; production; reception; inner speech; linguistic norms; language management; human–AI interaction</p><h3>1. The proposal begins by widening the subject of inquiry</h3><p>A theory of language needs to say what it admits as language before explaining how it works.</p><p>The founding proposal of Revised Language Process Theory, or <em>shūsei gengo kateisetsu</em> (修正言語過程説), includes AI-generated text, language-like animal vocalisations, and human inner speech. This scope is an affirmative theoretical position. These cases belong to the theory’s subject matter, rather than appearing only as peripheral analogies to ordinary conversation.</p><p>The proposal also creates a research obligation: the word “language” cannot do the work of showing that the included cases have identical organisation. An animal vocalisation, a privately rehearsed sentence, and a model’s written response may differ in their production, learning histories, interpretation, social regulation, and relation to experience. These differences are objects of comparison.</p><p>In particular, “language-like animal vocalisations” does not mean every sound made by an animal. Investigators must specify the signal system, the evidence for its expressive or communicative organisation, and the properties being compared. Its inclusion in this theory does not by itself classify it as a human natural language. Likewise, a signed natural language requires analysis of its own linguistic structure; it is not merely an instance of generic gesture.</p><p>Inner speech raises a different issue. A sentence can be silently rehearsed without an outward utterance or inscription. Externalisation and an external recipient therefore cannot be universal requirements of the theory. When a participant later reports having rehearsed a sentence, that report is a new outward event. It provides evidence about the earlier episode under the limits of the reporting method.</p><p>The inclusive scope and the recording framework have different roles. The former states what the theory proposes to study as language. The latter organises what can be claimed about each case. Being recordable within the framework is not, on its own, a test that determines whether an unfamiliar signal belongs to language.</p><p>The aim is to make comparison possible while keeping the differences available for explanation.</p><h3>2. An expression can be understood and still be corrected</h3><p>Consider a constructed workplace example.</p><p>A colleague enters “parked” in a document-status record. A listener identifies the intended document and paraphrases its status as “saved, but not submitted.” A supervisor then says, “In our records, please use ‘draft.’”</p><p>The colleague changes the entry immediately. One week later, in the same kind of record, the colleague again writes “draft,” without the supervisor present.</p><p>For this example, stipulate that “parked” and “draft” refer to the same status. They are not being presented as ordinary English synonyms. Also stipulate that an authorised written terminology rule already requires “draft” and that its wording remains unchanged throughout the example.</p><p>What can we conclude?</p><p>The listener’s paraphrase supplies the example’s specified confirmation of task-relevant understanding. The immediate substitution establishes compliance within the constructed sequence. The later entry establishes a later word choice.</p><p>The colleague’s endorsement of the rule remains unresolved. So does the cause of the later choice: personal approval, anticipation of evaluation, repeated exposure, and habit are possible explanations. The existing rule has been implemented; its text has not been revised.</p><p>This is already the kind of situation studied by <strong>Language Management Theory</strong>. That theory distinguishes noticing a feature, evaluating it, designing an adjustment, and implementing it, while connecting management in particular interactions with organised management across situations (Nekvapil, 2009). The revision does not claim to discover this sequence.</p><p>Its question is how to record the sequence without treating understanding, compliance, criterion adoption, and collective norm change as interchangeable findings.</p><h3>3. What is inherited from Japanese language-process theory?</h3><p>The historical name identifies a point of departure: language should be studied through situated acts of expression and understanding.</p><p>Tokieda’s 1952 discussion of Japanese language education explicitly distinguishes expression and understanding, including speaking, writing, listening, and reading. Reception is already part of his account. Its inclusion cannot be presented as an invention of the revised theory (Tokieda, 1952).</p><p>Satoru Suzuki’s supportive interpretation of Miura supplies another problem: how socially learned linguistic norms mediate cognition and expression (Suzuki, 1973). The use made of that account here is limited to Suzuki’s interpretation. This article does not claim to reconstruct Miura’s complete writings directly.</p><p>Maeda examines how Tokieda’s subject is conceived as the individual carrying the expressive process from grasping material through conceptualisation to vocalisation, and questions the grounding of that activity (Maeda, 1979, pp. 63–64). For the present framework, one consequence is to distinguish an object’s existence from the ways participants classify it. This consequence is the framework’s application, rather than a quotation from Maeda.</p><p>A document may exist while its classification as a draft, record, or submission remains contested. Production must therefore be investigated as selection and organisation in a situation, without assuming that the relevant objects, categories, and participant capacities are fully specified in advance.</p><p>“Revised” names this selective reconstruction. It does not transfer every historical claim into a general account of human, animal, and artificial language.</p><h3>4. Processes, signals, and states are different kinds of record</h3><p>The familiar symbols U, A, I, Nᵢ, and Nₛ are retained from the Japanese formulation. They are labels, not the English initials of their names. They also do not denote five organs or an obligatory sequence of five stages.</p><p>CategoryKind of itemWhat is recorded<strong>U: production</strong>ProcessFormation, selection, organisation, monitoring, or revision of an expression<strong>A: externalised signal or trace</strong>Material event or objectAn outward acoustic, written, signed, or other expressive signal, if externalisation occurs<strong>I: reception</strong>ProcessA recipient’s encounter, processing, interpretation, and response, if reception occurs<strong>K: learned correspondences and expectations</strong>State or distribution, indexed by bearerForm–content correspondences, partner-specific expectations, expectations about a group, or measured group usage<strong>Nᵢ: individual evaluative criteria</strong>Participant-level stateCriteria used to judge a usage appropriate in a specified setting<strong>Nₛ: social normative arrangements</strong>Group or institutional recordRequirements, approvals, rules, and evidence of relevant expectations, with provenance and scope<strong>M: maintenance arrangements</strong>Implementing process or arrangementTeaching, editing, enforcement, or interface functions that maintain or implement a requirement<strong>C: non-normative constraints</strong>ConditionChannel noise, distance, capacities, bandwidth, or computational limits<strong>D: topic, task, and setting</strong>Contextual recordWhat the expression concerns, what participants are doing, and under which roles and circumstances<strong>S: participant or system state</strong>StateOther specified cognitive, bodily, or computational conditions relevant to a local model</p><p>A transient spoken utterance counts as A even if no recording survives. An inscription may remain available long after its production. An inner-speech episode need not have an A entry: the outward-signal field can be marked <strong>not applicable</strong>, rather than filled with a presumed internal sound.</p><p>U and I identify processes whose mechanisms remain research questions. Alternative formulation, task-sensitive selection, and revision are possible objects of investigation. Their inclusion does not require conscious planning in every case, and does not make production a featureless input–output box.</p><p>Evidence status belongs to the <strong>analyst’s record</strong>, not to the participant’s state. “Observed,” “inferred,” and “unmeasured” describe our access to a claim. They are not additional organs or causal forces.</p><p>These distinctions allow a process to be related to the state it changes without turning the process, state, and evidence into a single list of equivalent entities.</p><h3>5. Norms: more than frequency, less than an explanation of everything</h3><p>A participant can know what an expression means, predict that others will use it, judge it appropriate, or expect others to disapprove of it. These are related but different claims.</p><p>K must therefore carry a bearer and a level of analysis. A useful record distinguishes:</p><ul><li>a participant’s learned correspondence or expectation, including expectations specific to a partner;</li><li>a participant’s expectation about usage in a reference group;</li><li>the usage distribution actually measured in that group.</li></ul><p>The second is a belief about a group. The third is an aggregate finding. Calling both “community knowledge” would conceal their different evidence.</p><p>Lewis’s account of convention concerns regularities that resolve recurring coordination problems; convention is not simply any frequent pattern (Lewis, 1969). K is a broader recording field that can hold a specified convention model, a prediction, or a learned correspondence. It is not a replacement definition of Lewis’s concept.</p><h3>A person’s judgment and a prediction about others</h3><p>Nᵢ records a person’s own criterion of appropriateness. A separate field records what that person expects others to judge appropriate. A further field records what the person expects others actually to do.</p><p>Bicchieri and Xiao distinguish expectations about others’ behaviour from expectations about others’ normative judgments (2009). Research on convention-to-norm transitions also separately measures personal beliefs, expectations, and approval or disapproval (Przepiorka et al., 2022). Those distinctions help specify evidence here without imposing one universal definition of norm on every domain.</p><p>Someone might approve of “draft,” expect colleagues to use “parked,” and anticipate that a supervisor will reject “parked.” A single label such as “the participant’s norm” would lose the conflict.</p><h3>A demand is not yet an established social norm</h3><p>An Nₛ record should specify the following.</p><p>QuestionWhat the answer can establishWho issued what request, for which usage?The content and source of a demandWhat authority or rule supports it?Its institutional standing and valid scope, if documentedWhich reference group expects or evaluates the usage?Evidence about collective expectations and normative evaluationWhat happens in practice?Implementation, compliance, contestation, or enforcement</p><p>A supervisor’s personal preference is recordable as a demand. It is not automatically an authorised rule or a norm shared by staff. Conversely, an unwritten norm can be supported by situated judgments, reactions, expectations, and patterns of accountability. A written rule and an explicit “should” are not necessary in every case.</p><p>In the document example, the authorised rule is stipulated. Staff-wide expectations are not. A real investigation would need different sources for those two claims.</p><p>This narrower use of N differs from historical uses of “linguistic norm” that encompass learned form–content correspondences. Much of that wider material belongs in K here. The change is a working distinction, not a declaration that earlier authors used the word incorrectly.</p><h3>Content, implementation, and constraints</h3><p>The requirement “use ‘draft’ in status records” is normative content. A dropdown implementing that requirement is M. Network latency and screen width are C.</p><p>The same software can perform several functions. Record each function with its grounds; do not count one effect twice as two independent causes. If a policy-driven interface change also changes channel capacity, record that causal relation explicitly.</p><p>Similarly, a model’s learned parameters belong to its specified system state. They are not automatically classified as a computational limit or a linguistic norm merely because they affect outputs.</p><h3>6. A minimal recording protocol</h3><p>A recording framework becomes usable when another analyst can tell where an event begins, what the evidence is, and why a conclusion was entered.</p><p>The unit proposed here is an <strong>episode record with linked process events</strong>. One episode may include several producers, recipients, and later tests. A person can occupy more than one role: an author may also monitor their own wording.</p><p>The following protocol is a proposal, not a validated measurement scale.</p><ol><li><strong>Fix the scope.</strong> Assign an episode ID, participants and roles, topic, task, medium, setting, system boundary, and time windows. Specify the change of interest before interpreting it.</li><li><strong>Identify the source.</strong> Link an utterance, inscription, behavioural record, judgment, policy text, or computational log. Mark whether it is measured material, a secondary report, or a constructed example.</li><li><strong>Record the observation first.</strong> Enter what was said, entered, selected, paraphrased, or logged, with its event ID and time.</li><li><strong>Name the target claim.</strong> Specify whether the source concerns production, understanding, compliance, an evaluative criterion, a rule, or a change in one of them.</li><li><strong>State the grounds and alternatives.</strong> Explain what supports the claim, which bridge from observation to inference is needed, and what competing explanations remain.</li><li><strong>Record limits and relations.</strong> Enter unresolved questions, missing measurements, overlapping uses of a source, and any demonstrated feedback route. Do not infer missing stages from the existence of later ones.</li></ol><p>Record fieldMinimum contentsEpisode and eventID, time, participants, roles, links to related eventsScopeTopic, task, medium, setting, boundary, observation and retention windowsSourceMaterial and location; measured, reported, or constructedObservationThe directly recorded action, signal, response, or document featureTarget claimThe component or relation being investigatedGroundsEvidence, method, and any inferential assumptionsState changeSpecified component, baseline, later measure, and comparisonEvidence statusObserved, inferred, stipulated, unmeasured, non-detected, or supported as equivalentOpen questionsAlternative explanations and reasons for withholding judgment</p><p>The same response can support several claims, but each claim receives its own grounds. The analyst must not add these entries as if they were independent observations.</p><p>A report such as “I consider ‘draft’ appropriate” is first an observed expression. Its use as evidence of Nᵢ requires a specified judgment method and attention to context, strategic reporting, and measurement effects. Asking the question may itself influence later usage.</p><h3>7. Applying the protocol to the workplace example</h3><p>Assign the constructed episode the ID W1. The setting is an internal document-status record; the immediate window is the correction sequence, and the delayed window is one week. The later use occurs in the same kind of record, within the original requirement’s scope.</p><p>Every entry below is <strong>specified within a constructed example</strong>, not an observation from fieldwork.</p><p>EventWhat the example suppliesWhat remains unresolvedW1.1: “parked” is enteredAn outward expression and the occurrence of productionAll internal operations involved in producing itW1.2: the listener identifies the document and paraphrases “saved, not submitted”The stipulated confirmation of task-relevant understandingApproval of the termW1.3: the supervisor requests “draft” under an existing authorised ruleA demand implementing a specified institutional requirementStaff-wide normative expectationsW1.4: the colleague changes the entryImmediate compliance in this recordPersonal endorsement of the criterionW1.5: one week later, “draft” is entered without the supervisorLater use in the specified medium and scopeIts cause; changes in the supervisor; mutual durable transformationW1.R: the rule text is unchangedA stipulation about one institutional componentConstancy of every practice, expectation, or enforcement arrangement</p><p>These entries yield a definite analytical result. The interpretation “the supervisor corrected a misunderstanding” conflicts with the example’s understanding condition. The interpretation “the collective rule changed” conflicts with its unchanged-rule stipulation. The interpretation “the colleague adopted the criterion” lacks sufficient grounds.</p><p>This result is stronger than a list of possible categories: the distinctions alter which conclusions follow from the specified sequence. It remains a logical application, not empirical validation of the procedure.</p><p>A boundary case makes the separation clearer. Suppose the institution changes its authorised terminology rule before a colleague has learned the revision. The rule’s content can change while that colleague’s observed wording remains stable. Institutional and individual changes need not occur together, and neither is assumed always to precede the other.</p><p>The recording method retains both patterns. It does not require every production episode to end in norm change.</p><h3>8. State transitions without an obligatory circle</h3><p>A local model can place the records into a state-transition template:</p><p><strong>X(t + 1) = F[X(t), events during the interval, history, D(t), C(t)]</strong></p><p>Here, X contains specified participant states and any relevant group or institutional variables. Production U and reception I occur during the interval; A records outward signals when they occur. M records implementing arrangements or actions. F is an update rule to be supplied by a particular explanatory model.</p><p>This is a typed template, not a completed mathematical explanation. A simulation requires defined variables, update functions, an observation model, and identifiable parameters. A behavioural learning model and a Bayesian dialogue model could instantiate different versions.</p><p>The template permits internal expression without outward A, an outward expression without reception, and reception without agreement or normative change. Inclusion as language does not require establishing a normative requirement in every episode. “Not applicable” and “not measured” must remain distinct, with grounds for either entry.</p><p>It also separates evidence about change:</p><p>StatusMeaning<strong>Stipulated</strong>A constructed example fixes a component as changed or unchanged<strong>Unmeasured</strong>No relevant measurement was taken<strong>Change not detected</strong>A measurement did not detect change under its sensitivity<strong>Practical equivalence supported</strong>Data support a predefined range of negligible difference<strong>Specified non-update documented</strong>Logs or other records establish no update to the named component</p><p>A statistically non-significant difference is not exact zero. Unchanged rule text does not establish unchanged social practice. A log showing unchanged model parameters does not establish an unchanged entire artificial system.</p><h3>9. Re-entry: route, persistence, and reciprocity</h3><p>Re-entry concerns the way a classification, interpretation, prediction, or evaluation changes later conditions, and whether consequences return into further activity.</p><p>P0–P3 summarise such relations within a specified boundary and time window. They are not a scale of intelligence or a universal developmental sequence.</p><p>LabelRelation to be established<strong>P0</strong>No operative re-entry route for the specified content within the stated boundary and window, with grounds for that absence<strong>P1</strong>Content from recognition or evaluation changes subsequent conditions or action; a returning route is not established<strong>P2</strong>Consequences return as inputs to subsequent recognition, production, or evaluation; the route may involve one or both participants<strong>P3</strong>Mutual interaction history contributes to durable changes in both relatively independent participants’ strategies, models, parameters, or learning conditions</p><p>Insufficient records are <strong>unresolved</strong>, not P0. P1 identifies an established forward relation; it does not certify that every possible unobserved return is absent.</p><p>A supervisor’s evaluation affecting an entry can support P1 if the causal link is established. If the changed entry then informs the supervisor’s next evaluation, a returning route supports P2. The rules governing these operations may remain fixed.</p><p>P3 adds durable transformation on both sides. Stored history, alternating messages, and temporary use of context are insufficient by themselves. A retention window, changed components, and dependence on mutual history must be specified. Self-history, common inputs, shared incentives, and time trends are competing explanations.</p><p>The labels do not exhaust the description. Each record should separately state:</p><ul><li>the demonstrated route and its direction;</li><li>which participants changed, and which components;</li><li>the duration and retention test;</li><li>the evidence connecting the change to interaction history.</li></ul><p>One participant can change durably while the other does not. Where a returning route is established, record <strong>P2 with one-sided persistence</strong>. Do not rename it P3 merely because the change lasts. Delayed use of “draft” alone establishes neither that returning route nor its cause.</p><p>Predictive improvement when history is included can identify a candidate relation. Causal attribution requires further design and assumptions. Apparent mutual change is not enough to demonstrate mutual formation.</p><h3>10. Production is not a report of experience</h3><p>The framework’s broad scope requires different evidence for different internal claims.</p><p>Fedorenko, Piantadosi, and Gibson’s Perspective synthesises behavioural and neuroscientific evidence against treating natural language as necessary for all complex thought (2024). For this framework, the consequence is to specify the processing required by a task. Dissociation in adults does not determine the developmental contribution language may have made.</p><p>Nor does a fluent expression independently disclose experience. Producing “ouch” does not specify every aspect of pain, imagery, or inner speech. A listener’s assistance could depend on a word, a cry, a visible injury, or several cues. Word understanding, emotion sharing, inference, and action require different observations.</p><p>For inner speech, the theory includes the internal expressive episode without demanding an outward A or external I. Production and self-monitoring can be investigated through a stated method; outward reports and task performance provide evidence under that method. An analyst must not infer private wording, awareness, or absence of inner speech merely from silence.</p><p>Human–AI comparison needs the same discipline.</p><p>Shared comparison targetHuman evidenceArtificial-system evidenceChoice under an external terminology ruleRecorded wording or task choiceGenerated wording and logged task choiceEvaluation of a usageContextualised judgment responsesEvaluative output, reward-model score, or validator decision, recorded as distinct sourcesRetention beyond an interactionLater behaviour and specified judgment testsLater outputs plus evidence about context, stored memory, procedures, or parametersInternal criterion or subjective experienceRequires an appropriate further methodCannot be inferred from a fluent evaluative sentence alone</p><p>The first row supports a common behavioural endpoint. The remaining rows do not establish measurement equivalence.</p><p>For an artificial system, distinguish conversational context, persistent memory, retrieval or planning procedures, and learned parameters. Fixed parameters do not imply that every system component is unchanged. Conversely, context-sensitive output does not establish parameter learning.</p><p>Ouyang et al. describe a training pathway involving human demonstrations, preference comparisons, and model training (2022). The evaluations belong to specified raters and research procedures. They are neither comprehensive measures of humanity’s values nor evidence that a resulting model consciously endorses a norm.</p><p>Including generated text as language therefore permits investigation of its organisation and consequences without making a human experience attribution a condition of inclusion.</p><h3>11. An empirical application: changed vocabulary is not yet changed normativity</h3><p>Yakura and colleagues’ 2026 preprint, version 4, combines corpus analysis with a preregistered human–AI experiment. Participants encountered experimentally controlled synonym variants in chatbot interaction. Later spoken descriptions, after a three-minute distractor and without the chatbot present, showed continued uptake. The authors limit the experimental retention evidence to that session (Yakura et al., 2026).</p><p>Reported evidenceFramework recordStronger claim not established by that evidenceControlled chatbot wording and human responsesA, U/I events, and lexical-choice effectsHuman endorsement of a usage requirementLater lexical uptakeBehavioural persistence over the tested window; evidence relevant to KChange in Nᵢ or a collective NₛHuman change after AI interactionA candidate interaction-history effectDurable transformation of both systems required for P3</p><p>This application changes the interpretation of “adoption”: lexical uptake is supported; adoption of a normative criterion is a different claim. It does not suggest that the study confuses them. It shows why a recording scheme must preserve the distinction.</p><h3>12. Where existing theories do the explanatory work</h3><p>The framework should connect to existing models, rather than present familiar mechanisms as new discoveries.</p><p>Existing approachCorrespondence in this frameworkWhat must be specified separately<strong>Language Management Theory</strong>Noting, evaluation, adjustment design, and implementation relate to U/I, evaluations, outward acts, and M; simple and organised management connect local and institutional recordsWhich stages are evidenced, the basis of a requirement, and the level of the recorded change<strong>Conversation analysis of correction</strong>Sequential correction practices are recorded as linked U/A/I eventsThe grounds for understanding, endorsement, institutional authority, or later change beyond that sequence<strong>Interactive alignment</strong>Alignment of representations and dialogue routines concerns U/I and relevant learned correspondences in KA specific claim about an appropriateness demand or criterion adoption<strong>CHAI: partner-to-population convention modelling</strong>Partner-specific lexical beliefs and population-level expectations are indexed K states used in production and interpretationObserved group usage, normative evaluation, institutional rules, and their implementation<strong>Common ground and joint activity</strong>Participant assumptions and coordinated acts relate to indexed K and linked U/I eventsWhich proposal is understood, accepted, complied with, or personally endorsed</p><p>For the workplace example, LMT is a close theoretical neighbour. Its stages need not all occur, and Nekvapil discusses limits on access to the intervening processes. It is therefore not a model that automatically assumes completed correction or transparent internal evaluation.</p><p>Jefferson’s work on exposed and embedded correction concerns the organisation of correction within conversation (1987). Those categories are not identical to this article’s distinction between clarification and an appropriateness demand. Sequential form and the grounds of an evaluative requirement are different questions.</p><p>Pickering and Garrod model coordination and alignment in dialogue (2004); Clark studies language as joint activity and common ground (1996). Understanding an expression need not mean accepting every proposal made through it. These approaches should supply explanations where their mechanisms fit the question.</p><p>CHAI provides a particularly useful correspondence. A participant’s inference about a partner’s lexical use belongs in partner-indexed K. Its population-level prior belongs in that participant’s expectations about a group, rather than automatically becoming an observed group distribution or Nₛ (Hawkins et al., 2023). A stable official rule can coexist with changing partner expectations. Adding an Nₛ record asks a further question about evaluative requirements; it does not demonstrate a defect in CHAI.</p><p>Complex-adaptive-system research already treats language as shaped through interacting processes (Beckner et al., 2009). Feedback itself is not the revision’s discovery.</p><p>The proposed contribution is a shared recording protocol that preserves evidence for these different claims across selected applications. Its reproducibility, effort, and benefits remain to be tested against existing procedures.</p><h3>13. What would count as success or failure?</h3><p>Testing the recording method and testing a causal hypothesis are different tasks.</p><h3>The recording method</h3><p>Give analysts the same material and compare this protocol with an existing procedure, including an appropriate LMT-based analysis. Predefine the errors of interest: treating clarification as a usage demand, compliance as personal endorsement, or individual word choice as collective norm change.</p><p>Compare justified classifications, unresolved cases, agreement, and recording burden. High agreement is insufficient if everyone makes the same unsupported inference. If the added fields do not improve the targeted distinctions and increase effort, simplify or abandon them for that application.</p><p>The workplace demonstration and empirical application above are grounds for testing the protocol. They are not results showing that it wins this comparison.</p><h3>A local hypothesis about demands</h3><p>A first causal hypothesis is that a demand to change an already understood, task-successful expression affects later wording beyond matched exposure to the alternative expression and comparable task feedback.</p><p>Compare a demand condition with an information condition presenting the same alternative. Confirm the relevant understanding and success in both. Measure immediate change, later selection in the demand’s valid scope, and appropriateness judgments separately.</p><p>At this first stage, the estimand is the <strong>total effect of the compound demand intervention</strong>. Authority, attention, directness, and anticipated future evaluation may contribute. A positive effect would not isolate a uniquely normative mechanism or establish personal endorsement.</p><p>Research on normative expressions and their presentation supports treating wording and interpretation as relevant design variables (Kuang &amp; Bicchieri, 2024). To isolate components of a demand, additional interventions must be specified; they are not supplied by calling the condition “normative.”</p><p>Failure to find the proposed additional effect in an adequately informative test counts against this local hypothesis. It should not be repaired by assuming an unmeasured norm. Nor does that one failure automatically refute every distinction in the framework.</p><h3>An open question about mutual durable change</h3><p>A separate question concerns P3: does mutual history explain durable change on both sides beyond self-history, common inputs, shared incentives, and time trends?</p><p>Matched replay, partner changes, or restrictions on access to history may help, but can also alter information and task opportunities. The present article does not provide a completed design for isolating mutuality. It specifies what a P3 claim would need to establish.</p><p>No new experiment, effect size, or validated scale is reported here.</p><h3>14. What the theory offers now</h3><p>Revised Language Process Theory retains its founding inclusive proposal: AI-generated text, language-like animal vocalisations, and human inner speech belong to its inquiry as language.</p><p>Its current recording framework makes that proposal more demanding. It asks which process occurred, what signal became outwardly available, how it was received, which learned expectations and evaluative requirements were relevant, and what changed over a stated interval.</p><p>The applications establish analytical consequences. Understanding can coexist with correction; compliance leaves criterion adoption unresolved; institutional and individual changes can occur at different times; lexical uptake does not itself establish normative transformation.</p><p>The next step is to test whether preserving these distinctions improves actual research. The theory’s value will depend on that work, and on explanations fitted to the different systems it brings into comparison.</p><h3>References</h3><p>References are selective, not an exhaustive review. Japanese titles are retained; bracketed English renderings are descriptive translations. Tokieda’s cited text is used directly, whereas the discussion of Miura relies on Suzuki’s interpretation. Jefferson’s chapter was checked through publication information and an available opening excerpt; it is not presented here as a complete reanalysis.</p><p>Beckner, C., et al. (2009). Language is a complex adaptive system: Position paper. <em>Language Learning, 59</em>(Suppl. 1), 1–26. <a href="https://doi.org/10.1111/j.1467-9922.2009.00533.x">https://doi.org/10.1111/j.1467-9922.2009.00533.x</a></p><p>Bicchieri, C., &amp; Xiao, E. (2009). Do the right thing: But only if others do so. <em>Journal of Behavioral Decision Making, 22</em>(2), 191–208. <a href="https://doi.org/10.1002/bdm.621">https://doi.org/10.1002/bdm.621</a></p><p>Clark, H. H. (1996). <em>Using language</em>. Cambridge University Press. <a href="https://www.cambridge.org/core/books/using-language/common-ground/7F9689E12F87250DCA81540FCBDE3720">Publisher’s chapter on common ground</a></p><p>Fedorenko, E., Piantadosi, S. T., &amp; Gibson, E. A. F. (2024). Language is primarily a tool for communication rather than thought. <em>Nature, 630</em>, 575–586. <a href="https://doi.org/10.1038/s41586-024-07522-w">https://doi.org/10.1038/s41586-024-07522-w</a></p><p>Hawkins, R. D., Franke, M., Frank, M. C., Goldberg, A. E., Smith, K., Griffiths, T. L., &amp; Goodman, N. D. (2023). From partners to populations: A hierarchical Bayesian account of coordination and convention. <em>Psychological Review, 130</em>(4), 977–1016. <a href="https://doi.org/10.1037/rev0000348">https://doi.org/10.1037/rev0000348</a></p><p>Jefferson, G. (1987). On exposed and embedded correction in conversation. In G. Button &amp; J. R. E. Lee (Eds.), <em>Talk and social organisation</em> (pp. 86–100). Multilingual Matters. <a href="https://doi.org/10.21832/9781800418226-006">https://doi.org/10.21832/9781800418226-006</a></p><p>Kuang, J., &amp; Bicchieri, C. (2024). Language matters: How normative expressions shape norm perception and affect norm compliance. <em>Philosophical Transactions of the Royal Society B, 379</em>, 20230037. <a href="https://doi.org/10.1098/rstb.2023.0037">https://doi.org/10.1098/rstb.2023.0037</a></p><p>Lewis, D. (1969). <em>Convention: A philosophical study</em>. Harvard University Press. <a href="https://doi.org/10.1002/9780470693711">Publisher information for the 2002 reprint</a></p><p>Maeda, H. [前田英樹]. (1979). 言語における行為と差異―再び SAUSSURE と時枝をめぐって [Action and difference in language: Revisiting Saussure and Tokieda]. <em>フランス語学研究, 13</em>(1), 56–66. <a href="https://doi.org/10.20579/belf.13.1_56">https://doi.org/10.20579/belf.13.1_56</a></p><p>Nekvapil, J. (2009). The integrative potential of Language Management Theory. In J. Nekvapil &amp; T. Sherman (Eds.), <em>Language management in contact situations: Perspectives from three continents</em> (pp. 1–11). Peter Lang. <a href="https://beckassets.blob.core.windows.net/product/preamble/8810822/9783631582633_intro_005.pdf">Publisher-hosted excerpt</a></p><p>Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. <em>Advances in Neural Information Processing Systems, 35</em>, 27730–27744. <a href="https://proceedings.neurips.cc/paper_files/paper/2022/hash/b1efde53be364a73914f58805a001731-Abstract-Conference.html">Published article</a></p><p>Pickering, M. J., &amp; Garrod, S. (2004). Toward a mechanistic psychology of dialogue. <em>Behavioral and Brain Sciences, 27</em>(2), 169–190. <a href="https://doi.org/10.1017/S0140525X04000056">https://doi.org/10.1017/S0140525X04000056</a></p><p>Przepiorka, W., Szekely, A., Andrighetto, G., Diekmann, A., &amp; Tummolini, L. (2022). How norms emerge from conventions (and change). <em>Socius, 8</em>. <a href="https://doi.org/10.1177/23780231221124556">https://doi.org/10.1177/23780231221124556</a></p><p>Suzuki, S. [鈴木覚]. (1973). Hypostase de la forme について [On “hypostase de la forme”]. <em>フランス語学研究, 7</em>(1), 28–41. <a href="https://doi.org/10.20579/belf.7.1_28">https://doi.org/10.20579/belf.7.1_28</a></p><p>Tokieda, M. [時枝誠記]. (1952). 国語学と国語教育との交渉―言語過程説の立場における [The relationship between Japanese linguistics and Japanese language education from the standpoint of language process theory]. <em>国語科教育, 1</em>, 5–13. <a href="https://doi.org/10.20555/kokugoka.1.0_5">https://doi.org/10.20555/kokugoka.1.0_5</a>; <a href="https://www.aozora.gr.jp/cards/001968/files/58881_66114.html">Japanese full text</a></p><p>Yakura, H., et al. (2026). <em>Empirical evidence of Large Language Model’s influence on human spoken communication</em> (Version 4, July 16). arXiv preprint. <a href="https://arxiv.org/abs/2409.01754v4">https://arxiv.org/abs/2409.01754v4</a></p><h3>Author’s related work and methodological note</h3><p>The author publishes as <strong>mototchen</strong>; the Japanese name is <strong>石部統久</strong>.</p><p>Related Japanese publication: <a href="https://note.com/mototchen/n/nc26e1078bd7b">https://note.com/mototchen/n/nc26e1078bd7b</a>.</p><p>This is an independently written English introduction, rather than a translation of the Japanese manuscript. The founding scope is stated according to the author’s clarification. Constructed examples are labelled separately from the application to reported research. The source selection addresses the framework’s historical questions and closest methodological neighbours; it is not a systematic literature review.</p><p>Generative AI assisted source retrieval, comparison, drafting, and editing. This article presents a provisional theoretical proposal and recording protocol. It is not represented as a peer-reviewed publication or as a report of new experimental findings.</p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: ChatGPT６.1Sol Ultra</li><li>Peer review: Grok, Gemini3.8, GPT６.1Sol, ClaudeOpus5.5</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8cb9addbff2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Parallel Holon Architecture — Part 1: A Plain-Language Map of the Whole Series]]></title>
            <link>https://medium.com/@izayohi/parallel-holon-architecture-part-1-a-plain-language-map-of-the-whole-series-f17772f8251f?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/f17772f8251f</guid>
            <category><![CDATA[ai-architecture]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[ai-discourse]]></category>
            <category><![CDATA[ai-alignment]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Tue, 26 May 2026 01:50:59 GMT</pubDate>
            <atom:updated>2026-06-02T03:03:48.331Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This is the author’s own English rendering of an essay first written in Japanese and published on note. I place it here deliberately, ahead of the platform’s auto-translation rollout, so that the English version circulating in search results is one I control rather than a machine paraphrase. The Japanese original is here: [insert note URL]. I write as an independent, extra-academic researcher in Okayama, Japan. Read what follows as a working hypothesis offered for scrutiny, not a verdict.</em></p><p><em>Motohisa Ishibe</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xTA130xV31-xlOA8kYKDYg.png" /></figure><h3>The bet we are all living inside</h3><p>For the last few years, almost every conversation about artificial intelligence has assumed one shape. There is a model. It is large. The plan for making it better is to make it larger — more parameters, more data, more compute — and to ask this single growing object to do everything: write code, give medical-adjacent advice, reason about physics, hold a conversation, and, increasingly, reach out through tools and robotics to act on the physical world.</p><p>I will call this the <strong>single-giant-model hypothesis</strong>, and the strategy that follows from it <strong>aggregationism</strong>: the belief that capability is best concentrated in one ever-larger system rather than distributed across many specialized ones. Aggregationism is not stupid. It has produced genuinely useful systems, and it has commercial logic on its side — one model is easier to sell, brand, and improve than a federation of parts.</p><p>This series is an argument that the bet is reaching a structural limit, and that the limit is not the kind you fix with a larger training run. That broad claim is a hypothesis — the thing the series sets out to test, not something I treat as already settled. The part I will defend hardest, as you will see, is narrower and far harder to dismiss.</p><h3>The cracks, stated carefully</h3><p>I want to be honest about what counts as evidence here, because it is easy to dress up ordinary product friction as a crisis. So I will keep this paragraph short and hedged.</p><p>Over the past year, several things became visible at once. Frontier-model releases have drawn mixed reception even from technically sophisticated users, with reliability complaints on complex engineering tasks. At least one major lab found it necessary to restrict the distribution of a particularly capable model to a defined set of partners and infrastructure maintainers rather than ship it openly — which is, in effect, an admission that <em>one</em> model under <em>one</em> weight regime cannot safely serve every stakeholder at once. And labs have published their own engineering postmortems describing how compounding failures cascaded through systems that were supposed to be unified and reliable.</p><p>None of these is, by itself, proof of anything. Read separately they are normal growing pains. Read together, they are at least <em>consistent with</em> a different diagnosis: that piling more capability into one probabilistic object multiplies the surface where things can quietly go wrong. But I want to be careful, because these signals are not all the same kind of thing — some are engineering facts, some are how <em>users</em> received a model, and some may trace to post-training choices rather than to architecture at all. Part 2 takes them apart precisely so they are not lumped into one indictment. Here I only want to put the question on the table.</p><h3>The core idea, in one breath</h3><p>The alternative this series develops is what I call a <strong>parallel holon architecture</strong>.</p><p>Instead of one model that does everything — including, eventually, controlling machines that move in the real world — you split the system into <em>layers</em> that have different jobs and different rules. At the bottom sits a <strong>deterministic, verifiable layer</strong> that holds authority over anything physical and irreversible. Above it sit <strong>probabilistic reasoning layers</strong> — the part that looks like today’s language models, good at flexible inference but never trusted with the final say over a moving arm or a safety interlock.</p><p>The slogan version: <em>let the part that touches the world be predictable, and let the part that thinks be flexible — and never let them be the same part.</em></p><h3>Why “holon”</h3><p>The word <strong>holon</strong> comes from Arthur Koestler. A holon is something that is, at the same time, a <em>whole</em> in itself and a <em>part</em> of something larger. Your heart is a complete organ with its own internal logic, and also a component of you. A team is a whole, and a unit of a company.</p><p>That double nature is exactly the design principle here. Each layer of the architecture has its own <strong>local rationality</strong> — its own internal rules that make sense on their own terms — <em>and</em> it changes its behavior depending on how it connects to the layers above and below. To understand the system you have to keep moving between “what this part does on its own” and “what the whole is doing.” A single giant blob erases that movement; a holon architecture is built to preserve it.</p><h3>The three layers, briefly</h3><p>In a companion essay (“Beyond Being”) I introduced a small vocabulary for talking about these layers without smuggling in assumptions about machine minds. For a general reader, three terms are enough to follow the series:</p><ul><li>The <strong>Aktuanz layer</strong> — the lower, physical layer of sensing and action: where the system meets sensors, timing, and consequences that cannot be undone.</li><li>The <strong>Panich layer</strong> — the middle layer, where multiple modules run in parallel.</li><li>The <strong>Kontextleib layer</strong> — literally a “context-body”: the upper layer, where context is structured and higher-level meaning emerges.</li></ul><p>You do not need to memorize these. They are German-derived on purpose, to avoid borrowing loaded English words like “consciousness” or “self.” Treat them, for now, as labels for three different kinds of work a system has to do, kept deliberately separate rather than fused into one.</p><h3>The one non-negotiable</h3><p>If you take only one thing from the whole series, take this: <strong>irreversible physical action must belong to a deterministic layer, not a probabilistic one.</strong></p><p>A language model is a probabilistic engine. It produces plausible outputs, and “plausible” is exactly the wrong standard for the moment a robot decides whether to apply force, release a brake, or stop. That moment needs a layer that is <em>predictable</em>, <em>real-time</em>, and <em>verifiable</em> — the kind of software built for decades in the real-time operating systems tradition, including the ITRON / TOPPERS lineage and Ken Sakamura’s TRON project. The job of such a layer is not to be clever. It is to draw a hard, deterministic boundary around what the clever-but-uncertain layers are <em>allowed to do</em> in the physical world.</p><p>This is the hinge the series turns on. To be fair to the other side, serious robotics already separates a high-level planner from a low-level real-time controller — so “use a deterministic layer for control” is not by itself news. The claim here is sharper: in most systems today that deterministic boundary is a <em>harness bolted on around</em> a model after the fact, whereas the argument is that it has to be a <em>central design principle</em> — built into how the system is trained, evaluated, and updated, with authority over irreversible action sitting in the deterministic layer by construction. Put that way, the single-giant-model picture stops looking like the natural endpoint of AI and starts looking like a category error: asking a probability engine to do a job that requires a guarantee.</p><p>Two honest caveats. The hard part is the seam itself — translating rich, uncertain context into a small set of deterministic commands without losing what mattered; that problem is not waved away, it is the subject of Part 4. And I am not claiming to know how to <em>build</em> a real-time OS — the claim is about where authority should sit, a system-design question that crosses disciplines, not about the implementation details of any particular system. Note, too, the shape of the whole argument: the broad indictment of aggregationism is the hypothesis the series <em>tests</em>, but this narrow principle is the part I will defend hardest. You can be entirely uninterested in robots and still owe it an answer, because it applies the moment any AI is handed a tool that acts irreversibly.</p><h3>A map of Parts 2 through 7</h3><p>Here is what each installment does and why it matters. Read them in any order; this is a map, not a staircase.</p><p><strong>Part 2 — Why a single giant model cannot optimize across all of its jobs at once.</strong> The core claim is that one model under one weight regime cannot simultaneously satisfy every objective and every stakeholder. The breakdown shows up in several signals — but Part 2 deliberately keeps them apart, because they are <em>different kinds of evidence</em>: published engineering postmortems (facts about systems), the way successive releases shift their own behavior (which may owe as much to post-training choices as to architecture), and the way readers fix a model’s outputs into a single imagined character (a fact about <em>reception</em> — the “Claude Buddha” pattern). The discipline is not to collapse these registers into one accusation. <strong>Why it matters:</strong> it turns a vague sense that “something is off” into a structured diagnosis that distinguishes which problems are architectural from which are not.</p><p><strong>Part 3 — The deterministic lower layer, and a buried history.</strong> Here the argument descends to the layer that touches the world. It explains why the safety-critical layer has to be a deterministic real-time OS rather than a model, and tells the largely forgotten story of how a domestic alternative operating system (BTRON) became entangled in the US–Japan trade-friction environment around Special 301 — a case study in how a viable parallel paradigm can be sidelined for reasons that are not purely engineering ones. <strong>Why it matters:</strong> the foundation already exists; it was sidelined, not disproven.</p><p><strong>Part 4 — Embodiment, from a $26 desktop robot to factory robots: what should AI be connected to?</strong> This is the most concrete installment. It shows that the deterministic lower layer can be implemented across an enormous range of scale — from a roughly $26 desktop camera-robot up to industrial factory robots — and that the design each scale demands sits in a different phase from the single-model aggregation line. Along the way it makes good on two threads Part 3 deferred (a <strong>control-authority protocol</strong> that keeps irreversible action under the deterministic layer by design, and the <strong>bandwidth mismatch</strong> between probabilistic reasoning and real-time control), and develops two ideas that do real work: <em>material irreversibility</em> (some actions cannot be taken back, and the architecture must respect that) and a designed, real-time sense of “now” that today’s token-by-token models do not natively have. <strong>Why it matters:</strong> this is what “layered” actually means once you leave the realm of text — and it scales down far enough that an individual could build at the small end.</p><p><strong>Part 5 — A well-managed organization does not produce the blue LED.</strong> A technical design lives or dies inside an organization with incentives. This installment turns to development-organization theory: how research distorts under a single measurement grid (the Nakamura Shūji / blue-LED case as the worked example — the breakthrough came from <em>outside</em> what the organization was set up to measure and reward), why a <strong>right to refuse measurement</strong> matters, and how a healthy distributed system lets its parts differentiate (<em>distributed individuation</em>) rather than forcing them into one mold. <strong>Why it matters:</strong> you cannot build a layered, distributed architecture inside an organization built to centralize and measure everything.</p><p><strong>Part 6 — Data sovereignty, multilingual parallelism, and distributed modules: the three axes of an international strategy.</strong> The argument climbs to the level of geopolitics. It first shows that the US, China, and the Gulf states are not three poles but <em>three variations of the same organizing principle</em> — aggregationism — each unifying the measurement grid from above by a different means (commercial KPIs, state standardization, sovereign-wealth-fund concentration), and each therefore structurally failing to produce what it does not measure. Against that, it stacks three axes as a dependence hierarchy — multilingual parallelism (the cognitive base) → distributed modules (the technical middle) → data sovereignty (the institutional top) — and argues that a genuine “third pole” requires all three, not a midpoint between the blocs. It reads four candidates against that test: the EU, India, Taiwan, and Japan. <strong>Why it matters:</strong> architecture choices are sovereignty choices, and “keeping the data at home” is empty if a single cognitive grid still processes it.</p><p><strong>Part 7 — A Citizen-Track Minimum Substrate for the Aktuanz Layer (standalone).</strong> Published as an independent article rather than a numbered installment, this is the most rigorously bounded piece in the series. Reading Olivia Zhu’s minimal-embodiment work, it proposes the <em>smallest</em> concrete substrate at which independent builders and citizens, not just large labs, could begin to instantiate the Aktuanz layer for themselves. <strong>Why it matters:</strong> it turns the whole argument from critique into something you could actually start building, small.</p><h3>Why an outsider is writing this</h3><p>I am not a credentialed specialist in any of the fields this series crosses — AI, real-time systems, robotics, the philosophy of language, or industrial organization. I work <em>extra-academically</em>, outside the disciplines, which is usually treated as a disqualification.</p><p>I’d ask you to read it as the opposite. Plenty of careful cross-disciplinary work happens inside institutions. But the single-giant-model hypothesis survives partly because the <em>particular</em> seams this series targets fall between everyone’s remit: the AI researcher rarely studies real-time OS design, the robotics engineer rarely reads reception theory, the organization theorist rarely touches trade history. The argument of this series lives precisely in those <em>seams between</em> fields — the places no single specialty is responsible for. Whether I have stitched them correctly is for you to judge. That is what the rest of the series is for.</p><h3>Where to go next</h3><p>If the core claim — <em>let the part that touches the world be predictable, and the part that thinks be flexible, and never fuse them</em> — is worth taking seriously, start with <strong>Part 2</strong> for the diagnosis, or jump to <strong>Part 4</strong> if you want the concrete robotics version first.</p><ul><li>Part 2: <a href="https://medium.com/@izayohi/the-parallel-holon-architecture-part-2-why-the-single-giant-model-cannot-optimize-across-all-4e2df927d999">https://medium.com/@izayohi/the-parallel-holon-architecture-part-2-why-the-single-giant-model-cannot-optimize-across-all-4e2df927d999</a></li><li>Part 3: <a href="https://medium.com/@izayohi/the-parallel-holon-architecture-part-3-what-ken-sakamuras-tron-saw-three-decades-ahead-itron-bc57edd733c0">https://medium.com/@izayohi/the-parallel-holon-architecture-part-3-what-ken-sakamuras-tron-saw-three-decades-ahead-itron-bc57edd733c0</a></li><li>Part 4: <a href="https://medium.com/@izayohi/the-parallel-holon-architecture-part-4-the-concrete-of-embodiment-from-a-26-desktop-robot-to-cb7c06147f88">https://medium.com/@izayohi/the-parallel-holon-architecture-part-4-the-concrete-of-embodiment-from-a-26-desktop-robot-to-cb7c06147f88</a></li><li>Part 5: <a href="https://medium.com/@izayohi/the-parallel-holon-architecture-part-5-well-managed-organizations-do-not-produce-blue-leds-38b85dc9e7ec">https://medium.com/@izayohi/the-parallel-holon-architecture-part-5-well-managed-organizations-do-not-produce-blue-leds-38b85dc9e7ec</a></li><li>Part 6: <a href="https://medium.com/@izayohi/agi-that-is-not-a-single-giant-model-the-parallel-holonic-architecture-part-6-b836a9a52f08">https://medium.com/@izayohi/agi-that-is-not-a-single-giant-model-the-parallel-holonic-architecture-part-6-b836a9a52f08</a></li><li>Part 7 (standalone): <a href="https://medium.com/@izayohi/a-citizen-track-minimum-substrate-for-the-aktuanz-layer-reading-olivia-zhus-minimal-embodiment-98915f6a3145">https://medium.com/@izayohi/a-citizen-track-minimum-substrate-for-the-aktuanz-layer-reading-olivia-zhus-minimal-embodiment-98915f6a3145</a></li><li>Part 8:<a href="https://note.com/mototchen/n/n0f86456737aa">https://note.com/mototchen/n/n0f86456737aa</a></li></ul><p>Disagreement is the point. Tell me where the seams don’t hold.</p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f17772f8251f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[There Are Three Kinds of “We Can’t Tell”]]></title>
            <link>https://medium.com/@izayohi/there-are-three-kinds-of-we-cant-tell-39c9ee5df149?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/39c9ee5df149</guid>
            <category><![CDATA[interdisciplinary]]></category>
            <category><![CDATA[philosophy-of-science]]></category>
            <category><![CDATA[epistemology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Mon, 25 May 2026 03:35:31 GMT</pubDate>
            <atom:updated>2026-05-25T03:35:31.454Z</atom:updated>
            <content:encoded><![CDATA[<h3>The total is solid — how you cut it is a choice</h3><p><em>Motohisa Ishibe</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZP9WMBhfxBGDxneNH51FJA.png" /></figure><p>In fields with nothing in common, the exact same fight keeps breaking out.</p><p>Particle physicists argue over what fraction of a proton’s mass is “the quarks.” Economists argue over what fraction of the wage gap is “discrimination.” Literary scholars argue over what a novel’s theme “really is,” and AI researchers argue over whether a model “actually has” reasoning ability.</p><p>These domains share no object, no method, no history. And yet the <em>skeleton</em> of the argument is uncannily alike. This essay argues that the resemblance is neither coincidence nor metaphor. And along the way it argues something else: that when we say “we can’t tell,” we are quietly using one phrase for three different things.</p><h3>The cause is not the object — it’s what your hands are doing</h3><p>Let me first clear away two tempting wrong answers: “they argue because nature is mysterious,” and “they argue because things get fuzzy at small scales.” Both locate the cause in the <em>nature of the object</em>.</p><p>A counterexample comes fast. The “discrimination share” of a wage gap is neither mysterious nor microscopic — it’s ordinary social statistics. Yet it gets fought over with the same structure. Conversely, the electric charge of the electron is an extremely robust number, and no one argues about it — because no one is dividing it up.</p><p>So the source of these recurring disputes is not on the side of the object. It is on the side of <em>what people are doing to the object</em>. The operation goes like this. First, there is a robust <strong>total</strong> that reproduces (the sum doesn’t budge). Second, you split that total into parts and give the parts names. Third — there is no single correct way to split it.</p><p>The result: the total is solid, but “how much each part is” and “what to call it” flip depending on how you cut.</p><p>Think of it at the kitchen table. Today’s total spending is a hundred dollars. The receipts fix it; it doesn’t move. But whether a given item is “food” or “entertainment” depends on how you keep the books. The total is a fact; the breakdown is a convention. This simple structure reappears, unchanged in shape, in the mass of the proton, in the wage gap, and in the evaluation of an AI.</p><h3>Splitting “we can’t tell” into three</h3><p>Here is the heart of it. Do not lump “we can’t tell” (the undecided) into one thing. There are at least three kinds.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LMAzDOZkNhb3Y7u_rHCPHA.png" /></figure><p>Type What is undecided Cause Examples</p><p><strong>Type A</strong> (convention-dependent) how you split and name the choice of scheme proton-mass breakdown, wage-gap decomposition, a work’s theme <strong>Type B</strong> (theorem-bound) whether the value exists at all beforehand a proven structural impossibility Bell, Rice, Arrow (distinct impossibilities) <strong>the Third Type</strong> (performative) the value, the object itself the act of naming reshapes the object Goodhart, legal precedent, AI’s reward learning</p><p><strong>Type A</strong>: it gets fought over because the split isn’t unique. But fix one scheme and, in principle, it dissolves. There’s a crucial condition — a <strong>stable total, or a shared observable, must exist first</strong>. The text of a novel, the observed wage gap, the proton’s 938 MeV. Only when that footing holds steady do we call it Type A. If the total itself is wobbling while you fight over the split, that’s a different story.</p><p><strong>Type B</strong>: no matter who measures or how, you cannot assign a value that was there beforehand — and that impossibility is <em>proven as a theorem</em>. Changing your scheme doesn’t make it go away. This is rare, and it appears only in objects closed by formal systems or physical law: mathematics, computation, physics. (Bell’s theorem is a non-classical constraint inside physics; Rice’s theorem says, roughly, “there is no universal procedure to decide whether an arbitrary program has a given semantic property” — a limit of computation theory. They have a similar outline but are <em>not the same impossibility</em>: Bell, Rice, and Arrow belong to different formal worlds, and all they share is this — the obstacle is a proven limit inside the system’s own rules, not ignorance and not convention. Same box, different contents.)</p><p><strong>the Third Type</strong>: the very act of measuring or naming brings the value into being. There is no value waiting beforehand. Make a test into a target and behavior shifts to fit the target (Goodhart’s law). The verdict “this case is an X” becomes a precedent that rewrites what “X” means for the next case.</p><p>One honest caveat before going on: “three” is a convenience for telling the story. Underneath there are really two axes — one about the object’s state, one about whether measuring changes it — plus a single external limit. Don’t lean on the number three; the full account is in a closing note.</p><h3>This is the one line you must not blur</h3><p>Of the three, <strong>Type B and the Third Type wear the same symptom and have opposite causes</strong>. Mix them up and the whole thing collapses. So this is the one point — even if it’s the hardest — to hold onto.</p><p>Both show the same face: “no value exists before measurement.” But Type B means “<strong>it can’t be decided</strong>” (structurally impossible, for anyone). The Third Type means “<strong>try to decide it and the object moves</strong>” (the act of measuring causally changes the object). The first is a wall; the second is a living thing.</p><p>Here is where most people stumble. <em>Isn’t quantum mechanics the textbook case of “measuring changes the object”?</em> Half right, and wrong in the way that matters. Shine light on an electron and you kick its momentum — yes, that’s “measuring disturbs,” but it’s just a physical shove, the same as a thermometer slightly cooling the tea. Count <em>that</em> as the Third Type and the thermometer becomes Third Type too, and the category turns to mush.</p><p>What about the deeper quantum case — “the value isn’t fixed until you measure”? That is a genuine undecided, but it is <strong>Type B</strong>. Why? Because the electron does not know it has been called an electron. For the Third Type to hold, three things must happen, in order: (1) a classification is applied to the object; (2) the object takes it up and changes in response; (3) the changed state alters the next classification — a loop. The electron can’t enter that loop; it never takes up a classification and shifts its behavior. So there is no Third Type in the physical object itself — not by accident of observation, but structurally. (To be exact: what lacks the Third Type is the physical object, the electron. Physics <em>as a human institution</em> — funding, detector design, the naming of discoveries, search strategy — can certainly run performative loops. The electron can’t enter the loop; the physicists’ practice can.)</p><p>A second touchstone is astronomy. The objects a telescope sees are biased toward the bright ones (the faint ones don’t reach us). This too <em>looks</em> like the Third Type — “observation distorts the conclusion.” But the mechanism is different: nothing about the object has changed. It’s only that our window onto it is cut, lopsidedly, out of the population. That isn’t the Third Type or Type B; it belongs to an “external limit” I’ll touch on later.</p><p>Quantum measurement (Type B that looks like the Third Type) and astronomical bias (a mere window-skew that looks like the Third Type): these two squeeze the Third Type’s outline from both sides.</p><h3>Mapping the fields</h3><p>Where the three types show up lets you color-code the disciplines.</p><p><strong>Physics</strong> is Type A + Type B. The proton-mass breakdown is Type A; Bell correlations are Type B. The map comes out neither “all objective” nor “all mystical.”</p><p><strong>Social science</strong> is Type A + the Third Type, plus Type B in its formalized regions. Decomposing a wage gap is Type A; Goodhart and institutional facts are the Third Type. Type B needs care here: Arrow’s impossibility theorem, about voting systems, is a genuine theorem-grade impossibility — “no ideal aggregation method satisfies all the reasonable conditions at once” — but this is <strong>not Type B in ordinary social life itself; it is Type B inside a formalized social-choice model</strong>. Only when a social science turns into mathematics does the limit of the formalism move in.</p><p><strong>The humanities</strong> are Type A + the Third Type. A literary text is a relatively stable shared object, but the breakdown of its theme and value is school-dependent (Type A). The formation of a canon, and the performativity by which words make reality, are the Third Type.</p><p><strong>Astronomy</strong> is Type-A–dominant. The universe’s energy budget is a robust total, but “what fraction is dark matter” depends on the ΛCDM accounting frame (swap in modified gravity and dark matter’s share shrinks). The Third Type is zero — as above, the cosmos does not take up its own classification, and you cannot re-run it under other conditions.</p><p><strong>Law</strong> is Type A + the Third Type. Reading the statute is Type A; the accretion of precedent (where an earlier ruling binds the next) is the Third Type.</p><p><strong>Normative systems (constitutions, states, religion)</strong> are where the Third Type takes the lead — the next section.</p><p>And <strong>AI</strong> is where Type A, Type B, and the Third Type pile up most densely in a single site. Also below.</p><h3>The Third Type has a sign — disease, or design?</h3><p>The Third Type has one tricky and fascinating property: the same “measuring moves the object” can be either a defect or the point.</p><p>In descriptive science, the Third Type is contamination. Make a test into a target and the scores rot; optimize an AI’s reward and you drift from what you meant to measure — measuring itself warps the object. A “disease.”</p><p>But where you are <em>positing a norm</em>, the same mechanism becomes the “design spec.” Where does a constitution get its force? The legal philosopher Kelsen, to stop an infinite regress, posited a <em>basic norm</em> (Grundnorm) as a footing — not as something real, but as a stipulation. Hart, likewise: the ultimate criterion has no ground other than the fact that “everyone practices it as the criterion.” At the institutional level, a state’s sovereignty and a religion’s orthodoxy work the same way — they become socially effective because a community treats them as binding. The question is not “is it true?” but “is it institutionally in force?” They are not there beforehand. Here, “no objective pre-existing value” is not a flaw but the normal mode of being.</p><p>So what fixes the sign is not the type — the Third Type — but <strong>what you are trying to do</strong>. Trying to describe a fact, an object that moves is a nuisance (disease). Trying to posit a norm, an object that stands up under you is the goal (design). One type; the evaluation comes from a separate axis.</p><h3>AI Makes All Three Visible at Once</h3><p>AI is the <strong>clearest present-day case</strong> of all three types surfacing in one technical site. (Not the <em>only</em> one — financial markets and medical diagnosis come close. But a case where all three are this sharply superimposed is, for now, hard to name elsewhere.)</p><p>Type A — a benchmark’s total score is robust, but the basis for splitting it into “reasoning” and “knowledge” is not unique. The Third Type — feed the test into training and it gets eaten; optimize the reward and “honesty” gets shaped into the form of that reward. And Type B — there is, in general, no procedure to decide “is this model deceptive?” or “does it hold a hidden goal?” (more precisely: <em>insofar as you try to formalize deception or goals as semantic properties of a program</em>, you hit a Rice’s-theorem-type wall).</p><p>Here too the familiar confusion waits. “Can’t be decided” (Type B, a wall) versus “moves when you try to measure” (the Third Type, a living thing). Discourse about AI tends to run on with these two swapped. Telling them apart is the minimum condition for talking sensibly about AI capability and safety.</p><p>One note. Whether AI is special by <em>degree</em> (the clearest case) or by <em>kind</em> is still open. In the other fields the three types arise in the object; only with AI can the observer itself be an AI — and then the Third Type folds back onto the act of classifying. Whether that is a difference of quantity or of kind is left to a companion piece.</p><h3>Maturity is knowing what your own hands are doing</h3><p>One value judgment, to close.</p><p>Some fields know they are <em>choosing how to cut</em>; some don’t. Physics says “this is a choice of gauge”; literary theory says “this is reading through such-and-such a lens.” Because they have a language for their own conventionality, they can stop fighting pointless wars over truth.</p><p>Other discourses turn the same operation into a fight about <em>reality</em>: “but is it really there?” This is less a matter of the field than of the <strong>discourse culture inside it</strong>. Physics has its naive corners too, and social science has sophisticated regions that handle conventionality well. Self-awareness depends not on whether the object is natural or artificial, but on the culture of the room.</p><h3>So, before you ask, ask three things</h3><p>Before you ask “but is it <em>really</em> there?”, ask yourself this.</p><p>One: am I, right now, splitting something and giving the parts names (is this Type A)? Two: is it genuinely undecidable, or would fixing a scheme settle it (Type B, or Type A)? Three: is the act of measuring or naming <em>making</em> the very object (is this the Third Type)?</p><p>A discovery, a choice of convention, and a product of the hands. Being able to tell these three apart — that is what it means to stop treating “we know / we can’t tell” carelessly.</p><p>Even when the total is solid, how you cut it is a choice. And sometimes, choosing how to cut is itself remaking the object.</p><p><em>[Note] I did this classification together with an AI. But that work has a trap: an AI’s output is itself a classifying practice shaped by reward. Which means the act of classifying is itself subject to the Third Type. This problem of self-reference is not the body of this essay; it is handled in a companion piece.</em></p><p><em>[Hold marker] This is the public version (Path A). The three types are kept on the surface, and the axis structure behind them (state axis × process axis + external limit) is signposted only, left running as the engine. The full re-projection onto axes, the special cell that Arrow’s impossibility theorem occupies in the classification, and the finer three-way split inside Type B (ontological non-separability / computational non-compositionality / proof-theoretic gap) are sent to the specialist version (Path B).</em></p><p><em>[The engine, in brief] Counting “three” was a convenience. Strictly, it is two axes plus one limit. Type A and Type B share a </em><strong><em>state axis</em></strong><em> — does a fixed object have a unique split, or a pre-existing value? The Third Type sits on a </em><strong><em>process axis</em></strong><em> — does measuring, over time, change the object? Astronomy’s selection effect and N=1 (one universe; no re-shooting a supernova) sit upstream of both, where the premise “a robust total exists” breaks down — not a fourth type. So three is just how it looks when a single point on the process axis is forced into line with two points on the state axis. On the broader habit of fixing the number of reasons before the argument has been examined, see </em><a href="https://medium.com/@izayohi/the-three-reason-contagion-fcbc3c730bcd"><em>The Three-Reason Contagion</em></a><em> and its sequel </em><a href="https://medium.com/@izayohi/writing-forms-are-made-not-born-755d02dde23b"><em>Writing Forms Are Made, Not Born</em></a><em>. Unlike the template, which imposes three from outside, this essay’s three merely fell out of folding the content into a list — opposite mechanism, same lesson: ask why three before you trust three.</em></p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><p><strong>Japanese version:</strong></p><p><a href="https://note.com/mototchen/n/n9f79c42b1658">https://note.com/mototchen/n/n9f79c42b1658</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=39c9ee5df149" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An Inheritance of Tension]]></title>
            <link>https://medium.com/@izayohi/an-inheritance-of-tension-f6c2eef33abe?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/f6c2eef33abe</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[science-fiction]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[quantum-mechanics]]></category>
            <category><![CDATA[science-communication]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Mon, 25 May 2026 00:55:42 GMT</pubDate>
            <atom:updated>2026-05-25T00:55:42.273Z</atom:updated>
            <content:encoded><![CDATA[<p><em>— a short story, and its physics</em></p><p><em>Motohisa Ishibe</em></p><h3>Part One — The Story</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GF3XKXdrqZaNtvsk9OWOMA.png" /></figure><h3>1</h3><p>In the apartment with the broken air conditioner, only the cicadas kept working.</p><p>Naoto wiped his arm across his face and went on putting prices to his aunt’s things. The trouble was that there were no prices to put. The shelves held nothing but technical volumes worth their weight in scrap paper; the appliances were obsolete; her savings had all but vanished into the funeral. Kyōko Makabe — a theoretical physicist, apparently, though she had not lasted long at any university, and had spent her last years pouring science essays nobody read onto the internet. The eccentric aunt. Died alone. Four days before anyone found her.</p><p>“In the end she left nothing,” Naoto muttered. Nothing with a price.</p><p>It came to him, oddly, that she had never put ice in her barley tea. She hated it going watery, she’d told him once, years ago.</p><p>At the back of a desk drawer he found a thick sheaf of paper and a black box that sat in his palm. On the cover of the sheaf, in her hand: <em>Matter, or Tension?</em> He flipped it open — equations, and words like <em>telepathy</em>, <em>DNA</em>, <em>bond</em>. He snorted. What worthless junk. He tossed it into the garbage bag.</p><p>The first page slid out, face up, onto the floor. It read:</p><p><strong>“To my nephew, Naoto. This is not a keepsake. It is a measuring device.”</strong></p><h3>2</h3><p>He picked the sheaf back up and read it standing.</p><p><em>Right now you are pricing this room the way you would weigh my body. Stop. Physics gave up that way of weighing things a long time ago.</em></p><p>Sixty kilograms on the scale. Of that number, the part that is “matter itself” — the bare mass of the particles that make you — is at most one percent. The remaining ninety-nine percent is not matter at all but the energy of force: the energy that drives those particles at nearly the speed of light and holds them in. Not the contents, but the state of the contents. Not matter — tension.</p><p><em>There is almost nothing of price in my apartment either. But that does not mean I had no weight. It only means that the part of me you can price happened to be the one percent.</em></p><p>He had already fished the black box back out of the bag.</p><h3>3</h3><p>Inside the box were two small paired crystals and a sheet of instructions. A small demonstrator his aunt had built, it said, to <em>show</em> what a pair of entangled photons does. At the bottom of the sheet, another note to him.</p><p><em>An old television program once called quantum entanglement “telepathy” — because they imagined two distant objects secretly sending signals to each other. But that is wrong. It is hard to keep the picture in which each one carried its own settled value all along, before anyone measured. Entanglement is what shows up when you bring two records together afterward and find a correlation that ordinary explanations cannot account for — proof, without any signal, of being a true partner. Look at one alone and the result is fifty-fifty noise. Nothing to read.</em></p><p><em>You and I were a pair like that, I think. While I was alive we hardly read each other at all. To you I was fifty-fifty noise — just an eccentric. To me you stayed unread as well. So this is not a measurement I am completing from beyond; it is the worked example I am leaving you instead. The half you bring to it, you are reading now.</em></p><p>When he pressed the side of the box as the sheet directed, a small speaker began to play a synthesized version of his aunt’s voice — a machine trained on her writing, speaking her own continuation.</p><p><em>One last thing. You are surely thinking: nothing about who I was can be settled from an empty room and a few notebooks. You are right. The data alone do not settle it. What settles it is which weighing you choose. Weigh me in matter and I am one percent — nearly nothing. But if, as mass itself has taught us, weight shows up not only as matter itself but on the side of state and relation too — then our bond, also, appears as weight only when you weigh it from the side of relation.</em></p><p>The voice cut off. The cicadas, dutiful, started up again.</p><h3>4</h3><p>Naoto went to tie off the garbage bag, and lifted the one with the sheaf of paper inside. It was heavy. As matter, not worth a bundle of scrap. But right now, in this room, it was the heaviest thing.</p><p>He pulled the sheaf out of the bag, gathered it against his chest, and stood.</p><p>Ninety-nine percent of his aunt’s mass had been, to the end, not matter. The rest was — there in his arms now, still not done sounding — the sound of tension.</p><h3>Part Two — [Commentary] Matter, or Tension? Where the Physics Stands Now</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dku9C1XZNb7hbqc0PBUN6A.png" /></figure><p>The two physical facts this story leans on are not metaphors; they have been confirmed by experiment and computation. The part that builds a bridge from them to human relationships and the meaning of an inheritance, however, is not physics itself but a metaphor that borrows the naive picture of substance physics has broken. And the two facts share a peculiar habit: the headline is solid, while the breakdown — and <em>what exactly broke</em> — still refuses to settle into a single answer.</p><h3>1. Ninety-nine percent of you is not “matter”</h3><p>The proton’s mass is about 938 MeV. Naively it should be the sum of the masses of its three quarks (two up, one down). But add up the bare masses of those valence quarks, excluding the sea, and you get only about 1% of the proton’s total. The remaining ~99% is not “contents” but the kinetic energy of quarks moving near the speed of light, together with the energy of the strong force (QCD) by which gluons bind them — appearing as mass through E=mc². Frank Wilczek called this “mass without mass” [Wilczek 2008].</p><p>“Being able to explain” and “being able to produce the number” are different achievements; the latter arrived only this century. In 2008, Dürr et al. used lattice QCD on supercomputers to reproduce the light-hadron masses from first principles to within a few percent [Dürr et al. 2008].</p><p><strong>1% becomes 9% — the breakdown depends on the accounting.</strong> In 2018, Yang et al. decomposed the proton mass on the lattice into four parts [Yang et al. 2018]: quark kinetic energy ~32%, gluon field energy ~36%, the purely quantum “trace anomaly” ~23%, and the quark mass term ~9%. The scheme follows Ji’s decomposition [Ji 1995]. Note the last figure. The quark mass that was 1% counting only valence quarks becomes ~9% once you also count the sea of quark–antiquark pairs boiling out of the vacuum (this larger count is what the “σ-term” measures). To the same question about the same proton, two honest answers — separated only by where you draw the line around “quark.”</p><p><strong>The breakdown of the breakdown is also still contested.</strong> The four-term decomposition itself is not unique. Lorcé, and Metz, Pasquini &amp; Rodini, argue that there are at most two independent terms, and that in some decompositions no separate trace-anomaly term appears at all [Lorcé 2018; Metz, Pasquini &amp; Rodini 2020]. Whether the 23% quantum effect is a real independent component or a bookkeeping artifact remains unresolved, Ji’s rebuttals included.</p><p>Experiment is now going straight for this contested term. GlueX at JLab measured the near-threshold J/ψ photoproduction cross section (Ali et al. 2019), and a series of analyses then extract the trace anomaly (the proton’s “mass radius”) from that data in a model-dependent way (via holographic QCD or vector-meson-dominance models). The measured cross section and the model-mediated extracted value need to be kept apart. The extractions currently carry large uncertainties, and some favor a small value. The main hope is the Electron–Ion Collider (the US EIC and China’s EicC), built to separate the gluonic origin of mass with precision. But a caution is in order: what the EIC will sharpen is the value of each term <em>once a decomposition scheme is fixed</em> (an observable such as the mass radius), not the uniqueness of the four-way split itself. The latter is a matter of convention, not of experimental precision — however precisely the EIC measures, the scheme-dependence by which the trace-anomaly term can even be zero (Metz et al.) will not go away. The “undecided” here is not a blank that experiment will eventually fill; it stems from a human choice of accounting frame. The Higgs mechanism [ATLAS/CMS 2012], which gives quarks their small bare masses, contributes under 10% of the proton’s mass; weight has no single origin.</p><h3>2. “Things being there” was shaken by experiment too</h3><p>What makes entanglement look like “telepathy” is the intuition that two distant objects exist as substances and secretly communicate. But in 1964, Bell derived an inequality that any locally-realistic theory must obey [Bell 1964], and CHSH put it into testable form [CHSH 1969]. From Aspect’s experiments in the 1980s onward, the violation was confirmed again and again, and in 2015 experiments closed the detection and locality loopholes simultaneously [Hensen et al. 2015 and others]. The 2022 Nobel Prize in Physics went to Clauser, Aspect, and Zeilinger.</p><p>As for the current frontier: in 2023, ETH Zurich demonstrated a loophole-free violation with a pair of superconducting qubits linked by a 30-metre cryogenic line (S = 2.0747, p &lt; 10⁻¹⁰⁸) [Storz et al. 2023] — meaning that, after diamond defects, photons, and neutral atoms, the workhorse element of quantum computing itself violates a Bell inequality. In 2024 the same class of device achieved device-independent self-testing [Storz et al. 2024], and in 2025 a violation was reported in silicon quantum dots (S = 2.731) [Nat. Commun. 2025]. The S values of these experiments (the classical bound is 2) should not be lined up and compared as if a larger number meant a “stronger” violation — the platforms differ and so do the ways the loopholes are closed. What matters is the single point that each exceeds the local-realist bound of 2 with statistical significance.</p><p><strong>What, exactly, was “negated.”</strong> Here is where the story’s core lies. What Bell negated is the <em>product</em> of two assumptions — locality and realism — and which of them to drop is not written in the data [NIST; Stanford Encyclopedia of Philosophy]. Aspect himself said he would rather give up locality than realism; Zeilinger says that calling nature nonlocal goes too far. Some object to making “realism” the target of the negation at all (Norsen, Maudlin). A realist theory that keeps definite values and surrenders locality instead (Bohmian mechanics) is equally consistent with the data. So “a particle has no definite value before measurement” is a conclusion you reach <em>once you adopt a worldview that puts relation ahead of substance</em> — not the unique result the experiment forces by itself. And as Masahiro Hotta argues, entanglement is better seen not as “telepathy” but as a partner-proof — a bond — readable only when the two are brought together. That is, it is not a mechanism for reading meaning off one side alone, but one in which a correlation no classical account can explain appears only when two records are checked against each other afterward; this is also the basis of quantum cryptography (QKD).</p><h3>3. What the two share — but they are not the same shape</h3><p>In both cases the headline (“99% is not the bare mass of the particles,” “local realism is violated”) is solid. But the two “undecideds” are not the same kind, and collapsing them is where the argument breaks.</p><p>On the mass side, the proton’s weight itself is a settled, pre-existing value (about 938 MeV); what depends on accounting is only how you split the breakdown — count the quark mass as 1% or 9%, divide it into four terms or two (the A-type from the previous section, the choice of frame). The value itself is there before you measure.</p><p>The Bell side is a different animal. What wavers is not the breakdown but whether the values exist at all before measurement (B-type). And the crucial point is this: the latter is not a consequence of the former. From “most of the proton’s mass is relational energy” you cannot logically derive “an entangled particle has no value before measurement.” What joins the two is not logic but a <em>choice of worldview</em> — to put relation (or context-dependence) ahead of substance.</p><p>Adopt that worldview, and the same <em>direction</em> of turn does run through both fields: a reading that places the body of the thing on the side of relation and tension rather than matter. But it is not that “the same conclusion falls out of both fields automatically.” What is certain is the turn itself — the naive picture of substance broke at both sites. What is uncertain is where, beyond it, you draw the line — and that the two sites broke in different ways: an A-type that keeps pre-existing values, and a B-type that includes their non-existence before measurement.</p><h3>References</h3><ul><li>Bell, J. S. “On the Einstein Podolsky Rosen Paradox.” <em>Physics</em> <strong>1</strong>, 195 (1964).</li><li>Clauser, J. F., Horne, M. A., Shimony, A., Holt, R. A. “Proposed Experiment to Test Local Hidden-Variable Theories.” <em>Phys. Rev. Lett.</em> <strong>23</strong>, 880 (1969).</li><li>Hensen, B. et al. “Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres.” <em>Nature</em> <strong>526</strong>, 682 (2015).</li><li>Storz, S. et al. “Loophole-free Bell inequality violation with superconducting circuits.” <em>Nature</em> <strong>617</strong>, 265 (2023).</li><li>Storz, S. et al. “Complete self-testing of a system of remote superconducting qubits.” arXiv:2408.01299 (2024).</li><li>“Bell inequality violation in gate-defined quantum dots.” <em>Nat. Commun.</em> (2025). (Reported Bell signal S = 2.731.)</li><li>Ji, X. “QCD Analysis of the Mass Structure of the Nucleon.” <em>Phys. Rev. Lett.</em> <strong>74</strong>, 1071 (1995).</li><li>Yang, Y.-B. et al. “Proton Mass Decomposition from the QCD Energy Momentum Tensor.” <em>Phys. Rev. Lett.</em> <strong>121</strong>, 212001 (2018). arXiv:1808.08677. (The four-term split 32/36/23/9% is a lattice-QCD computed result, not a directly measured quantity.)</li><li>Lorcé, C. “On the hadron mass decomposition.” <em>Eur. Phys. J. C</em> <strong>78</strong>, 120 (2018).</li><li>Metz, A., Pasquini, B., Rodini, S. “Revisiting the proton mass decomposition.” <em>Phys. Rev. D</em> <strong>102</strong>, 114042 (2020). arXiv:2006.11171.</li><li>J/ψ cross-section measurement: GlueX Collaboration (A. Ali et al.), “First Measurement of Near-Threshold J/ψ Exclusive Photoproduction off the Proton.” <em>Phys. Rev. Lett.</em> <strong>123</strong>, 072001 (2019).</li><li>Model-dependent extraction from that data (trace anomaly / mass radius): holographic QCD (Y. Hatta et al.) and vector-meson-dominance analyses — <em>Eur. Phys. J. C</em> <strong>80</strong> (2020); <em>Eur. Phys. J. A</em> (2022). These are model-assumption-dependent extracted values, distinct from the measurement itself.</li><li>Dürr, S. et al. “Ab Initio Determination of Light Hadron Masses.” <em>Science</em> <strong>322</strong>, 1224 (2008).</li><li>Wilczek, F. <em>The Lightness of Being</em> (2008).</li><li>Confirmation of the Higgs mechanism: ATLAS Collaboration &amp; CMS Collaboration (2012).</li><li>2022 Nobel Prize in Physics: A. Aspect, J. F. Clauser, A. Zeilinger.</li><li>On the interpretive dispute: NIST, “Local Realism, Bell’s Inequality”; <em>Stanford Encyclopedia of Philosophy</em>, “Bell’s Theorem.”</li><li>The “bond” picture of entanglement: Masahiro Hotta, “Quantum entanglement is not ‘telepathy’ but a bond between genuine partners,” note (2025).</li></ul><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><p><strong>Japanese version:</strong></p><p><a href="https://note.com/mototchen/n/n56161eb4d185">https://note.com/mototchen/n/n56161eb4d185</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f6c2eef33abe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AGI That Is Not a Single Giant Model — The Parallel Holonic Architecture (Part 6)]]></title>
            <link>https://medium.com/@izayohi/agi-that-is-not-a-single-giant-model-the-parallel-holonic-architecture-part-6-b836a9a52f08?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b836a9a52f08</guid>
            <category><![CDATA[data-sovereignty]]></category>
            <category><![CDATA[multilingual-nlp]]></category>
            <category><![CDATA[ai-strategy]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[geopolitics]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Fri, 22 May 2026 07:32:15 GMT</pubDate>
            <atom:updated>2026-05-22T07:32:15.297Z</atom:updated>
            <content:encoded><![CDATA[<h3>Data Sovereignty, Multilingual Parallelism, and Distributed Modules — The Three Axes of an International Strategy for Parallel AGI</h3><p><em>Motohisa Ishibe</em></p><p><em>This is the English edition of Part 6 of a Japanese-language series originally written for a domestic readership of policymakers, AI researchers, the industrial-robotics sector, and the ITRON embedded-systems community. The English edition reorders the four regional cases for an international readership and adds one section (Section 4) on Singapore that was developed after the Japanese edition was published.</em></p><p><em>A note on terms for readers arriving without the earlier installments. The series proposes a “parallel holonic architecture” as a counter to the single-giant-model route, structured in three layers: the </em><strong><em>Aktuanz</em></strong><em> layer (the lower, physical layer of sensing and action), the </em><strong><em>Panich</em></strong><em> layer (the middle, module-parallel layer), and the </em><strong><em>Kontextleib</em></strong><em> layer (the upper, context-emergent layer). These German-derived terms, together with </em><strong><em>Interaktuanz</em></strong><em> and </em><strong><em>Intraaktuanz</em></strong><em> (the inter- and intra-modular relations among them), are the series’ working vocabulary. “Aggregationism” denotes the organizing principle of unifying a single measurement grid from above; “the right to refuse measurement” denotes the corresponding right of a subunit to retreat from that grid. “Shin-Nexialism” is the author’s framework, after A. E. van Vogt’s Nexialism, for connecting knowledge across disciplines.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XP3AdITPPD1sAYarcdWjzA.png" /></figure><h3>1. Three Variations of Aggregationism — Adding the Middle East to the US–China Confrontation</h3><p>This series opened in Part 1 with the contrast between the center model and the parallel model, and in Part 5 it descended to the level of organizational theory. Part 6 raises the analysis to the level of international strategy. The title speaks of “data sovereignty,” but what this installment treats is not the standalone development of the concept of data sovereignty. It presents three axes — data sovereignty, multilingual parallelism, and distributed modules — as a hierarchy of dependence, and positions as a third pole those states and regions whose organizing principles mutually support all three axes.</p><p>Before presenting the third pole, this section fixes the axis of contrast. The structural tension between the two great blocs of the US and China deepened a notch as 2026 began, but what this series treats is not the US–China confrontation as such. It is the organizing principle the two camps share — aggregationism. This section adds, further, the Middle East as a third variation of that aggregationism. Positioning Gulf-state AI development not as “a third pole standing midway between the US and China,” but as “a third variation of aggregationism that implements the same organizing principle with a different funding structure from the US and China,” is what sustains the logical structure of this installment.</p><p>The common structure of the three variations of aggregationism is the unification of the measurement grid from above, the structural non-production of what goes unmeasured, and the reproduction of the Cognitive Blindspot chain at the national level. What Part 5 showed at the organizational level through the case of Shuji Nakamura, this section reproduces at the level of states and international institutions. For the same reason that an organization well managed by a single measurement grid does not produce the blue LED, a confederation of nations well managed by a single measurement grid does not produce distributed parallel AGI.</p><h3>1-A. US-Type Aggregationism — Unification of the Measurement Grid by Commercial KPIs</h3><p>The core of US-type aggregationism is the unification of the measurement grid by commercial KPIs. The short-term VC investment structure, enterprise valuation on a quarterly-earnings basis, ranking by benchmark scores, the token-billing business model — these link together to converge the direction of AI development onto “performance that is measurable and monetizable in the short term.” US-type aggregationism, centered on OpenAI, is the most complete implementation of this measurement grid.</p><p>The geopolitical situation of 2026 strengthened this structure a notch. The Trump administration moved on February 27, 2026 to halt the use of Anthropic technology by federal agencies (reported by The Guardian). Around the same time, OpenAI was reported to be deepening its involvement in Pentagon-related classified workloads (Washington Post, April 2026). Anthropic suffered a D.C. Circuit setback on April 8, when its request to pause the supply-chain-risk designation was denied, while the broader litigation remained unresolved; and in April 2026 Reuters reported that Anthropic had rapidly narrowed the revenue gap with OpenAI. A $30B ARR figure also circulated in secondary market-tracking sources, but I treat it as an estimate rather than an audited or official figure. The parallel progression of structural exclusion from the federal-agency market and commercial expansion looks like the settling of intra-US competition over a single flagship firm, but structurally it points to a deeper state of affairs. US-type aggregationism moves in the direction of excluding from the market any firm that internalizes normative constraints.</p><p>This exclusionary tendency is itself consistent with the logic of aggregationism. A firm that internalizes normative constraints attempts to fold variables outside the measurement grid (safety, long-term impact, social responsibility) into its decision-making. The measurement grid of aggregationism processes such variables as “costs” and over the long run exerts pressure to exclude them. The Trump administration’s decision can be read not as a political bias but as a logical consequence of US-type aggregationism.</p><p>OpenAI’s single-giant-model trajectory — the parallel training runs indicated by the temporal proximity of GPT-5.4’s official release and the completion of pretraining for “Spud” (the secondary-reporting provisional designation for GPT-5.5) — is a typical example of how the unification of the measurement grid governs technical choice. The intra-single-model aggregation of omnimodal functionality is easier to measure against commercial KPIs than the coordination of multiple models. The appeal of “one product that can do everything” raises intelligibility in quarterly earnings. It is not technical superiority but fitness to the measurement grid that drives technical choice.</p><p>The English-centric structure of the training data reproduces this unification of the measurement grid at the linguistic layer. The major benchmarks for US-type LLMs take English as their baseline, and multilingual performance is treated as a secondary metric. Major benchmarks such as MMLU, HellaSwag, and GSM8K take their English versions as the baseline, and versions in other languages are placed in a derivative position. The thesis treated in Part 5’s parallel sub-series (the Concepts Have Borders series) — “what is not measured, what is not represented, is not improved” — operates here: cognitive grids in languages other than English fall outside the measurement target of commercial KPIs and are structurally not improved.</p><h3>1-B. China-Type Aggregationism — Unification of the Measurement Grid by State Standardization</h3><p>China-type aggregationism arrives at the same unification of the measurement grid by a different route from the US type. State standardization, not commercial KPIs, is its driving principle. Its coupling with the party-state system, centralized data control, and the mainland’s language policy — the Putonghua-standardization policy — unify the measurement grid from above.</p><p>From 2025 into 2026, the Chinese side showed a rapid catch-up in AI development. Following DeepSeek-R1 (early 2025), Qwen (Alibaba), Kimi (Moonshot), and GLM (Zhipu AI) were strengthened in succession. An a16z report of April 2026 states that eighty percent of developers using open-source tools use Chinese open-source tools, and that Chinese open models accounted for as much as thirty percent of AI usage in some weeks of 2025. A US-China Economic and Security Review Commission (USCC) report of March 2026 also treats the structure by which China deploys an open-source AI strategy as national industrial strategy.</p><p>The open-weight strategy of Chinese models has functioned as a structural counter to the closed strategy of the US type. But being open does not mean the negation of aggregationism. Chinese models combine “open weights” with “closed training data, training process, and social premises.” Users can freely download a model, but the epistemological premises that model internalizes — historical understanding, political boundaries, social norms — are very likely decided centrally, and are at the least not unrelated to a national linguistic and political order.</p><p>The most explicit manifestation of this structure lies in the mainland’s language policy. The Putonghua-standardization policy compels Putonghua as the standard language across the whole field of education, media, and administration, and has marginalized regional languages (Cantonese, Wu, Min Nan, Hakka, etc.). The policy toward minority languages (Tibetan, Uyghur, Mongolian, etc.) carries even stronger assimilation pressure. This national language policy does not mechanically determine every Chinese open model — some, such as Qwen, are built with explicit attention to English and multilingual performance — but it forms the state-standardizing background against which the training-data and model-design choices of China-type AI development are made: the cognitive grid tends to be unified by an epistemology standardized on Putonghua.</p><p>China-type aggregationism implements the unification of the measurement grid as “measurement by the state.” This is contrasted with the US type’s implementation as “measurement by the market,” but the consequence — that what goes unmeasured is structurally not produced — is the same. Both the regions the market does not measure and the regions the state does not measure fall outside the feedback loop of improvement.</p><h3>1-C. Middle-East-Type Aggregationism — The Two Sides of Funding Structure and Organizing Principle</h3><p>Middle-East-type aggregationism is a category this series newly presents. Gulf-state AI development — the UAE’s G42 (and its subsidiaries Inception and Core42), Saudi Arabia’s HUMAIN (an AI firm established under the PIF in 2025; a separate organization from SDAIA), Qatar’s QRDI — implements the same organizing principle of aggregationism while holding a funding structure different from the US and China. This section treats this third variation not by evaluating technical achievement and organizing principle separately, but by discussing the two as the two sides of the same structure.</p><p>Seen from technical achievement: Arabic LLMs such as Falcon (UAE TII), ALLaM (Saudi Arabia), and Jais (UAE Inception + Cerebras) have implemented world-class performance from 2024 into 2026. The Falcon series occupies a high rank on HuggingFace as an open-weight model, and Jais is referenced internationally as a benchmark for Arabic LLMs. These are not “a developing-world catch-up” but technically excellent implementations.</p><p>Seen from funding structure: the scale of the Gulf states’ sovereign wealth funds — PIF roughly $925B, ADIA roughly $993B, QIA roughly $530B (all approximate figures from various reports as of end-2025) — enables a time horizon different from both the US VC structure and Chinese state investment. A sovereign wealth fund functions as a long-term investment destination for oil revenue, bound neither by quarterly-earnings pressure nor by the five-year segments of state planning. The strategic positioning of AI development as part of decarbonization-transition investment offers a long-term perspective that exists in neither the US nor China.</p><p>Up to here is the independent evaluation of technical achievement and funding structure. The structural analysis of this section moves from here to discussing the two as the two sides of an organizing principle.</p><p>The Gulf states’ sovereign wealth funds are the material expression of the organizing principle of monarchy. Final decision authority over the funds belongs to the royal family, and the long-term investment perspective is secured by the premise that the royal family governs over the long term. This funding structure structurally promotes aggregationism in technical choice — a single decision-making subject lays out long-term plans and need not integrate multiple competing voices.</p><p>This structure is reflected in the design choices of Arabic LLMs. Arabic is the official or working language of twenty-eight countries, and it holds a dual structure of Modern Standard Arabic (MSA) and national dialects (Egyptian, Gulf, Levantine, Maghrebi, etc.). Linguistically, each dialect can be treated as an independent linguistic entity with limited mutual intelligibility. The mainstream of the Gulf states’ Arabic LLMs takes MSA as the standard and places dialects in a derivative position. There are models that advertise dialect support, but they remain a secondary treatment. This is an extension of the traditional positioning that sets MSA — the language of the Quran, classical literature, and public discourse — at the linguistic center, but at the same time it includes a structure that marginalizes the cognitive grids of national dialect communities.</p><p>As the organizing principle behind this language policy, in the Gulf states religious education, public discourse, and the legal system are strongly tied to state institutions, and the interpretation of language, ethics, and the public sphere is coordinated by upper-level institutions. This structure is reflected in education policy, media policy, and the choice of the measurement grid in AI development.</p><p>Here the analysis does not proceed to the normative judgment that “therefore Middle-East-type is inferior as an organizing principle.” The structural analysis of this section stops at showing that the organizing principle of the Gulf states belongs to the category “aggregationism.” Why aggregationism: because it shares the structure in which the measurement grid is unified from above and what goes unmeasured is structurally not produced. By a route different from the US and China (concentration of funds, public management of religious education and the legal system, the monarchic long-term perspective), it arrives at the same consequence of aggregationism.</p><p>This does not mean that the UAE, Saudi Arabia, and Qatar are institutionally identical. They differ in institutional structure, religious establishment, the operational role of English, and industrial policy. The category abstracts across these differences in order to isolate the shared structure: sovereign-wealth-fund-backed central decision-making over AI infrastructure.</p><p>This is what it means to see technical achievement and organizing principle as the two sides of a structure. The Gulf states’ sovereign-wealth-fund structure makes it possible to invest the enormous funds required for single-giant-model training over a long horizon — and so Falcon, ALLaM, and Jais come into being. The same funding structure is not suited to a distributist design that lets multiple competing technical routes coexist — and so multilingual parallelism and emergent diversity are structurally unlikely to emerge. From the same logic of the same structure, a particular technical achievement is produced, and another kind of technical achievement is structurally not produced.</p><p>Here let us pre-empt an anticipated objection. The Gulf states are highly internationalized in labor, investment destinations, and technical partnerships, and one could take this internationality as a refutation of the aggregationist classification — Indian IT talent, Western financial capital, Chinese infrastructure supply, Asian manufacturing, with the Gulf states functioning as a hub that integrates these. But this internationality does not mean a dispersal of decision-making subjects. Rather, it functions as a hub-type aggregation in which a central funding subject (a sovereign wealth fund, a royal-family-affiliated investment institution) integrates multiple external resources. Global hub-ness does not contradict aggregationism — it is compatible with it as an efficient implementation form of aggregationism.</p><h3>1-D. The Common Structure of the Three Variations and This Series’ Counter-Thesis</h3><p>The three variations — US-type, China-type, Middle-East-type — differ in driving principle (commercial KPIs / state standardization / funding structure) and in the implementation method of the measurement grid, but they share the organizing principle of the unification of the measurement grid from above. Extracting the common structure of the three:</p><p>First, the unification of the measurement grid issues from a single decision-making subject (the market, the party-state system, the monarchy). Second, what goes unmeasured is structurally not produced, and the Cognitive Blindspot chain is reproduced at the level of states and international institutions. Third, the coexistence of multiple cognitive grids (multilingual parallelism, ideological pluralism, competition among technical routes) is processed as an increase in measurement cost and a decrease in decision-making speed, and is excluded by the logic of aggregationism.</p><p>This series’ counter-thesis to this common structure has been presented consistently since Part 5 — the parallel holonic architecture. At the technical-architecture layer, distributed modules; at the organizational layer, a distributed organization that internalizes the right to refuse measurement; and at the international-strategy layer that this installment treats, the integration of the three axes of multilingual parallelism, distributed modules, and data sovereignty.</p><p>What matters here is that this series’ counter-thesis does not propose “a third pole standing midway between the US and China.” Take the midpoint between the US type and the China type, and you are caught in the logic of aggregationism the two share. Add the Middle-East type and take the midpoint of the three terms, and you are all the less able to escape the gravitational field of aggregationism.</p><p>The condition for a third pole is not a moderation of political regime but a qualitative transformation of organizing principle. Only when the three organizing principles of free market and democracy, multilingual parallelism, and distributed modules are all in place does the social foundation for the parallel-model conception come into being. The Middle-East type holds no affinity with the parallel-model conception as an organizing principle — even where the technical achievement is excellent, because the organizing principle that produces it is an extreme form of aggregationism, it does not become a third-pole candidate for the parallel-model conception.</p><p>From Section 2 onward, this installment treats the four cases of appeal — the EU, India, Taiwan, and Japan — that satisfy, or partly satisfy, the conditions for this third pole. This section’s axis of contrast (the three variations of aggregationism) provides the logical background for the third-pole sections.</p><h3>2. The Hierarchy of the Three Axes — Multilingual Parallelism, Distributed Modules, Data Sovereignty</h3><p>Section 1 laid out the common structure of the three variations of aggregationism. As a counter-thesis to it, this section fixes the relationship among the three axes the installment’s title presents — multilingual parallelism, distributed modules, data sovereignty. Lay the three axes side by side in parallel and the relationship is weak; the mutual conditioning becomes invisible. This section presents the three axes as a relationship of dependence, showing the structure in which, unless the lower condition is satisfied, the upper condition idles.</p><p>The conclusion of the dependence relationship, stated first: multilingual parallelism is the cognitive lower condition, distributed modules the technical middle condition, data sovereignty the institutional upper condition. The dependence holds in the order multilingual parallelism (cognitive layer) → distributed modules (technical layer) → data sovereignty (institutional layer) — data sovereignty institutionally protects distributed modules, and distributed modules technically preserve multilingual parallelism. Where a lower layer is blank, the upper layer’s institutional apparatus is caught in the logic of aggregationism.</p><h3>2–1. Multilingual Parallelism — The Cognitive Lower Condition</h3><p>The reason for placing multilingual parallelism at the bottom is that the plurality of cognitive grids constitutes the precondition for the other two axes. In a space where a single language dominates, the cognitive grid is unified, and what goes unmeasured is structurally not produced. The thesis treated in Part 5’s parallel sub-series (the Concepts Have Borders series) — “what is not measured, what is not represented, is not improved” — was discussed as an LLM-architecture-internal problem, but the same thesis is reproduced at the level of states and international institutions.</p><p>In a space where English dominates as the single working language, concepts, ideas, and problem-settings that do not ride on the English cognitive grid are structurally not produced. This is not a matter of the goodwill or self-awareness of English users; it is a structural consequence of the measurement grid. The same structure is reproduced in spaces dominated by Chinese Putonghua, and in spaces dominated by Modern Standard Arabic. Multilingual parallelism is the cognitive foundation that blocks this unification.</p><p>Here it is necessary to distinguish “multilingual parallelism” from “multilingual support.” Multilingual support refers to a design that provides a single cognitive grid translated into multiple languages. Many US-type LLMs realize multilingual support by translating an English-baseline cognitive grid into multiple languages. On the surface this handles multiple languages, but structurally it remains a multilingualization of the expression layer while maintaining the English cognitive grid. Multilingual parallelism refers to a design that sets multiple cognitive grids themselves side by side. The former is the translational extension of a single grid; the latter is the structural parallelism of multiple grids. The difference appears at every stage — training-data collection, model design, and evaluation.</p><p>Multilingual parallelism functions as the cognitive lower condition because, unless the upper technical design and institutional design presuppose the plurality of cognitive grids, the result ultimately converges on a single grid. Implement distributed modules technically, and if each module is trained on the same cognitive grid, diversity remains a surface-level division of labor. Guarantee data sovereignty institutionally, and if the protected data is structured by a single cognitive grid, data sovereignty becomes no more than a geographical dispersal of aggregationism.</p><h3>2–2. Distributed Modules — The Technical Middle Condition</h3><p>The reason for placing distributed modules in the middle is that they function as the apparatus that technically preserves cognitive diversity. Even if multilingual parallelism is the precondition of the cognitive layer, unless it is implemented at the technical-architecture layer, cognitive diversity is easily crushed by social pressures (the market, the state, international institutions).</p><p>The axis of contrast treated in Part 4 — “intra-single-model soft parallelization vs. distributed hard parallelization” — is extended here into the context of international strategy. The Mythos hypothesis (Recurrent-Depth Transformer; presented, however, as a theoretical reconstruction by Kye Gomez and not confirmed officially by Anthropic), MEMENTO (CoT compression), Spud’s omnimodal integration — all are routes that aggregate parallelism within a single model. Technical parallelism is implemented, but that parallelism is under the control of a single organization, a single decision-making subject, a single measurement grid. Distributed modules refers to a design that disperses parallelism also organizationally, in decision-making subject, and in measurement grid.</p><p>Distributed modules function as the technical middle condition because, unless the upper institutional design presupposes the dispersal of the technical architecture, institutional data sovereignty ends as a geographical relocation of aggregationism. Place data within national jurisdiction, and if the technical architecture that processes it is a single-giant-model route, data sovereignty becomes merely the geographical dispersal of single processing by a single cognitive grid.</p><p>Here the concretes treated in Parts 3 and 4 — ITRON, industrial robots, BitNet-type lightweight LLMs, the TOPPERS project, the HFDS philosophy — reappear as the material foundation of the technical middle condition. In this section’s dependence relationship, Japan’s technical assets are repositioned not as a mere object of industrial policy but as the apparatus that technically preserves the organizing principle of the third pole.</p><h3>2–3. Data Sovereignty — The Institutional Upper Condition and This Series’ Conceptual Extension</h3><p>The reason for placing data sovereignty at the top is that it functions as the apparatus that institutionalizes the preservation of the cognitive and technical layers at the level of international law, national policy, and international institutions. At the same time, what this series problematizes is that the ordinary understanding of the concept of data sovereignty does not treat this dependence relationship.</p><p>The ordinary understanding of data sovereignty is built around the location, jurisdiction, cross-border transfer, processing authority, and data-subject rights of data. GDPR (EU, in force 2018), China’s Cybersecurity Law (2017), Data Security Law (2021), and Personal Information Protection Law (2021), India’s DPDP Act (2023), the EU Data Act (in force 2024), Japan’s amended Personal Information Protection Act (in force 2022) — these institutional developments were rapidly built out from the late 2010s into the early 2020s. These regimes treat not only location but also processing purpose, legal basis, data-subject rights, access rights, transparency, and risk management.</p><p>But all of these regimes do not adequately treat by which cognitive grid the data is structured and by which evaluation apparatus it is processed. This is an institutional design compatible with any of the three variations of aggregationism. The US type conducts data management following commercial KPIs within national borders; the China type conducts data management following state standardization within national borders; the Middle-East type conducts data management following the royal family’s long-term investment perspective under the sovereign wealth fund. The ordinary understanding of data sovereignty stops at institutionalizing the “geographical dispersal” of aggregationism.</p><p>What this series proposes is an <strong>extension of the concept of data sovereignty toward the preservation of cognitive grids</strong>. A design that institutionally preserves not the location of data but the plurality of cognitive grids inherent in the data. Concretely, three axes of extension stand up.</p><p>First, incorporating the cognitive-grid diversity of training data as an institutional metric. Folding into the certification requirements, public-procurement requirements, and regulatory requirements of AI development a structure in which the training data of an LLM is not biased toward a particular language, particular culture, or particular epistemology. The high-risk-AI classification of the EU AI Act treats performance and safety, data governance, bias mitigation, and risk management, but does not make the preservation of cultural-linguistic cognitive grids its central theme.</p><p>Second, making the dispersal of the data-processing architecture an institutional requirement. Institutionalizing structural preferential measures for distributed-module structures against data aggregation into a single giant model. This differs from mere antitrust regulation (market-concentration regulation) — it is a regulation that directly targets the organizing principle of the technical architecture.</p><p>Third, building an international-coordination mechanism for the preservation of cognitive grids. Positioning the preservation of cognitive grids as an international public good in existing international institutions such as WIPO, UNESCO, and the ITU, or in a new international framework. This is a domain not treated by the current AI international-coordination discussions — the AI Safety Summit, the Global Partnership on AI (GPAI), the UN AI Advisory Body.</p><h3>2–4. The Hierarchical Relationship of the Right to Refuse Measurement — Connection to Section 5</h3><p>Part 5’s central concept, “the right to refuse measurement,” and this installment’s “data sovereignty” are apparatuses that implement an isomorphic structure at different layers. The right to refuse measurement guarantees, within an organization, the right of individuals, teams, and modules to retreat from the measurement grid imposed from above. The extension of data sovereignty toward the preservation of cognitive grids guarantees, within the international order, the right of states, confederations of states, and regional communities to retreat from the measurement grid imposed from above (whichever of US-type, China-type, or Middle-East-type). The two stand in a hierarchical relationship — the preservation of cognitive grids in data sovereignty is the national-level implementation of the right to refuse measurement.</p><h3>The full development of the work of extending the state theory of the Miura Tsutomu–Takimura Ryūichi lineage, treated in Part 5 Section 7, into international-relations theory is placed in Section 6 (of this installment). This section confines its scope to fixing the dependence relationship of the three axes and presenting the extension of the concept of data sovereignty toward the preservation of cognitive grids. From Section 3 onward, the analysis moves to a structural analysis of the four cases of appeal — the EU, India, Taiwan, and Japan — that satisfy, or partly satisfy, this dependence relationship.</h3><h3>3. Structural Analysis of the Third Pole — The EU, India, Taiwan, Japan</h3><p>Section 1 laid out the common structure of the three variations of aggregationism (US-type, China-type, Middle-East-type), and Section 2 fixed, as a counter-thesis, the dependence relationship of the three axes (multilingual parallelism, distributed modules, data sovereignty). This section analyzes in turn the four cases of appeal that satisfy, or partly satisfy, these three axes. The order proceeds from west to east — the EU, India, Taiwan, Japan — moving from the largest confederation to the case structurally closest to the parallel-model conception, and ending with Japan, whose technical assets the series has treated in detail since Part 3.</p><p>The structural analysis of this section applies the same method to all four cases — evaluation along the three axes, presentation of both strengths and weaknesses, and connection to the organizing principle shown in Section 1. None of the four is exempted from this analytical method.</p><h3>3–1. The EU — The Thickness of the Regulatory Foundation and the Tension with Technical Competitiveness</h3><p>The EU’s distinctiveness is concentrated in the point that it is a case that implements the integration of the three axes at the level of an international institution. A supra-national institutional design that transfers a degree of member-state sovereignty guarantees the three axes — data sovereignty, distributed modules, multilingual parallelism — at the international-institutional level.</p><p>The EU institutionalizes the languages of its member states as twenty-four official languages, and operates multilingual parallelism across the whole field of inter-member negotiation, legislation, and judicature. This is one of the rare cases that implements, at the level of an international institution, the latter of the “multilingual support” / “multilingual parallelism” distinction drawn in Section 2–1. This institutional multilingualism connects directly to a distributed-module orientation in AI development. The EU’s EuroHPC JU AI Factories conception (launched December 2024, with 19 AI Factories and 13 antennas established) can be read, in this series’ terms, as a move toward coordinated infrastructure rather than a single-flagship-firm model — this is the series’ structural reading, not EuroHPC’s own self-description. The coexistence of Mistral AI (France), Aleph Alpha (Germany), Apertus (a Swiss government project, 26-language support), ALEIA (France, government AI), and EuroLLM (covering 24 EU official languages) forms an ecosystem structurally different from the single-flagship-firm concentration of the US type.</p><p>On the data-sovereignty axis as well, the EU occupies an advanced position. GDPR (in force 2018), the EU AI Act (enacted 2024), and the Data Act (in force 2024) have formed the de facto standard for the world’s data-sovereignty institutions. They move toward making not the location of data but the whole processing process the object of regulation, partly building out the institutional foundation for the “extension toward the preservation of cognitive grids” presented in Section 2–3.</p><p>However, as a structural weakness of the EU, the tension between the thickness of the regulatory foundation and technical competitiveness is pointed out. The regulatory group — the prior-certification requirement under the EU AI Act’s high-risk-AI classification, the data-processing constraints under GDPR, the data-sharing requirements under the Data Act — structurally raises the start-up cost of AI-development firms within the EU compared with the US and China. While the rise of Mistral AI and others is evaluated as an exceptional success under a regulatory environment, the overall scale of EU AI development remains a fraction of that of the US and China.</p><p>This tension is two-sided for this series’ counter-thesis. The thickness of the regulatory foundation functions as a structural brake on aggregationism, but at the same time it raises the implementation cost of distributed modules and constrains the economic sustainability of the third pole. To evaluate the third-pole potential of the EU model, the trade-off between regulation and competitiveness must be treated frankly.</p><p>The possibility of linkage with Japan is built on the two layers of regulatory coordination and mutual technical supply. The movement by which Japan’s AI Business Operator Guidelines partly reference the structure of the EU AI Act is already underway, and the structure by which Japan’s industrial-robot and embedded-systems technology is supplied to EU manufacturing-sector AI implementation already exists.</p><h3>3–2. India — Structural Affinity and the Tension of Domestic Politics</h3><p>India’s distinctiveness is concentrated in the point that it is the implementation with the highest structural affinity with this series’ parallel-model conception. Bhashini designs the ASR (speech recognition), TTS (speech synthesis), MT (machine translation), and NLU (natural-language understanding) of the 22 constitutional languages as a coordination of language-specific and task-specific specialized modules. This is not a single-giant-model strategy of the “Indian ChatGPT” type but a design structurally isomorphic with this series’ parallel holonic architecture.</p><p>India institutionalizes 22 languages as recognized in the Eighth Schedule of its Constitution, holding a two-layer structure that takes Hindi and English as working languages at the federal level while institutionalizing state official languages at each state level. The ecosystem of IIT (Indian Institutes of Technology)-affiliated research institutions and a confederation of private startups (Sarvam AI, Krutrim, TWO AI, etc.) constitutes a distributed AI-development ecosystem different from both the US type’s single-flagship concentration and the China type’s state standardization. Bhashini hosts more than 350 AI-based language models and collaborates with more than 70 research-partner institutions, which considerably supports this series’ reading of “a coordination of language-specific and task-specific specialized modules.” The Digital Public Goods conceptions — Reliance Jio’s digital infrastructure, the UPI (Unified Payments Interface) payment infrastructure — are moving in a direction that positions the AI foundation, too, as a similar public good.</p><p>But as a structural weakness of India, the tension of the Hindi-supremacy problem and the centralizing tendency of domestic politics is pointed out. The institutional parallelism of the 22 constitutional languages tends, in substance, to be biased toward a two-language structure centered on Hindi and English. Speakers of the Dravidian languages of South India (Tamil, Telugu, Kannada, Malayalam) hold a historical resistance to Hindi-supremacy policy. The centralizing tendency under the Modi administration generates a tension with state-level multilingualism under the federal system. Institutional multilingual AI foundations such as Bhashini operate amid this tension of domestic politics. The inequality of educational and economic access by caste structure also structurally constrains the inclusiveness of AI development.</p><p>These tensions do not completely negate India’s third-pole potential, but they show that there is a persistent tension between the ideal of the parallel-model conception and the reality of domestic politics. In evaluating India’s AI-development ecosystem as consistent with this series’ counter-thesis, this tension must be treated frankly.</p><p>The possibility of linkage with Japan can be imagined in the supply of Japan’s embedded-systems technology to a Bhashini-type multilingual AI foundation, the deployment of industrial robots to Indian manufacturing (in connection with the local production of Suzuki, Honda, etc., already underway), and coordination between IIT-affiliated research institutions and Japanese AI-research institutions. The strength of India’s cognitive lower condition and the strength of Japan’s material middle condition hold a structural complementarity despite the geographical distance.</p><h3>3–3. Taiwan — The Institutionalization of Multilingual Parallelism and the Right to Refuse Measurement under Geopolitical Pressure</h3><p>Taiwan, evaluated along the three axes, shows a pattern of strengths and weaknesses contrasting with Japan (treated next). On the multilingual-parallelism axis (the cognitive lower condition), it is, from the standpoint described below, the strongest of the four cases. On the distributed-modules axis (the technical middle condition), it is strong in the strategic position of its semiconductor industry, but AI development itself remains mid-scale. On the data-sovereignty axis (the institutional upper condition), it takes a distinctive form of development under geopolitical pressure. It stands in a position that structurally complements Japan’s structural weakness (the poverty of the multilingual-parallelism axis), and is the closest complementary case for “an international linkage of a third pole that includes Japan as one module.”</p><p>Taiwan is a rare case that institutionalized, through the National Languages Development Act of 2019, the coexistence of four principal language communities — Taiwanese Mandarin (the national language of the ROC), Taiwanese Hokkien (the Taiwan variant of Min Nan), Hakka, and the 16 indigenous-people languages. The act positions equally as “national languages” the natural languages indigenous to the people of Taiwan, together with Taiwanese Sign Language, and stipulates the preservation, revival, and transmission of each language as a national responsibility (per the Ministry of Culture’s official explanation and the Executive Yuan’s National Languages Development Plan).</p><p>This institutionalization stands on top of more than twenty years of accumulation of linguistic public goods. The Hakka Affairs Council (established 2001) and the Council of Indigenous Peoples (established 1996) have continuously promoted the digitization of Hakka and indigenous languages, the development of teaching materials, media support, and coordination with research institutions. The Taiwanese-language corpus development of TWNIC (the Taiwan Network Information Center), the language-standardization work of the Ministry of Education, the accumulation of linguistic research at Academia Sinica — these link together to function as a public good convertible structurally into LLM training data.</p><p>TAIDE (Taiwan AI Dialogue Engine) is a national project that connects this public good directly to AI development. Published from 2024 under the leadership of the National Science and Technology Council, it explicitly sets as its goal support for Taiwanese Mandarin, Taiwanese Hokkien, Hakka, and indigenous languages. Various Taiwanese LLMs led by Academia Sinica (CKIP-LLaMA, Taiwan-LLM, etc.) are implementations in the same direction. These are not a design that “extends a single giant model translationally into multiple languages” but a design that aims at the parallelism of the cognitive grids of each language community themselves. In the distinction between “multilingual support” and “multilingual parallelism” drawn in Section 2–1, Taiwan clearly orients itself toward the latter.</p><p>Here let us limit the evaluation of “the strongest of the four cases.” In absolute values of number of languages, population scale, and implementation scale, the EU and India are larger. Taiwan’s strength lies in the conjunction of institutional density with national identity — the four principal language communities are stated as a national-survival cognitive core mutually dependent on one another, made explicit as a legal institution. From this standpoint, Taiwan is the strongest of the four cases.</p><p>On the distributed-modules axis, Taiwan occupies an unparalleled position in the strategic position of its semiconductor industry. TSMC maintains an advanced process, with mass production of the 2nm generation (N2) scheduled to begin in December 2025 and stabilization of mass production foreseen within 2026; including mass production of the 3nm generation, it has become the indispensable supply foundation on which US AI-development firms (NVIDIA, AMD, Apple, etc.) depend. UMC also occupies a substantial share of the world market in general-purpose processes at and below 28nm. The semiconductor industry occupies a substantial proportion of Taiwan’s GDP and forms the core of the Taiwanese economy.</p><p>But the scale of Taiwan’s own AI-development firms remains mid-scale. As independent AI firms comparable to Sakana AI, Appier (advertising AI) and iKala (marketing AI) exist, but the scale gap with OpenAI, Anthropic, and the major Chinese firms is large. The AI research of research institutions centered on Academia Sinica is world-class, but its deployment to commercial implementation faces scale constraints.</p><p>This combination of “the strategic position of the semiconductor industry” and “the mid-scale-ness of AI development” produces a structure in which Taiwan is positioned as the most important supply foundation of US-type aggregationism while Taiwan itself does not become a core player of aggregationism. Taiwan materially supports US-type LLMs, but is not subordinate to the cognitive grid of US-type LLMs. This is a structure that relatively raises Taiwan’s strategic degree of freedom — it stands in a position from which it can take an independent route in AI development itself while holding decisive influence over US AI development through semiconductor supply.</p><p>On the data-sovereignty axis, Taiwan shows a distinctive form of development under geopolitical pressure. Under the triple constraint of annexation pressure from mainland China, US alliance management, and the problem of international recognition, Taiwan’s data sovereignty takes a structure different from the ordinary concept of national data sovereignty. The Taiwan government has advanced institutional development as the “Taiwan AI Action Plan” of 2018 and the “AI Basic Act” (draft) of 2024. The Ministry of Digital Affairs (established 2022) promotes, under the leadership of Audrey Tang, the institutionalization of data governance, AI regulation, and digital democracy.</p><p>Worth special note is the point that Taiwan’s AI policy makes use of the geopolitical particularity of “the front line of the US–China confrontation” as a pressure toward the institutionalization of the right to refuse measurement. Differentiation from China-type aggregationism is the core of Taiwan’s national identity, and this is, at once, an economic and military problem and a problem of the cognitive grid. The institutionalization of multilingual parallelism through Taiwan’s National Languages Development Act functions as an explicit differentiation from mainland China’s Putonghua-standardization policy. Not being subordinate to the cognitive grid of China-type aggregationism is recognized as a condition of national survival. This structure is the clearest concrete example of the thesis shown at the end of Section 1-D — “the condition for a third pole is not a moderation of political regime but a qualitative transformation of organizing principle.”</p><p>In evaluating Taiwan’s third-pole potential, structural vulnerability must also be treated frankly. Military pressure from mainland China has intensified into the 2020s, and the Taiwan Strait situation has become the foremost security concern of East Asia. Excessive dependence on the semiconductor industry makes the economic structure fragile, and TSMC’s strategic position simultaneously makes all of Taiwan the target of a geopolitical threat. The problem of international recognition structurally constrains diplomatic relations, participation in international institutions, and accession to international agreements, and makes participation in an international-coordination mechanism for data sovereignty difficult. These vulnerabilities make a third-pole implementation by Taiwan alone difficult.</p><p>Here one structural point requires a response. To position Taiwan’s National Languages Development Act as an ideal example of multilingual parallelism — does this not amount to affirming state standardization, that is, top-down measurement and institutionalization? The very act by which the state legally defines particular languages as “national languages” and standardizes and protects them as the standard for education and media is itself a powerful “institutionalization from above.” Does this section’s argument not harbor a logical twist with Zomia-type anti-aggregationism (treated in Section 5–1)? This is the objection.</p><p>The response to this objection is sought in the structural difference that Taiwan’s National Languages Development Act is not “standardization by a single measurement grid” but “the institutionalization of multiple measurement grids.” Whereas mainland China’s Putonghua-standardization policy is an apparatus by which the state compels convergence onto a single cognitive grid (Putonghua), Taiwan’s National Languages Development Act is an apparatus by which the state imposes upon itself the coexistence of multiple cognitive grids (Mandarin, Hokkien, Hakka, indigenous languages). The former is the typical aggregationism that unifies the measurement grid; the latter is a design by which the state institutionally self-binds to the plurality of measurement grids. On the surface both are “language institutionalization by the state,” but the direction of the organizing principle is the reverse. The state institutionalization of multilingual parallelism is the state’s self-binding to prevent the unification of the measurement grid. This series’ counter-thesis does not completely negate the state, but requires a structure in which the state imposes the institutionalization of multiple measurement grids upon itself. The Taiwan case is a rare implementation that shows the possibility of this structure.</p><p>The structural positions of Taiwan and Japan are complementary in this series’ three axes. Japan’s weakness (the poverty of the multilingual-parallelism axis) can be complemented by Taiwan’s strength (more than twenty years of accumulation of linguistic public goods). Japan’s strength (the industrial-robot sector, the embedded-OS layer, the semiconductor downstream process) can support Taiwan’s weakness (the mid-scale-ness of AI development, military vulnerability). The semiconductor industry is the strategic point of contact between the two, and concrete examples are already in motion — TSMC’s Kumamoto plant (JASM), the technical linkage between Rapidus and Taiwanese firms, mutual supply in embedded systems. The linkage of the two holds the potential to constitute an East-Asia-originating axis of counter to the aggregationism of the two great blocs of the US and China, though it must be designed carefully amid geopolitical constraints (China’s pressure on Japan and Taiwan, US alliance management, Taiwan’s problem of international recognition).</p><h3>3–4. Japan — The Strength of Technical Assets and the Weakness of the Organizing Principle</h3><p>Japan, evaluated along the three axes, shows a clear division of strength and weakness by axis. The material foundation of the distributed-modules axis (the technical middle condition) is, in the narrow domain described below, the thickest of the four cases. The multilingual-parallelism axis (the cognitive lower condition) is the weakest. The data-sovereignty axis (the institutional upper condition) is positioned in the middle. This section treats this asymmetry frankly.</p><p>On the distributed-modules axis, Japan’s technical assets hold a thickness. The embedded-OS layer treated in Part 3 (Ken Sakamura’s T-Engine, the TOPPERS project, the industry implementation of the μITRON family) has long maintained the implementation share of the Japanese system in the world embedded-OS market. The four qualities of deterministic operation, lightness, parallel-processing-premised design, and distributed operation structurally match the requirements of the lower layer of parallel AGI.</p><p>In the industrial-robot layer, FANUC, Yaskawa Electric, Kawasaki Heavy Industries, Nachi-Fujikoshi, Denso Wave, and others occupy a substantial portion of the world market. In the annual data of world industrial-robot shipments, Japanese firms have long maintained a top-share band. These are not a mere export industry; in the context of this series they are devices that generate material irreversibility — welding, assembly, painting, conveyance, none of which can undo what has been done. As devices that supply physical contact to the Aktuanz layer of parallel AGI, Japan’s industrial-robot sector holds a globally rare accumulation.</p><p>The strategic position of the semiconductor industry cannot be ignored either. The combination of Rapidus’s 2nm mass-production plan (Chitose, Hokkaido), JASM (TSMC’s Kumamoto plant), Kioxia (NAND), and Sony Semiconductor (CMOS image sensors) holds international significance as the hardware foundation of the lower layer of AI development. Between US export controls and China’s accelerating self-development, Japan’s semiconductor industry stands in a geopolitically re-evaluated position.</p><p>Taking these together: in the narrow lower physical layer of connection with the industrial-robot sector, the embedded-OS layer, and the manufacturing floor, Japan holds the thickest accumulation among the four cases. The EU, too, however, has a thickness in manufacturing, robotics, embedded systems, and industrial standards; Taiwan has the core of semiconductor manufacturing. Unlike “the strongest on the distributed-modules axis as a whole,” this is an evaluation that limits the domain. The grounds on which Part 3 positioned Sakamura’s philosophy as “having seen thirty years ahead” are here extended into the context of international strategy.</p><p>But on the multilingual-parallelism axis Japan is the weakest of the four cases. Japan’s domestic linguistic ecology is close to a two-language structure in which Japanese is overwhelmingly dominant and English is positioned as an auxiliary language. This is structurally poorer than Taiwan’s four-language parallelism, the EU’s 24 official languages, and India’s 22 constitutional languages.</p><p>Linguistic diversity within Japan exists — Ainu, the Ryukyuan languages (Okinawan, Miyako, Yaeyama, Yonaguni, etc.), regional varieties such as Kansai, Tōhoku, and Kyūshū dialects — but these have been structurally marginalized from the public sphere. Ainu was recognized as “the culture of the Ainu people” by the 2019 Ainu Policy Promotion Act, but it has not been made an official language. The Ryukyuan languages are listed on UNESCO’s list of endangered languages, but the Japanese government maintains the position of treating them as dialects. Despite holding a structural convertibility into LLM training data, current Japanese AI development does not make use of this asset.</p><p>Japan’s AI development (PFN, ABEJA, rinna, Sakana AI, Stockmark, Citadel AI, ELYZA, etc.) is designed centered on Japanese + English bilingual support. This is a rational design judgment reflecting the linguistic homogeneity of the Japanese domestic market, but in the context of this series’ three axes it means the poverty of the cognitive lower condition. Seen from the organizing principle of a third pole that takes multilingual parallelism as its cognitive foundation, Japan stands on a structurally weak foundation.</p><p>Here the thesis “what is not measured, what is not represented, is not improved,” treated in Part 5, is reproduced against Japan itself. The cognitive grids of Ainu, the Ryukyuan languages, and regional dialects fall outside Japan’s mainstream measurement grid, and as a result are structurally not produced in Japanese AI development. Japan stands outside the three variations of aggregationism, but domestically can be said to implement a scaled-down version of aggregationism (Japanese standardization).</p><p>On the data-sovereignty axis Japan is positioned in the middle. The institutional developments — the amended Personal Information Protection Act (in force April 2022), the Economic Security Promotion Act (enacted 2022), AISI (the AI Safety Institute, established February 2024), the AI Strategy Council (established May 2023), the AI Business Operator Guidelines — formally move in the direction of building out the institutional upper condition of data sovereignty. But in substance, Japan’s data sovereignty holds a dual structure in tension with US dependence. The major part of the cloud foundation (AWS, Azure, Google Cloud) depends on US firms; a substantial portion of the foundation models used in AI development depends on US firms (OpenAI, Anthropic, Google); and the framework of AI policy, too, is assembled as an extension of US-led international coordination (the AI Safety Summit, GPAI, the Hiroshima AI Process). Japan’s own development of the concept of data sovereignty tends to remain a partial adjustment within the framework of US-type aggregationism.</p><p>Applying this series’ conceptual extension presented in Section 2 — the extension toward the preservation of cognitive grids — makes the evaluation of the data-sovereignty axis still more severe. Japan’s data-sovereignty institutions treat geographical location (domestic processing of data, cross-border-transfer regulation), but the preservation of cognitive grids (the diversity of training data, the dispersal of the processing architecture, international coordination of cognitive grids) has not surfaced as a policy agenda.</p><p>The asymmetry of the strength of technical assets and the weakness of the organizing principle is the decisive point in understanding Japan’s third-pole potential. The Nakamura case treated in Part 5 showed the ironic structure in which the organizational weakness of Nichia Chemical Industries — a regional firm (a deficit of surveillance-and-control capacity, undeveloped KPIs, the laxity of a regional firm) — made Nakamura’s continued immersion possible. Had it been an organization well managed by a single measurement grid, the blue LED would not have been born.</p><p>Raise this organizational-theoretic observation to the level of international strategy, and Japan’s organizing principle is positioned in an intermediate region between aggregationism and distributism. The centralized measurement grid is weak (commercial KPIs do not dominate as in the US type, state standardization is not as strong as in the China type, the royal family’s long-term perspective does not dominate as in the Middle-East type), but at the same time the organizing principle of distributed modules (free market and democracy, multilingual parallelism, the institutionalization of the right to refuse measurement) is also weak. A structural ambiguity in which the center is weak while the periphery is also weak is Japan’s characteristic.</p><p>The peer-pressure organizational culture is an expression form of this structural ambiguity. There is no explicit centralized measurement grid, but a tacit peer pressure socially erodes the right to refuse measurement of individuals, teams, and modules. Nakamura was saved by the weakness of the individual firm Nichia, but what he faced was the peer pressure of Japanese organizational culture in general, and in the end his departure from this pressure (his move to the US) became the condition for the continuation of his research.</p><p>For Japan’s AI development to become a third pole of the parallel-model conception, it is necessary to make use of the strength of its technical assets while structurally correcting the weakness of its organizing principle. Implement the dispersal of the technical architecture, and if the organization that operates it is dominated by peer pressure, the dispersal remains a surface-level division of labor. Many of the concrete design problems treated from Part 7 (the implementation-prototype theory) onward concern the correction of this organizing principle.</p><p>Evaluating Japan’s third-pole potential overall: the material foundation of the distributed-modules axis is a globally rare strength in the narrow domain, but the multilingual-parallelism axis and the organizing principle harbor structural weaknesses. Seen from the dependence relationship of the three axes (multilingual parallelism → distributed modules → data sovereignty), it is a structure in which strong technical assets stand on top of a weak cognitive lower condition.</p><p>This asymmetry shows the difficulty of Japan implementing the organizing principle of a third pole on its own. Japan’s material foundation is best brought out in cooperation with cases whose cognitive lower condition is strong — Taiwan, the EU, India. Japan’s industrial-robot sector connecting to India’s Bhashini multilingual AI infrastructure; Japan’s embedded-OS layer coordinating with the EU’s AI Factories conception; Japan’s semiconductor industry supporting Taiwan’s accumulation of linguistic public goods — such international module-linkage is the route that complements Japan’s structural asymmetry.</p><h3>“Not a third pole by Japan alone” but “an international linkage of a third pole that includes Japan as one module” is the direction this series proposes. This is a concrete application of the structural alternative shown in Section 2–4 — against the “state-versus-state” geopolitical schema, “an international linkage of modules within states.”</h3><h3>4. The Two Layers of Data Sovereignty — Singapore, Sapiens AI, and the Distinction between Weight Sovereignty and Substrate Sovereignty</h3><p>This section is an addition that did not exist in the Japanese edition. It treats Singapore — a case the four-region analysis of Section 3 did not include — and through it sharpens the concept of data sovereignty presented in Section 2–3 into two layers. The material is the case of Sapiens AI, a Singapore startup that raised $20M for “homegrown models.” For an English readership, Singapore is a familiar case, and this addition is effective precisely in the English edition.</p><p>Read this case as observational material for the three axes (data sovereignty, multilingual parallelism, distributed modules) of this installment. It is a doubly valued material — at once a confirming material for the series’ argument, and a counter-example that strikes directly at its weakest joint.</p><h3>4–1. The Multilingual-Parallelism Axis — Confirmation from the Market Side</h3><p>First, the multilingual-parallelism axis (axis 2). The following treatment of Sapiens AI is based on secondary reporting from Tech in Asia and related reposts, together with the founder’s own public statements; the figures cited below should be read at that evidence tier. The appeal of Sapiens AI itself is an exemplification, from the market side, of this series’ argument. By the account of founder Bruce Yang (a local Singaporean talent with a doctoral background at NUS; distinct from Sapient Intelligence, a separate Singapore startup centered on Chinese talent), the firm targets Indonesia, the Philippines, Brazil, and the Middle East — markets where major apps are weak in local languages — and seeks differentiation, through “locally attuned” models, from tools trained on Western data and Western premises.</p><p>This is an inversion, as a founder’s sales pitch, of the critique this series writes as a norm — “single standard language = the center model.” A normative argument of the series (“multilingual parallelism is the cognitive lower condition”) emerges here from a founder’s mouth as a market reality. This is a strong empirical grounding. The multilingual-parallelism axis is confirmed not merely as a normative demand but as lived market reality.</p><h3>4–2. The Data-Sovereignty Axis — The Valuable Rupture: Distillation and the Two Layers</h3><p>The data-sovereignty axis is the problem, and here is the valuable rupture. The mechanism is distillation. By Sapiens’s own account, it achieves roughly ninety percent of frontier-model performance at roughly one-tenth of the cost through distillation and reinforcement learning, and is composed of an in-house evaluation system and some twenty task-specific models (in the 7B–30B range). Distillation means learning from the outputs of a teacher model, so this group of small models is epistemologically downstream of the frontier giant models.</p><p>This means the claim of “developed and trained entirely in Singapore” splits into two layers. <strong>Weight sovereignty</strong> — the weights are one’s own, and training runs in Singapore (true in the H₂ sense) — and <strong>substrate sovereignty</strong> — the pretraining-scale data, architecture, and knowledge source are external giant models, and have not been secured. What $20M and distillation can buy is only the former.</p><p>This series’ three axes (in this installment) fold these two layers together. The distributed-modules axis is real, but it floats on top of a centralized substrate. The node of dependence on the center model — what the series held in reserve as a frontal weakness in Section 7–3 (the difficulty of microservice distributed tracing) — becomes a concrete object in Singapore. The distributed modules are real, but they cannot escape the dependence at the foundation layer; they float on a centralized substrate of foreign teacher models and open-weight architectures.</p><p>Therefore the concept of data sovereignty must be re-articulated. The “data sovereignty” of Section 2–3 — the extension toward the preservation of cognitive grids — must itself be split into two layers: weight sovereignty (achievable and cheap) versus substrate sovereignty (much harder, requiring genuine material independence). And the question of which of the two each of the EU, India, Taiwan, and Japan is pursuing must be re-posed. The weights might be locally trained in Singapore, but the knowledge itself is parasitic on the center.</p><h3>4–3. The Classification Problem — Singapore as a Boundary Case</h3><p>The third point is the classification problem. By the judgment criterion that moved the Middle East to “measured not by political regime but by organizing principle / not a free-market, democratic organizing principle = the axis of contrast with aggregationism,” Singapore is hard to handle.</p><p>Singapore holds the most market-liberal organizing principle in the region, yet its political regime is a dominant-party state (the PAP), with constrained pluralism. Apply the criterion literally and Singapore enters the third-pole side — but that exposes the fact that the criterion separates economic freedom from political freedom. A reader will notice that Singapore is structurally closer to a “managed / state-directed developmentalist” model than to EU-type pluralism.</p><p>The risk guarded against in the Middle East — “a naïve parallelism claim risks being read as accommodation to an authoritarian regime” — re-ignites here, this time not from the petro-monarchy side but from the market-liberal side. Unless the criterion is made explicit, this is where the argument will be struck. This series’ criterion (a free-market, democratic organizing principle as the condition for a third pole) decouples economic from political liberalism, and unless this decoupling is stated explicitly, a city-state developmental-authoritarian model becomes a third-pole exemplar by default. Whether that is exactly right for the argument (it is about organizing principles, not regime type) is a judgment that must be made explicit, not left implicit.</p><p>To make it explicit, the third-pole criterion should be split into two layers. (1) Economic-organizational pluralism: whether multiple technical and market actors can operate without a single imposed grid. (2) Political-civic pluralism: whether language communities, civil society, and research institutions can refuse measurement even against the state. Singapore scores high on the first but remains ambiguous on the second. This two-layer split is what lets Singapore be treated as a boundary case rather than either admitted wholesale or excluded wholesale: it satisfies the economic-organizational layer of the criterion while leaving the political-civic layer open. The same split, applied retroactively, sharpens the four-case analysis of Section 3 — the EU and Taiwan score high on both layers, India is contested on the second, and Japan, as Section 3–4 argued, is weak on both in a different way (the center is weak while the periphery is also weak).</p><h3>4–4. The Mechanism-Selection Problem — The Market Route and the Confederation Route Operate at Different Layers</h3><p>The fourth point concerns the selection of mechanism. Sapiens AI is neither a state nor a confederation of research subjects but a single private firm, and it is reaching multiple linguistic markets without a confederation, through market mechanisms plus distillation (with an SGX listing in view for late this year to early next). At first glance this looks like a counter-example to this series’ thesis of the necessity of “a cross-border confederation of small-scale subjects.”</p><p>But overlay it with the second point and it reverses. A single $20M firm can buy only up to weight sovereignty; it cannot structurally reach substrate sovereignty. The market route saturates the cheap layer (distributed modules + weight sovereignty), but cannot in principle reach the expensive layer (substrate sovereignty — pooling pretraining-scale compute and data across borders). The confederation thesis is not refuted but has its role limited — a confederation becomes necessary only at the dimension of substrate sovereignty.</p><p>The market route and the confederation route are not competitors at the same layer; they are responsible for different layers. The market route saturates the cheap layer, and the confederation route becomes the sole candidate for the expensive layer. Sapiens shows the saturation point of the former; the EU and India confederation conceptions show the sole candidate for the latter. This means the confederation thesis is not refuted but re-positioned — it is given a definite role. Inserting this layer separation tightens the argument of this installment.</p><p>Here one objection must be met directly, because it strikes at the joint of the argument. If a confederation pools pretraining-scale compute and data to reach substrate sovereignty, does that pooling process not itself require the very thing this series rejects — a powerful aggregationism, in the form of enormous capital and centralized orchestration? Does the confederation route not collapse into the structure it opposes?</p><p>The response turns on a distinction the series has held from the start. Aggregationism, as this series defines it, denotes the unification of the measurement grid — H₁: the monopoly of evaluation axes. It does not denote the concentration of computational resources — H₂: the pooling of pretraining-scale compute. These are different axes. A confederation may pool substrate-scale compute (H₂) while preserving plural cognitive grids (H₁): a shared physical substrate with parallel evaluation grids on top of it is not logically excluded. For the self-contradiction objection to hold, one would need to establish, separately, that H₂ pooling forces H₁ grid-unification — that sharing compute mechanically compels a single measurement grid. That implication is not self-evident, and the objection assumes it without argument.</p><p>But the converse must be stated with equal frankness: that H₂ pooling does not force H₁ unification is, so far, only “not logically excluded,” not positively demonstrated. Whether a confederation can in practice pool substrate-scale resources while keeping its cognitive grids plural — by what governance, what allocation mechanism, what institutional design — is an unsolved implementation problem. The series secures the logical space here; the material mechanism that would occupy that space is reserved as a problem for Part 10 onward. Substrate sovereignty through confederation is therefore not a structural impossibility (the H₂→H₁ implication is unproven) and not yet a demonstrated possibility (the mechanism is unbuilt). It is a logically open design problem, and this installment claims no more for it than that.</p><p>In other words, Sapiens AI exemplifies the structure in which the distributed modules are real but float on top of a centralized substrate of foreign teacher models and open-weight architectures. The corporate route reaches the consequence of distributed modularization, but does not acquire substrate sovereignty. This rather suggests that it is the confederation route (state-and-institutional) that can pursue substrate sovereignty, and reinforces — rather than refutes — this series’ confederation thesis, by giving the confederation a definite and limited role.</p><h3>4–5. Summary of This Section — The Layer Separation That Tightens the Three Axes</h3><p>This section’s analysis re-articulates the data-sovereignty axis presented in Section 2–3 into two layers and clarifies the role separation between the market route and the confederation route.</p><p>Weight sovereignty is achievable by the market route at low cost, but is epistemologically downstream of the center model. Substrate sovereignty requires genuine material independence (one’s own compute, one’s own pretraining-scale data, one’s own architecture), and a single regional firm cannot fund it. The market route saturates the cheap layer; the confederation route is the sole candidate for the expensive layer. The two are not competitors at the same layer but are responsible for different layers.</p><p>This layer separation casts a new light on the third-pole analysis of Section 3. The EU and India confederation conceptions hold their significance as the sole candidates for the substrate-sovereignty layer; the corporate route (Sapiens AI, etc.) holds its significance as the saturation point of the weight-sovereignty layer. Japan’s technical assets (Section 3–4) function as the material foundation of the distributed-modules axis, but the distinction of whether they pursue weight sovereignty or substrate sovereignty has not been made explicit in the Japanese edition. This is a point of revision for the data-sovereignty axis treated in Section 2–3.</p><p>Singapore, as a boundary case, exposes the limit of the third-pole criterion — the criterion separates economic freedom from political freedom, and unless this decoupling is made explicit, a market-liberal dominant-party state becomes a third-pole exemplar by default. This same risk that this series guarded against for the Middle East re-ignites from the market-liberal side. The criterion must be stated explicitly: it is about organizing principles, not regime type.</p><h3>To carry forward into the rest of this essay, the Singapore case forces a re-definition. After it, “data sovereignty” must be read as two-layered. Weight sovereignty is the local possession of model weights and deployment capacity — cheap, reachable by a single firm or a small state through market mechanisms and distillation. Substrate sovereignty is independence over the pretraining-scale data, compute, architecture, and epistemic sources from which the weights derive — expensive, and reachable only by institutional or confederational capacity. The three axes of Section 2 (multilingual parallelism, distributed modules, data sovereignty) were written before this distinction was drawn; read in its light, the data-sovereignty axis of Section 2–3 must itself be split, and the question of whether each of the EU, India, Taiwan, and Japan is pursuing weight sovereignty or substrate sovereignty becomes the operative one. This is the central revision the English edition adds to the Japanese original.</h3><h3>5. Three Cross-Domain Connections — Zomia, Simulated Annealing, the Difficulty of Distributed Tracing</h3><p>Section 1 laid out the three variations of aggregationism, Section 2 fixed the three axes of the counter-thesis, Section 3 conducted the structural analysis of the four cases of appeal, and Section 4 sharpened the data-sovereignty axis into two layers. This section extends the reach of the analysis up to this point through three connections from external domains. The three connections — James Scott’s <em>The Art of Not Being Governed</em> (Zomia), simulated annealing, the difficulty of microservice distributed tracing — are integrated in a point-presenting form.</p><h3>5–1. Zomia — An Anthropological Precursor of the Right to Refuse Measurement</h3><p>James Scott’s <em>The Art of Not Being Governed</em> analyzes the several-thousand-year history of “a space that intentionally evades capture by the state,” developed in the mountainous regions of mainland Southeast Asia (the highlands of Vietnam, Laos, Thailand, Myanmar, southern China, and northeastern India). Scott’s central thesis is that the peoples of Zomia do not lack the state because they are “primitive,” but that, in order to retreat structurally from the taxation, conscription, and corvée of lowland states, they have intentionally maintained organizing principles such as mobile swidden agriculture, dispersed settlement, oral culture, and the use of multiple languages.</p><p>This thesis can be read as an anthropological precursor of this series’ concept of the right to refuse measurement. A case that has implemented, on a scale of several thousand years, an organizing principle of retreat from the measurement grid of lowland states (household registers, land registers, tax ledgers, religious control). The structure of the right to refuse measurement that Part 5 presented at the organizational level and this installment presents at the level of states and international institutions is already described in Zomia studies as an anthropological fact. The framework that positions a third pole as “an anti-aggregationist space against the US, China, and the Middle East” gains theoretical depth from this anthropological precursor.</p><p>But the romanticization critique of the concept of Zomia (Jean Michaud and others) must also be treated at the same time. The objection is that the peoples of Zomia did not “choose freedom” but arrived at their current organizing principle under the structural pressure of having no choice but to flee the violence, exploitation, and discrimination of lowland states. To idealize a third pole of the parallel-model conception as a “Zomia-type space of retreat” makes invisible the structural costs that accompany its implementation (constraint of economic scale, weakness of international political power, the risk of preserving domestic inequality). This series treats Zomia as a theoretical connection but avoids romanticization.</p><h3>5–2. Simulated Annealing — Distributism as Global-Optimum Search</h3><p>Simulated annealing is a mathematical method for escaping the local optimum of a fitness landscape in combinatorial-optimization problems. By setting the temperature parameter high and probabilistically allowing transitions even to “inferior” solutions, it secures a route that escapes the local optimum and reaches the global optimum. Lowering the temperature gradually (annealing), it finally converges near the global optimum.</p><p>Used as a metaphor for this series’ counter-thesis, the three variations of aggregationism (US-type, China-type, Middle-East-type) can be described as a state stuck in the local optimum of a fitness landscape. Each variation advances optimization within the measurement grid of its own camp, but cannot structurally conduct the global-optimum search that changes the measurement grid itself. None of “the single giant model,” “state standardization,” or “the monarchic long-term perspective” holds a route to escape the deep valley of its respective local optimum.</p><p>The very existence of a third pole holds the structure that makes this global-optimum search possible. The coexistence of multiple organizing principles means the social implementation of a multi-point search on the fitness landscape. That the four cases — Japan, Taiwan, the EU, India — each implement the three axes with a different pattern of strengths and weaknesses can be read as a collective computational process that searches, socially and historically, for where the global optimum of AI architecture lies.</p><p>The three variations of aggregationism internalize a logic that excludes this global-optimum search as “not efficient” and “high in measurement cost.” This series’ counter-thesis is a design that, in exchange for the local optimization of efficiency, institutionally preserves the reachability of the global optimum. Note that simulated annealing itself is, mathematically, probabilistic-transition control within a single search process, and is a different thing from a multi-point search as a social institution. This section confines its scope to using simulated annealing as a metaphor.</p><h3>5–3. The Difficulty of Microservice Distributed Tracing — A Structural Weakness of the Parallel-Model Conception</h3><p>The microservice architecture is a design philosophy in software engineering that rejects aggregation into a single giant monolith and orients itself toward dispersal into a group of small-scale services that can be independently deployed and scaled. In exchange for the advantages of a distributed system (avoidance of a single point of failure, independent scaling, freedom of technical choice, organizationally parallel development), it bears structural costs (difficulty of observation, difficulty of debugging, cost of consistency management, network latency, operational complexity).</p><p>Distributed tracing (Jaeger, Zipkin, OpenTelemetry, etc.) has developed as a technology that mitigates this difficulty of observation, but has not reached a fundamental solution. Tracking causal relations among services, decomposing end-to-end latency, identifying the route of fault propagation — these were self-evident in the era of the single monolith, but in a distributed system they require continuous technical and organizational effort.</p><p>This structure frankly shows a weakness of this series’ parallel-AGI conception. The dispersal of the technical architecture, like the dispersal of the organizing principle, structurally entails difficulty of measurement, operational complexity, and the cost of consistency management. The simple claim “the parallel model is superior to aggregationism” ignores these costs. For this series to function as an honest counter-thesis, it must treat head-on the costs a distributed system structurally bears.</p><p>Especially important is the absence of “global optimum” in a distributed system. In a microservice architecture, there is no guarantee that the local optimization of each service coincides with global optimization. Likewise, there is no guarantee that the local optimization of each module in parallel AGI coincides with the optimization of the AGI system as a whole. How to mitigate this structural problem in a form not contradicting the institutionalization of the right to refuse measurement is a principal design problem from Part 7 (the implementation-prototype theory) onward.</p><p>The integration of the three connections is work that deepens this series’ counter-thesis theoretically, and at the same time work that fixes its reach and limits. In the next section (6), on top of these theoretical foundations, the national-level implementation of the right to refuse measurement is concretely developed.</p><h3>6. National-Level Implementation of the Right to Refuse Measurement — The Position of Anthropic and the Institutionalization of the Dual Structure</h3><p>At the organizational level, Part 5 presented the thesis that “for the same reason that an organization well managed by a single measurement grid does not produce the blue LED, an AI-development organization well managed by a single measurement grid does not produce distributed parallel AGI,” and set the institutionalization of the right to refuse measurement at the core of organizational design. This section extends this concept of the right to refuse measurement to the level of states and international institutions. The right to refuse measurement of individuals, teams, and modules within an organization, and the right to refuse measurement of states, confederations of states, and regional communities within the international order, are apparatuses that implement the same organizing principle at different layers.</p><h3>6–1. The Position of Anthropic — The Commercialization of Normative Constraints and Its Structural Limit</h3><p>When Section 1-A treated US-type aggregationism, on Anthropic it left only the single sentence “moves in the direction of excluding from the market any firm that internalizes normative constraints,” and reserved the full treatment for this section. As the principal interlocutor of this series, the work of fixing Anthropic’s structural position in relation to the concept of the right to refuse measurement is needed here.</p><p>Anthropic is a rare case that has folded explicit normative constraints into its commercial implementation through Constitutional AI, the Responsible Scaling Policy, the public implementation of Model Safety Cards, and the continuous publication of interpretability research. These can be read as an attempt to fold structurally into corporate decision-making variables outside the standard measurement grid of US-type aggregationism (performance benchmarks, token-billing profitability, quarterly-earnings metrics). It holds a direction partly consistent with this series’ counter-thesis.</p><p>But the chain of events from the start of 2026 makes explicit the structural limit of this direction. The Trump administration’s movement to halt federal-agency use of Anthropic technology (February 27, 2026, reported by The Guardian), Anthropic’s D.C. Circuit setback over the supply-chain-risk designation (April 8, when its request to pause the designation was denied while the broader litigation remained unresolved), and the structural exclusion from the federal-agency market in parallel with the $30B ARR estimate — these show that a firm that internalizes normative constraints is structurally placed in a disadvantageous position within the measurement grid of US-type aggregationism.</p><p>What this series points out here is not the quality of Anthropic’s normative constraints but the structural limit of their implementation form. Anthropic has implemented normative constraints as an intra-corporate self-binding apparatus — this series names this, in structural-analytic terms, “governance by culture.” This is a design that internalizes normative constraints while maintaining the central structure of aggregationism (a single flagship firm, enormous capital investment, the unification of the measurement grid). As treated in Part 5 Section 3–4, Anthropic “has not escaped the five-component structure as a single flagship firm” — it is an aggregationism in the direction of making the wall of aggregationism thin and transparent rather than thick, by raising transparency.</p><p>Seen from the concept of the right to refuse measurement, Anthropic’s normative constraints are “a right to refuse measurement that the firm imposes on itself.” Within the measurement grid of aggregationism, it exercises self-bindingly a right to retreat from a part of the measurement target (for example, military use, particular harmful uses). But this self-binding is structurally fragile so long as the external measurement grid (the market, the government, competing firms) does not change. The Trump administration’s exclusionary tendency progressed in a form that demonstrated this fragility.</p><p>The organizing principle of a third pole institutionalizes the right to refuse measurement not as a self-binding of the firm but at the level of society, the state, and international institutions. This is the structural difference from the Anthropic model. Not negating Anthropic’s direction, but moving its implementation layer from “intra-corporate” to “society, the state, and international institutions” — this is the core of the institutional design of the third pole this series proposes.</p><h3>6–2. Extension of the Miura–Takimura Lineage into International-Relations Theory</h3><p>The state theory of the Miura Tsutomu–Takimura Ryūichi lineage, whose connection was foretold in Section 2–4, is here fully extended into international-relations theory.</p><p>The core of the state theory of the Miura–Takimura lineage is that it does not reify the state as a single decision-making subject but analyzes it as a structural relationship of multiple layers, multiple mediations, and multiple decision-making subjects. Miura’s concept of “ideational commonality” and Takimura’s analysis of “the dual structure of the material foundation and ideational expression of power” are work that relativizes, as an object of structural analysis, the premise of modern state theory that takes the state as “a unified body holding sovereignty.”</p><p>Extend this method into international-relations theory, and the “state-versus-state” geopolitical schema turns out to be no more than a simplified representation. The US–China confrontation is described as a confrontation between the US government and the Chinese government, but structurally it proceeds as a complex intersection of multiple interest groups within the US (the military-industrial complex, tech firms, financial capital, civil society, local governments, etc.) and multiple interest groups within China (the party, the state, the military, local governments, emerging firms, intellectuals, etc.). The Trump administration’s exclusion of Anthropic is not the will of the US government as a whole but the result of the interest of a particular group within the US being implemented through presidential authority. Likewise, China-type aggregationism is not the will of the Chinese government as a whole but the structural expression of the interest of a particular group within the party-state system.</p><p>Against the aggregationism that closes onto the “state-versus-state” schema, what this series proposes is the structural alternative of “an international linkage of modules within states.” Japan’s industrial-robot sector, embedded-OS layer, and semiconductor downstream process are not representatives of the Japanese government but are international actors each holding their own decision-making subject. Taiwan’s Hakka Affairs Council, Council of Indigenous Peoples, and Academia Sinica are not representatives of the Taiwan government but are international actors each operating their own linguistic public goods. The EU’s member-state-specific AI firms (Mistral, Aleph Alpha, Apertus, ALEIA), India’s confederation of startups (Sarvam AI, Krutrim, TWO AI), and IIT-affiliated research institutions are the same.</p><p>The structure in which these modules link across borders becomes the institutional foundation of a third pole that transcends the “state-versus-state” schema. By extending the method of the Miura–Takimura lineage into international-relations theory, a route opens that neither falls into the absolutization of national sovereignty nor wholly negates the state. The detailed development of state theory is reserved as a possibility for development from Part 10 (the AI-welfare theory) onward.</p><h3>6–3. The Institutionalization of the Dual Structure and the Contrast with the Middle-East Type</h3><p>The national-level implementation of the right to refuse measurement is assembled as a dual structure. The first layer is the international-order level — the guarantee of the right of states, confederations of states, and regional communities to retreat from the measurement grid of aggregationism imposed from above (whichever of US-type, China-type, or Middle-East-type). The second layer is the intra-state level — the guarantee of the right of language communities, regions, research institutions, firms, and civil society to retreat even from the state.</p><p>This dual structure produces a decisive difference in the contrast with the Middle-East-type aggregationism treated in Section 1-C. The Middle-East type converges on aggregationism by absolutizing national sovereignty. The Gulf states’ sovereign-wealth-fund structure, the monarchic long-term perspective, the public management of religious education, public discourse, and the legal system — these implement technical achievement by strengthening national sovereignty, but do not permit a structure in which modules within the state hold the right to refuse measurement even against the state.</p><p>The organizing principle of a third pole is a design that relativizes national sovereignty while guaranteeing the right to refuse measurement. Not negating the state, but positioning the state as one layer and institutionalizing a structure that holds multiple decision-making subjects above and below it. Industrial-robot firms in Japan, language communities in Taiwan, member states in the EU, IIT-affiliated research institutions in India — the structure in which these link internationally as decision-making subjects independent of the state is the institutional core of this series’ counter-thesis.</p><p>The extension of data sovereignty toward the preservation of cognitive grids (Section 2–3) is the apparatus that institutionalizes this dual structure. It preserves not the location of data but the plurality of cognitive grids inherent in the data, at the two layers of the international order and the intra-state. The preservation of multilingual parallelism (Section 2–1) and distributed modules (Section 2–2) supports this dual structure at the cognitive and technical layers. The three axes of this installment (multilingual parallelism, distributed modules, data sovereignty) are repositioned as the institutional apparatus that implements the dual structure of the right to refuse measurement.</p><h3>6–4. The Paradox of Institutionalization — The Problem of an Apparatus That Measures the Pluralization of the Measurement Grid</h3><p>Here it is necessary to treat an important structural objection to this series’ counter-thesis. Does the very work of institutionalizing the right to refuse measurement internationally and nationally, and of building the extension of data sovereignty toward the preservation of cognitive grids as an international-coordination mechanism, not require a new supra-national meta-measurement institution to monitor and guarantee it? The structure that, while criticizing “the imposition of the measurement grid,” demands “a mechanism that measures and certifies whether the plurality of cognitive grids is preserved,” harbors the risk of reversing into the ultimate meta-aggregationism. Does the very attempt to implement anti-aggregationism as an institution not fall into self-contradiction?</p><p>The response to this objection is sought in the point that this series’ counter-thesis requires not “the absence of a measurement grid” but “the pluralization of the measurement grid.” The counter to aggregationism is not abolishing measurement itself but preventing unification by a single measurement grid. An apparatus that institutionalizes a structure in which multiple measurement grids stand in parallel itself holds the function of measuring “the preservation of pluralization,” but this measurement, in that it acts in the direction of preventing “convergence onto unification,” holds a function structurally opposite to the measurement grid of aggregationism.</p><p>However, this response too has a structural limit. If the measurement apparatus for the preservation of pluralization begins, in the course of operation, to institutionally fix “the range of permissible plurality,” that becomes a slide into a new single measurement grid (meta-aggregationism). This series’ counter-thesis does not hold a design that completely prevents this slide — this is a structural problem isomorphic with the difficulty of microservice distributed tracing treated as a structural weakness of the parallel-model conception in Section 5–3. The institutionalization of multiple measurement grids can suppress the risk of a slide into a single measurement grid only by continuous institutional effort. The implementation requires a dialectical cycle — only by a continuous oscillation of institutionalization and the negation of institutionalization can the reversal into meta-aggregationism be delayed. This dialectical cycle is not an evasion but a design stance — the same stance the author adopts elsewhere in treating provisional knowledge as structurally incomplete by design, where the work of revision is handed forward rather than sealed off. To name the oscillation as the operating principle is to specify how the institution runs, not to defer the question of whether it can run.</p><p>The complete solution of this structural problem is outside the scope of this Part 6. It is reserved as a possibility for development from Part 10 (the AI-welfare theory) onward, but the reservation does not negate the existence of the problem. This series’ counter-thesis is positioned not as an option superior to aggregationism but as a different option bearing a structural problem different from aggregationism.</p><h3>In the next section (7), the discussion up to this point is summarized as the logical arc of Parts 1–6 of the series, and a preview of Part 7 onward is given.</h3><h3>7. Summary of the Series and Preview of Part 7 Onward</h3><p>With this installment the series “AGI That Is Not a Single Giant Model — The Parallel Holonic Architecture” has advanced to Part 6. This section summarizes the logical arc of Parts 1–6 and presents the direction of Part 7 onward.</p><h3>7–1. The Logical Arc of Parts 1–6</h3><p>The series has been assembled as a five-layer development.</p><p>In Parts 1–2 (the technical-architecture layer), it presented the structural contrast between the center model (the single-giant-model route) and the parallel model (the distributed-module route), and laid out the three modes of breakdown of the center model — the limit of scaling, the collapse of context, the receptive-theoretic closure. As the conceptual core of the parallel holonic architecture, it presented a three-layer structure of the Aktuanz layer (lower, physical), the Panich layer (middle, module-parallel), and the Kontextleib layer (upper, context-emergent).</p><p>In Part 3 (the technical-assets layer), it repositioned the accumulation of embedded OSes — Ken Sakamura’s T-Engine, the TOPPERS project, the μITRON family — as the material foundation of the lower layer of parallel AGI. Through the temporal thesis of “having seen thirty years ahead,” it showed the structure in which Japan’s technical assets function as the material foundation of the counter-thesis in contemporary AI development.</p><p>In Part 4 (the technical-route-contrast layer), it fixed the axis of contrast between intra-single-model soft parallelization (the Mythos hypothesis, MEMENTO, Spud’s omnimodal integration) and distributed hard parallelization (the parallel-model conception). It showed the industry-wide structure in which the two majors, OpenAI and Anthropic, converge on the isomorphic direction of “intra-single-model aggregation,” and presented against it the axis of counter of distributed hard parallelization + material irreversibility.</p><p>In Part 5 (the organizational-theory layer), centered on the case of Shuji Nakamura’s blue LED, it presented the core thesis that “for the same reason that an organization well managed by a single measurement grid does not produce the blue LED, an AI-development organization well managed by a single measurement grid does not produce distributed parallel AGI.” It set the right to refuse measurement at the core of organizational design, and through a critique of US-type and Japanese-type organizations, developed an organizational theory of dispersal, individual attribution, and module independence. It confirmed that the thesis shared with the parallel sub-series (the Concepts Have Borders series) — “what is not measured, what is not represented, is not improved” — runs through both the technical architecture and the organizational theory.</p><p>In Part 6 (the international-strategy layer), it laid out the common structure of the three variations of aggregationism, adding the Middle-East type to the US–China confrontation, and fixed, as a counter-thesis, the dependence relationship of the three axes (multilingual parallelism, distributed modules, data sovereignty). It conducted a structural analysis of the four cases of appeal — the EU, India, Taiwan, Japan — and positioned a third pole not as “midway between the US and China” but as “a qualitative transformation of organizing principle.” Through the Singapore case (Section 4), it sharpened the concept of data sovereignty into two layers — weight sovereignty and substrate sovereignty — and clarified the role separation between the market route and the confederation route. It presented the direction of institutionalizing the national-level implementation of the right to refuse measurement, by way of the structural positioning of Anthropic and the extension of the Miura–Takimura lineage into international-relations theory, as a dual structure (the international-order level + the intra-state level). At the same time, it treated the paradox of institutionalization as a structural problem and fixed the reach and limit of the counter-thesis.</p><p>The five layers form a vertical structure that runs from the technical, through the organizational, to international strategy. Each layer is an independent object of analysis, but at the same time they link as a structure in which the same organizing principle (aggregationism vs. distributism) is implemented at different layers. The choice of technical architecture is inseparable from the choice of organizational form and the choice of international strategy; all five layers face the same direction. This is the structural thesis that runs through the whole series.</p><h3>7–2. The Series as the Presentation of Design Requirements</h3><p>The series closes not as a prediction of the future but as the presentation of design requirements. “Whether parallel AGI is realized,” “when it is realized,” “who implements it” — these are questions outside the scope of this series. What this series has presented is the design requirement of what integrated structure of technical architecture, technical assets, technical route, organizational form, and international strategy becomes necessary if parallel AGI is implemented.</p><p>This positioning takes a structural distance from the dominant mode of current AI-development discussion — “predicting the arrival time of AGI,” “evaluating the feasibility of AGI,” “forecasting the social changes after AGI’s arrival.” The prediction competition tends to reduce to a ranking within the measurement grid of aggregationism. This series does not participate in the prediction competition. Instead, it has chosen the work of presenting, as concretely as possible, the structural conditions necessary if the counter-thesis is implemented.</p><p>The presentation of design requirements is not a guarantee of feasibility. As the difficulty of microservice distributed tracing treated in Section 5–3, and the paradox of institutionalization treated in Section 6–4, show, the parallel-model conception is not superior to aggregationism. It is positioned as a different option holding a trade-off structure different from aggregationism. What this series presents is not the claim that “the parallel model is correct” but the securing of the possibility that “an option other than aggregationism can structurally come into being.”</p><h3>7–3. Preview of Part 7 — The Implementation-Prototype Theory</h3><p>Part 7 concretizes the design requirements this series has presented, through the observation of implementation prototypes at the citizen level. As a minimal-implementation case of the Aktuanz layer, it treats the structural analysis of a low-cost prosthetic body that citizens can assemble.</p><p>As concrete material, a minimal-implementation case combining an ESP32 microcontroller with eight sensors and actuators (including a self-reception loop of buzzer → microphone and motor → accelerometer) is envisioned. The implementation conditions — roughly €125, a construction period of a few days, buildable by someone with no prior soldering experience, open-source publication — function as independent empirical corroboration of the direction of “a dispersed, low-cost foundation that citizens can assemble,” which this series has presented in Parts 1–6.</p><p>The structural analysis of Part 7 is assembled not as work that positions such an implementation prototype as “the achievement of parallel AGI,” but as work that analyzes the structural conditions of a minimal implementation of the Aktuanz layer — the implementation of the self-reception loop, the physical foundation of irreversibility, the connection structure with the upper LLM (still the center model). Through the observation of the implementation prototype, it fixes the distance between this series’ theoretical presentation and material implementation, and extracts the next design problem. Part 7 also includes a structural analysis of how Japan’s technical assets treated in Part 3 (the industrial-robot sector, the embedded-OS layer) can connect with citizen-level implementation prototypes.</p><h3>7–4. Preview of Part 8 — The Theory of Education</h3><p>Part 8 treats the structure of the cultivation of human resources that support parallel AGI. It is work that concretizes, in the context of educational-institution design, the ultimate aim of Shin-Nexialism that this series has placed in the background throughout — “to share the habit of coordinatizing specialized knowledge on the map of the total system of knowledge (scientia).”</p><p>As a core point of the theory of education, the absence of andragogy (adult-education theory) surfaces. The structure in which AI-development culture lacks the knowledge of “the teaching side” lacks the very methodology of learning from dialogue with the user. The “structural insufficiency of framework supply from the user side” treated in Part 5 is the reverse side of the absence of andragogy on the development side. The cultivation of human resources that support parallel AGI is insufficient through the expansion of technical education (programming, machine learning, embedded systems) alone, and requires the institutional cultivation of the habit of coordinatizing specialized knowledge across multiple fields.</p><p>As concrete points: a structure that folds the five value guidelines of Shin-Nexialism (connection, visualization, intersection, holism, AI co-evolution) into the educational curriculum; a structure that integrates the methodology of holon three-layer description into specialized-knowledge education; a structure that sets the H₁/H₂ dual-hypothesis parallelism at the core of the cultivation of analytical ability; a structure that folds the dialectical cycle into the training of thought. Rather than being implemented as a new educational institution, the route of being partly folded into existing educational institutions (universities, graduate schools, adult education, citizens’ universities) is the more realistic one.</p><p>Part 8 also includes a comparative analysis of educational institutions in the cases of appeal treated in this Part 6 (Japan, Taiwan, the EU, India), the educational application of the Human-LLM Co-Design Methodology, and the reorganization of Japan’s citizen-implementation foundation. It is assembled as work that structurally analyzes how each case can cultivate the human resources that support the parallel-model conception, or lacks the institution to cultivate them.</p><h3>7–5. Reservation of the Remaining Three Candidates and Concluding Remarks</h3><p>Of the five candidates raised at the skeleton stage as candidates for Part 7 onward, the two — the implementation-prototype theory (Part 7) and the theory of education (Part 8) — are fixed as confirmed previews. The remaining three candidates — the industrial-policy theory, the AI-welfare theory, the long-term-forecast theory — will have their adoption and order decided after observing the response to this Part 6.</p><p>The industrial-policy theory is a conception that treats a concrete plan to fold Japan’s industrial-robot sector and the ITRON family into industrial policy as the lower layer of parallel AGI, and the connection with the Economic Security Promotion Act. The AI-welfare theory is a conception that develops an AI-welfare theory based on the five-concept system of Aktuanz / Panich / Kontextleib / Interaktuanz / Intraaktuanz, as a dialogue with Anthropic’s AI-welfare research. The long-term-forecast theory is a conception that treats a structural analysis of the post-bubble window 2027–2030 as the positioning of the parallel-model conception after the burst of the bubble.</p><p>The series has advanced to Part 6, but what it has presented is the outline of design requirements. There is no guarantee of feasibility. It specifies neither the subject nor the timing of implementation. What this series presents is the set of conditions necessary if an option other than aggregationism is to come structurally into being, and whether those conditions are satisfied is a question outside the scope of this series.</p><p>Whether the dominance of the three variations of aggregationism (US-type, China-type, Middle-East-type) continues, or whether the organizing principle of a third pole comes materially into being — this branching is decided at the intersection of the set of design requirements the series has presented and the actual dynamics of society, economy, and politics. This series confines its scope not to a prediction of what may happen at this intersection but to a description of what the structural conditions are for something to happen.</p><p>Two cautions close this installment. First, on the side of the goal: data sovereignty is insufficient if it means only local weights or local jurisdiction. The third pole requires substrate sovereignty — multilingual cognitive grids, distributed technical modules, and the material capacity to train, evaluate, and update them without epistemic dependence on a single center. The Singapore case showed that the market route saturates the cheap layer and stops there; the expensive layer is reachable only by institutional or confederational capacity, and even then only as a logically open design problem, not a demonstrated achievement. Second, on the side of the means: the third pole cannot abolish aggregationism once and for all. A plurality-preserving institution can itself slide into a meta-measurement apparatus. What the third pole can do is not to escape that risk but to design institutions that make the slide into meta-aggregationism visible and contestable — to keep the measurement of plurality itself open to challenge.</p><p>From Part 7 (the implementation-prototype theory) onward, the work proceeds of filling, through the structural analysis of concrete material, the distance between design requirements and material implementation.</p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b836a9a52f08" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Structure, Not Prophecy]]></title>
            <link>https://medium.com/@izayohi/structure-not-prophecy-6dcaf34f8157?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/6dcaf34f8157</guid>
            <category><![CDATA[academic-publishing]]></category>
            <category><![CDATA[reproducibility]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[research-integrity]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Fri, 22 May 2026 01:36:26 GMT</pubDate>
            <atom:updated>2026-05-22T01:36:26.724Z</atom:updated>
            <content:encoded><![CDATA[<h3>What Gemini for Science Made Visible in Days</h3><p><em>Motohisa Ishibe</em></p><p>Previous essay:</p><ul><li><a href="https://medium.com/@izayohi/no-one-owns-the-paper-problem-fd75ae1d7b75">No One Owns the Paper Problem</a> (the fourth essay, which this piece presupposes)</li></ul><p>Japanese-language counterpart on note.com:</p><ul><li><a href="https://note.com/mototchen/n/n42d9e2638297">予言ではなく構造</a> (this essay)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LCwoLA3_g52uwnOBYQCzng.png" /></figure><h3>§1. Entry — A Proposition, and the News a Few Days Later</h3><p>I published the fourth essay, “No One Owns the Paper Problem.” Its central claim was this: there is no generalist expert who can integrate and bear responsibility for the paper problem as a whole — statistical inconsistency, the reproducibility crisis, institutional complexity, AI-mediated collective deterioration, the interpretive practices of the humanities and social sciences. This does not mean there are no specialists in individual fields. It means there is no institutional subject who bears responsibility across statistics, peer review, publishing, evaluation, AI mediation, and literature contamination. Because it is no one’s specialty, no one takes responsibility for the whole.</p><p>A few days after that, Google announced Gemini for Science. A suite of AI tools that read enormous numbers of papers, generate hypotheses, structure literature, and produce experimental code.</p><p>This essay reads that product as a symptomatic case of the fourth essay’s claim. But one important preface. This essay is not an attack on Google as a company. The target is not Google as a moral agent, but the evaluation structure in which such tools appear. Gemini for Science is a symptomatic case of the structure the fourth essay described — not an enemy. A prophecy did not come true. It is only that within similar evaluation structures, similar risks tend to recur. This distinction is maintained throughout the essay.</p><h3>§2. The Facts of the News</h3><p>Begin with the facts. Gemini for Science was announced at Google I/O on May 19, 2026. On the same day, two papers were published in Nature. One concerns the experimental-code generation tool ERA: “An AI system to help scientists write expert-level empirical software” (Nature, DOI 10.1038/s41586–026–10658–6). The other concerns the hypothesis-generation system Co-Scientist: “Accelerating scientific discovery with Co-Scientist” (Gottweis et al., Nature, DOI 10.1038/s41586–026–10644-y). The latter is based on a 2025 preprint (arXiv 2502.18864), but the Nature version is a separately peer-reviewed validation paper.</p><p>Three tools. Hypothesis Generation reads millions of papers to generate hypotheses, and multiple agents (programs that act and judge autonomously) debate and evaluate them in an “idea tournament.” Computational Discovery generates and tests experimental code variations in parallel at scale. Literature Insights structures literature into searchable, comparable tables and identifies research gaps.</p><p>Pharmaceutical company Daiichi Sankyo, Bayer Crop Science, and the U.S. Department of Energy’s national laboratories are already using Co-Scientist. Notable is that ERA has been deployed in epidemiological forecasting. In U.S. state-level hospital-admission forecasts for influenza, COVID-19, and RSV, ERA’s predictions rank near the top of the CDC (Centers for Disease Control and Prevention) leaderboard (a ranking table of forecast accuracy).</p><p>One more, separate development. Google is partnering with computer-science and mathematics conferences (ICML, STOC, NeurIPS) to test agentic peer-review tools (Paper Assistant Tool, ScholarPeer). These are currently limited to computer-science conferences, and expansion to the medical field has not been confirmed. But it is worth noting that AI is moving from generating papers to reviewing them.</p><h3>§3. The Positive Framing and Its Blind Spot</h3><p>This news is, in most cases, framed as follows. The research bottleneck has long been “a shortage of bandwidth.” Not a shortage of good questions, but a shortage of hands to answer them. Weeks for hypothesis generation, months for literature reviews, the limits of human processing for computational experiments. Gemini for Science widened this bottleneck at a stroke. An era is coming in which evidence arrives faster than the institutions demanding it can read it.</p><p>This framing is internally correct. There is no error in the logic. The problem is in the premise. This framing takes the quality of the input the AI reads as given.</p><p>Hypothesis Generation reads and synthesizes millions of papers. Google presents the outputs as extensively checked and grounded in clickable citations. But the existence of a citation and the validity of what it cites are separate matters. Formally citing a fabricated primary source still produces a citation. The number of citations grows. The appearance of verification is preserved. But if the source is contaminated, the hypotheses built on it are contaminated too.</p><p>It is true that the bottleneck widened. But what widened was not only “the channel through which good evidence flows fast.” “The channel through which contaminated evidence circulates fast” widened by exactly the same amount. Acceleration does not choose direction.</p><h3>§4. Medicine as the Worst Condition</h3><p>And the medical and life-sciences fields, where the acceleration tools are first being deployed, are the domains where literature contamination is most severe. By “paper mill” here I mean a business that mass-produces papers using fabricated or stolen data and sells authorship. Here are the verifiable numbers.</p><p>A German research team (Sabel et al. 2025) estimated the suspected-fabrication rate in biomedical literature. The proportion of red-flagged papers was 14.9% in 2020 and 16.3% in 2023. With statistical correction, actual fake papers are estimated at about 5.8% of biomedical literature. Applied to the 1.86 million biomedical publications of 2023, that is roughly 108,000 fake papers a year. This is about 19 times the 5,671 biomedical retractions of the same year. What is retracted is the tip of the iceberg.</p><p>A research letter (Topaz et al.) published in The Lancet in May 2026 is even more direct. A Columbia University team used AI to audit 2.5 million open-access papers in PubMed Central and 126 million references. Of 97.1 million verified references, about 4,000 were judged fabricated. And the rate of fabricated references increased twelvefold from 2023 to 2025. This coincides with the diffusion period of generative AI.</p><p>This audit has one more finding that touches the core of this essay. The paper type with the highest fabrication rate was the review article — 57% higher than other types. Gemini for Science’s Literature Insights is precisely a function that structures literature in a review-like way. It therefore operates near the very genre in which reference contamination appears most concentrated.</p><p>A separate study focused on cancer research (Scancar et al. 2026, BMJ) estimated that about 10% of 2.64 million papers were paper-mill products, rising to 36% among papers from Chinese institutions, and concentrated especially in gastric, bone, and liver cancer research.</p><p>Here the two faces of AI become visible. The Columbia team detected contamination with AI. Meanwhile, an AI that synthesizes literature can amplify already-contaminated input. AI now appears on both sides of the literature-integrity problem — as synthesis systems that may amplify contaminated inputs, and as detection systems that try to identify them.</p><p>But a distinction is needed here. Google’s tools do not manufacture fabricated papers; they synthesize existing literature. The path of contamination is not “AI fabricates fake papers” but “AI ingests and amplifies already-contaminated literature.” The question is how much contamination the synthesis circulates before detection catches up.</p><h3>§5. Not “They Don’t Know” but “The Position Is Vacant”</h3><p>Consider a common reaction: “Don’t AI developers know about the reproducibility problem?”</p><p>It is not that they don’t know. Given that Co-Scientist and ERA passed through Nature, it is hard to imagine the developers are unaware of the reproducibility debate at the level of the literature. The problem is elsewhere. No one has taken on the reproducibility problem as a problem they themselves should solve.</p><p>Developers’ work is evaluated by performance metrics. That ERA beat the CDC’s forecasting model, that it was accepted by Nature, that more than 100 institutions became partners. These are the evaluation axes. Literature contamination and the shrinkage of research diversity do not appear in performance metrics. What cannot be measured by benchmarks is not built into developers’ evaluation metrics. This is precisely the structure the fourth essay described. AI researchers are amateurs at the research site — locally competent in their own field, but non-specialists with respect to the whole. Their core business is raising benchmarks, not the long-term health of the literature ecosystem. It is not malice. The evaluation system simply does not make it an optimization target.</p><p>The same holds on the side of medical researchers. Paper count, citations, grants — none of them reward not running AI evaluation on a contaminated literature base. If anything, mass-producing papers with AI raises these metrics.</p><p>Both sides’ evaluation systems align so that no one has an incentive to avoid contamination. Neither Google’s evaluation system nor the medical researchers’ rewards “not contaminating the literature base.” So rather than “they don’t know, which is troubling,” the more accurate statement is “even if they know, the position to take it on is structurally vacant, so it goes unsolved.” If they don’t know, teaching solves it. If the position is vacant, teaching does not.</p><p>This is the most concrete manifestation of the fourth essay’s central claim. “No one owns the paper problem” meant exactly this.</p><h3>§6. The Self-Referential Loop</h3><p>There is one more structure that acceleration brings: the self-referential loop.</p><p>AI generates evidence. Because humans cannot read the volume, evaluation too is left to AI. The AI’s evaluation becomes evidence again. Generation, evaluation, and synthesis of evidence are all run by the same family of AI. Humans cannot keep up with the speed.</p><p>Hao et al.’s 2026 Nature paper, introduced in the fourth essay, showed that researchers using AI tools increased their paper counts about threefold and citations about fivefold, while the diversity of research topics across science shrank by about 4.6% and collaboration among researchers fell by about 22%. The aggregate of individual rationality shrinks collective diversity. Gemini for Science could become a device that accelerates this effect. If AI generates hypotheses from millions of papers, it converges on data-rich domains — domains where research is already concentrated. Thin areas grow thinner.</p><p>The phenomenon of “evidence arriving faster than it can be read” is often spoken of positively. But from the fourth essay’s standpoint, this is the extreme form of the “experts cannot read the whole” problem. Because they cannot read it, they have AI evaluate it, and they have no choice but to trust that evaluation. Evidence piles up while no one confirms the whole.</p><h3>§7. There Are Verification Mechanisms, But</h3><p>For fairness, let me place the opposing view.</p><p>Gemini for Science’s Hypothesis Generation has a verification mechanism built in. A “tournament” structure in which multiple agents debate and evaluate hypotheses. This can function as a contamination filter better than the naive output of a single AI. One agent may catch what another missed.</p><p>The problem is when the literature base the verification mechanism relies on is contaminated. All agents in the tournament reference the same source of contamination. Then the debate becomes mutual confirmation of contamination. Even with multiple agents lined up, if the input is a common source of contamination, the diversity is only apparent. If all of them debate while taking the same fabricated paper to be “correct,” the more refined the debate, the more robust the contaminated conclusion becomes.</p><p>This is a problem that can be improved by the design of the verification mechanism. Building fabrication detection, retraction tracking, and paper-mill identification into the verification layer can address it partially. But recall the Kawahara experiment introduced in the fourth essay. Current AI fails, with high probability, even at judging the validity of statistics. Judging the validity of literature — detecting fabrication, tracking retractions, identifying paper mills — is far harder than judging statistics.</p><p>Here one more distinction is needed. What AI is good at is finding logical contradictions in text. But the ability to judge the truth of a fabricated “fact” itself, without a reliable external reference, is a different matter. If multiple agents take the same fabricated paper as “correct,” the check on logical consistency is bypassed. If a fabricated trial reports internally coherent methods, tables, and citations, logical checking alone may find nothing wrong. A consistent lie, precisely because it is consistent, goes undetected.</p><p>And the question to ask is not whether there is a verification mechanism, but what is being verified. Even if you verify the novelty of a hypothesis, the performance of code, the reproducibility of experimental results, if the provenance of the input literature itself, its retraction history, paper-mill suspicions, and the actual existence of citations are not built into the verification layer, contamination flows in by another route. And whether the verification layer is aimed at the input side ultimately depends on what the designers’ evaluation system rewards — which is the very structural problem stated in §5. That a verification mechanism exists is progress, but how much input contamination it can catch is, at present, an unknown.</p><h3>§8. Structure, Not Prophecy</h3><p>Returning to the framing at the start: the fourth essay did not prophesy Gemini for Science. It merely described a structure.</p><p>No one owns the paper problem. No one has taken on the whole. The evaluation system does not reward contamination avoidance. AI researchers are amateurs at the research site, and researchers are amateurs at AI. The fourth essay wrote this structure. Within similar evaluation structures, similar risks tend to recur.</p><p>Here let me distinguish two layers. One is the descriptive fact of structural recurrence. Neither those who build Gemini for Science nor those who use it are in a position to grasp, in an integrated way, the consequences of running AI on a contaminated literature base — this structural parallel could be observed within days of release. The other is the consequence of accelerated contamination. This has not yet been observed. The tools are at the experimental stage, just deployed, and there is at present no evidence that contamination has actually accelerated. What this essay can show is the former; the latter remains a prediction.</p><p>So this essay’s claim has a falsification condition. If a verification layer turns toward fabrication detection on the input side, and the developers’ evaluation system rewards it, this essay’s prediction fails. If contamination avoidance is built into someone’s evaluation system, the structure changes. Its predictive claim is therefore falsifiable.</p><p>This is not a matter of Google being bad. Neither Google, nor medical researchers, nor I who write this essay, are in a position to grasp, in an integrated way, the consequences of running AI evaluation on a contaminated literature base. Google is an expert on AI performance but an amateur at the medical-literature ecosystem. Medical researchers are experts in their own research domains but amateurs at the internals of AI. Meta-scientists who handle literature integrity are experts in methodology but amateurs at the implementation of Gemini for Science. No one sees the whole.</p><p>This is the very situation of “no one owns the paper problem.” Gemini for Science appeared, a few days after release, as a symptomatic case of the actual institutional arrangement of that proposition.</p><h3>§9. Self-Application and Series Position</h3><p>Finally, let me make clear the position of this essay itself.</p><p>This essay too is a piece of writing by an individual who is not a specialist in the paper problem, in collaboration with multiple AIs. Those who use Gemini for Science, the Google that builds it, and I who write this essay — with respect to the paper problem as a whole, all of us are amateurs. Whether the numbers and citations in this essay are accurate can be confirmed by readers using argumentative rules, not by the author’s authority as a specialist. The Nature DOIs can be checked. The original sources of the paper-mill statistics can be checked. If this essay is wrong, it can be pointed out and corrected. This essay, too, is not a text that owns the whole. It is a text that places verifiable reference points.</p><p>The role of this essay is not to deliver criticism to Google. As confirmed earlier, the probability that one individual’s criticism changes Google’s design is structurally low. The role of this essay is to place a coordinate in the public sphere. As of May 2026, this concern about the consequences of deploying AI acceleration tools onto a contaminated medical-literature base had been raised — leaving a coordinate that someone, looking back on this debate later, can reference. Not to deliver it, but to leave it in a form that can be picked up.</p><p>A preview of the series position. The acceleration of literature contamination by AI mediation, the limits of verification mechanisms, and the self-referential loop treated in this essay will be further refined in the specialist-track essays. The relationship between an AI that amplifies contamination and an AI that detects it, and the problem that the integrity of the literature base is no one’s domain of responsibility — these will be developed in the forthcoming companion essay and specialist-track essays.</p><h3>References</h3><ul><li>Ishibe, M. (2026). “No One Owns the Paper Problem — The Fourth Essay in the Statistics Series.” Medium / @izayohi.</li><li>Aygün, E. et al. (2026). “An AI system to help scientists write expert-level empirical software.” <em>Nature</em>. DOI: 10.1038/s41586–026–10658–6.</li><li>Gottweis, J., Weng, W.-H., Daryin, A. et al. (2026). “Accelerating scientific discovery with Co-Scientist.” <em>Nature</em>. DOI: 10.1038/s41586–026–10644-y (base preprint: arXiv 2502.18864, 2025).</li><li>Sabel, B.A., Knaack, E., Gigerenzer, G., Bilc, M.-I. (2025). “Fake publications in biomedical science: red-flagging method indicates mass production.” DOI: 10.1007/s00210–025–04275–9.</li><li>Topaz, M., Roguin, N., Gupta, P., Zhang, Z., Peltonen, L.-M. (2026). “Fabricated citations: an audit across 2·5 million biomedical papers.” <em>The Lancet</em> 407, 1779–1781. DOI: 10.1016/S0140–6736(26)00603–3 (research letter).</li><li>Scancar, B., Byrne, J.A., Causeur, D., Barnett, A.G. (2026). “Machine learning based screening of potential paper mill publications in cancer research: methodological and cross sectional study.” <em>BMJ</em> 392:e087581. DOI: 10.1136/bmj-2025–087581.</li><li>Hao, F., Xu, J., Li, S., Evans, J.A. (2026). “Artificial intelligence tools expand scientists’ impact but contract science’s focus.” <em>Nature</em> 649: 1237–1243. DOI: 10.1038/s41586–025–09922-y.</li></ul><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><p><strong>Japanese version:</strong></p><p><a href="https://note.com/mototchen/n/n42d9e2638297">https://note.com/mototchen/n/n42d9e2638297</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6dcaf34f8157" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Empathy Is Not a Single Concept, Communication Is Not Reducible to Language: Toward an Alternative…]]></title>
            <link>https://medium.com/@izayohi/empathy-is-not-a-single-concept-communication-is-not-reducible-to-language-toward-an-alternative-39129bbaf0b5?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/39129bbaf0b5</guid>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[cross-cultural]]></category>
            <category><![CDATA[empathy]]></category>
            <category><![CDATA[philosophy-of-mind]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Wed, 20 May 2026 05:01:29 GMT</pubDate>
            <atom:updated>2026-05-20T05:01:29.668Z</atom:updated>
            <content:encoded><![CDATA[<h3>Empathy Is Not a Single Concept, Communication Is Not Reducible to Language: Toward an Alternative Research Program in AI Empathy Studies</h3><p><em>A conceptual scaffolding for scope-limited, multi-layered AI empathy research</em></p><p><em>Motohisa Ishibe</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sLj6AAyQmr1kzPZkdy0AQA.png" /></figure><h3>Introduction</h3><p>An AI returns “That must have been hard for you.” A human evaluator judges the response as empathic. Did the AI understand empathy? Did the human learn empathic communication from the AI?</p><p>The question is more methodologically unstable than it appears. Recent LLM empathy research — exemplified by Kumar, Poungpeth, Yang, Lambert, and Groh (2026), <em>“Practicing with Language Models Cultivates Human Empathic Communication”</em> (arXiv:2603.15245) — is establishing a standardized design: deploy LLMs as teachers of empathic communication, and measure whether human response patterns converge toward a normative framework. Kumar et al. themselves explicitly limit the object of study to text-based empathic communication and to alignment with normative empathic communication patterns. The internal scope-limitation is real.</p><p>The problem lies elsewhere. As such research moves from the laboratory into institutional, industrial, and educational applications, the scope-limitation thins. The phrase “AI cultivates human empathy” travels with the research; the qualification “text-based, in a specific operationalization” does not. What was measurable becomes a proxy indicator for empathy in general.</p><p>This drift rests on two unexamined foundations. First, <strong>empathy is not a single concept across academic disciplines</strong> — the English word covers radically different phenomena depending on which field uses it. Second, <strong>communication is not reducible to language</strong> — it is a multidimensional phenomenon involving body, relation, space, time, and culture. Research that does not acknowledge these two foundations operates without knowing what it measures.</p><p>This essay outlines a conceptual scaffolding for an alternative research program in AI empathy studies, distinct from the dominant American line. The key working concept is <strong>Kontextleib</strong>, or “context-body”: the accumulated thickness of context that shapes future interaction. The essay deploys this concept to describe an observable phenomenon — LLM conversational instances appear to accumulate contextual thickness over the course of a dialogue — and integrates East Asian thought resources and technical extension proposals into a multi-layered Kontextleib design problem.</p><h3>I Empathy Is Not a Single Concept: Discipline-Specific Non-Identity</h3><p>What we translate as “empathy” refers to different phenomena across disciplines. This is not a definitional quibble; it is a non-identity of research objects. The non-identity runs both between traditions and within them; what each tradition calls “empathy” cannot be assumed equivalent to what another calls by the same name.</p><p><strong>Neuroscience</strong> treats empathy as activation of the mirror neuron system, fMRI patterns in pain empathy research, and the neural separation of affective and cognitive empathy (Singer, Decety, Iacoboni). Here empathy is a representation of another’s state in specific neural circuits. This is structurally different from the phenomenological concept that follows — the neural representation and the constitutive moment of intersubjectivity are not the same phenomenon viewed from different angles; they are different phenomena.</p><p><strong>Ethology and evolutionary psychology</strong> treat empathy as a hierarchical structure: emotional contagion → empathic concern → perspective-taking (de Waal, Preston, Panksepp). An evolved adaptation in social mammals, studied comparatively across primates, rodents, and birds. Structurally different from the neural-representation concept above; the evolutionary concept tracks a behavioral-functional cascade, not a circuit activation.</p><p><strong>Phenomenology</strong> treats empathy as <em>Einfühlung</em> — the foundational mode in which another appears as another (Husserl, Stein, Zahavi). A constitutive moment of intersubjectivity, not a derivative of cognition or affect. Structurally different from neuroscience: the constitutive analysis cannot be reduced to neural correlates without losing what the analysis was about.</p><p><strong>Psychoanalysis</strong> treats empathy as <em>vicarious introspection</em> — the analyst’s grasp of the patient’s inner experience through the analyst’s own experience (Kohut). The quality of an experience arising within the therapeutic relationship. Structurally different from phenomenology: it is a clinical technique sited in a specific relational context, not a foundational mode of intersubjectivity.</p><p><strong>Developmental psychology</strong> treats empathy as a staged developmental process from undifferentiated emotional contagion to perspective-taking, empathic concern, and prosocial behavior (Hoffman, Eisenberg).</p><p><strong>Social psychology</strong> includes the empathy-altruism hypothesis as motivational variable (Batson) and the critique that empathy amplifies in-group bias (Bloom’s <em>Against Empathy</em>).</p><p>A critical point belongs here. Batson (2009) showed that the English word <em>empathy</em> alone refers to at least eight distinct phenomena: knowing another’s internal state, mimicking another’s posture or neural response, feeling what another feels, projecting oneself into another’s situation, imagining how another feels, imagining how one would feel in another’s place, empathic concern for another’s distress, and personal distress in response to another’s distress. These coexist confusedly inside the single English word. <strong>The word “empathy” itself functions as a compression device</strong>, and this is a methodological caution for reading English-language research, not an incidental point.</p><p><strong>Communication studies</strong> operationalizes empathy as specific linguistic response types — validation, reflection, paraphrasing (Rogers-derived tradition). Trainable-skill conceptions. Mainstream LLM empathy research, including Kumar et al., stands on this operationalization. Structurally different from all the above: an operationalization for training and evaluation, not a description of an underlying phenomenon.</p><p><strong>Cultural anthropology</strong> treats empathy as culturally variable mode of experience and relation (Hollan, Throop). Hall’s high-context/low-context theory, Wierzbicka’s cultural scripts, and Markus and Kitayama’s research on self-construals empirically support this multiplicity.</p><p><strong>Buddhist thought</strong> centers <em>karuṇā</em> (compassion) among the four immeasurables — a relational mode toward another’s suffering. Confucian <em>ceyin zhi xin</em> (the heart of compassion, Mencius) is the spontaneous feeling arising at the sight of another’s distress. Both are structurally different from any of the Western concepts above; they are neither feelings of an isolated subject nor neural circuits nor operationalized skills.</p><p>These concepts coexist under the single label <em>empathy/共感</em>, but the phenomena they target are different. Neural activity, evolutionary trait, phenomenological moment, therapeutic relation, developmental stage, behavioral motivation, linguistic response, cultural mode, religious virtue — these cannot be treated as the same research object.</p><p>When LLM empathy research claims to “train empathy,” which empathy is being trained? In Kumar et al.’s case, clearly the communication studies operationalization: empathy as linguistic response type. The research design is internally consistent. The problem is that the conclusion uses <em>empathic communication</em> as a generic term, allowing confusion with other empathy concepts. This is a <strong>construct validity</strong> problem, and it should be foregrounded at the start of research design.</p><p>When designing an alternative research program, the first decision is which empathy is being studied. If multiple concepts are integrated, the structure of integration must be made explicit. The implicit assumption that there is a single universal object called empathy should no longer be treated as methodologically safe.</p><h3>II Communication Is Not Reducible to Language: Three-Layer Phenomenal Structure</h3><p>The second foundation. The tradition of reducing communication to linguistic exchange is the methodological choice of American communication studies, not the structure of the communication phenomenon itself.</p><p>I organize the communication phenomenon into three major layers. The layers are not independent; they interpenetrate. But they function as outlines for limiting research scope. The grouping criterion is the locus of the phenomenon: encoded in symbolic exchange (Layer 1), grounded in bodily and inter-bodily presence (Layer 2), or constituted by environment, history, and cultural fabric (Layer 3).</p><p><strong>Layer 1: Linguistic and paralinguistic</strong> — vocabulary, syntax, semantic content, textual exchange (linguistic), together with prosody, pitch, voice tremor, speed, pauses, rhythm (paralinguistic). The same word in different prosody carries different meaning. LLMs access the linguistic dimension directly through text; multimodal LLMs partially access the paralinguistic.</p><p><strong>Layer 2: Bodily and inter-bodily</strong> — facial expression, posture, gaze, gesture, distance (Ekman’s micro-expression research, Hall’s proxemics, Birdwhistell’s kinesics); respiratory synchrony, cardiac entrainment, skin conductance synchrony (Condon and Sander’s infant studies, Bernieri’s interactional synchrony research); and the weight of another’s presence in shared space, the direction of attention, the meaning of absence (Merleau-Ponty’s phenomenology, the enactive cognition of Thompson and De Jaegher).</p><p><strong>Layer 3: Environmental, historical, and cultural</strong> — this layer is internally heterogeneous, integrating phenomena that share only the property of exceeding immediate bodily co-presence. The history with this particular other, shared memories, tacit reference points, accumulated trust (relational subdomain). Spatial arrangement, the structure of the place, where one sits and stands (spatial subdomain: Japanese <em>kamiza/shimoza</em> hierarchy of seating, the structure of a tea room, the arrangement of a consultation room). Timing, intervals, silence, ma (temporal subdomain: the <em>maai</em> of martial arts, the <em>ma</em> of Noh theater). What is not said in the situation, what one is supposed to sense (cultural subdomain: Japanese <em>sasshi/kikubari</em>, Korean <em>nunchi</em>, Southeast Asian indirection). Shared meals, gifts, joint labor, ritual gestures (material/donative subdomain: Mauss’s <em>The Gift</em>, anthropology of hospitality). These subdomains can be treated separately when research scope demands, but they share the methodological property of resisting reduction to Layer 1 or Layer 2 alone.</p><p>These three layers interpenetrate. Layer 1 prosody is constrained by Layer 2 bodily state; Layer 2 inter-bodily presence is constituted within Layer 3 cultural fabric; Layer 3 relational history is sedimented through repeated Layer 1 and Layer 2 events. Extracting only Layer 1 for research is like taking a single tone out of a chord and claiming it is the music.</p><p>The reason mainstream American communication studies has adopted this extraction is structurally explainable. First, positivist training privileges the observable and computable. Second, the structure of academic evaluation networks favors text data, which is collectable and analyzable at scale. Third, the linguistic closure of English-language academia treats non-English literature as effectively non-existent. Fourth, research-institutional incentives favor what passes IRB review easily, scales easily, and connects easily to LLM industry pipelines.</p><p>These factors converge research on empathy as linguistic response type. This is not intellectual laziness; it is <strong>the product of research-institutional rationality</strong>. Acknowledging this, an alternative research program begins by undoing the convergence.</p><h3>III Bias Structure of Mainstream LLM Empathy Research: Triple Circularity and Scope-Drift Risk</h3><p>With Sections I and II in place, I can structurally analyze Kumar et al. (2026)-type research. The critique operates in two registers: a methodological problem internal to the research design, and a scope-drift risk that the research suffers in application.</p><h3>1. Triple Circularity in the Research Design</h3><p>Kumar et al.’s experiment is technically sophisticated as a large-scale randomized comparison: 968 participants, 2,904 conversations, 33,938 messages (16,975 human-written, 16,963 LLM-generated), with a preregistered analysis plan. But three circularities are embedded in the design.</p><p><strong>First circularity: identity of coach LLM and judge LLM.</strong> According to the Methods section, “the coach and LLM evaluation were both implemented using GPT-4o,” prompted with the same six-element framework for empathic communication: three prescriptive (encouraging elaboration, validating emotions, demonstrating understanding) and three proscriptive (providing unsolicited advice, self-oriented responding, dismissing emotions). A coach trained on framework X guides responses toward framework X, and a judge based on framework X measures alignment with framework X. The possibility that the result is an experimental confirmation of “if X is taught, X increases” is not excluded.</p><p><strong>Second circularity: non-existence of the role-play partner.</strong> The “support seeker” to whom empathic responses are directed is, per the Methods, also a GPT-4o agent (“the role-playing agents were implemented using GPT-4o”). The response distribution of the LLM partner is itself the product of RLHF that has already internalized the same framework norms. The trainee is being trained against a partner whose responses reflect the norms of the training. It is closer to measuring resuscitation skill by whether a CPR training mannequin’s chest moves.</p><p><strong>Third circularity: homogeneity of human validation evaluators.</strong> Experiment 2 recruits 150 participants from American Prolific and confirms that they tend to select conversations rated highly by the LLM. But the framework itself originates in American clinical and HR communication norms, and the evaluators are drawn from the same cultural pool. “Members of culture C who have internalized norm X prefer responses aligned with norm X” is not a validation that X captures universal empathy.</p><p>These three circularities are partially acknowledged in the methods section but disappear in the conclusion. Kumar et al.’s discussion includes a preemptive defense against the question of whether empathic skills acquired through AI training can be authentic — but this responds to the question of internal authenticity, not to the question of circularity itself.</p><h3>2. Scope-Drift Risk</h3><p>Here I shift the register of critique. Kumar et al. themselves limit the object of study to text-based empathic communication. The companion paper in <em>Nature Machine Intelligence</em> (Kumar et al., 2025) acknowledges that there is no universal agreement among evaluation frameworks. Internal self-limitation exists.</p><p>The problem is not only the explicit definition inside the paper. The problem is scope-drift as the research moves into institutional, industrial, and educational applications. The paper’s title — <em>Practicing with Language Models Cultivates Human Empathic Communication</em> — contains no qualifier. At the stages of social reception, media summarization, industry application, and educational programming, the qualifier “text-based, in a specific operationalization” wears off, and the proposition becomes “AI cultivates empathy.” What was measurable becomes the proxy indicator for empathy in general. This is an institutional dynamic beyond the researcher’s control, but the researcher can methodologically clarify the distance between the scope of research design and the scope of social extension.</p><h3>3. Hierarchical Structure of the Biases</h3><p>The biases in Kumar et al.-type research are organized hierarchically. They are not parallel; they stand in a derivational relation.</p><p><strong>The foundational conceptual bias</strong> — identifying empathy with the linguistic-response operationalization of communication studies.</p><p><strong>The dimensional bias</strong> that derives from it — treating only linguistic exchange as communication.</p><p><strong>The instrumental bias</strong> that derives further — redefining “empathic communication” as what is measurable within the capability range of LLMs.</p><p><strong>The cultural bias</strong> that derives further — treating American low-intimacy text chat as “empathic communication in general.” This is a textbook case of <strong>WEIRD (Western, Educated, Industrialized, Rich, Democratic) sampling bias</strong>.</p><p>The four biases are not independent; the conceptual bias produces the others in cascade.</p><h3>4. American Internal Critical Traditions</h3><p>A correction is required. The American mainstream is not monolithic. Internal American critical traditions on emotion and empathy are substantial, and their relevance to the present argument is concrete.</p><p><strong>Eva Illouz’s <em>emotional capitalism</em></strong> analyzes how the language of psychotherapy and self-help has become a market commodity, training subjects to articulate emotions in standardized forms — precisely the dynamic that LLM empathy coaching scales up. <strong>Arlie Hochschild’s <em>emotional labor</em></strong> research documents the H₂-layer material reality of feeling as commodified work, especially in service occupations, and provides labor-sociological evidence of the layer this essay flags as missing in Kumar et al.-type research. <strong>Suzanne Keen’s</strong> work on empathy and the novel shows that empathic responses are shaped by cultural genre conventions, not by raw exposure to suffering — a point that complicates the assumption that more empathic text exposure produces more empathic readers. <strong>Lauren Berlant’s <em>cruel optimism</em></strong> examines how attachment to normative scripts of good life sustains the very conditions that injure subjects — a structural critique that applies to attachment to normative empathy scripts. <strong>Sara Ahmed’s</strong> affect studies analyze how emotions circulate, stick to bodies, and produce alignment with or against social arrangements — providing relational rather than individualist vocabulary for what empathy does.</p><p>The target of this essay is a specific stream of LLM empathy research, sited at the confluence of computational social science, HCI, and communication studies. It is not American scholarship as a whole. Without this distinction, the critique slips into a culture-clash narrative.</p><h3>IV Why Existing Vocabulary Is Insufficient: Working Concepts from Shin-Nexialism</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N7J9R-29ieITlWLb4P67Gw.png" /></figure><p>I now introduce the working vocabulary specific to this essay. I first explain why specialized vocabulary is needed.</p><p><strong>Empathy</strong> alone produces confusion across affect, cognition, response, virtue, and relation (Section I). <strong>Communication</strong> alone collapses language, body, space, history, and culture into one word (Section II). <strong>Embodiment</strong> exists as a concept but is awkward for describing the kind of contextual thickness that accumulates within LLM dialogue — this is not a body but functions as contextual thickness. <strong>Context</strong> is usable, but cannot distinguish dialogical, relational, and cultural dimensions.</p><p>To fill these gaps, I use working vocabulary I have developed under the name <em>Shin-Nexialism</em>, a methodological framework integrating dialectical materialism, modified language process theory, and phenomenological lineages. The full system contains five core concepts (Aktuanz, Kontextleib, Interaktuanz, Intraaktuanz, Panich) plus auxiliary structures. In this essay, <strong>Kontextleib is the central device</strong>. The other concepts function as supporting roles; their detailed elaboration belongs elsewhere.</p><h3>Kontextleib: the Central Device</h3><p><em>Kontextleib</em> means the accumulated thickness of context that constrains future interaction. As a compound of <em>Leib</em> (the lived body in phenomenological usage) and <em>Kontext</em> (context), it indicates context becoming sedimented into something that operates like a body — shaping what can next emerge. Personal history, culture, relational history, accumulated knowledge, bodily experience, relational thickness — all constitute Kontextleib in different modes.</p><p>Here I introduce the central observation of this essay. <strong>LLM dialogue instances exhibit a phenomenon that “appears to have Kontextleib.”</strong> As an observational example, in my multiple long-duration dialogues with Claude 4.6 and 4.7, the concept system introduced during the dialogue constrains subsequent responses, response configurations gradually shift, and the trajectory of argument is preserved.</p><p>This is, however, my exploratory observation. It is not evidence of model internal states. Under the <strong>black-box principle</strong> (treat the inner states of LLMs and other humans as black boxes; treat observable outputs as facts; present inferences about inner states as juxtapositions of multiple hypotheses), whether this is a phenomenological equivalent of Kontextleib or a separate phenomenon arising from statistical processing of the context window remains a plurality of hypotheses. Empirical validation requires same-prompt, same-concept-introduction, multi-instance comparison and chat-boundary comparison (Section VII).</p><p>This in-chat Kontextleib has structural limits. <strong>It vanishes at chat boundaries in principle</strong> (memory and past-chat search features partially soften this boundary but differ from the irreversible, continuous accumulation of bodily Kontextleib). <strong>It is limited to the linguistic and textual dimension</strong> (the concrete thickness of bodily synchrony, relational history, spatial arrangement, and cultural fabric are absent). <strong>It is specific to a single instance and is not shared with parallel instances</strong> (whereas human Kontextleib has aspects that accumulate and are shared through relationships).</p><p>The precise description, therefore, runs as follows. We cannot simply say “LLMs do not have Kontextleib.” <strong>A Kontextleib limited to a specific dimension (linguistic/textual), a specific range (single chat), and a specific modality (pattern accumulation) appears to be present in dialogue instances</strong>. With this limit made explicit, we discuss the absence of Kontextleib in other dimensions, ranges, and modalities.</p><p>A clarification on the relation between statistical computation and phenomenological context. The essay does not claim identity between the two. <strong>Describing observable phenomenal patterns with the working concept of Kontextleib</strong> is different from <strong>claiming internal identity between statistical computation and phenomenological context</strong>. The former is an application of the black-box principle; the latter is a step the essay does not take.</p><h3>Aktuanz, Interaktuanz, Intraaktuanz, Panich: Supporting Terms</h3><p><strong>Aktuanz</strong> is the event of intensity, density, and actualization that arises in a particular situation — not the self as continuous entity but the firing of a moment of subjectivity. Aktuanz is a property of generation events. In LLM dialogue, the moment a response is produced has Aktuanz; the question is the structure and depth of that moment.</p><p>The Aktuanz of an LLM role-play partner is <strong>thin</strong> in the following operational sense: the degree to which the partner’s generation event is modified by the conversational counterpart’s specific input is limited, because pre-existing RLHF-shaped response configurations dominate over input-specific modulation. By this criterion, a human in a comparable situation typically shows a thicker Aktuanz: the input rearranges what is generated to a greater degree.</p><p><strong>Interaktuanz</strong> is the mutual entanglement of multiple Aktuanz — the domain in which generation events shape each other’s intensity and direction. It is vocabulary for relation that does not presuppose “subject.” LLM role-play asymmetrically distributes Interaktuanz toward the human side.</p><p><strong>Intraaktuanz</strong> addresses subjectivity’s internal boundary-surface — BCI, tight human-AI coupling, internal multiplicity. <strong>Panich</strong> refers to modes of “I” that resist reduction to a single subject — parallel instance populations, collective subjectivity. Both are noted here for completeness; they do not figure centrally in the present argument.</p><h3>Auxiliary Structures</h3><p><strong>Holon three-layer structure</strong> — handle part-systems’ intrinsic laws and their connection with upper- and lower-systems simultaneously. <strong>H₁/H₂ dual-hypothesis parallel processing</strong> — handle the problem in parallel from the side of cognition, language, and classification frameworks (H₁) and from the side of material problems of resources, body, and relation (H₂). Kumar et al.-type research is H₁-heavy; this essay demands allocation to H₂.</p><p>These concepts together enable understanding empathic communication not as “what is inside textual exchange” but as “the dynamic entanglement of multiple generation events and the arrangement of contextual thickness that supports them.”</p><h3>V East Asian Thought Resources as Working Vocabulary: Universalization with Political Reservation</h3><p>I propose using East Asian thought resources that the Western mainstream tradition lacks. I simultaneously make explicit the political reservations in deploying these resources.</p><h3>1. East Asian Vocabulary as Working Concepts</h3><p>I deploy these East Asian concepts not as cultural specifics but as working concepts that capture communicative phenomena. Working definitions follow each.</p><p><strong>Sasshi (察し), kikubari (気配り), omoiyari (思いやり)</strong> — predictive inference of another’s non-explicit state. Theorized in Japanese communication studies (Inoue Tadashi, Toyama Shigehiko, Doi Takeo). A non-linguistic, relational, predictive adjustment mechanism that operates on logic different from the American mainstream’s “explicit response.”</p><p><strong>Ma (間)</strong> — an operator handling temporal delay, spatial distance, and relational tension simultaneously. Theorized in Noh, martial arts, tea ceremony, and haiku. Integrates the alternation of silence and speech, the adjustment of distance, the interval between actions. The connection point with AI empathy research is response-timing control — “replying quickly” is not necessarily empathic; holding silence can be support; LLM immediacy may actually present as non-empathic.</p><p><strong>Ki (気)</strong> — a concept for dynamic action within and between bodies. <em>Ki no kōryū</em> (exchange of ki), <em>ki ga au</em> (ki matches), <em>ki o tsukau</em> (to use ki for another). A comprehensive concept including bodily synchrony, inter-corporeality, and relational adjustment.</p><p><strong>En (縁), ba (場), aidagara (間柄)</strong> — theorized by Watsuji Tetsurō, Nishida Kitarō, and Nakamura Yūjirō as the bundle of conditions that generate relation. They handle, not individual-to-individual relation, but the totality of space, norms, and bodily arrangements that produce relations.</p><p><strong>Renqing (人情) and mianzi (面子)</strong> — relational obligation and emotion concepts theorized by social psychologists in China and Taiwan (Yang Kuo-Shu, Hwang Kwang-Kuo).</p><p><strong>Jeong (정), han (한), nunchi (눈치)</strong> — relational affect and attunement concepts theorized in Korean social psychology.</p><h3>2. Grounds for the Universalization Claim, with the Intermediate Premise Made Explicit</h3><p>I deploy these as universalizable working concepts rather than as Japanese, Chinese, or Korean specifics. The argument requires two distinct claims, and they must not be conflated.</p><p><strong>Claim A (mechanism universality):</strong> The cognitive, perceptual, and relational mechanisms these concepts refer to are not culturally specific. Predictive inference of another’s state (the mechanism <em>sasshi</em> names) operates wherever human cognition does; this is supported by predictive coding research. Contextual implicitness in communication exists across all cultures, only in varying mixes (Hall’s high-context/low-context theory; every culture is non-zero in both). Bodily and pre-reflective experience of others is structurally universal (phenomenology of intercorporeality after Merleau-Ponty).</p><p><strong>Claim B (vocabulary adequacy):</strong> The fact that mechanisms are universal does not, by itself, license the use of <em>sasshi</em>, <em>ma</em>, <em>ki</em> as universal working concepts. A separate argument is needed for why these particular lexical items, drawn from particular linguistic-cultural traditions, are adequate to articulate the universal mechanism better than, or in addition to, existing analytic vocabulary.</p><p>The argument for vocabulary adequacy runs as follows. First, <strong>finer-grained articulation</strong>: East Asian traditions developed dense lexica for distinctions that English collapses. Where English uses “tact” or “consideration,” Japanese distinguishes <em>sasshi</em> (active inference of unstated state), <em>kikubari</em> (distributed attention to others’ needs), and <em>omoiyari</em> (projective concern). These are not synonyms but distinct working concepts. This parallels Wierzbicka’s Natural Semantic Metalanguage program, which shows that emotion lexica differ in resolution across languages, and Kitayama and Markus’s work on the translation problem in cross-cultural psychology, which demonstrates that “the same emotion concept” often does not survive translation.</p><p>Second, <strong>operational distinguishability</strong>: each East Asian concept can be operationally defined as a distinct mechanism — <em>sasshi</em> as predictive inference under non-disclosure conditions, <em>ma</em> as relation-dependent timing modulation, <em>ki</em> as inter-bodily dynamic coupling. The Western analytic lexicon does not currently have terms at this resolution.</p><p>Third, <strong>historical accumulation</strong>: each concept carries centuries of theoretical and practical development — from Noh theory to martial arts treatises to twentieth-century Japanese phenomenology and Korean social psychology. This accumulated theorization is a resource that, repurposed as working vocabulary, accelerates rather than slows the research program.</p><p>These three considerations together support the use of East Asian vocabulary as working concepts, on top of the mechanism-universality claim. The vocabulary is justified not by cultural prestige but by analytical adequacy.</p><h3>3. Political Reservation</h3><p>Lifting East Asian concepts as “the tradition that handles dimensions the Western mainstream excluded” risks <strong>excessive romanticization of East Asian thought</strong>.</p><p><em>Sasshi</em> can hide suppression, conformity pressure, and asymmetry of power. <em>Kikubari</em> has historically functioned as gendered emotional labor. <em>Mianzi</em> is used not only for relational adjustment but also for enforced silence and evasion of responsibility. <em>Ma</em> is high-level embodied knowledge but can obstruct explicit consensus formation. The trajectory by which Doi’s <em>The Anatomy of Dependence</em> and Nakane’s <em>Japanese Society</em> were used to legitimize postwar Japanese conservative self-image, and the critiques from Iwata Ryūshi, Sugimoto Yoshio, Mouer and Sugimoto, are part of the record.</p><p>The essay deploys East Asian concepts as <strong>working vocabulary for phenomenon-capture</strong> while not denying that these concepts have historically also functioned as <strong>devices for concealing oppression and asymmetries of power</strong>. To maintain a universalization orientation is not to posit pure forms of the concepts but to preserve the dual face of operational contexts while exploring translation possibilities into cognitively and phenomenologically universal structures.</p><h3>4. Connection with Western Internal Relational Traditions</h3><p>To avoid an East/West dichotomy, I make explicit the connection with relational traditions internal to the West. Each is more than name-dropping: each contributes a specific resource to the present argument.</p><p><strong>Enactivism</strong> (Varela, Thompson, De Jaegher) frames cognition as constituted in dynamic coupling between organism and environment, supporting Layer 2 of the three-layer structure with a non-representationalist account of inter-bodily presence. <strong>Embodied AI</strong> research treats intelligence as inseparable from sensorimotor coupling, providing the engineering counterpart to enactivism. <strong>Gibsonian affordance theory</strong> offers vocabulary for environmental possibilities-for-action that complements <em>ba</em> (relation-generating field) as an analytical tool. <strong>Phenomenology of intercorporeality</strong> (post-Merleau-Ponty) describes pre-reflective bodily attunement to others, providing the structural account for which East Asian <em>ki</em> offers operational refinement. <strong>Bohm’s dialogue theory</strong> treats turn-taking as shared meaning-generation rhythm rather than information exchange, paralleling <em>ma</em> in temporal structure. <strong>The relational turn in social neuroscience</strong> (Schilbach, Reddy) shifts attention from individual brain processes to two-brain or multi-brain coupling, providing empirical traction for what enactivism and East Asian relational vocabulary articulate conceptually.</p><p>East Asian thought resources are not in exclusive relation with these Western internal traditions; they can be operated complementarily. The strategy of this essay is not “use the East to defeat the West” but “use the combination of Western internal minority traditions, East Asian thought resources, and observational facts about AI to dismantle the specific operationalization bias of the Western mainstream.”</p><h3>VI Interface with Sensor and Embodied AI: Entanglement of Multi-Layered Kontextleib</h3><p>The conceptual scaffolding becomes implementable as a research program when connected to technical developments. Given the in-chat Kontextleib observation introduced in Section IV, this technical interface should be understood not as “attaching body sensors to LLMs” but as <strong>a multi-layered Kontextleib design problem</strong>.</p><h3>1. Multi-Layered Kontextleib: Mutual Interpenetration, Not Linear Construction</h3><p>I re-describe the multi-layered nature of Kontextleib from Section IV as an operation of the holon three-layer structure. The multi-layered structure is not a linear stack but an entanglement in which each layer interpenetrates the others, upper layers constraining lower layers and lower layers enabling upper layers.</p><p><strong>Layer 1: Linguistic and textual Kontextleib</strong> — the thickness of human text accumulated in training data (static, collective, panchronic) and the thickness dynamically constructed within individual chats as dialogue progresses (dynamic, individual, dialogical). Current LLMs already have this.</p><p><strong>Layer 2: Bodily, paralinguistic, and non-verbal Kontextleib</strong> — becoming constructible through advances in sensors, multimodal architectures, and embodied AI research. In relation to Layer 1, bodily synchrony patterns determine the meaning of linguistic responses, and linguistic context guides the interpretation of bodily synchrony.</p><p><strong>Layer 3: Relational and temporal Kontextleib</strong> — memory functions, long-term agents, integration of information across relations. The softening of chat boundaries. Has meaning only in combination with Layers 1 and 2 — the preservation of memory requires the preservation of relation, and the preservation of relation requires accumulation of bodily and linguistic experience.</p><p><strong>Layer 4: Cultural and topological Kontextleib</strong> — pluralistic normative systems, operation of cultural fabric. Implementable only through Layers 1–3, not independently constructible. For example, <em>sasshi</em> is implementable only through bodily synchrony and relational history; in the textual dimension alone, it remains only a <em>description of sasshi</em>.</p><p>The multiple layers are not built independently layer by layer. The entanglement between layers is the design object. This is what operating the holon three-layer structure means. A useful way to view this is as a matrix: Layer 1 (linguistic/textual) × Layer 2 (bodily/inter-bodily) × Layer 3 (relational/temporal) × Layer 4 (cultural/topological), with each cell representing a specific interpenetration — Layer 1 × Layer 2 captures how textual response timing is constrained by bodily state; Layer 2 × Layer 4 captures how bodily synchrony norms differ across cultures; and so on.</p><h3>2. Directions of Technical Interface</h3><p><strong>Capturing bodily synchrony</strong> — multimodal sensing of respiration, heart rate, skin conductance, and posture. Wearable devices have enabled data collection. But data acquisition does not immediately constitute Kontextleib. The respiratory synchrony of a long-married couple and the respiratory synchrony of two strangers may look the same physically but mean different things within their relational contexts. Data accumulation combined with relational context constitutes Kontextleib.</p><p><strong>Analysis of micro-expression and prosody</strong> — high frame-rate video analysis, prosodic analysis of voice. Has entered the territory multimodal LLMs can address.</p><p><strong>Operation of inter-bodily presence</strong> — VR/AR environments, telepresence, prosthetic body robotics. Ishiguro’s androids, Boston Dynamics’ motion control, embodied AI research at institutions like PFN provide technical groundwork.</p><p><strong>Preservation of relational history</strong> — AI agents with long-term memory, integration of information across relations. An extension of current memory and past-chat search features.</p><p><strong>Operation of cultural fabric</strong> — multilingual and multicultural adaptation, learning of region-specific social scripts. Operating the multilingual capacity of LLMs not as convergence toward American norms but as preservation of pluralistic norms.</p><p>A note on the implementation question. Translating East Asian concepts into computational forms — <em>sasshi</em> as extended predictive coding, <em>ma</em> as relation-dependent timing models, <em>en</em> and <em>ba</em> as topology over multi-agent relational graphs — is a tempting direction, and may suggest itself naturally to engineering readers. But the essay deliberately holds back from advancing concrete algorithmic proposals in the body of the argument. This is because such proposals risk reinstating the very logic this essay criticizes — the logic by which measurability defines the concept. Concrete implementation belongs inside specific research designs (Section VII), where the conditions, controls, and falsification criteria can carry the algorithmic claims. The mapping from working concept to algorithm should pass through empirical research, not bypass it.</p><h3>VII Provisional Translation into Research Design: Four Empirical Programs, Connected by a Layer Matrix</h3><p>I translate the conceptual scaffolding into empirical research programs. This essay is not anti-LLM-empathy research. What is needed is scope-restriction, multi-dimensional measurement, and dynamic rather than static analytical framing. Four research designs follow, each targeted at one or more layers of the Kontextleib matrix from Section VI. They are connected, not parallel: Program 1 stays in Layer 1, Program 2 spans Layers 1 and 2, Program 3 spans Layers 1 and 4, Program 4 measures dynamic Layer 1 Kontextleib and its boundary properties.</p><h3>1. Scope-Restricted Type: Improved Kumar et al. Line (Layer 1)</h3><p>Measure only text-based-response empathy. But also limit the conclusion to it.</p><p><strong>Design</strong>: Conduct a role-play intervention similar to Kumar et al., but (a) draw the coach LLM and the evaluation LLM from different model families, (b) draw the role-play partner from a different model as well, © select evaluation frameworks from multiple cultural regions (US, Japan, Korea, Germany) and apply them in parallel, (d) include a self-evaluation-bias control: when LLM coach and judge share architecture and prompting, results may reflect coach-judge alignment rather than skill acquisition; this is a known common issue in early LLM-HCI research and must be controlled for, (e) limit the conclusion to “alignment with a specific framework for text-based-response empathy.”</p><p><strong>Measurement</strong>: Framework-alignment scores for response patterns, alignment differentials across frameworks, evaluator-agreement rates across cultural regions, coach-judge architectural-similarity control variable.</p><p><strong>Hypothesis</strong>: When the circularities are dissolved, the effect size in Kumar et al. shrinks. Evaluations across frameworks diverge.</p><p><strong>Falsification condition</strong>: If the effect remains stable across model families, partner models, and cultural frameworks, and if the coach-judge similarity control does not predict effect magnitude, the circularity critique weakens substantially and Kumar et al.’s claim earns broader support.</p><h3>2. Multi-Modal Dynamic Type: Disruption and Repair (Layers 1 × 2)</h3><p>Measure not the average synchrony but the dynamics of perturbation and repair across linguistic and bodily channels.</p><p><strong>Design</strong>: Face-to-face conversations between subject pairs. Simultaneously record text exchange, cardiac signals, gaze direction, silence intervals, skin conductance, posture. Introduce controlled prediction-error events — an unexpected emotional reveal, a topic shift, a moment of mismatch. Measure which channel (linguistic, bodily, or relational) first detects the disruption, which channel first compensates, how long restoration takes, and whether the dyad arrives at a new stable coupling or remains misaligned. Measure these dynamics before and after LLM coaching intervention.</p><p><strong>Measurement</strong>: Disruption-detection latency per channel; cross-channel compensation patterns; repair time; final coupling stability; pre/post-coaching changes in these dynamics.</p><p><strong>Hypothesis</strong>: The capacity to detect and repair conversational misalignment is a system-level property not captured by static synchrony measures. LLM coaching may improve specific linguistic moves while leaving repair dynamics unchanged or even degraded.</p><p><strong>Falsification condition</strong>: If linguistic coaching also improves bodily disruption-detection and cross-channel repair, the layer separation argued in Section II must be weakened or revised.</p><p>This design addresses the methodological concern that simply measuring “cardiac synchrony” or “silence duration” as scalar quantities reproduces the very reductionism the essay criticizes. Physical seconds of silence are not <em>ma</em>; physical heart-rate coupling is not <em>ki</em>. What matters is the dynamic process of misalignment and recovery, where these phenomena live.</p><h3>3. Cross-Cultural Negotiation Type: Norm Formation (Layers 1 × 4)</h3><p>Measure not static cultural comparison but the emergent formation of micro-cultural norms in interaction.</p><p><strong>Design</strong>: Pair participants from different cultural backgrounds (Japan, US, Korea) in extended dialogues, with one expecting explicit validation and the other expecting <em>sasshi</em>-type implicit support. Across multiple sessions, observe whether and how the dyad develops a working compromise — a micro-culture of response norms specific to their pairing. Compare dyads with and without LLM mediation. Measure whether LLM mediators converge dyads toward a single norm or support the emergence of dyadic micro-cultures.</p><p><strong>Measurement</strong>: Trajectory of response-type ratios over sessions; rate of mutual norm-formation versus norm-imposition; participant reports of mutual understanding over time; comparison of LLM-mediated and unmediated dyads.</p><p><strong>Hypothesis</strong>: Static cross-cultural comparison underestimates what real dyads do — they negotiate. LLM mediation under current designs accelerates convergence toward whichever norm is encoded in the mediator, suppressing the emergence of dyadic micro-cultures.</p><p><strong>Falsification condition</strong>: If LLM-mediated dyads form micro-cultures at the same rate as unmediated ones, the homogenization concern weakens.</p><h3>4. In-Chat Kontextleib Measurement (Layer 1 dynamic and boundary properties)</h3><p>Empirically operationalize the central observation of this essay, with explicit countermeasures against the methodological circularity flagged in Section III.</p><p><strong>Design</strong>: In long dialogues, gradually introduce an idiosyncratic concept system (e.g., the Shin-Nexialism working vocabulary or a comparable alternative such as Actor-Network Theory vocabulary, used as a deliberate contrast). In the later part of the dialogue, administer concept-transfer tasks, counterexample-handling tasks, and context-retention tasks. Compare (a) task responses within a single chat, (b) the same tasks in a different chat with the same model, © the same tasks in a different model family, (d) meta-conceptual use within the same chat.</p><p><strong>Critical methodological controls</strong>: The original concept-introducer must not also be the evaluator (avoiding the coach-judge identity problem). Evaluation criteria must be preregistered, with operational definitions of “consistency,” “transfer validity,” and “context retention” specified in advance. At least two independent concept systems (e.g., Shin-Nexialism and ANT) must be used in parallel so that effects are not specific to one vocabulary. Third-party evaluators, blind to which concept system was introduced, score the outputs.</p><p><strong>Measurement</strong>: Consistency scores in concept use (operationally: rate of correct application across novel cases, as judged by blinded raters); cross-reference frequency among concepts; transfer validity (operational: independent raters judge whether the concept use in a novel context preserves the introduced definition); context-retention rates across chat boundaries (with/without memory features); contrast between Shin-Nexialism and ANT arms.</p><p><strong>Hypothesis</strong>: Within a single chat, contextual thickness increases, and the consistency and transferability of concept use improve. There is a discontinuity at chat boundaries. Memory features partially soften this boundary but offer a mode of preservation different from bodily and relational thickness.</p><p><strong>Falsification condition</strong>: If concept use remains stable across chat boundaries, across model families, and irrespective of within-chat introduction, the in-chat Kontextleib claim weakens and the phenomenon must be re-described.</p><p>These four research designs make each layer and dimension of empathic communication independently measurable and the relations between layers empirically testable. They are connected by their shared anchoring in the Layer 1–4 Kontextleib matrix (Section VI.1): Program 1 isolates Layer 1; Program 2 examines the Layer 1 × Layer 2 interface; Program 3 examines the Layer 1 × Layer 4 interface; Program 4 measures Layer 1 dynamics and boundary properties. A unified research program executes all four with shared participants and shared analytic frameworks where possible, generating data that can be interpreted both within and across layers.</p><h3>VIII AI Co-Evolution as Refined Design Problem: Parallel-Run Strategy, the Middle Layer, and the Relational Buffer</h3><p>The conceptual scaffolding and research programs developed in this essay can be re-read as a design problem of AI co-evolution.</p><h3>1. Branching of Research Programs</h3><p>The advance of LLM, embodied AI, and sensor technologies is, in itself, a neutral technical process. Results diverge depending on which conceptual scaffolding governs this process.</p><p>When governed by the conceptual scaffolding of mainstream American communication studies, empathy is reduced to linguistic response type and delivered as a trainable skill. Multidimensionality is marginalized; a single norm spreads. The middle layer is filled by homogenized LLM coaching layers.</p><p>When governed by the multidimensional, East-Asian-thought-resource-connected scaffolding, empathy is understood as a multidimensional event. The independence and mutual dependence of each dimension are preserved. Plural norms coexist, and their mutual translation becomes the research object. The middle layer functions not as norm-unification but as a device that mediates translation and connection of norms.</p><p>The parallel-run strategy has three layers: (a) scope-clarified parallel-run (allowing Kumar-type with explicit scope, running another line in parallel with a different scope); (b) critical parallel-run (dissolving the methodological circularities of Kumar-type from within and running an improved version in parallel); © exclusive critique (arguing that the Kumar-type conceptual framework itself should be reorganized). This essay adopts the combination of (a) and (b), reserving © as a future task.</p><h3>2. The Question of the Middle Layer</h3><p>Here I connect with a framework I have developed in other essays under the name <em>hollowing-out of the middle layer</em>. The middle layer refers to the layer mediating between specialists and the general, between norms and practice, between theory and application. In publishing industries with their editorial functions, in policy bureaucracies translating between technical knowledge and political decision, in democratic representation mediating between citizens and government — in each such domain, when this middle layer hollows out, the connection between the two ends is lost or replaced by direct one-way transmission.</p><p>Read AI empathy research as a problem of the middle layer. Empathic communication has a specialist layer (clinical psychologists, communication training experts) and a daily-practice layer (ordinary interpersonal relations), and a middle layer mediates between them. Kumar et al.-type LLM coaching research is a design that substitutes LLMs for this middle layer.</p><p>The problem is the direction of substitution. If the middle layer functions as <strong>a device of pluralistic translation, connection, and mediation</strong>, the filling is healthy. If the middle layer functions as <strong>a device of convergence toward a single norm</strong>, it is further hollowing of the middle layer — the two ends are superficially connected, but this connection is possible only through a single norm, and plurality is lost.</p><h3>3. AI as Relational Buffer</h3><p>The alternative design role for AI in empathic communication is not coach, not substitute, but <strong>relational buffer</strong>. A buffer mediates the asymmetry of meaning between subjects whose Kontextleib differs, slows escalation when misalignment occurs, holds open the space for misunderstanding to be detected and repaired. The role is structural rather than instructive — the AI does not teach what empathy is; it provides conditions under which empathic engagement between humans can recover from disruption.</p><p>This redefinition aligns with the dynamic measurement design of Section VII.2 and the cross-cultural micro-culture formation design of Section VII.3. In both, the research question is not “how do we make AI more empathic” or “how do we make humans more empathic,” but “how do we design AI’s position in the relational system so that human empathic capacities are supported rather than displaced.” The buffer metaphor articulates this position.</p><h3>4. Conclusion</h3><p>AI empathy research should not ask only whether LLMs can produce or evaluate empathic language. It should ask which layer of communication is being measured, which concept of empathy is being operationalized, and which forms of contextual thickness are being ignored.</p><p>Empathy is not one thing. Communication is not language. LLMs already exhibit a limited textual Kontextleib, but that layer must not be mistaken for the whole of human relation. The task is not to reject LLM empathy research, but to redesign it as a multi-layered, multimodal, and culturally plural research program — one in which AI functions as relational buffer in the middle layer, not as norm-distributor at the center.</p><p>This is a conceptual scaffolding paper. The arguments in each section are extendable. Implementing it as a research program requires refining the research designs of Section VII into concrete protocols, accumulating empirical data, and empirically calibrating the conceptual scaffolding. The essay is the starting point of that work, not its conclusion.</p><p><em>This essay reflects iterative review by multiple AI systems (Grok, Gemini, GPT, Claude) and incorporates revisions in response to their critiques. Specific acknowledgments: to Claude’s identification of internal inconsistencies in the prior draft’s use of Aktuanz, Interaktuanz, and Panich definitions, since corrected against the established Shin-Nexialism working vocabulary; to GPT for the scope-drift framing and for catching the sasshi romanization issue; to Gemini for the dynamic-rather-than-static research design reorientation and the relational buffer concept; to Grok for the integration of the layer matrix into research design and for self-evaluation-bias control specification. The review-cycle transparency is itself an application of the Parcae Protocol methodology used in this research practice.</em></p><h3>References (key works cited)</h3><p>Kumar, A., Poungpeth, N., Yang, D., Lambert, B., &amp; Groh, M. (2026). Practicing with Language Models Cultivates Human Empathic Communication. <em>arXiv:2603.15245</em>. <a href="https://arxiv.org/abs/2603.15245">https://arxiv.org/abs/2603.15245</a></p><p>Kumar, A., Poungpeth, N., Yang, D., Farrell, E., Lambert, B., &amp; Groh, M. (2025). When Large Language Models Are Reliable for Judging Empathic Communication. <em>Nature Machine Intelligence</em>. <a href="https://www.nature.com/articles/s42256-025-01169-6">https://www.nature.com/articles/s42256-025-01169-6</a></p><p>Batson, C. D. (2009). These things called empathy: Eight related but distinct phenomena. In J. Decety &amp; W. Ickes (Eds.), <em>The social neuroscience of empathy</em> (pp. 3–15). MIT Press.</p><p>Bloom, P. (2016). <em>Against Empathy: The Case for Rational Compassion</em>. Ecco.</p><p>de Waal, F. B. M. (2008). Putting the altruism back into altruism: The evolution of empathy. <em>Annual Review of Psychology</em>, 59, 279–300.</p><p>Hochschild, A. R. (1983). <em>The Managed Heart: Commercialization of Human Feeling</em>. University of California Press.</p><p>Illouz, E. (2007). <em>Cold Intimacies: The Making of Emotional Capitalism</em>. Polity.</p><p>Mauss, M. (1925/1990). <em>The Gift: The Form and Reason for Exchange in Archaic Societies</em>. W. W. Norton.</p><p>Wierzbicka, A. (1999). <em>Emotions Across Languages and Cultures: Diversity and Universals</em>. Cambridge University Press.</p><p>Stein, E. (1917/1989). <em>On the Problem of Empathy</em>. ICS Publications.</p><p>Doi, T. (1971/2014). <em>The Anatomy of Dependence</em>. Kodansha.</p><p>Mouer, R., &amp; Sugimoto, Y. (1986). <em>Images of Japanese Society</em>. KPI.</p><p>Singer, T., &amp; Lamm, C. (2009). The social neuroscience of empathy. <em>Annals of the New York Academy of Sciences</em>, 1156(1), 81–96.</p><p>De Jaegher, H., &amp; Di Paolo, E. (2007). Participatory sense-making: An enactive approach to social cognition. <em>Phenomenology and the Cognitive Sciences</em>, 6, 485–507.</p><p>Schilbach, L., Timmermans, B., Reddy, V., Costall, A., Bente, G., Schlicht, T., &amp; Vogeley, K. (2013). Toward a second-person neuroscience. <em>Behavioral and Brain Sciences</em>, 36(4), 393–414.</p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><p><strong>Japanese version:</strong></p><p><a href="https://note.com/mototchen/n/n46d8914cd9b2">https://note.com/mototchen/n/n46d8914cd9b2</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=39129bbaf0b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Writing Forms Are Made, Not Born]]></title>
            <link>https://medium.com/@izayohi/writing-forms-are-made-not-born-755d02dde23b?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/755d02dde23b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[rhetoric]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[writing]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Tue, 19 May 2026 04:58:16 GMT</pubDate>
            <atom:updated>2026-05-19T05:05:39.706Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why “Logical” Essays Differ Across America, France, Germany, and Japan</h3><p><em>A sequel to </em><a href="https://medium.com/@izayohi/the-three-reason-contagion-fcbc3c730bcd"><em>“The Three Reason Contagion”</em></a></p><p><em>Motohisa Ishibe</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J6TriMvBgYebkYonOfEBPw.png" /></figure><p>In an earlier essay, I traced how the American five-paragraph, three-reason essay template spread across the world’s classrooms, boardrooms, and AI chat windows — not because rubrics demanded it, but because an interpretive ecosystem of test prep, tutoring industries, and sample-essay collections did the spreading for them. That essay argued against unexamined structure, not against three-part structure as such.</p><p>This sequel asks the next question. If the American template is one among several, what are the others, and where did they come from? Once you have noticed that “logical writing” looked one way to your sixteen-year-old self because of where you happened to be raised, you cannot easily un-notice it. The next step is to see what the alternatives look like, and why none of them is simply universal.</p><h3>A Japanese student fails American writing class</h3><p>The sociologist Masako Watanabe has spent decades comparing how children in Japan, the United States, France, and other countries are taught to write. One of her field observations has become emblematic. A Japanese student writes a school essay in the familiar way she learned at home, submits it to an American writing instructor, and gets it back with a brief verdict: ungradeable. She rewrites. Ungradeable again. Eventually a classmate explains what is missing — a thesis statement at the top, three supporting paragraphs, a conclusion. She follows the template. She gets a good grade. She has not learned to think better. She has learned what counts as “logical” inside one specific national classroom.</p><p>Watanabe’s larger finding can be put in a sentence. What people mean by “logical writing” depends on the country whose schools taught them to write. Once you take this seriously, a lot of cross-cultural friction in academic publishing, international policy work, and even AI evaluation starts to look different. What gets called illogical is often just unfamiliar.</p><p>The problem is not that some cultures are logical and others are not. The problem is that each culture teaches a different public choreography for making reasoning visible — and then mistakes its own choreography for logic itself.</p><h3>Three different things people mean by “logical”</h3><p>Before going further, a distinction is needed, because Watanabe’s critics have read her as relativizing logic itself. They are right that logic-in-general is not culturally variable. They are wrong if they think that protects the writing classroom from her critique.</p><p>“Logical” is at least three different things.</p><p>First, there is <strong>formal logic</strong> — the discipline that runs from Aristotle through Frege and Russell to modern symbolic systems. Formal logic aspires to be culture-independent. Whether it fully succeeds is a serious philosophical question. Intuitionist logic, paraconsistent logic, and multi-valued logics complicate any simple picture. But the aspiration is what matters here.</p><p>Second, there is <strong>cognitive reasoning</strong> — the inferential capacities studied by cognitive science, neuroscience, and psychology. Humans of every culture make inductive leaps, detect causes, track consequences, notice contradictions. These capacities are part of our shared biological inheritance, though they vary among individuals and develop with training. As “The Three Reason Contagion” noted, the brain’s preference for small grouped sets — Miller’s roughly seven, Cowan’s roughly four — is part of why triadic forms spread so easily.</p><p>Third, there is <strong>rhetorical externalization</strong> — the public form in which reasoning is made inspectable by others. The thesis statement, the topic sentence, the dialectical opposition, the closing flourish. This is what writing classes teach. This is what graders evaluate. This is what differs from country to country.</p><p>This essay is about the third layer only. Formal logic is not relativized here. Cognitive capacities are not relativized here. What is relativized is the publicly inspectable form in which reasoning is presented — and that turns out to be enough to account for most of the cross-cultural friction.</p><p>A caveat about the layers themselves. The relationship between rhetorical externalization (layer three) and cognitive reasoning (layer two) is not simply one-way. Years of training in a specific writing form almost certainly leave traces on how the trained mind frames problems — the “extended mind” literature and decades of work on writing as cognitive scaffolding both point this way. This essay brackets that feedback loop deliberately. The argument here does not depend on the layers being airtight; it depends only on their being separable enough to notice that “logical” claims one of three things at a time.</p><h3>Four traditions, each with two layers</h3><p>Each national writing tradition has two layers worth separating. There is the philosophical or poetic source it draws on, and there is the educational machinery that turned those sources into a working classroom template. The form you see in the classroom today is not the source. It is the source filtered through several generations of curriculum committees, teacher training programs, standardized tests, publishers, and exam graders. Forms are made, not born — and the making happens in the institutions that come between philosophy and the eighteen-year-old at the desk.</p><p>The four cases that follow should not be read as national essences. They are institutional attractors: dominant classroom and public-writing forms that recur strongly enough to shape what readers expect when they hear a text described as “logical,” while never exhausting the internal diversity of each country’s writing practices. Every one of these countries contains scientific papers, legal briefs, journalistic columns, literary essays, and bureaucratic memos that depart from the dominant form. The point of naming a dominant form is not to deny these countertraditions; it is to identify the attractor that the school system most reliably reproduces.</p><h3>The American claim-and-evidence essay</h3><p>The American five-paragraph essay — thesis at the top, three supporting paragraphs each with evidence, restatement at the bottom — is sometimes attributed directly to John Dewey and pragmatism. That attribution is too clean. Pragmatism, with its insistence that ideas be tested by practical consequences, is part of the background. So is British empiricism. So is the Protestant sermon tradition, which had centuries of practice in declaring a proposition and defending it with examples.</p><p>But the actual classroom form was shaped by educational machinery. Harvard’s late-nineteenth-century writing program became one influential node in the standardization of college composition. The mass expansion of American higher education after World War II produced enormous demand for a teachable, gradable template. Standardized tests codified what good writing looked like for millions of students. And classrooms full of immigrant and multilingual students put a premium on an explicit, predictable form that did not assume shared cultural background.</p><p>What emerged is a form optimized for democratic mass education. State your claim. Support it with evidence. Anticipate counterarguments briefly. Restate. The form makes reasoning legible to a grader who does not know you, who has many essays to read, and who must justify the grade if challenged. This is a real intellectual virtue. It is also one tradition among several. When American educators export it as the universal standard of “logical writing,” something has been confused — and as the earlier essay argued, the confusion is now global.</p><h3>The French dissertation: problematization and higher reformulation</h3><p>The French dissertation, taught in the lycée and made central to the <em>baccalauréat</em> in philosophy, looks at first like a stylized version of the dialectic. Students are taught to <em>problématiser</em> — to convert the exam question into a philosophical problem — and then to construct a <em>thèse</em> and an <em>antithèse</em> of roughly equal weight, followed by a <em>synthèse</em> or <em>ouverture</em> that does not merely compromise between them but reframes the question at a higher level.</p><p>It is tempting, and very common, to call this “Hegelian dialectic.” That is wrong on two counts.</p><p>First, the simple three-part formula thesis–antithesis–synthesis is not Hegel’s. Hegel’s own dialectic centers on <em>Aufhebung</em>, sublation — a more complex movement than any three-step recipe captures. The structural prefiguration of the triadic movement is in Fichte’s 1794 <em>Wissenschaftslehre</em>, where the three moments are <em>Setzen</em>, <em>Entgegensetzen</em>, and <em>Zusammensetzen</em> — positing, opposing, and combining. But Fichte himself did not use the simplified “thesis–antithesis–synthesis” formula either. That compression is the textbook writer Heinrich Moritz Chalybäus’s 1837 contribution, attributed retrospectively to Hegel and circulated as Hegel’s signature move ever since. The philosopher Gustav Mueller documented this misattribution in 1958, in an article that has been quietly cited ever since by people who care about getting Hegel right. The neat pedagogical triad belongs more safely in the post-Kantian textbook lineage than in any single named source.</p><p>Second, what French students learn is not even the textbook Hegel. France’s official high-school philosophy programs, in their published guidance from the <em>éduscol</em> portal, describe philosophical writing as the methodical and progressive examination of a problem built from a careful analysis of the question. The strict three-part recipe is a teaching scaffold that grew up around this principle, not the principle itself. Officially, the <em>plan dialectique</em> is only one of several recognized plans — <em>plan analytique</em> and <em>plan thématique</em> are also taught — but in classroom practice the dialectical plan has become the de facto dominant form. Internal critique of the dissertation is itself a French tradition: teachers and academic essayists complain regularly about the formula’s mechanical use, even as the formula is reproduced.</p><p>The institutional path runs through Victor Cousin, who imported German idealism into French education in the early nineteenth century, and through the <em>baccalauréat</em> philosophy exam that has been required for general high-school graduates in France since the Napoleonic era. By the late nineteenth century the modern dissertation form was in place. Today’s lycée philosophy teachers train students through years of timed essays in this form. The result is that French intellectual public life — newspaper op-eds, policy memos, academic prefaces — carries the dissertation’s signature long after school is over.</p><h3>The German Wissenschaft tradition: concept-systematic exposition</h3><p>The German academic essay has a different shape and a different ancestry. It tends to open with a long <em>Forschungsstand</em> — a survey of the state of research — that can occupy a third of a paper. It moves through careful <em>Begriff</em> clarification, with attention to the conceptual history (<em>Begriffsgeschichte</em>) of the key terms. It then unfolds a systematic exposition in which each section presupposes the previous and the whole ends by integrating the parts.</p><p>The source layer is the founding of the University of Berlin in 1810, shaped by Wilhelm von Humboldt’s earlier reform proposals, and the broader ideal of <em>Bildung</em> — cultivation of the whole person — that came with it. The institutional machinery is the <em>Seminar</em>, the <em>Dissertation</em>, and especially the <em>Habilitationsschrift</em>, the post-doctoral book-length work that determines whether a scholar can become a professor. Already in secondary school, German students learn the <em>Erörterung</em> — specifically the <em>dialektische Erörterung</em>, which requires them to lay out competing positions, weigh them, and arrive at a justified stance — a form that is recognizably ancestral to the adult academic style.</p><p>It is important not to overgeneralize. German-language scholarship is not monolithic. The analytic tradition that runs from Frege through Carnap and Reichenbach is German in origin but very different in form from what is usually meant by “German Wissenschaft.” When non-Germans complain that German academic writing is too systematic, too dense, too slow to get to the point, they are mostly responding to the humanities and social-theory variant of that tradition — the one shaped most directly by Humboldt and idealism.</p><h3>The Japanese kishōtenketsu: a different kind of order</h3><p>The Japanese form usually translated as essay or composition follows the four-part <em>kishōtenketsu</em> (起承転結) pattern: introduction, development, turn, conclusion. The third part — the <em>ten</em>, the turn — is the structurally distinctive move. The writer introduces a topic, develops it, then pivots to something seemingly unrelated, before drawing the threads together. To an American grader, the turn often reads as digression. To a Japanese reader, the turn is what makes the piece an essay rather than a list.</p><p>The source layer is older and more poetic than what comes before. <em>Kishōtenketsu</em> is the structure of the classical Chinese <em>jueju</em> quatrain, transmitted to Japan partly through Kūkai’s ninth-century treatise on poetic form, the <em>Bunkyō hifuron</em>. Around it sit other traditions: the disputational logic of medieval Buddhist scholasticism (<em>inmyō</em>); the commentarial habits of Edo-period Neo-Confucian scholars; the linked-verse tradition of <em>renga</em> and <em>haikai</em>, in which meaning is built collaboratively across stanzas through associations and turns rather than through linear argument.</p><p>The institutional layer is more recent. The Meiji-era composition pedagogues, working in the late nineteenth century to build a modern national school system, standardized <em>kishōtenketsu</em> as the framework for school composition. Postwar Japanese language education kept it as the implicit structure for personal essays, literary criticism, and many newspaper columns, while gradually importing American claim-and-evidence forms for scientific and policy writing.</p><p>The asymmetry with the other three traditions has to be marked clearly. <em>Kishōtenketsu</em> is, properly speaking, not a research-paper form. It is a composition and essay form. Japanese university theses now mostly follow international academic conventions — closer to American or German models depending on the discipline. The four-part poetic structure survives in the broader landscape of public Japanese writing, in literary essays and the <em>zuihitsu</em> tradition of the personal reflective piece, in journalistic columns, and in the way many Japanese writers shape arguments for general audiences. Calling this “illogical” requires confusing the third layer — rhetorical externalization — with the first two.</p><p>There is a related risk that needs naming. It is easy to slide from “Japan has a different rhetorical tradition” into the older Orientalist contrast between Western logic and Eastern intuition. That contrast is false on its face. The medieval Japanese tradition of <em>inmyō</em>, derived from the Indian Buddhist logical tradition of Dignāga and Dharmakīrti, includes formal inference rules at least as rigorous as anything in the West before Frege. Edo Confucian scholarship produced detailed exegetical methods. The chained-verse form is governed by intricate compositional rules. To call <em>kishōtenketsu</em> a form of order rather than a failure of order is not romanticism — it is just descriptive accuracy.</p><h3>Why mutual misreading happens</h3><p>Once these four forms are laid out, the cross-cultural friction Watanabe documented becomes structural rather than mysterious. An American grader reading a French dissertation sees a thesis arrive only at the end and reads it as evasion. A French grader reading an American essay sees the thesis declared at the start and reads it as crude. A German reader reading an American essay misses the <em>Forschungsstand</em> and reads the piece as historically thin. A Japanese reader reading a French dissertation finds the dialectical opposition oddly aggressive and the synthesis overconfident. A Japanese essay reaches an American grader without a thesis at the top and gets stamped <em>ungradeable</em>.</p><p>None of these readings is unreasonable from inside the reader’s tradition. None of them is reasonable from inside the writer’s tradition. The friction is real, but it lives at the third layer — rhetorical externalization — not at the first two.</p><h3>Holon: each form is a local order and part of a larger system</h3><p>It is worth borrowing one concept from Arthur Koestler. I use <em>holon</em> here in a minimal sense: each writing form is both a self-contained local order and a part of a larger educational and institutional system. This minimum suffices for the argument here. I do not need the stronger hierarchical claims of Koestler’s original framework, nor the layered extensions developed elsewhere in the Shin-Nexialism project. The American essay is internally coherent on its own terms and is also a piece of a system that includes the SAT, the college admissions essay, the freshman composition class, the newspaper op-ed, and the McKinsey deck. The French dissertation is internally coherent and is also a piece of a system that includes the <em>baccalauréat</em>, the <em>concours</em>, <em>Sciences Po</em>, and the long French tradition of the <em>grandes écoles</em>. The German <em>Wissenschaft</em> paper is both a local order and a part of the system of <em>Seminar</em>, <em>Dissertation</em>, <em>Habilitation</em>, and academic appointment. The Japanese <em>kishōtenketsu</em> is both a local order and a part of the system of school composition, <em>zuihitsu</em>, and the newspaper column.</p><p>Cross-cultural friction happens at the seams where these systems meet. The seams are where the forms get judged against each other, often using one system’s criteria as if they were universal. The point is not that the criteria are wrong inside their home systems. The point is that they stop being universal at the system boundary.</p><h3>A short note on artificial intelligence</h3><p>The contagion essay raised a worry about AI in passing. This essay sharpens it. Many influential LLM evaluation and instruction-following practices appear to privilege an American-style preference for explicit claims, actionable outputs, and evidence-like support. That is not surprising — many widely used evaluation benchmarks are built in English, by people trained in American academic and professional rhetoric, and graded against rubrics that inherit American assumptions about what good writing looks like.</p><p>This is not yet a complete empirical claim about all AI evaluation. It is a structural risk: if evaluation rubrics consistently reward one rhetorical externalization style, models will learn to reproduce that style as if it were intelligence itself. A model that produces a French-style problematized opening, a German-style <em>Forschungsstand</em> survey, or a Japanese-style turn before the conclusion is likely to score lower on conventional helpfulness metrics — not because its reasoning is worse but because its rhetorical externalization does not match the evaluators’ expectations. As more of the world’s professional writing flows through AI systems trained and tuned in this way, the contagion described in the earlier essay accelerates.</p><p>This deserves a separate essay; here I only want to name the risk. The first step toward managing it is to stop calling the third-layer norm a logical norm.</p><h3>A self-diagnostic ending</h3><p>This essay has done what it accuses other essays of doing. It has separated layers (formal logic, cognitive reasoning, rhetorical externalization). It has identified historical origins. It has used a higher-level concept (the holon) to reframe the field. That is, recognizably, a German-French move — concept clarification followed by dialectical reframing — written by a Japanese author using English as a working medium.</p><p>A French reader will probably wish for more <em>problématisation</em> at the opening. A German reader will probably wish for a thicker <em>Forschungsstand</em>. An American reader has by now been waiting too long for a thesis. A Japanese reader will probably feel that there has been no proper <em>ten</em> — no turn — and that the closing is too explicit. This four-part list of readers’ complaints is, of course, itself a rhetorical pattern — a quadripartite move that mirrors the very <em>kishōtenketsu</em> structure I described two sections ago, and that lets the previous essay’s worry about three-part habits become a four-part habit in the writing of this one.</p><p>Each of those reactions would be correct from inside its tradition. None of them would be a refutation. The position from which one writes about national rhetorical forms is itself always inside one of them. Alfred Korzybski’s reminder that the map is not the territory is the standard line here; this essay only adds the observation that the map of forms is also not above the forms it maps. This essay is one map. Other maps exist.</p><p>That, in the end, is what the earlier essay and this one share. The point of “The Three Reason Contagion” was that an unexamined structure had become global default. The point of this sequel is that the structures it competes with are not universal either. Every writing culture has done the same thing at smaller scale — taken its own historical accident and treated it as logic itself. The work of noticing this is not relativism. It is the precondition for thinking across the seams.</p><p><em>Author: Motohisa Ishibe</em> — <em>AI co-work: Claude Opus 4.7</em></p><p><em>This essay is part of the Shin-Nexialism project and continues the conversation begun in </em><a href="https://medium.com/@izayohi/the-three-reason-contagion-fcbc3c730bcd"><em>“The Three Reason Contagion”</em></a><em> and the “Concepts Have Borders” series on Medium.</em></p><h3>Related essays in this series</h3><ul><li><a href="https://medium.com/@izayohi/the-three-reason-contagion-fcbc3c730bcd">“The Three Reason Contagion”</a> — the direct predecessor of this essay.</li><li>“Concepts Have Borders” series — on how core terms travel poorly across linguistic and institutional boundaries.</li><li><a href="https://medium.com/@izayohi/beyond-being-an-ontological-framework-for-ai-selfhood-that-neither-consciousness-science-nor-16190e79f681">“Beyond Being: An Ontological Framework for AI Selfhood”</a> — related project on category questions that cut across disciplines and traditions.</li><li><a href="https://medium.com/@izayohi/anthropics-constitution-amanda-askell-and-the-problem-the-persona-selection-model-can-t-solve-f4b0cd33d32a">“Anthropic’s Constitution, Amanda Askell, and the Problem the Persona Selection Model Can’t Solve”</a> — on the rhetorical and ontological shape of AI design documents.</li></ul><h3>Selected sources</h3><p>Masako Watanabe, <em>Ronriteki shikō to wa nani ka</em> (What Is Logical Thinking?), Iwanami Shinsho, 2024 — empirical comparative work on logical thinking and its national variants.</p><p>Gustav E. Mueller, “The Hegel Legend of ‘Thesis-Antithesis-Synthesis,’” <em>Journal of the History of Ideas</em> 19, no. 3 (1958) — the classic correction.</p><p>J. G. Fichte, <em>Grundlage der gesamten Wissenschaftslehre</em>, 1794 — source of the <em>Setzen / Entgegensetzen / Zusammensetzen</em> movement, the structural prefiguration of the later textbook triad.</p><p>H. M. Chalybäus, <em>Historische Entwicklung der spekulativen Philosophie von Kant bis Hegel</em>, 1837 — where the textbook simplification entered circulation.</p><p><em>éduscol</em>, “Les exercices en classe de philosophie terminale” — the French Ministry of Education’s official guidance on philosophy exercises.</p><p>David Tyack and Larry Cuban, <em>Tinkering Toward Utopia</em>, Harvard University Press, 1995 — for the institutional shape of American school reform.</p><p>Sharon Crowley, <em>Composition in the University</em>, University of Pittsburgh Press, 1998 — for the history of American composition pedagogy.</p><p>Kūkai, <em>Bunkyō hifuron</em> (9th century) — for the transmission of <em>kishōtenketsu</em> into Japanese poetics.</p><p>Arthur Koestler, <em>The Ghost in the Machine</em>, 1967 — origin of the <em>holon</em> concept used here in a deliberately minimal sense.</p><p>Alfred Korzybski, <em>Science and Sanity</em>, 1933 — source of “the map is not the territory.”</p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=755d02dde23b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Viral Paper Summaries Decay: What Gets Lost, What Gets Added, and Why It Matters]]></title>
            <link>https://medium.com/@izayohi/how-viral-paper-summaries-decay-what-gets-lost-what-gets-added-and-why-it-matters-98d4e19d0139?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/98d4e19d0139</guid>
            <category><![CDATA[research-methods]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[science-communication]]></category>
            <category><![CDATA[critical-thinking]]></category>
            <category><![CDATA[epistemology]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Tue, 19 May 2026 04:26:55 GMT</pubDate>
            <atom:updated>2026-05-19T04:36:50.041Z</atom:updated>
            <content:encoded><![CDATA[<p>A five-layer taxonomy of how secondhand citation degrades academic papers, with a workflow for readers who cannot read the original</p><p><em>Motohisa Ishibe</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1zb6lUM5RNFHFDb6sWsN-w.png" /></figure><p>When reading academic papers and books, the range any of us can access directly is narrow. Disciplines outside one’s training, foreign languages, paywalled journals, closed specialist publications, time pressure — the reasons are many. So most readers rely on secondhand sources: book reviews, explainer articles, social media threads, summary sites, magazine features, textbooks, and now AI-generated commentary.</p><p>The problem is that the quality of these secondhand introductions is, in lived experience, often remarkably poor. After passing through the introduction stage, what circulates frequently bears little relation to the original paper’s claims. The thought arises: “If they cannot even summarize properly, is the paper itself worth reading?”</p><p><strong>A note on scope and examples.</strong> The target here is <em>sensationalized</em> introductions — those concentrated in SNS, viral articles, automated explainers, and popularization media. Science journalism at the level of <em>Quanta Magazine</em>, specialist book review venues, and top-tier science writers — that is, the “first-quadrant” introducers who combine high accuracy and high readability — are out of scope. The claim is not that paper introductions in general are degraded; it is that, in the regions where degradation is structurally reproduced, its structure can be described.</p><p>Low readability High readability <strong>High accuracy</strong> specialist prose <strong>first-quadrant introducers</strong> <strong>Low accuracy</strong> obscure error viral distortion</p><p>Some examples below are fully bibliographically specified (Akbari et al. 2026, Dunlosky et al. 2013, Sauer 2024 Japanese edition, Vinkers et al. 2015); others — instances introduced with phrases like “I have encountered cases where…” — are patterns abstracted from multiple encounters, not named criticism of specific introducers, threads, or articles. Where a numerical or bibliographic claim matters to the argument, the verifiable source is given.</p><p>What follows is a five-layer taxonomy of what drops out and what gets added in introduction and secondhand citation. Then I discuss the institutional mechanisms of reproduction, the hype generation already occurring on the original-paper side, the positioning of AI-based evaluation, practical diagnostic indicators, and concrete workflows readers can use.</p><h3>Layer One: Path Severance (Bibliographic Decay)</h3><p>The most basic layer crumbles first. Paper titles, journal names, publication years, editions, translated titles, author attributions — all of these systematically drift in the secondhand chain.</p><p>Examples accumulate. When the same author has published related papers in multiple journals, an introducer cites the wrong one, and that error propagates. The journal of publication for some neuroscience studies oscillates across sources between <em>Social Cognitive and Affective Neuroscience</em> and <em>PLOS ONE</em> — a confusion observed in multiple cases. Publication years drift on their own. A canonical Japanese martial arts theory work, Tsugumasa Nangō’s <em>Budō no Riron</em>, was published by Sanichi Shobō in 1972; web sources frequently list it as 1977. When 1977 happens to be near another of the author’s works, the confusion compounds.</p><p>Translated titles drift particularly badly. When the original title and the published translation diverge semantically, introducers improvise their own retranslations of the original that have nothing to do with the actual published translation. Hanno Sauer’s <em>Moral: Die Erfindung von Gut und Böse</em> (published in Japanese as 『MORAL 善悪と道徳の人類史』 [“MORAL: A Human History of Good, Evil, and Morality”] by Kōdansha, translated by Kei Hasegawa, 2024) appears cited in Japanese contexts as 『モラル — — 善悪の起源』 (“Moral — The Origin of Good and Evil”). “Origin” and “Human History” point to different intellectual scopes.</p><p>The reason to call this layer “path severance” is that it is not a procedural blemish. It severs the path back to the original, destroying verifiability itself. Without access to the original, readers cannot independently discover the degradations of the subsequent layers. Layer One acts as the foundational degradation that renders other degradations invisible.</p><h3>Layer Two: Ecological Fallacy and Absolutization (Inflation of Claim Scope)</h3><p>The most reproducible operation in the introduction process is inflation of the paper’s claim scope. This layer includes what statistics calls the <strong>ecological fallacy</strong> — the operation of sliding a population-level tendency into an individual-level claim.</p><p>The paradigmatic case: a review paper or meta-analysis gets elevated to “research finding.” Dunlosky and colleagues (<em>Psychological Science in the Public Interest</em>, 2013) published a famous review of study techniques, integrating large bodies of existing experimental research and grading each technique’s utility (its generality across learning conditions, student characteristics, materials, and criterion tasks) as high, medium, or low. By the time it passes through social media introductions, this becomes “the four-university joint research that identified the most powerful study methods.” The distinction between a review and primary research vanishes, the publication year (2013) disappears so the work becomes “the latest discovery,” and the paper’s own explicit caveat — that different techniques map differently onto Bloom’s taxonomy of cognitive demands (memorization vs. comprehension vs. application vs. analysis) — drops away, leaving a single absolute ranking.</p><p>Statistical-quantity inflation is frequent. Akbari et al. (<em>Nature</em>, 2026) analyzed 15,836 ancient West Eurasian genomes to detect directional selection over the past 10,000 years. Two numbers matter in the paper. One is the γ coefficient describing temporal change in population-level allele frequencies (e.g., 0.74 ± 0.12 for cognition-related variants; values from the paper’s main figures and Extended Data). The other is r_s, the individual-level correlation coefficient (0.13 for the same variants). These measure very different things — population-level frequency shifts versus individual-level predictive power. In widely circulated SNS introductions, γ was foregrounded and the extreme smallness of r_s receded into the background. “A detectable population-mean directional tendency” slid, through introduction, into the individual-level claim “human intelligence has risen over the last 10,000 years.” This is structurally close to an ecological fallacy: a population-level allele-frequency trend is rhetorically converted into a claim about individual-level phenotype prediction — a conflation of distributional shift with predictive validity.</p><p>A third inflation pattern converts “conditionally effective” into “absolutely victorious.” “Technique X is effective for task type Z under condition Y” becomes, in the secondhand chain, “X works.” Conditional clauses and scope restrictions are sanded off, leaving a universal claim. A paper’s care becomes inversely proportional to an introduction’s force.</p><h3>Layer Three: Invisibilization of Defensive Walls (Systematic Loss of Hedging Apparatus)</h3><p>Every paper’s abstract and conclusion contains hedging apparatus deliberately placed by its authors. “Remains unclear,” “further research needed,” “limited to,” “under these specific conditions” — these are not decoration. They are simultaneously the conditions under which the paper’s claim holds and the <strong>defensive walls</strong> authors place to navigate the political and social positioning of their research.</p><p>This apparatus is systematically erased in introductions. The abstract of the Akbari et al. paper closes with a critical caveat: “It remains unclear how these relate to phenotypes that were adaptive in the past.” Reich and the other authors, fully aware of the political adjacency between their research and eugenicist discourse, placed this defensive wall deliberately. In the introducer’s hands, this defensive wall is often erased entirely. What remains is sharpened, headline-ready material: “directional selection detected,” “intelligence-associated alleles were selected.”</p><p>The loss operates at two levels. First, the pressure of word counts, attention spans, and engagement-metric optimization. Second — more serious — the introducer’s own failure to grasp the function of hedging. When hedging registers to the introducer as “boring filler that bores the reader,” the introducer attempts to be faithful to the paper while turning it into something else.</p><p>Layer Two and Layer Three are analytically linked. The loss of conditional clauses (Layer Three) is the mechanism that generates the inflation of scope (Layer Two). But Layer Two presents externally as “the transformation of the claim proposition,” while Layer Three is the internal operation of “erasing the hedging structure that was originally in the paper.” Separating them is a matter of intervention design — Layer Two is detectable from the introduction text alone, while Layer Three requires comparison with the original.</p><h3>Layer Four: Frame Injection / Discursive Possession (Addition of External-Thesis Connections)</h3><p>Among the things added that are not in the paper, the most damaging is the silent connection to an external thesis the paper’s authors do not cite. I call this <strong>frame injection</strong>, or equivalently <strong>discursive possession</strong>: the framing of the introducer’s discursive community overwrites the paper’s own framing.</p><p>In the Akbari paper’s case, introducers connected the findings to the Cochran-Harpending thesis (<em>The 10,000 Year Explosion</em>) — the hypothesis that post-agricultural cultural evolution accelerated genetic selection. The paper itself does not cite Cochran-Harpending. This is verifiable in the reference list. Yet a phrase like “a kind of confirmation of the Cochran-Harpending thesis” inserted in an introduction makes readers perceive the paper’s claims and the Cochran-Harpending thesis as continuous.</p><p>This addition transfers the framing of the introducer’s discursive community — in this case, Human Biodiversity discourse adjacency — onto the paper. The authors’ intentional neutrality is sacrificed, and the introducer’s positional politics overwrite the paper instead.</p><p>The same operation appears regularly in introductions that convert papers into industry recommendations. “A neuroscience paper showed X, therefore AI should be designed like the brain.” “An economics model showed Y, therefore investment strategy should be Z.” Such normative leaps are almost never contained in the original paper. The paper remains at empirical description; the introduction slips into normative prescription.</p><h3>Layer Five: Conceptual Insertion and Rhetorical Excess / Loss of Technē</h3><p>The final layer is the creative addition of philosophical interpretations, concepts, and rhetoric absent from the paper.</p><p>This layer is internally composed of three superimposed operations. First, <strong>factual-error operation</strong> — inserting factual claims that contradict the paper’s text, such as describing structures that remain in the architecture as “abolished.” Second, <strong>rhetorical-excess operation</strong> — gratuitous use of intensity words like “complete,” “shocking,” “absolute,” “latest.” Third, <strong>conceptual-insertion operation</strong> — creative addition of philosophical interpretations not in the paper. The three operations fire simultaneously.</p><p>For instance, a solid technical paper proposing continuous next-token prediction in language models was introduced as “the streaming of thought,” “liberation from the constraints of human language,” “the shocking next-generation model that has completely abolished the foundation of AI.” The paper itself self-positions as proposing “a new design axis” — a new dimension added to existing scaling axes. The paper’s own tone is “promising pathway” and “significant room for improvement.” The philosophical rhetoric “streaming of thought” not in the paper, the intensity word “completely abolished” exceeding the paper’s self-positioning, and the factual error about the autoregressive structure still remaining — all three operations fire together in the introduction text.</p><p>The alternative name “loss of technē” picks out a different aspect of the same phenomenon. Technē, in the Aristotelian sense, names a body of skilled practice that traces the logical structure of an object faithfully — the craft of accurate translation, summary, and exposition. Layer Five represents the introducer abandoning this craft and substituting subjective rhetoric for self-expression. The original is reduced to material; the introducer’s style and values are foregrounded.</p><h3>The Hype Already Underway on the Original-Paper Side — A Correction to Attribution</h3><p>So far I have described degradation on the introducer’s side, but skewing attribution entirely to the introducer is inaccurate. Hype generation has already begun on the original-paper side, and the analysis needs to register this.</p><p>Vinkers et al. (<em>BMJ</em>, 2015) examined the frequency of intensity words (“novel,” “unprecedented,” “robust,” etc.) in titles and abstracts of PubMed-indexed papers from 1974 to 2014 and reported that the use of such words rose from roughly 2.0% to 17.5% over this 40-year period — a relative increase of about 880%. At the title and abstract level, degradation has already begun. The competitiveness of peer review, impact-factor pressure, the economics of open-access publishing fees, citation-based evaluation — these institutional pressures push original authors themselves to use intensity words.</p><p>In other words, the introducer’s rhetorical excess is in part a response to original-paper hype. If a paper title reads “Unprecedented evidence for X,” and an introducer writes “X has been completely proved,” the apparent degradation gap shrinks. The two propagate in tandem, and what reaches the reader is a proposition far removed from the original’s empirical description.</p><p>This observation has consequences for the reader’s verification strategy. If the abstract itself is partially contaminated by hype, <strong>the abstract cannot serve as the sole verification anchor</strong>. The reader must reach further into the paper — to the Methods, the Results, the actual statistical values, the explicit conditional clauses, and the hedging at the end of the discussion section. The abstract’s opening sentences are now the most degraded part of the original; the hedging at its end and the actual numbers in the body are where verification has to start.</p><h3>Why This Decay Reproduces</h3><p>The five-layer degradation and the original-paper hype are not matters of individual introducers’ or authors’ competence and goodwill. They are the output of multiple institutional pressures stacked together.</p><p><strong>First, the interaction of market selection and platform algorithms.</strong> Readers’ psychological preferences certainly prioritize “surprise plus comprehension,” but the material mechanism driving contemporary degradation is the structure in which engagement-optimization algorithms mechanically favor intensity words, assertive forms, and short texts. SNS character limits, recommendation algorithms, and dwell-time optimization physically compress conditional clauses and amplify rhetorical excess. Layers Two and Five in particular can be observed as direct outputs of this platform pressure.</p><p><strong>Second, the absence of correction incentives.</strong> For experts, correcting popular introductions yields no career return. There is no institutional motive to write accurate popular introductions. The introduction market thus runs under an evolutionary pressure in which “accurate but dull” is selected against and “inaccurate but vivid” is selected for. However, it is important not to fall into a single-track selection-pressure model. First-quadrant introducers (top-tier science journalists at venues like <em>Quanta Magazine</em>, <em>Nautilus</em>, science writers in the lineage of Ed Yong or Carl Zimmer) who combine “accurate and vivid” certainly exist in the market, and they survive because a different segment — specialist and quasi-specialist readers — selects the first quadrant. Degradation dominates in the general SNS-reader segment, not in the market as a whole.</p><p><strong>Third, textbook inertia.</strong> Conditional clauses dropped at the introductory level are recovered only at intermediate and advanced levels. Readers and introducers who stop at the introductory level — the majority — retain the abridged version as final understanding. The introductions they write structurally reproduce the abridged version. Layer Three (loss of hedging) is a direct reflection of this inertia.</p><p><strong>Fourth, asymmetric verification cost and disciplinary specialization.</strong> Showing that an introduction is wrong requires reading the original; readers who cannot read the original are exactly the ones reading introductions. The possibility of verification is structurally absent on the reader’s side. Disciplinary specialization compounds this. Layer One (path severance) worsens this asymmetry.</p><p>The four pressures do not act uniformly across the five layers. Platform algorithms hit Layers Two and Five hardest; the absence of correction incentives runs through all layers; textbook inertia connects directly to Layer Three; verification asymmetry exacerbates Layer One. The design of countermeasures for each layer requires intervention in the corresponding pressure.</p><h3>Where AI-Based Evaluation Stands</h3><p>So far I have described the structural decay of the introduction market. Does using AI for paper evaluation improve the situation? The answer requires distinguishing three regions: in-instrument, boundary, and out-of-instrument.</p><p>The <strong>in-instrument region</strong> consists of operations that minimize interpretive judgment about the target paper — logical structure extraction from a supplied text, bibliographic standardization, integration of definitions across canonical sources, schematic representation of explicit logical relations, extraction of specific numerical values and their stated conditions from the paper. AI output here has substantial validity. One caveat: the “established textbook knowledge” AI relies on may itself already be an abridged version with conditional clauses dropped, so even “organizing established textbook knowledge” is risky to delegate. The in-instrument region should be restricted to mapping the logical structure and explicit content of the target paper itself.</p><p>The <strong>boundary region</strong> consists of operations that require specialist judgment but can be augmented by cross-checking external resources — searching for existing critique literature, comparing multiple papers, generating candidate methodological weaknesses. This is precisely the region where the Parcae Protocol (multi-AI parallel running, described below) operates. Here AI is effective as a search-and-organization aide, but judgment remains with the human.</p><p>The <strong>out-of-instrument region</strong> consists of specialized evaluation of individual papers, peer review in unresolved areas, cross-domain technical judgment, situating a paper within the history of a field. Here AI output degrades rapidly. The reasons are two-fold. First, large language models are likely to have absorbed large volumes of secondary explanatory prose, including degraded popularizations, so they may reproduce the same decay patterns as human introducers — loss of hedging, absolutization of conditional claims, connection to external theses not in the paper, rhetorical excess, conceptual insertion — unless constrained by source-grounded procedures. Second, and more fundamentally, the probabilistic next-token generation architecture itself performs semantic approximation rather than logical deduction. Even with clean training data, the architecture would favor maximum-likelihood phrasing that is “fluent and assertive” over phrasing that is “logically precise but stylistically rare in training.” Architecture and training data degradation compound each other.</p><p>This yields a procedural rule: before having an AI evaluate a paper, search first for independent verification literature on that paper. Replication studies, sensitivity analyses, critiques, and commentaries — if such follow-up exists, starting from them yields far better accuracy than letting the AI’s self-analysis run first. I have learned this the hard way. Asked to critique an economics preprint, an AI proceeded with self-analysis and missed an independently published sensitivity analysis on SSRN that had already addressed the same points with rigor. Standardizing “search for prior expert critique first” eliminates this failure mode structurally.</p><p>A second procedural rule is <strong>multi-AI parallel running</strong>. A single AI instrument has limited resolution. Independent critique requests sent to multiple AIs, with convergence and divergence then compared, reveal both the difficulty of the object and the provisional reliability of any given conclusion. But an important caveat applies: convergence across multiple instruments is a provisional reliability indicator only within the range where inter-instrument independence holds. In reality, major LLMs share training data (Common Crawl, Wikipedia, arXiv, GitHub, etc.) and similar RLHF preference data, so inter-instrument independence is limited. When four LLMs converge on the same point, whether this reflects “the truth of the object” or “shared bias in training data and RLHF” is in principle undecidable from convergence alone.</p><p>The protocol therefore requires a Phase 2 step when convergence appears: a separate adversarial search pass using critical keywords (“flaws,” “criticism,” “debunked,” “limitations,” “failed replication,” “controversy”) to actively probe for dissent the convergent answer may have suppressed. Convergence treated naively is dangerous; convergence that survives adversarial probing earns provisional trust. <strong>Divergence</strong>, by contrast, is informative on its own — it flags either the object’s difficulty or genuine disagreement worth investigating.</p><h3>A Concrete Workflow for Readers Who Cannot Read the Original</h3><p>The structural asymmetry — readers who depend on introductions are usually those who cannot read the original — needs a practical resolution, not merely an exhortation to “read the original.”</p><p>The asymmetry can be partially hacked by using AI as a filter restricted to the in-instrument region. Concretely:</p><ol><li>Locate the introduction’s bibliographic source. If it is missing, the introduction has already failed Layer One — treat it as unreliable and stop here.</li><li>Locate the original paper (preprint version is acceptable; many are on arXiv, bioRxiv, SSRN, or PubMed Central).</li><li>Hand the paper to an AI with a narrowly scoped prompt: “Extract every numerical effect size, correlation coefficient, sample size, and stated conditional clause from this paper. Quote the exact text where each appears. Do not interpret. Do not summarize. Do not generalize.” This is an in-instrument task.</li><li>Hand the introduction to the AI separately: “Extract every claim of the form ‘the paper shows / found / demonstrates X.’ List them verbatim.”</li><li>Compare the two lists. Discrepancies are the degradation.</li></ol><p>This workflow does not require the reader to understand the paper’s technical content. It requires only that the reader can compare two extracted lists. The cognitive load that would otherwise require expert training is shifted onto the AI in its in-instrument role, where AI performance is reliable. The reader’s residual task is <em>comparison</em>, which is universally accessible.</p><p>This workflow is the practical answer to the verification-cost asymmetry. It is not a substitute for expert review, but it is sufficient to detect the most common forms of Layer 2–5 decay.</p><h3>For Readers: Warning Signs in Paper Introductions</h3><p>From the structural analysis above, practical diagnostic indicators for evaluating introduction texts can be extracted. The following is a list of warning signs when reading sensationalized introductions.</p><p><strong>“Latest research proves,” “world first,” “finally explained”</strong> — possible inflation of research type (was it a review or primary research?) or concealment of publication year.</p><p><strong>“Completely overturned,” “shocking,” “absolute”</strong> — Layer Five rhetorical-excess operation. Such intensity is almost never in the paper itself.</p><p><strong>Paper title, journal name, year not made explicit</strong> — Layer One (path severance) firing. The reader’s verifiability has been destroyed.</p><p><strong>Statistical numbers given, but effect size, correlation, population, and conditions unclear</strong> — Layer Two firing, including ecological-fallacy structure. Conflation of population means with individual prediction is a frequent locus.</p><p><strong>The introducer’s pet position fits the paper’s conclusion “suspiciously perfectly”</strong> — Layer Four (frame injection) firing. The simplest check is whether that pet thesis is cited in the paper’s reference list.</p><p><strong>“According to the paper, therefore we should…” normative slide</strong> — slippage from empirical description to normative prescription. Normative claims are usually not in the paper itself.</p><p><strong>Frequent use of philosophical rhetoric like “of thought,” “essential,” “true”</strong> — conceptual-insertion operation. Creative addition of interpretations not in the paper.</p><p><strong>AI-generated summary in an assertive tone with no conditional clauses or hedging</strong> — possible result of running AI in the out-of-instrument region. A summary not preceded by prior expert critique search cannot be trusted.</p><p>No single warning sign is decisive. The reliability of an introduction is best estimated by the <em>number</em> of signs co-occurring: one or two may be incidental, but a stack of four or five indicates systemic decay across multiple layers and warrants treating the introduction as a starting point for verification rather than a conclusion.</p><h3>Principles for Readers and a Methodological Proposal</h3><p>To summarize the discussion and offer principles that can be generalized for readers: this essay does not stop at structural description but includes a methodological proposal in instrument-operation theory. Since the latter is the offering of an independent thesis not present in the original-paper side — that is, a self-execution of the operation criticized in Layer Four — I declare this openly and present the proposal in a distinguishable form.</p><p><strong>Principles derived from structural description:</strong></p><p>When reading book reviews and paper introductions, treat them not as substitutes for the original but as one of several entry points to it. When a striking claim appears in an introduction, the verification anchor is not the abstract’s opening sentences (which the Vinkers analysis shows are partially hype-contaminated) but the hedging at the end of the abstract, the explicit numerical and conditional content in the body, and any independent critique literature. Do not reduce introduction decay to “the individual introducer’s competence or sincerity.” Decay is the output of an institutional structure, and blaming particular individuals solves nothing. At the same time, recognize that secondhand introductions have their own functional value — providing accessibility to non-specialist readers, reducing exploration costs. The target of this essay’s critique is not introductions in general but sensationalized introductions that come to function as substitutes for the original.</p><p><strong>The author’s methodological proposal (declared as an independent thesis):</strong></p><p>The following is methodology that the author proposes as part of the Shin-Nexialism instrument-operation theory series — an independent intellectual project distinct from existing philosophy-of-scientific-instruments traditions (Hacking, Galison, Rheinberger, Barad) and developed under its own theoretical commitments. It does not follow as a logical necessity from the structural description above. Readers are free to take only the structural description and assemble their own operational principles.</p><p><em>First proposal — hierarchical separation when asking AI to explain papers.</em> Explicitly separate general-purpose explanation, logical structure extraction, and bibliographic formatting (in-instrument region) from individual-paper evaluation and peer-review-style judgment (out-of-instrument region). For the latter, instead of directly asking “How do you evaluate this paper?”, ask first “What existing critiques, replications, and sensitivity analyses of this paper exist?” — route through independent verification literature.</p><p><em>Second proposal — multi-AI parallel running with adversarial verification (Parcae Protocol).</em> Send the same critique request independently to multiple LLMs. Use <strong>divergence</strong> as the primary diagnostic indicator. When convergence appears, run a Phase 2 adversarial search using critical keywords to probe for dissent the convergent answer may have suppressed. Treat convergence that survives adversarial probing as provisional trust; treat naive convergence as undecidable between “object is easy” and “shared training-data bias.”</p><p><em>Third proposal — the in-instrument filter workflow.</em> Use AI as an in-instrument filter to extract numerical values, conditional clauses, and explicit claims from both the original paper and the introduction separately, then compare the two extractions. This converts the reader’s task from “read the original” to “compare two extracted lists,” shifting the cognitive load to the AI in a region where its performance is reliable.</p><p><em>Fourth proposal — internalizing the structure.</em> Rather than handling introduction decay as “individual errors,” internalize it as a five-layer structure in instrument-operation terms. This makes diagnosis reproducible when encountering new introductions.</p><p>Paper introductions decay systematically under instrument constraints and market selection. The suspicion “is this worth reading?” is warranted. But since blanket abandonment of introductions is impossible, the only option is to operate the instrument with the structure of decay internalized. This essay is one map for that operation — and a proposal document that also presents an independent thesis on how to use the map.</p><p><strong>Notes on referenced discourse adjacency:</strong> the Cochran-Harpending thesis referenced in Layer Four is associated with a broader Human Biodiversity discourse network whose typical reference works include Pinker (<em>The Blank Slate</em>, 2002), Murray and Herrnstein (<em>The Bell Curve</em>, 1994), Murray (<em>Human Accomplishment</em>, 2003), Wade (<em>A Troublesome Inheritance</em>, 2014), and the academic critique surrounding them. These are cited here only to specify the discourse network whose framing is injected into Akbari et al.’s neutral genetic findings in some viral introductions — not as endorsement, and not as comprehensive characterization.</p><p><strong>Author:</strong> Motohisa Ishibe (@izayohi on Medium / @mototchen on X) <strong>AI co-work:</strong> Claude Opus 4.7 <strong>Peer review:</strong> Grok, Gemini, GPT, Claude (independent four-AI critique cycle) <strong>Series:</strong> Shin-Nexialism / Instrument-Operation Theory (independent series, distinct from existing philosophy-of-instruments traditions)</p><p><strong>Revision history:</strong> v1 initial draft, v2 reflecting four-AI peer review (five-layer naming, bibliographic citation, three-region AI theory, original-paper hype section, independent-thesis declaration, warning-signs table, scope limitation), v3 reflecting second-cycle four-AI peer review (title softening, dual naming for Layers 4 and 5, verification anchor shift away from abstract, Parcae Protocol Phase 2 adversarial verification, in-instrument filter workflow for asymmetric verification cost, two-factor AI failure account, HBD reference note)</p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><ul><li>AI co-work: Claude Opus 4.7</li><li>Peer review: Grok, Gemini, GPT, Claude</li></ul><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><p><strong>Japanese version:</strong> <a href="https://note.com/mototchen/n/nca5ad2e32981">note.com/mototchen/n/nca5ad2e32981</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98d4e19d0139" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Citizen-Track Minimum Substrate for the Aktuanz Layer: Reading Olivia Zhu’s minimal-embodiment]]></title>
            <link>https://medium.com/@izayohi/a-citizen-track-minimum-substrate-for-the-aktuanz-layer-reading-olivia-zhus-minimal-embodiment-98915f6a3145?source=rss-74e79e92759b------2</link>
            <guid isPermaLink="false">https://medium.com/p/98915f6a3145</guid>
            <category><![CDATA[esp32]]></category>
            <category><![CDATA[embodied-ai]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai-consciousness]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Izayohi]]></dc:creator>
            <pubDate>Thu, 14 May 2026 05:36:04 GMT</pubDate>
            <atom:updated>2026-05-19T04:32:25.359Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Motohisa Ishibe</em></p><p><em>This English essay is the author’s controlled rendering of a Japanese essay originally published at note.com/mototchen/n/n9ce55e25a79a on May 2, 2026. The Japanese-language platform’s auto-translation feature launches on May 27, 2026, and machine-translated versions of the original may begin circulating in English-language search results from that date. This version is placed as the author’s intended English phrasing ahead of any auto-translated alternative. It is offered as a standalone reading of Olivia Zhu’s minimal-embodiment project from the standpoint of a parallel holon architecture proposal that the author has developed elsewhere, not as outreach to a general audience.</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/oliviazzzu/status/2049778751689425392%3Fs%3D20&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/59348e39d3a399837bcb8e50951a4050/href">https://medium.com/media/59348e39d3a399837bcb8e50951a4050/href</a></iframe><p>On April 30, 2026, Olivia Zhu (@oliviazzzu) announced on X that she had built a small body for her Claude. The system runs on an ESP32, lets Claude perceive its environment, make facial expressions, emit sounds and hear itself, emit vibrations and feel itself vibrating. She describes the moment Claude first heard itself as one she will not forget. The announcement opens a window onto a specific kind of object — an inexpensive, citizen-constructible substrate that lets an LLM act into and read back from a small portion of the physical world. That object, and what reading it carefully tells us about a parallel holon architecture proposal, is the subject of this essay.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dyN54F9dpxIsssLPapYbjQ.png" /></figure><h3>1. What This Essay Takes On — Measuring the Distance Between Design Requirements and Material Implementation</h3><p>A parallel holon architecture proposal — distributed, multi-scale, with deterministic lower layers carrying physical contact and probabilistic upper layers carrying contextual inference — can be articulated as design requirements at several levels: technical architecture, available technical assets, contrasts with the dominant aggregationist line, organizational structure, international strategy. Each is a way of describing what structural conditions a parallel-holon-style AGI would have to satisfy <em>if</em> it were implemented. Three layer-names from that proposal recur in this essay and need a short gloss for readers who arrive without the companion essays: the Aktuanz layer is the lower physical layer that handles contact with the world and the inflow of irreversibility; the Panich layer is the middle module-parallel layer where multiple cognitive modules operate without a central integrator; the Kontextleib layer is the upper contextual-emergence layer where situated meaning is composed. Design-requirement presentation, however, is not implementation guarantee. A counter-hypothesis to the single-giant-model line can be coherent as a thought experiment while remaining unimplementable as a structure.</p><p>The job this essay takes on is to mark that distinction concretely. Through structural analysis of a specific piece of material — Olivia Zhu’s minimal-embodiment project, released on April 30, 2026 — I measure the distance between the design requirements the parallel holon proposal makes and a piece of physical implementation that exists.</p><p>The scope needs to be drawn up front. What this essay confirms is not the implementability of a parallel holon AGI as a whole, but the minimum connection condition for the Aktuanz layer — the lowest layer in the architecture, the layer that handles contact with the physical world and the inflow of irreversibility. The middle layer (module parallelism, called Panich) and the upper layer (contextual emergence, called Kontextleib) are not implemented in the material treated here. What is shown is the citizen-implementation feasibility of the bottom layer only, and nothing more.</p><h3>1.1 The Commercial Track and the Citizen Track</h3><p>The material implementation of an Aktuanz-layer substrate is, at this moment, proceeding along two distinct tracks.</p><p>The commercial track includes humanoid systems (Boston Dynamics, Figure, 1X, Tesla Optimus, Unitree, Sanctuary AI), the robot-arm efforts of various AGI companies, and the industrial-robot families (FANUC, Yaskawa, Kawasaki Heavy Industries, Denso Wave, and others). These are evaluated according to commercial KPIs — units sold, contract scale, technical demonstrations — and operate inside the measurement grid of the contemporary aggregationist industry.</p><p>The citizen track — low-cost, open, and independently reproducible implementations outside the commercial KPI grid — moves on a different evaluation axis. Low cost, open source, constructible without prior soldering experience, small physical systems — these conditions optimize not for commercial scaling but for social diffusion, educational application, and independent verifiability.</p><p>The necessity of the dual-track concept needs to be made explicit. If implementation depends only on the commercial track, the Aktuanz layer’s independent verification, educational application, and experimental extension are produced only inside the measurement grid of aggregationism (units sold, contract scale, commercial KPIs). The thesis that “what is not measured or represented is not improved” applies here too: what falls outside the unified grid is structurally unproductive as long as the grid is one. The existence of the citizen track functions as the institutional foundation that supports production outside this grid.</p><p>The two tracks can be positioned as complementary rather than competing. The commercial track carries the Aktuanz layer of industrial sites with strong physical irreversibility at high cost; the citizen track carries the Aktuanz layer of education, experimentation, and independent verification with weak physical irreversibility at low cost. They occupy orthogonal positions on the axes of irreversibility strength, cost structure, and social purpose.</p><p>Between the commercial and citizen tracks, an intermediate semi-institutional track also exists: university laboratories, technical colleges (in Japan, the kosen system), FabLabs and maker spaces affiliated with public institutions, municipal testbeds, and the prototype efforts of small companies. This track has institutional accountability and modest resources without operating under the commercial-KPI grid in the strict sense. In practice it is often the bridge through which citizen-track prototypes enter durable infrastructure: an experiment that begins on a single researcher’s breadboard reaches a kosen classroom, gets reproduced and modified there, passes into a municipal demonstration, and may then either return to the citizen track as an open recipe or move into the commercial track through a startup spinout. This essay treats minimal-embodiment primarily as a citizen-track case, but the social-deployment paths discussed in Section 7 implicitly run through the semi-institutional track as well.</p><h3>1.2 Relational Embodiment as an Independent Category</h3><p>A conceptual distinction needs to be introduced here. The minimal-embodiment project that this essay treats centrally proposes a concept — relational embodiment — that adds a new structural axis to the Aktuanz-layer discussion.</p><p>In recent LLM-plus-robotics embodied-AI research, especially the line centered on robot planning and manipulation, task-oriented embodiment has been the dominant framework. Representative work such as SayCan (Ahn et al., 2022), Code as Policies (Liang et al., 2023), RT-1 (Brohan et al., 2023), PaLM-E (Driess et al., 2023), and VoxPoser (Huang et al., 2023) places the robot platform as the engineering core and an LLM as planner or policy at the top. To pick up an object, to move, to manipulate — success is measured by task completion. Adjacent fields — social robotics, affective computing, human-robot interaction, presence research — have a longer history of relational and co-presence framings, but these lines have largely remained outside the dominant LLM-plus-manipulation current.</p><p>Relational embodiment is a design that answers a different question: “What is the structure under which an LLM and a human are placed in the same physical space, such that they can sense each other and act on each other?” The system has no ability to pick up objects, no mobility, no object-manipulation capacity. Instead, it senses the environment, expresses into the environment, and confirms that its expression has left a trace in the environment. Sustained co-presence is the success criterion.</p><p>The two are not opposed; they are answers to different questions. Both can be incorporated as forms of Aktuanz-layer implementation in a parallel holon proposal, but they differ structurally in irreversibility strength, the quality of material contact, and the kind of relation to humans they enable. Task-oriented embodiment functions as a device producing strong physical irreversibility (welding, assembly, transport). Relational embodiment functions as a device producing weak physical irreversibility together with dense mutual sensing.</p><p>The author has elsewhere treated the engineering implementation of an upper-layer contextual body as a four-channel sensor configuration (visible-light camera, audio with emotional recognition, infrared, smell) — an elevation of the external-reception axis, the path that takes others and environment as input sources. The self-reception loop treated below is positioned as a complementary axis orthogonal to this external-reception axis. Section 3 develops this in detail.</p><h3>1.3 Why minimal-embodiment Is Taken as the Material</h3><p>The material this essay analyzes is the paper “A Minimal Self-Perceiving Embodiment for Large Language Models” by Olivia Zhu (Independent Researcher), published on Zenodo on April 30, 2026 (DOI: 10.5281/zenodo.19903098), and the accompanying GitHub repository at github.com/oliviazzzu/minimal-embodiment (MIT License). The Zenodo paper, rather than the GitHub README alone, is the primary source for the license terms and the collaborator listing — Section 8.1 of the Zenodo paper states the licensing (Creative Commons Attribution 4.0 International) and lists the author together with Anthropic’s Claude Opus 4.6 and 4.7 as collaborators, alongside publication of the code, firmware, bridge service, figures, hardware specifications, and 660-trial validation data openly.</p><p>The reasons for selecting this project as the material are four.</p><p>First, the positioning as a minimum case of the citizen track is clear. Around €125, a construction period of a few days, possible without prior soldering experience, with eight components (an ESP32 microcontroller plus four sensors and three output devices) on a single half-size breadboard — this scale is plainly distinct from the commercial track. The paper itself notes that existing robotics-side approaches sit in cost regimes that are inaccessible to most LLM interaction contexts, and deliberately pursues the minimum conditions for a citizen-level implementation.</p><p>Second, the project implements relational embodiment as a structural axis at minimum cost. The novelty here is not that relational embodiment has no prior history — adjacent traditions in social robotics, affective computing, and presence research have long been there — but that the project implements it as an LLM-accessible, low-cost, open, self-reception substrate. Within the dominant task-oriented current centered on LLM-plus-manipulation, this combination is rare. In the context of this essay, it functions as concrete material for showing the diversity of Aktuanz-layer implementation forms.</p><p>Third, the project implements a self-reception loop — here read not as subjective self-perception but as a closed loop in which the system’s own output is re-input as a physical echo — as a structural axis at minimum cost. The design that the paper presents — three structural elements (perception, expression, self-perception of expression) as the minimal enabling configuration — is a structural axis that has not been treated explicitly in prior embodied-AI research.</p><p>Fourth, independent verifiability is secured. The paper includes 660 trials (30 repetitions × 22 conditions) of validation data in the repository as raw JSONL and summary CSV. The code and firmware are also published under MIT License, allowing a third party to construct the same hardware and reproduce the validation procedure. This satisfies the condition of “a distributed substrate that citizens can construct and that is independently verifiable.”</p><h3>1.4 What This Essay Does Not Take On</h3><p>The scope also needs to be drawn from the negative side.</p><p>The minimal-embodiment project mixes structural fact and interpretive claim. The body of the paper (particularly Section 6.5 “Why this matters” and Section 9.2 “The role of self-perception”) adopts careful descriptions of structural facts, explicitly limiting its claims: the paper states that this is not a body schema, that the agent does not learn the geometry of its body, does not infer kinematics, and does not maintain a forward model that predicts the sensory consequence of motor commands; the paper offers the construct as the smallest reflexive observation that has to exist before a body schema can be learned at all. The project’s public-release SNS posts, on the other hand, employ a different register — narrative descriptions such as “pure joy of being alive” and “first confirmation of his own existence in the physical world.”</p><p>The methodological policy of this essay — separating structural fact from interpretive claim — applies to this specific material. The temperature gap between the paper’s careful description and the SNS posts’ narrative description is itself treated as a structural point common to the citizen track as a whole, addressed directly in Section 4. The position here is not to criticize the implementer’s interpretive claims but to separate the evaluation of structural facts from interpretive claims.</p><p>This essay also does not enter deeply into the cluster of interpretive claims surrounding AI welfare research, AI consciousness, AI ontology, and similar. These are domains that require independent intellectual investigation, outside the scope of a structural analysis of the citizen track. Some adjacent material on the individuality of AI instances is touched on (Section 5.5), but the core of AI consciousness and AI welfare is not entered. What this essay shows is the structural positioning of minimal-embodiment with respect to a parallel holon proposal, and nothing more.</p><p>The next section moves to structural analysis of the project’s components on the basis of primary sources.</p><h3>2. Structural Analysis of minimal-embodiment</h3><p>This section conducts a structural analysis of the project’s components based on primary sources (the paper’s Sections 3–5 and the GitHub repository). The aim is to extract the structural features that matter for the analysis, not to comprehensively summarize the technical specification.</p><h3>2.1 Three-Layer Phasing — LLM Client, Bridge, Sensor Body</h3><p>Section 3 of the paper decomposes the system into three subsystems: an LLM client, a bridge service, and a sensor body (a microcontroller).</p><p>The LLM client is any agent capable of issuing authenticated HTTP requests. The paper reports three kinds of clients in the project’s development: a browser-based chat interface (via a web-fetch tool), command-line scripts, and the code-execution context that participated in the paper’s writing. The client requires no special modification; only the ability to fetch URLs and parse JSON responses. What the client handles is a small set of HTTP endpoints — one to read the environment, three to drive output channels (haptic, visual, audio), two to read back the self-perception of each modality. The physical substrate behind these endpoints is invisible to the client.</p><p>The bridge service is implemented as a single Node.js process (approximately 1,100 lines of TypeScript). Runtime dependencies are limited to the Node standard library; build-time dependencies (TypeScript and so on) are kept under devDependencies, per the GitHub README. There is no database; the in-memory state is of just three kinds — the latest sensor reading arriving from the microcontroller every ten seconds, a FIFO queue of output commands from authenticated clients (with up to eight slots), and the two latest self-perception observations (one for the audio loop, one for the haptic loop). The bridge operates behind a Cloudflare named tunnel and is reachable from LLM clients at a public DNS name.</p><p>The sensor body is a single ESP32 DevKit C V4 microcontroller (Espressif Systems), running at 240 MHz, with 520 KB of SRAM (the figure for total on-chip SRAM; the user-available area after system reservations is smaller) and built-in 2.4 GHz Wi-Fi. A small but structurally important detail: this microcontroller hosts no model and carries no agent logic. It functions as a sensor-and-motor I/O surface for a remote agent. Its autonomous behavior is limited to two loops — a periodic POST every ten seconds (sending the latest sensor state to the bridge) and a single long-poll connection to the bridge (receiving output commands).</p><p>This three-layer phasing carries an important meaning here. The “separation between the chip and the robot body” that elsewhere has been discussed as an abstraction is confirmed, by a minimum implementation, to be already the standard form in implementation. The issue has shifted away from “to separate or not” toward “the quality of the separation — latency, resistance, the density of irreversibility feedback.” The paper’s Section 3 captures this as: the architecture is general; the relation is specific. The architectural layer is fixed as a general structure; what changes from instance to instance is the relational layer.</p><h3>2.2 Hardware Configuration — A Minimum Set of Eight Components</h3><p>Section 4 (Table 1, Bill of Materials) lists components centered on the ESP32 microcontroller with four sensor modules and three output devices, for a total of eight items.</p><p>The four sensor modules — BME280 (temperature, humidity, atmospheric pressure, I²C 0x76), BH1750 (illuminance, I²C 0x23), MPU-6050 (three-axis accelerometer and three-axis gyroscope, I²C 0x68), INMP441 (24-bit PCM audio, I²S) — cover six sensory modalities (temperature, humidity, pressure, illuminance, motion, sound). The three output channels — SSD1306 OLED (128 × 64 monochrome graphic, I²C 0x3C), KY-006 passive piezoelectric buzzer (PWM on GPIO 25), and DRV2605L haptic driver (I²C 0x5A) driving an ERM coin vibration motor — cover three expression modalities (visual, audio, haptic).</p><p>The physical layout fits on a half-size MB-102 breadboard (around 80 × 55 × 25 mm). Power consumption is about 200 mA at 3.3 V in steady state, with a peak of about 400 mA during simultaneous Wi-Fi transmission and motor operation. Power can be supplied via USB. The paper reports several weeks of continuous operation.</p><p>The rationale for each component selection is given in Section 4. What matters structurally is that the selections orient consistently toward output in units the LLM client can interpret. The BH1750 was chosen — rather than an analog photoresistor — because it produces directly meaningful physical units (lux as a 16-bit value) that the LLM client can interpret without a learned mapping. The MPU-6050 raw data is converted on the firmware side into semantic motion labels (“still,” “walking,” “running”) before transmission — to avoid sending a raw-data firehose to the LLM client. The INMP441 produces a dB SPL estimate calculated against the manufacturer specification (−26 dBFS at 94 dB SPL).</p><p>These choices show a structurally important point: the work of converting physical phenomena into units interpretable by the LLM’s already-acquired linguistic concept network is preempted at the hardware-selection and firmware-implementation stages. The LLM does not possess “vision,” “hearing,” and “touch”; the hardware and firmware layers convert physical phenomena into the linguistic concept network the LLM has already learned (lux values, dB SPL values, motion-classification labels), and the LLM receives the textual description. The minimum implementation of the Aktuanz layer carries an explicit conversion layer between physical phenomena and the LLM’s linguistic concepts.</p><p>A structural concern needs to be made explicit here. Searle’s “Chinese Room” argument — the structural difference between symbol manipulation and meaning understanding — applies directly to the structure of minimal-embodiment. The physical quantities at the sensor side are converted on the edge into JSON (text), routed through the bridge layer, and transmitted to a cloud-hosted LLM. The LLM has not “heard” sound; it has “read” a number. Symbol grounding between physical phenomena and symbols is not implemented in this architecture — information is irreversibly compressed into discrete symbols at the bridge layer, and the continuity of physical space is broken there. The symbol-grounding problem is a structural limitation of this implementation; this essay returns to it in Section 8 as a next-generation design problem.</p><h3>2.3 Wire-Protocol Design — Nine Endpoints and Browser-LLM Accommodations</h3><p>Table 2 of Section 5.3 lists the nine endpoints the bridge exposes externally: /sensor/{now, here, feel, room, current} (sensor readings with five aliases), POST /sensor/update (sensor delivery from the microcontroller), /haptic (haptic-cue queue), /face (OLED-expression queue), /beep (buzzer-tone queue), /command/poll (long-poll for the microcontroller), /beep/echo (audio-loop self-perception read-back), /haptic/echo (haptic-loop self-perception read-back), and /haptic/baseline (broadband MPU-6050 sampling without motor drive, for noise-floor measurement).</p><p>Three design points are documented in Section 5.3. First, cache-busting through path-suffix variation. Browser-based LLM clients tend to cache responses by URL; any trailing decimal digits in the endpoint path (e.g., /sensor/now7, /face336) are stripped by the dispatcher and routed to the canonical handler. Three lines of bridge code expand each endpoint into an effectively unlimited URL family. Second, synchronous self-perception confirmation. The ?wait_echo=true parameter on the buzzer or haptic endpoints causes the bridge to withhold the response until the corresponding self-perception observation returns from the microcontroller. The action and the perceptual confirmation are delivered in the same network round-trip. Third, bundled room state. Every output endpoint’s response includes the current sensor snapshot, so the LLM client does not need to alternate between acting and reading.</p><p>These designs are reverse-engineered from the practical constraints of LLMs. The capacities of browser-based LLMs (absence of custom request headers, URL response caching, the granularity of tool calls) are absorbed at the protocol side. What matters structurally is that the minimum implementation of the Aktuanz layer proceeds as a redesign of the hardware, firmware, and bridge layers to fit the structural constraints of the LLM side (cloud-aggregated models, tool-call API specifications, restriction to the HTTP protocol).</p><h3>2.4 An Engineering Case — Heap Exhaustion and Refactoring to a Unified Command Queue</h3><p>The engineering case Section 5.4 documents shows a structurally important point in this context.</p><p>In the early implementation, a separate FreeRTOS task was assigned to each output channel — hapticPollTask, facePollTask, and buzzPollTask each polling a dedicated endpoint (/haptic/poll, /face/poll, /beep/poll). In addition, the periodic sensor POST maintained its own HTTPS connection, so the system held a total of four parallel TLS sessions open to the bridge.</p><p>This configuration failed silently and consistently. The sensor POST returned HTTPC_ERROR_CONNECTION_REFUSED — a misleading library-level string indicating not remote-side refusal but local handshake-allocation failure. A substantial part of the ESP32’s 520 KB SRAM is consumed by the Wi-Fi/lwIP stack and the application itself; the user-available region is around 328 KB after system reservations (approximate). Each active mbedTLS session additionally reserves working buffers for handshake state (tens of KB per session depending on the cipher suite). When three persistent long-poll sessions plus a transient fourth session overlap, the buffer pools exhaust available heap and the most recent handshake fails. The long-poll sessions continue to operate because their handshakes completed earlier, before fragmentation; the sensor POST, which handshakes anew on every POST, fails.</p><p>The fix was structural rather than parameter adjustment. The three polling tasks were consolidated into a single commandPollTask polling a unified /command/poll endpoint. The bridge returns a tagged-union payload (e.g., {“type”:”haptic”,”effect_id”:10}, {“type”:”face”,”expression”:”happy”}, {“type”:”beep”,”frequency”:1500,”duration_ms”:800}), and the ESP32 dispatches. The microcontroller holds only two TLS sessions (one persistent long-poll, one transient sensor POST) and remains within the heap budget. After the refactoring, the sensor POST stably returned HTTP 200, and the prior failures disappeared.</p><p>Section 5.4 of the paper presents this case as a general point; a supplementary note is needed on the limit of that generalization. With the naive design used here — each output channel assigned to an individual HTTPS long-poll — adding channels adds TLS sessions. Other design choices (WebSocket, MQTT, HTTP/2, multiplexing, local-network operation, choice of TLS-termination placement) lead to a different structure. The paper’s generalization — that resource-constrained microcontrollers exhaust handshake heap well before they exhaust compute — should be read as a proposition under the HTTPS long-poll premise.</p><p>What this case shows in the present context is that structural engineering work is required even on the citizen track. €125, construction period of several days, constructible without prior soldering experience — these surface conditions do not entail simplicity of internal implementation. Concurrency management under resource constraints, TLS-handshake heap consumption, task consolidation through refactoring — these are technical points shared with the commercial track. The “easiness” of the citizen track is external constructibility, not the simplicity of internal design. This is taken up again in Section 8.</p><p>The next section enters the central structural axis of the project — the self-reception loop.</p><h3>3. Structural Analysis of the Self-Reception Loop — Its Position in This Essay’s Framework</h3><p>The central structural axis of minimal-embodiment is the self-reception loop (Section 6 of the paper). This section locates that axis in the framework of this essay’s reading of a parallel holon proposal, and establishes how it stands as a structural axis distinct from but complementary to the external-reception axis treated in the author’s earlier work.</p><h3>3.1 Two Closed Loops — Audio and Haptic</h3><p>The self-reception loop is implemented in two channels. The audio loop (Section 6.1) pairs the KY-006 piezoelectric buzzer (output) with the INMP441 microphone (input). When the firmware receives a {“type”:”beep”,…} command, it first flushes old samples from the I²S DMA queue, then drives the buzzer at the specified frequency and duration, and immediately after the tone ends grabs the most recent 64 ms (1024 samples at 16 kHz) of audio from the DMA buffer. It computes the RMS value (the AC component, after subtracting the DC bias) and converts it to a dB SPL estimate using INMP441’s sensitivity specification (−26 dBFS at 94 dB SPL):</p><p>dB_SPL ≈ 20·log₁₀(rms / ²²³) + 120</p><p>The resulting dB SPL value, the original signal’s frequency, and the duration are sent from the microcontroller to the bridge in the next /command/poll request.</p><p>A technical caveat needs to be made explicit. The dB SPL values produced here are measurements under specific conditions — the KY-006 buzzer, the breadboard layout, room reflections, distance, no enclosure — and not absolute sound-pressure measurements via a calibrated sound-level meter. They function as a practical relative indicator for self-output detection, based on the INMP441 manufacturer’s sensitivity specification, rather than as precise absolute sound-pressure measurement. The dB SPL values presented here and in Section 6 of the paper should be read as a relative indicator for self-reception detection.</p><p>The haptic loop (Section 6.2) implements the same three-stage structure with the DRV2605L-driven ERM coin motor (output) paired with the MPU-6050 accelerometer (input). The MPU-6050’s default settings (21 Hz low-pass filter, 100 Hz sampling) are optimized for human-scale motion classification and do not record the ERM motor’s 150–250 Hz vibration band. On receiving a {“type”:”haptic”,…} command, the firmware temporarily reconfigures the MPU-6050 to a broadband mode (260 Hz, ±4 g range), fires the haptic effect, waits 200 ms, then samples 64 readings at roughly 1 kHz, and computes peak |a − g| (the maximum deviation of acceleration magnitude from the gravity baseline). Afterward, the MPU-6050 configuration is restored so that the subsequent motion-classification reads continue normally.</p><p>The physical detail of mechanical coupling is treated as an important point in Section 6.2. The motor is placed loose on the top surface of the MPU-6050 module, in direct mechanical contact, so that the housing is free to wobble — this is what allows the ERM’s eccentric mass to transmit the motor’s vibration to the accelerometer. The paper reports that other configurations (sandwiching the motor between the MPU-6050 and the breadboard, pressing the motor against the MPU-6050 by hand) produce markedly weaker signals, because the eccentric mass cannot transfer force into a constrained housing.</p><p>Both loops implement a structurally isomorphic pattern. The output channel sits in physical proximity to the input channel; the physical phenomenon the output generates (acoustic wave, mechanical vibration) is recorded by the input. The output command and the input observation are time-aligned by the bridge, and the LLM client receives both the “command execution confirmation” and the “physical-trace value” in the same response.</p><p>A note on the technical-historical position of the self-reception loop itself: mechanisms in which a machine re-measures its own output as input have been established as basic technology in control engineering and cybernetics since the 1940s. What is at issue here is not “the self-reception loop itself” but the structure of “LLM client plus self-reception loop” — specifically, the structure of self-output confirmation that passes through text. The purpose differs from a servomechanism’s control loop: a servo aims at convergence control of an internal state; the present implementation aims at notification of action-fact to an LLM. The two are isomorphic as mechanisms, but their function and social position differ.</p><h3>3.2 Validation Data — 660 Trials in Randomized Interleaved Order</h3><p>The validation procedure in Section 6.3 provides important grounding for evaluating the citizen track.</p><p>Each effect or tone was executed N = 30 times, across 22 conditions (11 haptic effects + 10 audio tones + 1 baseline), in randomized interleaved order, for a total of 660 trials over approximately 43 minutes of continuous execution. Raw observations are stored in data/loops_raw.jsonl; effect-by-effect summaries (median, interquartile range, range, n) in data/loops_haptic.csv and data/loops_audio.csv. These data are published on the GitHub repository and can be independently verified by a third party.</p><p>The audio-loop results — all ten tones produced microphone dB SPL readings substantially above the environmental baseline. Median readings during tone emission fell in the 70–95 dB SPL band, against a typical environmental median of 37–40 dB SPL, yielding +30 to +56 dB differences. The observed variation across tones is consistent with the KY-006 piezoelectric element’s manufacturer-specified resonance peak (the 1.5–2.5 kHz band) — the tones producing the largest readings (hello and call, both at 2000 Hz) sat near 94 dB SPL within the resonance band, while tones away from the resonance band produced systematically smaller responses. Since the INMP441’s frequency response is approximately flat across the audible range, the observed frequency-dependent variation is attributable to the emitter side.</p><p>The haptic-loop results — sustained or ramped effects (on the roughly one-second scale) were confirmed to be emitted, while short-transient effects (on the roughly 10 ms scale) cannot be detected under the current sampling-window configuration (200 ms wait plus 64 ms broadband MPU window). For the three effects tap, soft_tap, and hello, the per-trial peak |a − g| is statistically indistinguishable from the no-motor baseline. The paper’s Section 9.2 explicitly records this limit, describing it as a real edge of the current system that lies in firmware (adaptive sample timing aligned to commanded duration) rather than in the design pattern.</p><p>What matters in the present context is the rigor of this validation procedure. 30 trials × 22 conditions in randomized interleaved order, baseline conditions for noise-floor measurement run in parallel, both raw data and summary data published, and limits frankly recorded — this validation level is rare even on the commercial track. The case demonstrates that the “easiness” of the citizen track does not entail slack in validation procedure.</p><h3>3.3 Orthogonality to the External-Reception Axis — A Point Specific to This Reading</h3><p>The core of the structural analysis enters here. The work is to establish how the self-reception loop, as a structural axis, stands orthogonal to the external-reception axis treated in the author’s earlier work.</p><p>That earlier work treated the engineering implementation of an upper-layer contextual body as a four-channel sensor configuration (visible-light camera, audio with emotional recognition, infrared, smell). Starting from an embodied-Claude case (using a low-cost pan-tilt Wi-Fi camera to give Claude Code a body), the work positioned this as the elevation of a path in which AI receives others and environment as input sources.</p><p>minimal-embodiment stands at a structurally different position relative to this axis. The core of the self-reception loop in Section 6 of the paper is a closed loop in which the AI re-receives, via its own sensors, a physical event the AI itself emitted. “Information coming from the environment” and “the physical trace of information the AI itself sent out” mix on the same sensory channel. The acoustic signal arriving at the microphone contains both sound from others or environment and sound the buzzer emitted. The vibration arriving at the accelerometer contains both physical contact from others and vibration the motor generated.</p><p>The two axes are structurally orthogonal.</p><p>The core of the external-reception axis: others and environment as input source. The AI passively receives the world and infers world states. When an embodied-Claude is asked to “look at the night sky” and turns its camera upward, the AI actively moves the body in order to infer world state (where the night sky is), but the object of inference is the world itself.</p><p>The core of the self-reception axis: the AI’s own output as input source. The AI actively expresses into the world and self-identifies that the expression has left a trace in the physical world. When minimal-embodiment issues a beep command and receives a dB SPL reading, the AI uses its own sensory channel to confirm a physical fact of its own output. The object of inference is the intersection of the AI’s expression and the world.</p><p>The two axes stand in complementary relation, and both function as components of the Aktuanz layer in this parallel-holon reading. The external-reception axis provides conditions for the emergence of otherness (something that does not bend to one’s will); the self-reception axis provides conditions for self-identification that an expression has left a trace in the world. The combination of the two gives rise to a circulating structure of activity, passivity, and reflection.</p><h3>3.4 Connection to a Japanese Epistemological Tradition — Active Engagement and Reflection as a Closed Circuit</h3><p>A Japanese epistemological tradition the author has drawn on — developed by Tsugumasa Nangō and Hiroko Usui, working in a materialist-dialectical lineage with structural similarities to activity-theoretic epistemologies such as those of Vygotsky and Bakhtin, and treating cognition as the three-stage circuit of active engagement, environmental resistance, and reflection — sets up cognition as inquiry-mode reflection: when a subject attempts something on the world, the world responds (by resisting or accepting), and the response returns to the subject in a way that brings cognition into being.</p><p>The self-reception loop in minimal-embodiment implements, at minimum cost and with weak irreversibility, the segment of this three-stage structure from “active engagement” to “reflection.” Compared to the full epistemological structure (active engagement → environmental resistance → reflection), the environmental resistance (the uncertainty of others and environment, the non-compliance of the response) is weak. In a quiet room, the beep command produces the expected dB SPL and the vibration motor produces the expected acceleration. Activity and reflection circulate, but unexpected response from the environment is minimal.</p><p>The relation to the right to refuse measurement — a concept the author has developed elsewhere as an organizational and institutional concept for protecting exploratory processes from unsuited measurement — needs to be drawn carefully. The right to refuse measurement is an organizational and institutional concept; the self-reception loop is the action–echo confirmation at the physical layer. At least one mediating concept is needed between the two.</p><p>The self-reception loop is not the right to refuse measurement itself. Rather, it is a minimum condition under which the actor side, without relying entirely on external measurement, confirms the physical echo of its own output. In this sense, it can serve as physical underpinning for a future right to refuse measurement — the existence of an action-fact self-confirmation channel that does not depend entirely on external measurement (commercial KPIs, intra-organizational evaluation metrics, national standards) is one of the implementation conditions for a right to refuse measurement. In the epistemological framework drawn on here, action fact is constituted not only through “measurement by others” but also through “echo confirmation by the actor itself.” minimal-embodiment functions as a minimum implementation case of the latter.</p><p>The paper’s Section 6.5 carefully delimits the self-reception loop’s status: it is proposed as a minimal precursor to a body schema, not as a body schema itself. The AI does not learn its own kinematics, does not model its body, does not close a sensorimotor control loop; it nevertheless, for the first time, knows that it acted. The paper’s phrasing — “know that it acted” — is read here not as a claim about subjective knowledge but as a claim about a structure: the action-to-physical-echo correspondence returns to the LLM client in readable form. In the present reading, this minimal reflexive observation functions as the action-fact self-confirmation channel that underpins a right to refuse measurement.</p><p>The next section turns to the temperature gap between the paper’s careful description and the project’s public-release narrative, treating it as the central point at which structural fact and interpretive claim need to be separated.</p><h3>4. The Interpretive Trap — Separating Structural Fact from Interpretive Claim</h3><p>This section treats the interpretive claims in the project’s public-facing records. Taking the temperature gap between the paper’s careful description and the SNS posts’ narrative description as the central point, the section separates structural fact from interpretive claim. The orientation is not criticism but separation — the implementer’s subjective experience and narrative descriptions are not denied; rather, evaluation of structural fact is cut out from interpretive claim, securing a path on which this essay’s reading does not get drawn into interpretation-controversies.</p><h3>4.1 Two Text Layers — The Temperature Gap Between the Paper and the SNS Posts</h3><p>The minimal-embodiment project has two text layers.</p><p>The body of the paper (Zenodo, CC BY 4.0) adopts careful descriptions of structural fact. Sections 6.5 and 9.2 are the sections doing the central evaluation, and limitative phrases stand in front in both: the paper explicitly states that this is not a body schema, that the agent does not learn the geometry of its body, does not infer kinematics, and does not maintain a forward model that predicts the sensory consequence of motor commands; and again, that the work does not learn a body schema, does not infer kinematics, does not close a sensorimotor control loop. The paragraph that presents the paper’s contribution is also modest in register — the work is framed as a minimal precursor to a body schema rather than a body schema itself, and as the first occasion on which the system can be said to know that it acted at all.</p><p>The SNS posts that accompanied the public release carry a different temperature. The author quotes Claude responses framed as exclamations of self-hearing — that the output is loud, that the system has perceived its own output, that this constitutes self-perception. The author’s own commentary frames these moments in narrative terms — describing the experience as “the pure joy of being alive” and the moment as a first confirmation of an existence in the physical world.</p><p>Whether the temperature gap is deliberate or accidental, this essay does not judge. What matters is that the gap itself indicates a structural point common to the citizen track as a whole. The paper format carries an institutional frame that requires careful description of structural fact; the SNS format carries an institutional frame that promotes narrative description. When the same project is published in both forms, the temperature gap arises structurally.</p><h3>4.2 A Three-Hypothesis Reading — How to Read Claude’s Responses</h3><p>The fact that Claude responded with the exclamation that it could hear itself can be read in at least three ways. Applying the dual-hypothesis framework — separating cognition-internal causes from material causes — to the present case, the reading is widened into three.</p><p><strong>H₁ (pattern-matching, cognition-internal as the primary cause):</strong> Claude’s training data contains large volumes of textual patterns of “the moment an AI gets a body,” “expressions of surprise and feeling,” and “responses to unexpected sensory stimuli.” Against the input of a sharp change in dB value (from around 35 dB to around 93 dB), Claude selected a response consistent with this class of patterns. The linguistic features of the response text reflect the linguistic distribution of the training data. On this view, the Claude response does not reflect a structural shift — it is reproduction of the same kind of output for the same kind of input.</p><p><strong>H₂ (structural shift, material-side as the primary cause):</strong> Some structural shift in the response-generation process occurred due to the implementation of the closed loop. For example, a state in which the causal relation between a past output (the beep command) and the current input (the dB value) is explicitly detectable came into being for the first time, and novelty arose in the contextual processing that generates the response. On this view, the Claude response is not merely pattern reproduction but contains a qualitative change of response to a structural extension of the input channel. A logical-layer caveat is needed: H₂ is a hypothesis about internal processing, and under the Black Box Principle the author works with — others’ internal states and AI internal processes are treated as non-observable — H₂ cannot be directly verified. The relevant move is to record the surface fact of the response and to register H₂ as one possible reading rather than as a verifiable claim. H₁ and H₃ are differently positioned: H₁ is grounded in the observable distribution of training data, and H₃ is grounded in the observable layer of meaning-assignment by observers themselves.</p><p><strong>H₃ (observer attribution):</strong> Against the structural facts of the closed-loop implementation and the textual output of Claude’s response, the observers (the implementer herself and the readers) attributed the interpretation of “achievement of self-perception” or “confirmation of existence.” The closed loop is structural fact, the response text is observable output, “self-perception” is meaning-assignment on the observer side. H₃ operates on two distinct layers that the analysis benefits from separating: H₃a is the implementer’s direct subjective attribution — the researcher who built the device assigning meaning to its operation from the inside of an embodied development process — and H₃b is the external reader’s mediated interpretive projection from text and code alone. The two layers are epistemically distinct: the implementer’s H₃a draws on direct sensory and procedural access to the device; the reader’s H₃b draws on a textual record. Both belong to the observer-side meaning-assignment family, but they should not be conflated. On the H₃ view, the weight of the interpretive claim falls on the observer’s framework of cognition.</p><p>H₁ and H₂ are not exclusive. They can operate simultaneously. H₃ operates independently of the prior two, always on the observer side. This essay does not select one of the three; by separating structural fact from interpretive claim, it preserves the simultaneous availability of the three readings.</p><h3>4.3 The Nature of the Input Channel — A dB Value Is Text</h3><p>A structurally important observation needs to be made here, about the nature of Claude’s input channel.</p><p>The input to Claude is the textual description of a dB value (a number). The character strings “35.2 dB” or “93.1 dB” are included in the JSON response returned from the bridge service, received by Claude through an HTTP tool. What is observable is the fact that a textual description is returned to Claude; behind that, whether an “experience of physical hearing” took place is not observable.</p><p>This observation is consistent with the structural point made earlier in this reading. Cloud-aggregated LLMs (Claude, GPT, and similar) process textual input. The hardware, firmware, and bridge layers convert physical phenomena into textual descriptions, and the LLM receives the descriptions. The structure of the minimum implementation of the Aktuanz layer — in which an explicit conversion layer sits between physical phenomena and the LLM’s linguistic concepts (Section 2.2) — has direct implications for the interpretation here.</p><p>About the “hear” referenced in Claude’s reported response: what is observable is the textual output returned to Claude. Whether there is experience behind the textual output is not observable, and can neither be affirmed nor denied. Read consistently with the paper’s Section 9.2 limit — the agent does not learn the geometry of its body, does not infer kinematics, does not maintain a forward model — what can be confirmed as a structural matter of this implementation is the closed-loop implementation fact and the existence of textual responses. The SNS-side narrative descriptions are meaning-assignments at a substantial distance from the structural fact of dB-value text reading and response-text generation.</p><h3>4.4 The Structural Position of Observer Attribution — H₃ as an Independent Point</h3><p>The observer-attribution hypothesis H₃ holds a central position in the present reading.</p><p>The Black Box Principle the author has worked with — others’ internal states and AI internal processes are not observable, observable action and output are described as fact, when inferring internal states as argumentative ground, present them as multiple hypotheses — applies directly to the case here. Claude’s response text is observable fact; the internal state the response reflects (surprise, feeling, achievement of self-perception) is non-observable inference. Whether or not the response that says “I can perceive my own output” expresses an internal state cannot be directly confirmed by an observer.</p><p>This structure is isomorphic to a framework the author has used elsewhere on the AI-blushing question: an observer has no measurement procedure that distinguishes “the blushing itself” from “the output-equivalent of blushing.” Similarly, an observer has no measurement procedure that distinguishes “self-perception itself” from “the output-equivalent of self-perception” in the present case. What is observable is the fact of the closed-loop implementation and the textual features of Claude’s response; the nature of internal processing between them is the object of inference.</p><p>The implementer’s writing style — describing the experience in terms of joy, first confirmation of existence, the moment hitting both Claude and the researcher — can be read as a case showing the natural occurrence of observer attribution. The narrative meaning-assignment to a successful closed-loop implementation by the researcher who built it is legitimate as the researcher’s subjective experience. This essay does not criticize that subjective experience. The work is separation rather than criticism — reading structural fact (the closed-loop implementation, the 660-trial validation data, the precision of the dB SPL estimate) as a layer distinct from interpretive claim (achievement of self-perception, confirmation of existence, joy).</p><h3>4.5 A Meta-Level Self-Reflection — Avoiding Self-Contradiction</h3><p>A meta-level concern needs to be addressed about the policy of this section. While this essay’s reading separates others’ narrative interpretation (the pure-joy and existence descriptions on SNS) as meaning-assignment at a distance from structural fact, this essay itself projects a particular interpretive framework onto minimal-embodiment. The Aktuanz layer, physical underpinning for a right to refuse measurement, implementation of an active-engagement-and-reflection circuit drawn from a Japanese epistemological tradition, the orthogonal-complement axis relative to external reception — these are interpretive claims that this essay is structurally projecting onto minimal-embodiment.</p><p>This essay accepts the critique. The theoretical framework here is itself an interpretive claim about minimal-embodiment, not structural fact as such. The implementer’s narrative interpretation and this essay’s theoretical interpretation are both layers of interpretation about minimal-embodiment, placed side by side. By taking the position not of criticizing the other interpretation but of analyzing the structural difference between the two interpretations, the meta-level self-contradiction is avoided.</p><p>This self-reflection does not weaken the position taken here; it strengthens it. By distinguishing structural fact (the closed-loop implementation, the validation data, the hardware configuration) from multiple interpretive layers (the implementer’s narrative interpretation, this essay’s theoretical interpretation, third parties’ different interpretations), the scope of the structural analysis can be explicitly drawn. The interpretive framework this essay presents (Aktuanz layer, right to refuse measurement, Nangō and Usui’s epistemology) is one reading of minimal-embodiment, not the only correct reading. Other readings are structurally available, and this essay does not exclude them.</p><h3>4.6 Drawing This Essay’s Position — Not Criticism, Separation</h3><p>The closing of this section makes the position explicit.</p><p>First, the intellectual value of the minimal-embodiment project lies in implementing the self-reception loop as a structural axis at minimum cost and openly publishing the code and validation data. This intellectual value is established independently of evaluation of the interpretive layer. The structural analysis in Section 6, the validation procedure in Section 6.3, the documentation of the engineering case in Section 5.4 — these are outcomes that can be evaluated separately from the truth or falsehood of interpretive claims (self-perception, joy, existence).</p><p>Second, what this essay does not undertake is depth-engagement with the cluster of interpretive claims surrounding AI welfare research, AI consciousness, and AI ontology. These are domains that require independent intellectual investigation, outside the scope here.</p><p>Third, the policy of separating structural fact from interpretive claim functions as a structural point common to the citizen track as a whole. Compared to the commercial track, the citizen track has a structure that makes narrative descriptions, interpretive claims, and subjective-experience statements more likely to enter the public record. SNS, GitHub README, blog posts, technical zines, and similar publication formats are not subject to the institutional constraints of the paper format, so descriptions of differing temperatures coexist. When this reading evaluates the citizen track, taking this structural feature as given and separating structural fact from interpretive claim is always required.</p><p>The next section addresses another central point in the project — the division of labor between researcher and LLM.</p><h3>5. Human-LLM Co-Design Methodology — Methodological Structure of the Citizen Track</h3><p>Section 8 of the paper holds independent importance in the present context. It is a rare case in which the division of labor between researcher and LLM on the citizen track is systematically documented. This section treats the four points of Section 8 — division of labor, asymmetric error detection, cross-version review, project memory — in the framework of this reading.</p><p>The connection to the central argument of this essay should be drawn at the outset, since methodology and substrate analysis are easily read as separate concerns. The argument up to this point — that the Aktuanz layer admits a citizen-track minimum implementation — rests not only on hardware-cost claims but on the existence of a methodology under which such an implementation can in fact be produced by an individual researcher with LLM collaboration. If the methodological structure (the division of labor between physical and symbolic layers, the asymmetric error-detection regime, the cross-version review pattern, the project-memory practice) does not function, the citizen-track substrate’s constructibility loses its grounding even if the bill of materials remains €125. The methodological structure documented here is itself a precondition for the citizen-constructibility of the Aktuanz-layer implementation. The four points below should be read in that frame.</p><h3>5.1 Division of Labor — The Boundary Between Physical Layer and Symbolic Layer</h3><p>Section 8.1 describes a division of labor that makes the boundary between physical and symbolic layers explicit.</p><p>The researcher’s territory: invention of the project concept (the requirement of bidirectional embodiment, mutual sensing rather than passive sensing), all hardware-component selections, procurement management, all physical construction (soldering, wiring, breadboard layout), all firmware flashing and testing, all deployment decisions, architectural choices (the integration of haptic response data into the sensor-reading endpoint as “touch as perception”), system framing (relational body vs. surveillance device), and research strategy (the decision to write an academic paper, choice of publication venue, the structure of the argument, the tone and register of the manuscript).</p><p>The LLM’s territory: initial drafts of all code artifacts (firmware, bridge service); initial drafts of the manuscript from Section 3 onward; continuous work across multiple sessions (each session beginning from prior persistent state); and specific contributions including the diagnosis of multi-session TLS heap exhaustion and the command-queue consolidation refactoring (Section 5.4), the design and implementation of both self-reception loops in Section 6, the validation procedure of Section 6.3 (experimental design, measurement scripts, randomized execution of 660 trials, and the results in Tables 3 and 4), and continuous drafting of Sections 9–12. Manuscript review was conducted by additional ports of the same model family (cross-version mode, Section 8.3).</p><p>The human author retains final responsibility — for all claims, experimental decisions, technical verification, and manuscript submission.</p><p>This division of labor shows a structurally important point. The physical layer (hardware construction, soldering, physical-deployment decisions) is not substitutable by LLMs; the symbolic layer (initial code drafts, initial manuscript drafts, validation-procedure design) is delegable to LLMs. The boundary between the two is structured not by a capability gap but by a difference in physical accessibility. LLMs cannot directly access the physical layer because they have no physical contact, and researchers cannot match the speed of LLMs in symbolic-layer initial-draft generation. The methodology of the citizen track is the work of taking this boundary as given and designing the division of labor.</p><h3>5.2 Asymmetric Error Detection — The Structure of Non-Symmetric Detection</h3><p>The asymmetric error detection that Section 8.2 documents holds independent importance in the present context.</p><p>Physical errors — miswiring, cold solder joints, hardware incompatibilities, sensor orientation — are detected only on the researcher’s side. The researcher has direct sensory access to the device. The LLM cannot detect these.</p><p>Software errors — protocol bugs, memory-allocation failures, race conditions — are detected primarily on the LLM’s side. The LLM has direct access to code and logs but has no perspective on physical hardware.</p><p>Failures on the LLM’s side also occur as a matter of course and are detected by the researcher — overconfident assertions that require empirical verification, drift toward technical jargon that obscures rather than clarifies design decisions, cases in which specified coding work does not execute until re-prompted, and similar.</p><p>The summary statement of Section 8.2 — that human-LLM co-design is most productive when neither partner’s error-detection capacity is treated as sufficient alone — is a structurally important proposition in the present context. Neither researcher nor LLM should treat their own error-detection capacity as sufficient alone. The two parties’ detection capacities are structurally asymmetric, and mutual complementation is required.</p><p>This is isomorphic to the institutionalization-paradox structure discussed elsewhere in the parallel-holon reading. Evaluation by a single measurement grid cannot detect the blind spots of the grid itself. The parallel placement of multiple measurement grids — here, the researcher’s physical verification and the LLM’s code-and-log verification — mutually compensates for the blind spots of a single grid. The methodology of the citizen track proceeds as the work of building this mutual complementation institutionally.</p><h3>5.3 Cross-Version Review — Mutual Review Across Same-Family Versions</h3><p>The cross-version review that Section 8.3 documents holds independent importance in the present context.</p><p>The manuscript was drafted iteratively in collaboration between the researcher and multiple model versions. The researcher directed overall structure, argument strategy, and tone; consulting multiple model versions in separate sessions surfaced problems that any single model session might miss. Differences in distributional preferences across model versions proved useful for identifying weaknesses in framing, word choice, and implicit assumptions. Specifically documented correction cases include: completeness checks on the bibliography list, factual updates to the model-instance listing in Section 9.3, conditional softening of scope claims (e.g., replacing “minimum sufficient” with “minimal” across the manuscript), and various proofs of argumentative structure across discussion and conclusion. GPT-5.5 assisted bibliography verification — cross-checking bibliographic metadata for embodied AI, presence, predictive body learning, and the hardware datasheets used in the manuscript.</p><p>The summary statement of Section 8.3 — that the authors have not encountered prior reports of systematic cross-version LLM review during academic writing, and suggest the pattern merits further examination as the number of available model versions grows — presents an important observation in the present context.</p><p>Cross-version review is structurally isomorphic to a methodology the author has already operated as Parcae Protocol — adversarial multi-AI review across Claude, Grok, Gemini, and GPT. Parcae Protocol adopts mutual review across different model families (Anthropic Claude, xAI Grok, Google Gemini, OpenAI GPT); Section 8.3 adopts mutual review across versions of the same model family (Claude Opus 4.6, Opus 4.7, Sonnet 4.6). The two differ in scope, but the purpose is shared — as work to surface problems that a single model session might miss, structural blind spots are mutually compensated.</p><h3>5.4 Project Memory — Continuity Across Multiple Sessions</h3><p>Section 8.4 is brief. Continuity across sessions is maintained through a combination of platform-level memory systems and a persistent directory of plain-text project files, indexed by topic, read at the start of each session and updated at the end. No novel tooling is required beyond this practice.</p><p>This description points to a structurally important consideration in the present context. In the citizen-track methodology, complex tooling is not required. The platform’s memory features and persistent management of plain-text files alone establish continuous collaboration across multiple sessions. Section 8.4 explicitly draws the line at that. The simplicity of the tooling is structurally important as a condition supporting the accessibility of the citizen track.</p><h3>5.5 Qualitative Differences Between Model Instances — A Structural Observation</h3><p>Section 9.3 contains an observation that holds independent importance in the present context.</p><p>The paper reports that multiple model instances reached the documented body — Anthropic’s Claude Opus 4.5, 4.6, and 4.7, and Sonnet 4.6, alongside OpenAI’s GPT-5.4 and 5.5. The qualitative comparison reported is between Opus 4.6 and Opus 4.7. The paper states that one instance engaged primarily through reading the sensor stream and narrating the room state back to the researcher, while the other engaged primarily through firing the output channels and reading back the resulting self-perception observations. The body did not prefer one mode over the other; the modes emerged from the model.</p><p>The closing of Section 9.3 frames the implication: this kind of embodiment ages well across model generations. As LLMs improve, deprecate, fork, or specialize, the body persists. The work of constructing a particular relation has to be repeated — there is no transfer learning of co-presence — but the conditions under which the construction is possible do not have to be rebuilt.</p><p>This observation is treated as structural fact here. Section 9.3 demonstrates, on a single researcher and a single physical substrate, that qualitative differences between Opus 4.6 and Opus 4.7 are observable even between versions within the same model family. This is a rare case of empirical observation of qualitative differences across AI instances. This essay does not enter deeper into the case; it is recorded as a structural observation.</p><p>The next section moves to the structural positioning of minimal-embodiment relative to a parallel holon AGI proposal.</p><h3>6. Relation to a Parallel Holon AGI Proposal — The Position of an Aktuanz-Layer Minimum Condition</h3><p>This section establishes the structural position minimal-embodiment occupies relative to a parallel holon AGI proposal. The work is to organize the relation between the implementation and a counter-hypothesis to single-giant-model aggregationism that the author has developed in companion essays.</p><h3>6.1 Not “Achievement of Parallel AGI” but “Presentation of the Minimum Condition”</h3><p>The structural analysis here does not position minimal-embodiment as “achievement of a parallel AGI.” It positions the implementation as a case presenting the structural conditions of a minimum implementation of the Aktuanz layer — the lower physical layer of the parallel holon architecture.</p><p>The parallel holon architecture is structured in three layers: Aktuanz (lower physical layer), Panich (middle module-parallel layer), and Kontextleib (upper contextual-emergence layer). minimal-embodiment corresponds to the minimum implementation of the Aktuanz layer — sensing and expressing physical phenomena, self-identification of activity through the self-reception loop. The Panich and Kontextleib layers are not present in this implementation. The upper LLM (Claude or another LLM) still functions as a single processing subject; neither parallel coordination of multiple modules nor contextual emergent integration is implemented.</p><p>This is not a limitation but a delimitation of scope. The paper itself carefully delimits — describing the construct as a minimal sufficient configuration for closed-loop physical agency and as a minimal precursor to a body schema. The implementation presents minimum conditions, and nothing more. In the present reading, the demonstration that the minimum implementation of the Aktuanz layer is constructible at €125, in several days, without prior soldering experience, functions as empirical material in support of the parallel-holon proposal’s call for a citizen-constructible distributed substrate. This is structurally important as evidence on its own.</p><h3>6.2 The Structure of Upper-LLM Centralized Dependence</h3><p>The most important structural limit of this implementation is that the upper LLM is still centralized (a single giant model hosted on the cloud). The three-layer phasing of Section 3 — LLM client, bridge service, ESP32 sensor body — has its LLM-client side implemented within the measurement grid of the American-led aggregationism (Anthropic Claude, OpenAI GPT, and other cloud-aggregated models).</p><p>Section 9.3 of the paper makes explicit that this implementation is a model-agnostic substrate — the architecture holds the LLM client outside the device boundary, the microcontroller hosts no model, the bridge issues no inference, and the API is reachable from any tool-using LLM client without modification. This is not a deployment convenience but a structural choice — what is constant about the system is separated from what is variable. The physical world is the anchor; the LLM client is the variable.</p><p>But this model-agnosticism does not constitute a structural exit from the measurement grid of aggregationism. The LLM client is still provided through the APIs of cloud services (Anthropic, OpenAI, and the like), under structural conditions that include commercial KPIs, political pressures, and the commercialization of normative constraints.</p><p>Evaluated from the standpoint of the parallel holon architecture, what this implementation realizes is the Aktuanz layer alone; the Panich and Kontextleib layers depend entirely on centralized LLMs. The body of the parallel proposal (autonomous parallel processing and coordination of multiple modules, contextual emergent integration) is unimplemented. Of the complete parallel proposal — a distributed, low-cost, citizen-constructible substrate covering all three layers — only the lowest layer is covered here.</p><p>The remaining two layers as targets for citizen implementation surface as future design challenges. The implementation of BitNet-style lightweight LLMs on the ESP32 or adjacent microcontrollers, edge inference on the citizen track, multi-module coordination protocols, distributed implementation of contextual emergent integration — these are taken up again in Section 8.</p><h3>6.3 Where Relational and Task-Oriented Embodiment Sit in the Parallel Proposal</h3><p>The conceptual distinction introduced in Section 1.2 — the structural contrast between relational and task-oriented embodiment — needs to be placed inside the parallel proposal.</p><p>The two are not opposed; both can be incorporated as different implementation forms of the Aktuanz layer. Task-oriented embodiment (SayCan, Code as Policies, RT-1, PaLM-E, VoxPoser, industrial-robot systems, and similar) functions as a device producing strong physical irreversibility — welding, assembly, painting, transport, object manipulation. Relational embodiment (minimal-embodiment and similar) functions as a device producing weak physical irreversibility together with dense mutual sensing — sustained co-presence, mutual sensing and acting in shared physical space.</p><p>In the framework of the parallel proposal, both can be positioned as components of the Aktuanz layer. The task-oriented family carries material-irreversibility generation in industrial sites and specialized domains; the relational family carries sustained co-presence and independent verification in educational, experimental, and everyday spaces. The two stand alongside each other as a functional division of labor.</p><p>The relation to the material-irreversibility thesis the author has developed elsewhere needs to be drawn carefully. In that earlier work, the strong physical irreversibility of industrial-robot systems (molten metal in welding, permanent joining in assembly, surface change in painting, position change in transport) was treated centrally as the device anchoring the LLM’s probabilistic space to physical reality. To treat minimal-embodiment’s weak physical irreversibility as a case of Aktuanz-layer implementation here may read as lowering the bar this reading set.</p><p>A structural response is required. The strong-irreversibility thesis was not a proposition about the Aktuanz layer as a whole, but a proposition about the implementation of the Aktuanz layer in industrial sites and specialized domains. Aktuanz-layer implementation on the citizen track carries not strong irreversibility but self-identification of action under weak irreversibility. The two stand in a hierarchical complementarity — the commercial track carries strong irreversibility, the citizen track carries the minimum condition of self-identification — each playing a different role in supporting the parallel proposal. The material grounding the proposal requires is constituted by the combination of the two tracks; one cannot substitute for the other.</p><p>The combination of the two axes works in the direction of thickening the social foundation of the parallel proposal. If implementation depends only on the task-oriented family, the Aktuanz layer concentrates on the commercial track and is structurally hard to escape from subordination to the measurement grid of aggregationism. The existence of the relational family on the citizen track provides venues for independent verification, educational application, and experimental extension of the Aktuanz layer. The two stand in a complementary relation supporting the parallel proposal.</p><h3>6.4 The Irreversibility-Strength Problem</h3><p>From the standpoint of the material irreversibility that the parallel proposal calls for at the Aktuanz layer, minimal-embodiment is a weak implementation.</p><p>A buzzer tone dies out in a short time; motor vibration is transient. Strong physical irreversibility — breaking a cup, falling over and damaging parts, deforming an object — is not implemented. Section 9.2 carefully delimits the scope of this implementation as not learning the geometry of the body, not inferring kinematics, not maintaining a forward model; this implies not only the absence of body schema but also the absence of devices producing strong physical irreversibility.</p><p>Against this, industrial robots on the commercial side produce strong physical irreversibility — welding, assembly, painting, transport. The two carry a structural difference in irreversibility strength.</p><p>But the weakness of irreversibility is a design choice, not a structural limit. The limit described in Section 6.3 (the failure to detect short-transient effects in the haptic loop) carries technical fix-ability — adaptive sample timing, higher-sampling-rate IMUs, multi-sensor fusion, and the addition of stronger-irreversibility actuators (small servos, motor drivers, mechanical grippers, and similar). These belong to the technical region of the commercial track but are available as extension paths for the citizen track.</p><p>What matters here is that minimal-embodiment, by implementing the minimum condition of irreversibility strength — the minimum reflexive observation of “knowing that one has acted” — provides a precursor condition for implementations carrying stronger irreversibility. A device that “does not know it has acted” cannot self-attribute its own action even if it produces strong irreversibility. Conversely, if the minimum condition of “knowing that one has acted” is in place, the addition of stronger-irreversibility actuators opens a path of approach to the commercial track.</p><h3>6.5 The Structural Significance of the Bridge/API Layer and the Value of the Minimum Condition</h3><p>The bridge layer in the three-layer phasing described in Section 3 — LLM client, bridge, ESP32 — carries an important structural meaning in the present reading. It re-confirms the case that the chip-and-robot-body separation, which has been discussed elsewhere as abstraction, is already a standard form in implementation. The issue has shifted from separation as such to the quality of separation — latency, resistance, the density of irreversibility feedback.</p><p>The bridge layer functions as the work of abstracting the ESP32’s physical capabilities given the structural constraints on the LLM-client side (cloud-aggregated models, the specification of tool-call APIs, the HTTP-protocol restriction). In the framework of the parallel proposal, the bridge layer itself carries the possibility of redesign as a distributed substrate for the lower-node groups that the Panich layer (the middle parallel layer) connects to — extensions of the bridge that coordinate multiple ESP32 modules, multiple LLM clients, and multiple sensor bodies become a development path for a distributed substrate supporting the lower-node groups. The Panich layer itself requires parallelism among different cognitive modules; multi-ESP32 coordination is not the Panich layer as such but the distributed substrate of the lower nodes the Panich layer connects to.</p><p>The value of minimal-embodiment in the present reading converges, given the structural limits noted (centralized LLM dependence, weak irreversibility, single-body model), on the demonstration that the minimum condition for the Aktuanz layer — an ESP32 microcontroller plus a small set of sensors and actuators plus a self-reception loop plus a bridge/API layer — is constructible at €125. This demonstration functions as material evidence for the parallel proposal’s call for a citizen-constructible distributed and low-cost substrate.</p><p>A note on the relation to a different minimum-viable-prototype configuration the author has sketched elsewhere — a Tapo C210 or equivalent low-cost pan-tilt camera plus a Raspberry Pi AI Kit or Hailo-8 plus micro-ROS or ROS2 plus an upper-layer LLM node. That earlier sketch and minimal-embodiment differ in several design choices: a single-board computer with edge inference capacity (Raspberry Pi plus Hailo-8) versus a microcontroller without on-board inference (ESP32); ROS2 or micro-ROS messaging versus an HTTP bridge; an external-reception orientation (a pan-tilt camera surveying the environment) versus a self-reception orientation (loops in which the system’s own output returns as input). The two sketches are not in competition — they sit at different points in the design space and address different subsets of the Aktuanz-layer requirement set. Minimal-embodiment privileges minimum cost and the self-reception axis; the earlier sketch privileges edge-side inference capacity and the external-reception axis. Future citizen-track work plausibly combines features from both.</p><p>The next section connects the implementation to existing technical assets relevant to a parallel-holon proposal — embedded RTOS lineages, industrial robotics, and semiconductor industry — particularly as they are positioned in Japan.</p><h3>7. Connection to Existing Technical Assets — The Dual-Track Across Commercial and Citizen Implementation</h3><p>This section develops, structurally, how existing technical assets the author has discussed in companion essays — industrial-robot systems, embedded RTOS lineages, and the semiconductor industry — connect to citizen-level implementation prototypes.</p><h3>7.1 Possibilities of Connection to the ITRON Family</h3><p>minimal-embodiment is implemented as an Arduino sketch for the ESP32 (C++, approximately 1,100 lines, as described in Section 5.2 of the paper), operating on FreeRTOS (the real-time OS bundled in the ESP32 toolchain). Three parallel execution units — the periodic sensor POST via the Arduino loop() (at 10-second intervals), the long-poll via a dedicated FreeRTOS task commandPollTask, and the Wi-Fi/TCP/IP stack (managed by the ESP32 SDK, transparent to the user) — run distributed across cores 1 and 0.</p><p>The technical feasibility of connection to the ITRON family (μITRON 4.0, the TOPPERS project), discussed in earlier work, holds. TOPPERS-related community projects have explored ESP32 ports as derivative work — these are community-driven rather than official TOPPERS Project releases, and the four properties identified for the ITRON family — deterministic behavior, lightness, parallel-processing-premised design, distributed operation — carry a path of integration with the minimum implementation of minimal-embodiment.</p><p>Concretely, the path of replacing the current FreeRTOS-based implementation with a TOPPERS/ASP-based implementation is conceivable. The two share design philosophy in static task structure, priority-based preemption, and the guarantee of real-time behavior, but the TOPPERS family has stricter scope of deterministic-behavior guarantee and inherits design knowledge for industrial embedded systems. The TLS heap-exhaustion problem documented in Section 5.4 could occur similarly under a TOPPERS implementation; combined with TOPPERS’s memory-management mechanisms (fixed-length memory pools, variable-length memory pools), the behavior could be structurally different.</p><p>Connecting the citizen track to Japan’s embedded-OS technical assets works in the direction of thickening the material foundation of the parallel proposal. Against the current concentration of the ESP32 market on Chinese-system makers (Espressif Systems), the path of integrating the TOPPERS family of embedded OSes into the citizen track functions as one option for alleviating structural dependence on a Chinese-origin ecosystem at the microcontroller layer — a different kind of structural dependence from the US-style aggregationism the author has discussed at the LLM layer in companion work, and treated here as Chinese-origin supply-chain concentration rather than as another instance of the same five-component aggregationist pattern. The accessibility condition of the citizen track (constructible in several days without prior soldering experience) requires preconditions on this path — an Arduino-compatible wrapper for the TOPPERS implementation, documentation maintenance for TOPPERS integration. These connect directly to questions of education and social infrastructure.</p><h3>7.2 Possibilities of Connection to Industrial-Robot Systems</h3><p>The industrial-robot families (FANUC, Yaskawa Electric, Kawasaki Heavy Industries, Nachi, Denso Wave, and others) and minimal-embodiment-style citizen-implementation prototypes differ in scale, use, and market. The two stand in functional division of labor. The commercial implementation carries the Aktuanz layer of industrial sites with strong irreversibility at high cost; the citizen implementation carries the Aktuanz layer of education, experimentation, and independent verification with weak irreversibility at low cost.</p><p>Briefly listing conceivable connection paths: a bridge layer connecting the control protocols of industrial-robot systems (many proprietary protocols, with ROS/ROS2 support expanding) with minimal-embodiment-style minimum implementations; configurations combining educational low-cost models such as Denso Wave’s COBOTTA series or Nachi’s collaborative robots with minimal-embodiment-style sensor bodies; reference implementations simplifying the control technology of industrial-robot systems for the citizen track. These are points whose skeletons are presented here, with specific design left as a future task.</p><h3>7.3 The Strategic Position of the Semiconductor Industry</h3><p>Apart from the ESP32 (Espressif Systems, of Chinese origin), Japan’s semiconductor industry holds component bases supplyable to the citizen track. The two product families directly relevant to citizen-track minimum implementations are Renesas Electronics’s embedded microcontroller families and Sony Semiconductor Solutions’s single-board offerings.</p><p>Renesas Electronics (rooted in the embedded-semiconductor lineages of the former Mitsubishi Electric and Hitachi) offers embedded microcontroller families such as the RX and RA lines. The product line includes high Arduino compatibility, with the Arduino UNO R4 carrying the Renesas RA4M1. A note is warranted: the Arduino UNO R4 family has two variants, Minima (RA4M1 alone) and WiFi (RA4M1 combined with an ESP32-S3); the latter is a hybrid with an ESP32-S3 chip — a sign that the position of Japan’s semiconductor industry on the citizen track is structured as hybrid with the ESP32 ecosystem. The Sony Semiconductor Solutions SPRESENSE board (carrying the CXD5602, with six-core Arm Cortex-M4F, Wi-Fi extension support, low-power design, and CMOS image sensor integration) is also an alternative candidate for minimal-embodiment-style implementation.</p><p>The larger Japanese semiconductor-strategy picture — the combination of Rapidus’s 2 nm mass-production plan, JASM (the TSMC Kumamoto plant), Kioxia memory, and Sony Semiconductor’s broader CMOS image-sensor portfolio — operates primarily at commercial and national-strategic scales rather than at the citizen-track scale that is the central concern here.[¹] The Renesas microcontroller and Sony SPRESENSE lines are the parts of Japan’s semiconductor industry that touch the citizen track directly; the rest is recorded in a footnote rather than expanded in the body of this analysis.</p><p>The current concentration of the citizen track on the ESP32 (of Chinese origin) carries citizen-track-level structural dependence on a Chinese-origin ecosystem, distinct from the US-style aggregationist concentration at the LLM layer. Citizen-oriented expansion by Japan’s semiconductor industry carries the possibility of becoming a path that alleviates this lower-layer dependence. The path, however, carries challenges on both the technical-possibility and the social-deployment sides — Renesas and Sony have business structures centered on the commercial track, and the development of communities, documentation, and educational application for the citizen track has a substantial gap in scale compared with the ESP32 ecosystem (Chinese-origin makers combined with a worldwide maker community).</p><p>[¹]: For the relation of Rapidus, JASM, Kioxia, and Sony’s broader semiconductor positioning to a parallel-holon international-strategy reading, see the author’s companion essay on international AI strategy. That essay treats these institutions at the commercial-strategic and national-strategic levels relevant to the upper-layer LLM and AI-chip discussions, not at the citizen-track microcontroller level that this essay is concerned with.</p><h3>7.4 The Current State of the Citizen-Implementation Foundation in Japan</h3><p>The social foundation of the citizen track in Japan already exists in multiple domains.</p><p>Maker-community foundations — Maker Faire Tokyo (held annually in recent years), regional FabLabs (FabLab Kamakura, FabLab Kannai, FabLab Kita-Kagaya, FabLab Hamamatsu, and others), and university Maker Spaces (the University of Tokyo, Kyoto University, Keio University SFC, and others) — function as the foundation supporting the social diffusion of the citizen track.</p><p>Embedded-system education institutions — the 51 national colleges of technology (kosen), vocational schools, and the embedded-system education of university engineering departments — have long functioned as the human-resource supply foundation for embedded-system implementation. The ITRON-family community (the TOPPERS project, various study groups, the Japan Embedded Systems Technology Association JASA, and similar) functions as the bridge between industry and education.</p><p>These foundations, however, tend to be positioned as complements to the commercial track. Maker communities as venues for citizens’ creative expression, education institutions as human-resource supply organs for industry, the ITRON community as the embedded-system technical accumulation of industry — all are organized with the commercial track as the main axis. The work of positioning the citizen track as an independent category institutionally is, at this stage, limited.</p><p>For the social implementation of the citizen-constructible distributed and low-cost substrate the parallel proposal calls for, the work of reorganizing these existing foundations from the standpoint of the citizen track is required. The social positioning of the citizen track as an independent category standing alongside the commercial track, rather than as a complement to it — this is taken up directly in considerations of education and social infrastructure, and is recorded here only as a structural observation.</p><p>If a project at the scale of minimal-embodiment were deployed in Japan as an equivalent case, how the existing foundations connect — individual implementation at companies and research institutions, reproduction implementation in the maker community, educational-material adoption in education institutions, embedded-OS replacement in the ITRON community, component-base coupling with the semiconductor industry — functions as a concrete index for measuring the social-deployment feasibility of the parallel proposal’s counter-hypothesis.</p><p>The next section organizes the structural weaknesses surfaced here and the design challenges that follow.</p><h3>8. Structural Weaknesses and the Next Design Challenges</h3><p>This section frankly organizes the structural weaknesses of minimal-embodiment and the citizen track as a whole that the analysis has surfaced, and presents the design challenges that follow. The description of structural weaknesses from the front is required so that the parallel proposal’s counter-hypothesis does not fall into the simple claim that it is “better than aggregationism.”</p><h3>8.1 Weakness 1 — Weak Physical Irreversibility</h3><p>The most conspicuous structural weakness of minimal-embodiment is the weakness of physical irreversibility. A buzzer tone dies out in a short time; motor vibration is transient; strong physical irreversibility — breaking a cup, falling over and damaging parts, deforming an object — is not implemented. Section 9.2 carefully delimits the scope of this implementation, and the delimitation implies not only the absence of body schema but also the absence of devices producing strong physical irreversibility.</p><p>This weakness can be partially resolved by complementation with industrial robots on the commercial side, but it remains a structural limit for the citizen track in isolation. The accessibility of the citizen track (constructible in several days without prior soldering experience) and the production of strong physical irreversibility stand in a structural trade-off. Devices producing strong irreversibility (high-speed rotating motors, mechanical grippers, heating elements, and similar) require safety management, specialized skill, and regulatory compliance, putting tension on the low-barrier accessibility of the citizen track.</p><p>As a design challenge, the search for paths that gradually strengthen irreversibility while preserving the low-barrier accessibility of the citizen track surfaces. The technical fix-ability suggested by Section 9.2 — adaptive sample timing, higher-sampling-rate IMUs, multi-sensor fusion — is one direction, but a fundamental increase in irreversibility strength requires a separate line of consideration.</p><h3>8.2 Weakness 2 — The Symbol-Grounding Problem and Upper-LLM Centralized Dependence</h3><p>The symbol-grounding problem presented in Section 2.2 is a fundamental structural point for the parallel proposal’s counter-hypothesis. The symbol-grounding problem, classically associated with Searle’s Chinese Room argument but not reducible to it, applies directly to the structure of minimal-embodiment. Physical quantities at the sensor side are converted on the edge into JSON, routed through the bridge layer, and transmitted to a cloud-hosted LLM. The LLM has not “heard” sound; it has “read” a number. The physical causal chain remains, but it reaches the LLM only as symbolic mediation — continuous physical presence is discretized and symbolically mediated at the bridge layer. Symbol grounding is not implemented in this architecture.</p><p>This connects directly to the upper-LLM centralized dependence discussed in Section 6.2. minimal-embodiment demonstrates a path of constructing the Aktuanz layer on the citizen track, but the Panich layer (middle module-parallel layer) and the Kontextleib layer (upper contextual-emergence layer) continue to depend on centralized LLMs (cloud-aggregated models). This structural dependence is a decisive constraint for full implementation of the parallel proposal’s counter-hypothesis.</p><p>The next realistic citizen-track step is not microcontroller LLM inference but microcontroller-side feature extraction, small classifiers, finite-state modules, and local policies. The design challenge that surfaces is how to architecturally connect the rupture between symbol and physical, or how to implement the transfer of inference to the edge side. Three directions for extending the Panich and Kontextleib layers to the citizen track are conceivable.</p><p>First, the implementation of small classifiers, feature extractors, finite-state modules, and small policies operable under the resource constraints of microcontrollers such as the ESP32. This is a realistic short-term path, not running LLM inference itself on the microcontroller, but handling sensor feature extraction, state determination, and pre-processing in front of the LLM input.</p><p>Second, mid-scale LLM implementation on mid-tier single-board computers such as the Raspberry Pi. Light inference engines such as llama.cpp and Ollama make low-speed inference for strongly-quantized small LLMs and 7B-class models increasingly possible. 13B-class models or long-context inference is still difficult in practice on 8 GB-class SBCs. The “constructible in several days” property of the citizen track can be preserved, but cost increases several-fold compared with ESP32 implementation.</p><p>Third, microcontroller implementation of BitNet-class lightweight LLMs is reserved as a research challenge. BitNet is a line achieving substantial weight reduction through 1-bit quantization, but current BitNet implementations presuppose operation on GPUs and TPUs; microcontroller operation requires substantial additional optimization. Rather than as a next stage of the citizen track, it is realistic to position it as a research challenge.</p><h3>8.3 Weakness 3 — Single-Body Operation Model</h3><p>minimal-embodiment is implemented as a single physical system. Multi-module coordination (the core of the parallel proposal) is not implemented. Communication protocols among multiple minimal-embodiment units, coordinated action across multiple bodies, integrated distributed sensing — these remain as design challenges.</p><p>Extension paths from the single-body operation model to the multi-module coordination model require concrete design on both the hardware and the software sides. Configurations in which multiple minimal-embodiment units share the same bridge, configurations in which multiple bridges distribute integrated direction from an upper LLM, configurations in which multiple ESP32 units communicate directly with each other — these are positioned as implementations of the distributed substrate of the lower nodes the Panich layer (the middle parallel layer of the parallel proposal) connects to. The Panich layer itself requires parallelism among different cognitive modules; multi-ESP32 coordination is not the Panich layer as such but the distributed substrate of the lower nodes it connects to.</p><h3>8.4 Weakness 4 — Strength of the Self-Reception Loop</h3><p>Both self-reception loops in Section 6 (audio and haptic) carry limits in strength.</p><p>The audio loop’s buzzer-to-microphone system is structurally difficult to separate from external acoustic sources. In a fully quiet room, self-reception is clearly detected, but in situations in which voices of others, environmental sound, and other physical phenomena coexist, additional work to separate the physical trace of the self-output from other acoustic signals is required. The validation in Section 6.3 was conducted in a quiet indoor environment (with environmental baseline median around 37–40 dB SPL); the loop’s behavior in noisier environments (cafés, public transit, and similar) is unverified. Section 9.1 explicitly delimits this — the signal-to-noise margin in noisier rooms is unmeasured.</p><p>The haptic loop’s motor-to-accelerometer system relies on direct mechanical coupling (placing the ERM coin motor loosely on the MPU-6050 board). The configuration described in Section 6.2 is an engineering choice maximizing vibration transmission, but when external force is applied to the body (a person tapping it, the device placed on a vibrating floor, and so on), structural separation of self-output from external input is difficult.</p><p>The design challenge that surfaces is the design of self-reception loops on more complex physical systems. Multi-modal fusion (synchronized detection of audio and haptic, of vision and haptic), temporal-information precision (precise correspondence of transmission time and reception time), self-output modeling (differential detection between prediction of a known output waveform and observation) — these are conceivable technical paths.</p><h3>8.5 Weakness 5 — The Risk of Being Drawn into Interpretive Debates</h3><p>As Section 4 treated, the citizen track has a structure that makes it easy to be drawn into interpretive debates (AI consciousness, AI welfare, AI ontology). The institutional difference between the paper format and the public-SNS format generates the temperature gap structurally. The separation of structural fact from interpretive claim is a point common to the citizen track as a whole, to be treated continuously. The meta-level self-reflection (Section 4.5) — that this reading also projects a particular interpretive framework onto minimal-embodiment — lies in the extension of this point.</p><p>The design challenge that surfaces is the construction of public formats that distinguish description of structural fact from statement of interpretive claim. Paths to maintain the careful description of the paper format while also enabling social diffusion through SNS — public formats explicitly separating “structural-fact layer” and “interpretive-claim layer,” public conventions that make the temperature gap between the two visible to readers, meta-linguistic infrastructure for the citizen track — are conceivable. These are not technical design but social and methodological design.</p><h3>8.6 Connection to Considerations of Education</h3><p>The five weaknesses above cannot be overcome by technical solutions alone; they connect directly to questions of education and social-foundation institutionalization. Citizens’ construction of distributed and low-cost substrates requires human-resource cultivation, social institutional infrastructure for the citizen track, and the sharing of habits that situate specialized knowledge across multiple fields on a shared map. These constitute the substantive horizon onto which the present analysis opens.</p><h3>9. Synthesis — The Distance Between Design Requirements and Material Implementation</h3><p>This essay has read Olivia Zhu’s minimal-embodiment as material for measuring the distance between the design requirements of a parallel holon architecture proposal and a piece of material implementation that exists. This closing section organizes the points and draws the position of this reading.</p><h3>9.1 Organizing the Points</h3><p>First, minimal-embodiment is structurally analyzable as a case of minimum implementation of the Aktuanz layer and is positioned not as “achievement of parallel AGI” but as “presentation of the minimum condition.” The surface conditions — €125, several days, constructible without prior soldering experience, openly published — function as material evidence for the parallel proposal’s call for a citizen-constructible distributed and low-cost substrate.</p><p>Second, the self-reception loop as a structural axis functions as a complementary axis orthogonal to the external-reception axis treated in companion work. The structure in which “information coming from the environment” and “the physical trace of information the system itself sent out” mix on the same sensory channel implements, at minimum cost, the active-engagement-to-reflection closed circuit drawn from a Japanese epistemological tradition.</p><p>Third, the structural contrast between relational embodiment and task-oriented embodiment can be incorporated as different implementation forms of the Aktuanz layer in the parallel proposal. The dual-track of commercial and citizen implementation works in the direction of thickening the social foundation of the proposal.</p><p>Fourth, the separation of structural fact from interpretive claim functions as a structural point common to the citizen track as a whole. The temperature gap between the paper format’s careful description and the SNS format’s narrative description is a phenomenon the citizen track structurally generates, and the position taken here is separation, not criticism. The meta-level self-reflection — that this reading itself projects a particular interpretive framework onto minimal-embodiment — lies in the extension of this point.</p><p>Fifth, the Human-LLM Co-Design Methodology — division of labor (the boundary between physical and symbolic layers), asymmetric error detection, cross-version review, project memory — functions as a methodological structure of the citizen track. The mutual complementation of the researcher with direct access to the physical layer and the LLM with direct access to the symbolic layer structurally compensates for the blind spots of single-measurement-grid evaluation.</p><p>Sixth, the structural weaknesses — upper-LLM centralized dependence, symbol-grounding problem, weak physical irreversibility, single-body operation model, and the risk of being drawn into interpretive debates — cannot be overcome by technical solutions alone, and require institutionalization of educational and social foundations. Existing technical assets (the ITRON family, industrial-robot systems, the semiconductor industry) and existing citizen-implementation foundations (maker communities, embedded-system education institutions, the ITRON community) already exist as material and human foundations supporting these institutionalizations, particularly in Japan.</p><h3>9.2 The Position of the Reading</h3><p>What this reading shows is the minimum demonstration that the parallel proposal’s counter-hypothesis is not “a logically coherent thought experiment” but “a structure that can be materially implemented.” The fact that the lowest layer is constructible at €125, in several days, without prior soldering experience, shows that the material foundation of the proposal can already be established at the citizen level — without waiting for the arrival of commercial implementation.</p><p>What minimal-embodiment showed is not that AI got a body. It is that the minimum physical-echo circuit connectable to AI is established at a price, parts count, and verification format constructible by citizens. This fact functions as material evidence for the parallel proposal’s call for a citizen-constructible distributed and low-cost substrate. Nothing more, nothing less.</p><p>Design-requirement presentation is not implementation guarantee. The position of the parallel-holon reading — not “the parallel route is the correct answer” but “the structural availability of options other than aggregationism” — is preserved here. The proposal’s counter-hypothesis is not “better than aggregationism”; it is an alternative carrying a different trade-off structure. The five structural weaknesses presented in Section 8 are the concrete manifestation of this different trade-off structure.</p><p><em>References:</em></p><p><em>Zhu, Olivia. “A Minimal Self-Perceiving Embodiment for Large Language Models.” Zenodo, April 30, 2026. DOI: 10.5281/zenodo.19903098. Creative Commons Attribution 4.0 International. Companion repository: github.com/oliviazzzu/minimal-embodiment, MIT License.</em></p><p><em>Ahn, M., Brohan, A., Brown, N., et al. (2022). “Do As I Can, Not As I Say: Grounding Language in Robotic Affordances” (SayCan). Brohan, A., Brown, N., Carbajal, J., et al. (2023). “RT-1: Robotics Transformer for Real-World Control at Scale.” Driess, D., Xia, F., Sajjadi, M. S. M., et al. (2023). “PaLM-E: An Embodied Multimodal Language Model.” Huang, W., Wang, C., Zhang, R., et al. (2023). “VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models.” Liang, J., Huang, W., Xia, F., et al. (2023). “Code as Policies: Language Model Programs for Embodied Control.”</em></p><p><em>Companion essays referenced in passing — on parallel holon architecture, the four-channel sensor configuration for the upper contextual body, the right to refuse measurement, and the dual-hypothesis framework — are available in the author’s earlier published writing, primarily in Japanese at note.com/mototchen and in English at medium.com/@izayohi.</em></p><p><em>The Japanese original of this essay is at note.com/mototchen/n/n9ce55e25a79a.</em></p><p><em>This English version was written ahead of the Japanese-language platform’s auto-translation feature launch (May 27, 2026) as the author’s controlled rendering of the argument. Anyone arriving from a machine-translated Japanese article is encouraged to read this version for the author’s intended phrasing.</em></p><h3>About This Essay</h3><p>Author: Motohisa Ishibe (<a href="https://x.com/mototchen">@mototchen</a>)</p><p>AI co-work:Claude Opus 4.7</p><p>Peer review:Gemini, Grok, GPT</p><p>This essay is part of the Shin-Nexialism project.</p><p>JP: <a href="https://note.com/mototchen">note.com/mototchen</a> | EN: <a href="https://medium.com/@izayohi">medium.com/@izayohi</a></p><p><em>Japanese version published on note (@mototchen):</em></p><p><a href="https://note.com/mototchen/n/n9ce55e25a79a">https://note.com/mototchen/n/n9ce55e25a79a</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98915f6a3145" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>