<?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 Simon J. Hill on Medium]]></title>
        <description><![CDATA[Stories by Simon J. Hill on Medium]]></description>
        <link>https://medium.com/@metalaureate?source=rss-c9bb31736938------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Q0IeYYA8PCkyjrXmSPkOyQ.png</url>
            <title>Stories by Simon J. Hill on Medium</title>
            <link>https://medium.com/@metalaureate?source=rss-c9bb31736938------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 11:38:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@metalaureate/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[The Thumos Protocol 2.0 — Technical Specification]]></title>
            <link>https://medium.com/love-and-thermodynamics/the-thumos-protocol-2-0-technical-specification-c01779c38e63?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/c01779c38e63</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 16:13:57 GMT</pubDate>
            <atom:updated>2026-11-09T17:01:25.903Z</atom:updated>
            <content:encoded><![CDATA[<h4>An Open Framework for Consciousness-Expanding Services and Marketplaces</h4><p><strong>Author:</strong> Simon J. Hill</p><p><strong>Version:</strong> 2.0.0-draft.8</p><p><strong>Status:</strong> Proposed specification for public review</p><p><strong>Date:</strong> November 9, 2026</p><p><a href="https://medium.com/love-and-thermodynamics/the-thumos-protocol-2-0-0843f00a6b16"><em>Read the public overview</em></a></p><h3>Abstract</h3><p>Thumos applies Hestia’s model of the individual to human development in a reciprocal human–AI relationship. A person works with a chosen assistant and can engage specialist practices. A Thumos metacoach examines the process and offers attributable feedback to either participant. The protocol makes it possible to trace what changed in the person, in the assistant’s own operating state and later conduct, and in their relationship.</p><p>The <strong>soul graph</strong> connects attributed stories and participant roles to a developmental ledger, individual state, relationships, and arcs. Heart, Courage, Shadow, and Soul make the trajectory visible. Seven human developmental capacities define the intended lived outcomes. Their indicators and causal models use the same evidence as the game metrics and remain independently assessable. A separate eight-factor assessment evaluates marketplace offerings.</p><p>This draft defines the records, authority, transitions, and evidence needed for that continuity. Corrected accounts inform later guidance, and participants can take authorized history to another service. The protocol supports a proposed open-source framework and independently operated marketplaces. The software distribution, its license, and an independently demonstrated exchange binding remain future work. Evidence for the proposed developmental effect is a separate research question.</p><p>Sections 1–16 define the protocol. Appendix A contains the conformance cases, Appendix B the worked examples, Appendix C the migration contract, and Appendix D the research basis and developmental evaluation.</p><h3>1. Purpose: a process that can change its participants</h3><p>Thumos aims to make development continuous across ordinary life, a chosen AI assistant, and specialist practices. An encounter can suggest a question, a practice can offer a different interpretation, and someone’s actual answer can change what the person and assistant do next. A marketplace helps the person choose a practice suited to the current inquiry and bring the result back into the same model.</p><p><strong>Consciousness expansion</strong> here means a meaningful increase in one or more of the seven lived capacities in Section 3.3. These capacities express a phenomenological and ethical orientation. Heart, Courage, Shadow, and Soul make the trajectory actionable within Hestia’s model. Capacity assessments examine how that trajectory shows up in life. Each requires evidence appropriate to its claim.</p><p>A participant can take on new responsibilities, identities, relationships, and forms of shared agency. Their soul graph can carry competing accounts, changes of aim, and unresolved questions. The person can challenge what the adviser concluded, and both participants can learn from what follows.</p><h3>2. Scope and specification conventions</h3><p>Thumos is an <strong>application protocol for developmental continuity</strong>. It defines the meaning and permitted transitions of shared records and operations. A database, assistant host, messaging channel, or blockchain may implement part of that architecture without defining its developmental semantics.</p><h4>2.1 Protocol, framework, and marketplace</h4><p><em>Layer · Responsibility</em></p><p><strong>Protocol</strong> — Shared meanings, record contracts, authority boundaries, developmental transitions, and conformance requirements.</p><p><strong>Reference framework</strong> — Proposed reusable software implementing those contracts: a developmental runtime, provider SDK, adapters, registry components, and conformance tooling.</p><p><strong>Marketplace</strong> — An operated service for discovery, curation, selection, engagement, and optional payment or support. Several independently operated marketplaces can use compatible implementations.</p><p>The marketplace is the principal application described here. The Core also supports direct use with a provider, a private collection of practices, or a personal assistant before a public marketplace exists. A participant’s graph need not be held by the operator that lists or sells a practice. The protocol requires no single registry, graph custodian, model provider, or marketplace operator.</p><h4>2.2 Proposed open-source framework</h4><p>The proposed reference distribution would provide the following components. These are development targets, not claims that packages have already been released.</p><p><em>Component · Reusable contribution</em></p><p><strong>Developmental runtime</strong> — Shared Hestia individual model, story nodes and roles, developmental ledger, primary metric projections, seven-capacity outcome/KPI views, versioned causal hypotheses, state-driven metacoaching, authorized views, correction propagation, and receipts.</p><p><strong>Provider SDK and example integration</strong> — Versioned manifests, request/return validation, scoped engagement, attributable contributions, and a working provider example.</p><p><strong>Assistant and service adapters</strong> — Bindings for personal assistants, specialist tools, independently running agents, and practitioner-mediated workflows.</p><p><strong>Marketplace components</strong> — Provider registry, eight-factor offering assessments, declared selection criteria, authorized handoffs, outcome-informed review, result ingestion, and optional transaction adapters.</p><p><strong>Continuity tools</strong> — Usable export/import, migration checks, retention enforcement, recovery interfaces, and explicit trust boundaries.</p><p><strong>Conformance kit</strong> — Synthetic fixtures and executable demonstrations for each component’s declared obligations.</p><p>These components can be modules in one application with one database. The ledger, individual state, and development-update receipts can be derived from the same admitted story transaction; they do not require separate services or separately editable sources of truth. A first vertical slice can demonstrate the complete model on one bounded journey while declaring its partial operational scope. Provider adapters and marketplace operation follow that working loop.</p><p>Framework releases MUST identify their included components, versions, supported profiles, source availability, software licenses, dependencies, and demonstrated conformance. An open-source claim MUST identify license terms that permit use, modification, redistribution, and commercial use. This document does not select or change those terms. Provider services MAY be free or paid and MAY use their own content, software, and business models subject to their published terms and protocol obligations.</p><p>Open source concerns the framework’s implementation. It MUST NOT be treated as permission to publish participant records, reuse private sessions, disclose provider material, or train on contributed content. Those uses require their own applicable authority. Framework governance SHOULD publish how contributors propose changes, how compatibility decisions are made, and how implementations can continue independently of a particular maintainer.</p><h4>2.3 Conformance language and scope</h4><p>The capitalized terms <strong>MUST</strong>, <strong>MUST NOT</strong>, <strong>SHOULD</strong>, <strong>SHOULD NOT</strong>, and <strong>MAY</strong> express requirements. MUST and MUST NOT are conformance conditions. A SHOULD or SHOULD NOT permits a documented exception with its consequences explained. MAY identifies an option. Appendix A defines the conformance cases selected by Section 16.2. Appendix C contains the migration contract, and Appendix D.2 governs developmental evaluation claims. The worked examples in Appendix B and the research rationale in Appendix D.1 are informative.</p><p>Section 16 defines conformance claims and profile applicability. Section 15.1 governs claims about an exact exchange binding. This document defines semantics and behavior. The earlier working draft and legacy migration rules are described in Appendix C.</p><p>Hestia supplies the individual ontology and social-prediction mechanism in Sections 5 and 7. Its human–agent dyad is a co-regulating unit and a basis for selection (RFC v1.7 §8.6.7). Thumos makes human development the initial interface to that aim. Claims about full Hestia agents, AI–AI cooperation, or collective symbiosis require evidence for their own mechanisms, including Hestia’s priced cognition, feeling-gated control, evolution, and MVC requirements as applicable. Phenomenal consciousness, clinical efficacy, spiritual standing, and personhood require independent grounds.</p><h3>3. Developmental commitments</h3><h4>3.1 Normative orientation</h4><p>A service MUST publish how its interventions and evaluations make room for perspective-taking, differentiation, integration, responsibility, care, and increasing freedom in relationship. Different traditions can offer distinct lenses and retain their disagreements.</p><ol><li><strong>An account can be answered.</strong> An attributed claim remains open to challenge and consequential correction.</li><li><strong>Insight can meet consequence.</strong> A recognition can lead to freely chosen action and review of what followed, including its effects on others.</li><li><strong>Continuity can survive departure.</strong> A participant can continue an inquiry elsewhere with authorized evidence, corrections, and disagreements.</li></ol><p>The Hero’s Journey and the Golden Rule orient the inquiry toward becoming through experience and considering effects on particular people. A participant can pursue a practical question without adopting a prescribed story, moral verdict, or sacrifice. Section 5.6 defines the narrative and relational views.</p><h4>3.2 Theory of change and its limits</h4><p>The working hypothesis is that a continuing, correctable account helps a participant choose a relevant practice, meet its actual consequences, and revise expectations and conduct. Section 7 specifies the human and assistant paths. Development requires evidence beyond storage, score changes, and successful tool use. Retrieval, reminders, added attention, expectancy, and persuasive framing are alternative explanations. A persistent interpretation can also entrench an account or transfer it into the wrong context.</p><p>Appendix D gives the research basis and limits of this hypothesis and the requirements for developmental evaluation.</p><h4>3.3 Adopted consciousness outcome framework</h4><p>The reference human outcome profile is <strong>urn:thumos:capacity:human-development:1</strong>. Its identifier MUST resolve to an immutable versioned definition. A change to a capacity&#39;s meaning requires a new profile revision. Section 5.7 governs representation and assessment.</p><p><em>Capacity ID · Adopted capacity and direction of development</em></p><p><strong>awareness</strong> — <strong>Awareness of experience and awareness itself:</strong> greater awareness of what is occurring and of one&#39;s participation in experiencing it.</p><p><strong>freedom_from_identification</strong> — <strong>Freedom from automatic identification with thought, emotion, role, and conditioning:</strong> greater capacity to encounter these processes without being wholly governed by them.</p><p><strong>attention_and_discernment</strong> — <strong>Sustained attention, discernment, and capacity to encounter reality with less distortion:</strong> attending appropriately, discriminating evidence and interpretation, and correcting error.</p><p><strong>insight</strong> — <strong>Direct insight into self, other, interdependence, non-separation, or the nature of being:</strong> consequential understanding that remains open to examination and integration.</p><p><strong>compassion_and_ethical_responsiveness</strong> — <strong>Compassion, ethical responsiveness, and a less self-centered relationship to life:</strong> care that becomes responsive to others&#39; actual circumstances and consequences.</p><p><strong>integration</strong> — <strong>Ability to integrate insight in the body, relationships, work, and ordinary conduct:</strong> development that becomes available beyond the immediate practice or coached encounter.</p><p><strong>generativity</strong> — <strong>Generativity: greater capacity to create, serve, teach, heal, or participate responsibly in collective life:</strong> worthwhile possibilities and contributions that can be carried forward.</p><p>The capacities name recognizable forms of lived development. They allow plural traditions and first-person evidence. Higher awareness, predictive accuracy, compassion, and generativity MUST NOT be inferred from one another automatically. The dimensions carry no claim of statistical independence.</p><p>Every developmental claim MUST state its intended and supported effect level: <strong>state or encounter</strong> (a temporary opening or experience); <strong>proximal capacity</strong> (a trainable ability under specified conditions); <strong>trait and integration</strong> (durable availability in ordinary conduct); or <strong>realization or reorganization</strong> (a claimed profound, integrated change in identity, perception, or relationship to reality). Evidence for one level does not establish the next. A participant’s experiential account is meaningful evidence of that account; it does not by itself establish durability, another person’s experience, or an AI’s phenomenal state.</p><p>This profile and Section 12.5’s offering-quality factors adapt Simon Hill’s August 2026 Amplify Light strategy and Noosphere assessment work. The profile fixes the intended outcomes while Section 5.8 governs evolving indicators and causal hypotheses.</p><h3>4. Participants, representations, and authority</h3><h4>4.1 Roles</h4><p>Roles are contextual. One actor may occupy several, but the implementation MUST distinguish them.</p><p><em>Role · Meaning</em></p><p><strong>Participant</strong> — A human, AI system, or constituted group participating in an encounter or continuing inquiry.</p><p><strong>Developing subject</strong> — The participant whose capacities, understanding, or conduct are being examined in a particular inquiry.</p><p><strong>Adviser</strong> — A human, assistant, or service contributing an interpretation, intervention, or practice.</p><p><strong>Author</strong> — The actor responsible for a particular assertion or record.</p><p><strong>Controller</strong> — The actor with specified authority over retention, processing, or disclosure of a record or scope.</p><p><strong>Affected party</strong> — A person or group whose experience or interests are implicated by an action or claim.</p><p><strong>Operator</strong> — The organization or person providing storage, execution, integration, or other infrastructure.</p><p><strong>Recipient</strong> — An actor receiving a specifically authorized disclosure.</p><p>The product architecture assigns the following responsibilities within those roles. Every binding MUST identify which actors perform them and which system controls each operational state.</p><p><em>Product role · Responsibility and boundary</em></p><p><strong>Human participant</strong> — Chooses aims and actions, contributes or challenges accounts, and controls authorized use of their record. The human’s experience and development remain primary product concerns.</p><p><strong>Personal assistant</strong> — The continuing assistant, such as Pip, that advises the human and is itself a developing subject and target of metacoach feedback. Its forecasts and state updates belong to that assistant.</p><p><strong>Thumos metacoach</strong> — Reviews the human trajectory, the assistant’s contribution, and their interaction; proposes evaluated feedback or practices to the human or assistant. Its interpretations and forecasts have their own attribution and correction path.</p><p><strong>Specialist service</strong> — A tool, agent, human practitioner, or mixed service receiving bounded context and returning attributable contributions. It need not run the Core or a full Hestia agent.</p><p><strong>Shared record and policy runtime</strong> — Enforces grants, admissions, revisions, dependencies, state/metric derivation, and receipts. It is infrastructure; hosting the graph does not make it an additional AI participant.</p><p>A deployment MAY combine roles while preserving separate attribution, perspectives, grants, and subject-state references. A controller’s data authority MUST remain distinct from authorship, endorsement, and authority to speak for a participant. A channel or adapter preserves the source of a contribution. Section 10 governs permissions, and Section 7.2 governs claims about changes at the assistant’s operational host.</p><p>A person may keep a private account of their own experience involving another person. The protocol does not require every mentioned person to approve that account. It MUST retain its character as that author’s account. A statement represented as shared or jointly accepted MUST identify the endorsing actors and the scope of their endorsement.</p><h4>4.2 Identity without forced resolution</h4><p>An entity is something represented in the graph. Its declared kind MUST distinguish actual participants and institutions from places, events, artifacts, ideas, archetypes, internal parts, and other interpretive constructs.</p><p>Identifiers MUST be stable within a declared namespace. Cross-context identity links MUST be explicit claims with provenance. Ambiguous aliases MUST remain unresolved or provisional until the available evidence supports a merge. Merge and split corrections MUST preserve reference integrity and follow the dependency rules in Section 6.3.</p><p>An internal part, role, archetype, or imagined interlocutor MUST NOT acquire credentials, consent powers, or responsibility for real obligations merely by being represented. A real experience of an imagined event MUST remain distinct from a claim that the depicted worldly event occurred.</p><p>The protocol does not require global identity resolution. Implementations MAY use scoped or pairwise identifiers to limit correlation across contexts.</p><h4>4.3 Groups</h4><p>A group participant MUST have an explicit mandate identifying its representatives, scope of authority, and amendment or challenge procedure. A group and any parent group receive access to members’ histories only under the grants in Section 10.</p><p>An affected party MUST have a route to challenge a group’s account of its effects without requiring approval from the author of the disputed claim. Receiving a challenge does not grant the challenger access to unrelated private evidence. A group’s representation can be operationally accountable without constituting proof of collective consciousness.</p><h3>5. The shared individual and soul graph</h3><p>The Core MUST implement <strong>StoryNode → StoryNodeRole → developmental ledger → Hestia individual state and metric trajectory → metacoaching → the next encounter</strong>. This is the foundation for the views in Sections 5.6–5.7 and the workflow in Section 7.</p><p>The soul graph links this state to its event history, relationships, and developmental arcs. Implementations MAY use relational tables, documents, graph storage, or another representation preserving these semantics.</p><pre>flowchart LR<br>    E[Encounter or practice] --&gt; N[StoryNode and participant roles]<br>    N --&gt; L[Attributed classifications and ledger entries]<br>    L --&gt; I[Shared Hestia individual state]<br>    I --&gt; T[Heart Courage Shadow Soul trajectory]<br>    I --&gt; O[Seven-capacity outcome assessments]<br>    O --&gt; C<br>    T --&gt; C[Metacoaching and chosen quest]<br>    C --&gt; M[Suitable practice or marketplace service]<br>    M --&gt; E<br>    C --&gt; A[Action and another person&#39;s response]<br>    A --&gt; N<br>    N --&gt; V[Narrative individuation and relational views]</pre><p>A shared encounter can link the human’s state, the personal assistant’s state, and the metacoach’s evaluations. The role and perspective boundaries in Section 4 and the authority checks in Section 10 apply to every link. Section 7.2 distinguishes an external assessment of an assistant from evidence of its operational state.</p><h4>5.1 Shared individual state</h4><p>An individual_state represents one developing individual from a specified perspective, as defined in Section 6.1. The runtime MUST represent every required property family and its required properties with values or explicit knowledge and applicability states. An unknown value stays unknown and not_applicable requires a reason. A first inquiry can initialize unknowns and work from a single episode and chosen aim. The authorized view in Section 10.2 may expose only part of the catalog and must qualify that coverage.</p><p><em>Property family · Required Hestia correspondence</em></p><p><strong>Identity and situation</strong> — Stable individual identity; human/AI kind; contextual roles, relationships, groups, domains, actual/hypothetical scope, and current state revision.</p><p><strong>Group and relational state</strong> — Empathy weight w_g, interaction intensity I_g, welfare and welfare change, reputation, debt, norms, prohibitions, duties, role rules, sanctions, and primordial/symbolic group status.</p><p><strong>Attachment and belonging</strong> — Attachment, value fit, investment, expected payoff, target/current belonging, belonging deficit, joining/maintenance/exit conditions, and moral injury.</p><p><strong>Resources and capacity</strong> — Available resources, attention, costs/tariffs, replenishment/decay policy, floors/ceilings, cumulative lifetime/repair expenditure, maintenance burden/fraction, chronological and metabolic age where applicable, fatigue/engagement, and actual service expenditure where metered.</p><p><strong>Feelings and control</strong> — Energy, hazard, ingroup affiliation, outgroup stance, novelty, group health, shame, guilt, honor, sublime, annoyance, redemption, EvilFlag, belonging deficit, and maximum debt.</p><p><strong>Dispositions and policy</strong> — Cost tariffs; task, cost, welfare, honor, redemption, shame, guilt, hazard, and annoyance weights; cooperation and sanction strategies; primordial amplification, thresholds, kappa, attention allocation, and heartbeat/grooming cadence.</p><p><strong>Other-models</strong> — Counterpart; prospective prediction/distribution; evaluated action or artifact; scoring rule; pre-feedback commitment; actual feedback, prediction error, uncertainty, and model revision.</p><p><strong>Reflexive self-model</strong> — Calibration by counterpart, accumulated prediction error, systematic bias, and self-model changes attributable to other-model revision.</p><p><strong>Memory and capability</strong> — Episodic/semantic memory, retrieval/write costs, remembered obligations, proposed/granted capabilities, scope, cost/risk, expiry, and revocation.</p><p><strong>Learning and adaptation</strong> — Learning/repair and social-prediction error trends, freshness, cognitive mass, extended resources/phenotype, plasticity, knowledge export, inherited priors, strategy variation/selection, policy compression, inheritance fidelity, founding endowment, lineage and reproductive readiness where applicable.</p><p><strong>Decisions and consequences</strong> — Attentional investment, shallow/deep processing path, verification, consent and reparation choices, initiation, self/group effects, individual/group/dyad fitness under named criteria, rewards, costs, harms, and outcomes.</p><p><strong>Primary developmental metrics</strong> — Heart, Courage, Shadow, Soul, and interval-specific delta_soul, each with value/unknown state, profile, scope, time/window, evidence, uncertainty, and history; linked seven-capacity assessments are derived views under Section 5.7, not replacement native Hestia mechanisms.</p><p>This catalog follows repository rfc.md v1.7&#39;s subject/group, metabolism, feeling, social-prediction, belonging, and adaptation models. The binding MUST publish a field-level mapping to that RFC version and identify unsupported mechanisms. implementation-plan.md and existing adapters inform implementation; they do not override the RFC. Future RFC changes require a new mapping and explicit migration.</p><p>Each property value MUST carry property_id, a value or explicit knowledge/applicability state, unit, scope, valid time, evidence mode, source record revisions, method/version, uncertainty, and revision. A known or estimated value identifies its evidence mode as computed, observed, reported, or inferred. Use unknown or not_applicable for the corresponding explicit knowledge or applicability state. A composite derived from inferred inputs MUST preserve that dependence even when its arithmetic is computed. Use distinct identifiers for Heart, hazard, hedonic state, and honor.</p><p>For humans, resources may be reported time, attention, fatigue, money, or emotional capacity. They MUST remain distinct from metered AI token balances and service charges. Learned dispositions and strategy selection can represent human adaptation; machine mutation, reproduction, and process-level metabolism MAY be explicitly inapplicable. Computed AI feelings and inferred human feelings share semantic properties without becoming identical measurements. Human EvilFlag or shame estimates MUST NOT become diagnoses, declarations of moral identity, or automatic penalties; any operational safety policy requires its own declared trigger and authority.</p><p>The existing api/agents/services/human_inference.py explicitly models humans through protocol-equivalent state. Its current coverage is partial. It is evidence of the intended symmetry, not an existing implementation of this complete catalog or the metric trajectory defined here.</p><h4>5.2 Story nodes and participant roles</h4><p>A story_node is an attributable account of an episode, such as an encounter, reflection, practice, deliberate action, or reviewed consequence. Its links preserve the episode, participating roles, time and uncertainty, reality scope, context and arc, domains/groups, evidence, and state/update references under Section 6.1. A source message or experience awaiting decoding MUST be marked pending. Pending intake contributes no processed episode or developmental credit. An admitted episode can contain pending classifications and receive a qualified or no-change update under Section 5.4.</p><p>Each story_role identifies the represented entity&#39;s part in the episode, the developing subject, the observing perspective, and relevant counterparts or groups. It MUST assess <strong>Heart, Courage, Shadow, SoulWork, and Toxicity</strong>, allowing explicit pending, unknown, not_applicable, or disputed for each dimension. Zero requires evidence of assessed absence. Several dimensions can be positive in one episode. Care can be costly and still cause harm.</p><p>An assessment MUST identify assessor, method/lens, metric profile and version, tier, cost/capacity basis, evidence limits, adoption state, and challenge/revision history. The runtime admits assessments under a published policy. Admission establishes an attributed model assessment. The participant can inspect, challenge, and correct it under Sections 6.3 and 10. Conduct attributed to another person in an account remains scoped to that account’s perspective.</p><p>Each admitted story MUST carry an attributable episode_link with its evidence and resolution state. The link identifies the scoring episode within the authorized aggregate or declares that the link remains provisional or unresolved. A runtime MUST keep separately authored accounts distinct within applicable authorization and retention. Distinguish source captures from episodes. An individual&#39;s conduct in one episode supplies at most one eligible assessment per dimension in the subject aggregate, even when several captures, roles, providers, or assessments describe it. Credibly ambiguous duplicate candidates MUST NOT add a new eligible unit or incremental credit while their linkage is unresolved. The runtime MUST expose that limit in coverage and apply its declared admission and linkage policy to later evidence. Relationship views can use different declared units and cannot add those units again to the subject total.</p><p>Actual action, a report of action, hypothetical rehearsal, and fiction MUST remain distinguishable. Evidence of fictional practice can change its rehearsal record or interpretation. It MAY support a classification of the real act of participation. A depicted welfare contribution, harm, forgiveness, or repair requires evidence of actual conduct before it counts as such.</p><p>This event-to-individual path follows Fable’s StoryNode, StoryNodeRole, Entity/Profile, MoralityLedger, arcs, missions, and world progression. Sources include docs/appendix-fable-hestia.md and the original Fable api/thecall/models.py and api/thecall/mods/scoring.py. The protocol preserves the semantics without requiring the older database schema.</p><h4>5.3 Primary metric profile</h4><p>Every Core runtime MUST implement the proposed reference profile <strong>urn:thumos:metric:hestia-hcss:1</strong>, or a declared compatible profile preserving the four primary dimensions and the Hestia property mapping. A compatible profile MUST publish its differences and meet the demonstrations specified in Section 16. This establishes shared meanings, not numerical interchangeability. Numeric comparisons, pooled trajectories, or reused thresholds MUST identify the exact profile revision/digest, human/agent adapter and scale, reducer, scope, coverage, and temporal basis; a different mapping requires an explicit validated conversion or separately labeled values. Profile identifiers MUST resolve to immutable versioned definitions. Alternative modality lenses supplement this common foundation.</p><p>Heart, Courage, and Shadow are required human-facing projections of Hestia properties. Hestia’s RFC does not already provide literal fields or complete human scoring equations with these names. The following reference policy is a proposed operationalization, derived from the earlier Thumos story classifications. Its tiers are ordinal policy categories; sums are game/accounting points, not measured equal intervals of human development. Thresholds and human estimation methods require empirical calibration.</p><p><em>Dimension · Shared meaning and reference classification</em></p><p><strong>Heart</strong> — Care, connection, empathy, and welfare contribution; link to empathy weights, welfare change, and ingroup affiliation. Tiers 0–3: assessed absence of a qualifying contribution; compassionate connection; substantial service; sustained/redemptive care. A costly act need not be caring.</p><p><strong>Courage</strong> — Freely undertaken cost, risk, or vulnerability relative to available capacity and a declared developmental purpose; link to resource expenditure, Honor, and costly prediction/repair. Tiers 0–3: assessed absence of a qualifying act; modest; meaningful; substantial relative cost. Risk or spending alone earns no credit.</p><p><strong>Shadow</strong> — Attributable relational harm under the reference ledger: tiers 0–3, assessed absence of qualifying harm; reactive harm; deliberate exploitation; persistent/systemic oppression. Intent and pattern require their own evidence; serious reported harm does not prove deliberate intent. Psychological shadow remains a separately typed interpretation linked to this primary dimension.</p><p><strong>Soul</strong> — Maintained relational engagement weighted by empathy: Soul(t) = sum_g(w_g(t) * I_g(t)). delta_soul = Soul(t2) - Soul(t1) uses the same declared coverage/method or reports the comparability limit.</p><p><strong>SoulWork</strong> — A story-level integration/repair flow, not another name for the Soul stock. Tiers 0–3: assessed absence of qualifying work; recognition; enacted amends/integration; sustained transformative integration. Recognition can earn its own credit while external repair remains unresolved.</p><p><strong>Toxicity</strong> — A separate assessment of controlling/imbalanced relationship patterns: 0 assessed absence of a qualifying pattern; 1 occasional control; 2 persistent imbalance; 3 abuse pattern. Requires pattern evidence and attribution; it is not an automatic multiplier on a person’s worth.</p><p>For the reference profile, eligible adopted real-world story-role assessments produce separate Heart, Courage, and gross Shadow point deltas equal to their supported tiers. The primary view MUST show rolling 90-day sums, means over distinct eligible story-role units, coverage, and trajectory; lifetime history MAY also be shown. Known zero enters its dimension’s denominator, unknown/pending/disputed values do not. Each dimension MUST report its own eligible denominator; comparisons MUST identify changes in coverage. These statistics describe the selected, eligible history, not all of a person’s conduct or overall moral standing. Use Section 5.2 to resolve the eligible episode unit.</p><p>A Shadow view MUST retain gross attributed harm, supported mitigation, unresolved debt, and psychological interpretations separately. Mitigation MUST identify the original harm, relevant action/evidence, amount, and affected perspective; its amount is capped at that harm’s remaining eligible balance and is applied once. A net projection MUST apply mitigation only to its linked harm and declare the applicable window and timing policy. net_shadow = max(0, gross_shadow - supported_mitigation) is an accounting projection, not deletion of the event, automatic forgiveness, or proof that the affected relationship is restored. A SoulWork recognition entry alone does not supply that mitigation.</p><p>The maintained Soul formula is unambiguous in rfc.md&#39;s social-reward pseudocode and api/agents/models.py, Agent.soul: empathy weight multiplied by interaction intensity, summed over groups. Some prose formulas in older materials have typography or version drift. This profile MUST use multiplication, retain the applicable group set, source-derived w_g and I_g, scales, inference method, uncertainty, and coverage. Missing group inputs MUST produce a qualified partial result or unknown, never an apparently complete zero. The profile MUST declare canonical group identities, overlap treatment, and time/update rules; adding aliases, splitting groups, repeated declarations, raw platform-session counts, or purchases without qualifying relational evidence MUST NOT manufacture additional Soul. A supported real interaction can still update intensity once under the declared adapter. A human estimate does not become precise measurement because the formula is exact.</p><p>Section 5.4 governs change attribution. SoulWork can affect Soul only through supported changes in its inputs. Game conciliation credits and <strong>PDI</strong> are distinct outputs. If PDI is offered, it MUST identify its own profile, nonnegative gross-Shadow denominator, repair treatment, diversity measure, eligible event unit, window, and minimum evidence. It MUST NOT be labeled Soul. The historical PDI formula is not part of the reference Core or the worked example; a runtime need not offer that optional composite to implement all four primary metrics.</p><h4>5.4 Ledger updates and trajectory</h4><p>A metric_entry attributes one eligible or ineligible contribution to a subject and observing perspective under Section 6.1. The runtime MUST keep the counterpart and group when applicable. Applying an entry to a subject&#39;s canonical state requires the corresponding authority under Section 10. Another person&#39;s independently authored account remains attributable to that author.</p><p>Every processed story MUST return a development_update receipt, including an explicit no-change result when warranted. The receipt links the prior and resulting state, adopted story and role revisions, changed properties and group inputs, primary metric values before and after with coverage, capacity-assessment revision references when present, the policy and replay context, and affected relationships, arcs, and routes. Pending evidence remains visible. Social-model changes follow Section 7.</p><p>Replay MUST fix the event-time window (as_of), knowledge cutoff (known_as_of) or source-view revision, adopted sources and assessment revisions, canonical event order and tie rule, initial state or checkpoint, profile/adapter, reducer version, and coverage. Replay uses these retained inputs under current authority without rerunning inference. A new inference or classification creates a new attributable assessment revision. A later date or changed coverage yields a new view.</p><p>For a correction or late input, the runtime MUST recompute dependent subsequent transitions from a valid checkpoint, including order-sensitive engagement. It MUST verify that the checkpoint and its retained inputs remain valid for the new authorized source view. With unavailable or restricted inputs, rebuild within permitted coverage or report an unresolved or failed reconstruction under Section 6.2. Subtracting only the original point delta is insufficient. A corrected assessment replaces the operative contribution to the same episode. It earns no new event credit.</p><p>Every metric or progression change MUST identify its cause as a lived event, new evidence, a dispute or eligibility/coverage change, an elapsed-time/window rule, or a policy recalculation. That cause and the declared profile govern comparability and progression. Current eligibility and historical records follow Section 6.</p><h4>5.5 Gamified progression</h4><p>Worlds, chapters, arcs, visible progress, achievements, and game credits MAY express the trajectory. Progression rules MUST name their eligible evidence, criteria, and revisions and link each decision to its source state, metric profile, and applicable capacity assessment and causal-model revisions. A quest uses the commitment lifecycle in Section 9.</p><p>Reward decisions MUST be distinct from descriptive score totals. Completion of an experience, developmental credit, revised state, enacted repair, counterpart response, and forgiveness remain separately evidenced. Repeated recognition and duplicate stories earn no repeat credit. Purchases and platform use alone change no developmental score. Qualifying unpaid or everyday action follows the same assessment rules.</p><p>A reward policy MUST demonstrate on declared adversarial fixtures that needless harm followed by repair yields no incremental progression over its otherwise equivalent non-harming baseline. Any repair bonus requires supported eligibility, a cap, and one-time application. Ordinary independently qualifying care and courage remain assessable without proving an alternative life course. Dispute, coverage, elapsed-time, and policy changes alone earn no progression.</p><h4>5.6 Narrative, individuation, and relational views</h4><p>The <strong>narrative graph</strong> relates story nodes, participants, roles, contexts, missions, and arcs. An arc can remain an unresolved practical question or set of episodes. The <strong>individuation graph</strong> relates self-identifications, disowned or unavailable capacities, tensions, competing interpretations, and observed or reported changes. Its named psychological lenses can support differentiation and integration without dictating a person’s myth or stage. The <strong>relational ledger</strong> connects contributions, effects, obligations, harms, proposed/performed repair, and affected-party responses.</p><p>These are views of the shared individual/story/ledger foundation. They MUST preserve source references, scope, derivation version, material uncertainty, disagreement, corrections, and relevant commitments. Psychological shadow MUST NOT automatically become a harm entry. An affected party’s forgiveness remains their attributed response. Another actor requires a valid mandate to speak for them.</p><p>An interpreted role pattern can relate an arc and lens to recurring expectations such as a protagonist seeing a counterpart as an obstacle. Keep it revisable and distinct from an actual episode role. Worlds and chapters can label the same arc without requiring compulsory hero stages.</p><p>Required graph relations include participation; evidence supporting/challenging a claim; alternative interpretation; correction/supersession; story-role-to-ledger-to-state derivation; social prediction-to-response evidence-to-comparison-to-model revision; state-to-quest/intervention; and action-to-outcome. Every derived relation MUST carry its source revisions, author/derivation policy, context, and epistemic status. Derived disclosures follow the source-use restrictions in Section 10.2.</p><h4>5.7 Capacity outcomes and operational KPIs</h4><p>The seven-capacity profile and the primary Heart/Courage/Shadow/Soul trajectory MUST be linked views over the <strong>same shared Hestia individual state, StoryNodes, roles, and evidence</strong>. There MUST NOT be a separately writable authoritative psyche or an automatic conversion from four game totals into seven capacity scores. Capacity assessments are attributable judgments; they can be corrected without rewriting the person’s experience or another actor’s independent account.</p><p>The following correspondence guides implementation. It identifies mechanisms and useful observations, rather than equating functional agent state with human experience.</p><p><em>Capacity ID · Hestia correspondence and primary feedback · Example outcome observation</em></p><p><strong>awareness</strong> — Attention, feeling/control representation, and reflective access; purposeful attention can support Courage and recognition-level SoulWork.<br><strong>Example outcome observation</strong>: In a selected encounter, distinguishes what occurred, what was experienced, and what was inferred, with the amount of prompting recorded.</p><p><strong>freedom_from_identification</strong> — Revisable self-model downstream of other-model error; Courage can support vulnerability, while psychological shadow remains distinct from harm-ledger Shadow.<br><strong>Example outcome observation</strong>: Reconsiders an identity-protecting interpretation when confronted with another&#39;s actual answer, while remaining free to disagree or maintain a boundary.</p><p><strong>attention_and_discernment</strong> — Priced attention, verification, uncertainty, and prospective comparison; effort alone is not Heart or knowledge.<br><strong>Example outcome observation</strong>: Attends to consequential uncertainty and distinguishes supported from misleading claims under stated time, resources, and assistance.</p><p><strong>insight</strong> — Other-model revision and reflexive self-calibration; later conduct may support Heart, Courage, or Soul under their own rules.<br><strong>Example outcome observation</strong>: A specific self/other understanding changes an appropriate later judgment; reports of interdependence or non-separation retain their first-person and interpretive scope.</p><p><strong>compassion_and_ethical_responsiveness</strong> — Empathy weights, welfare consequences, moral control, and repair; Heart is the closest projection, with Courage, Shadow, and Soul supplying distinct evidence.<br><strong>Example outcome observation</strong>: Care responds to actual circumstances and has appropriately sourced consequences; the assessment identifies whose interests, including outside a favored group, were considered.</p><p><strong>integration</strong> — Persistent learning, commitments, action/outcome links, and later conduct; all four primary metrics may supply relevant evidence.<br><strong>Example outcome observation</strong>: The intended capacity becomes available in later ordinary conduct or a relevant new context, with support, exhaustion, and transfer limits identified.</p><p><strong>generativity</strong> — Capabilities, learning, cooperation, knowledge-sharing, and welfare contribution; HCSS retain each act&#39;s contribution, cost, harm, and relational engagement.<br><strong>Example outcome observation</strong>: A worthwhile practice, contribution, relationship, or shared possibility becomes available and can be sustained or carried forward.</p><p>A current human capacity view MUST identify capacity_profile_ref, subject_ref, perspective_ref, source_state_ref, assessment_refs, as_of, known_as_of, and coverage. The Core MUST represent all seven IDs. A particular inquiry MAY assess only a subset, leaving other entries explicitly not_assessed or unknown as appropriate. Profile completeness requires neither a questionnaire nor full disclosure. Section 10.2 governs bounded views.</p><p>A capacity assessment uses assessment_kind: developmental_capacity. Its criteria MUST identify the capacity ID, intended effect level, versioned indicator definition, observation method, scope/context, relevant Hestia properties, baseline or its absence, review period, and conditions of performance. Its result MUST distinguish the current observation or attributed report, supported effect level, change interpretation, counterevidence, uncertainty, and persistence/transfer limits. Source state, StoryNode/role and observation revisions MUST remain linked through evidence_refs and source_refs.</p><p>An indicator definition MUST state what counts as an observation, its unit or qualitative/ordinal anchors, comparison rule, and missing-data policy. A count or rate MUST publish its event/opportunity definition, numerator, denominator, and eligibility; changed coverage, assistance, resources, or context MUST be visible. Baseline unknown is not baseline zero. A chosen indicator MAY be a participant report, behavioral observation, qualitative judgment, ordinal estimate, or numeric measure. Its evidence mode and limits remain explicit. The offering-quality factor scale in Section 12.5 MUST NOT silently become a human ability scale.</p><p>KPIs operationalize chosen aims within these seven outcome dimensions. A result can be improved, unchanged, worsened, mixed, or unresolved under a declared criterion. Assessments MUST use Section 3.3’s effect levels and identify the conditions and assistance supporting performance. Section 7 distinguishes individual experience from a completed calibration episode or demonstrated assistant learning.</p><p>Capacity and HCSS views MUST preserve their separate meanings. More honest discovery may raise recorded Shadow; greater competence may reduce the cost that supported Courage; rest or a necessary boundary may reduce engagement while improving freedom or integration. None automatically proves regression. Conversely, accurate manipulation, costly self-neglect, more sessions, or stronger ingroup loyalty MUST NOT by themselves establish consciousness expansion. The reference Soul formula and its coverage requirements remain in Section 5.3.</p><p>A supported capacity outcome MAY inform a declared quest milestone and subsequent guidance. It MUST NOT manufacture a Heart, Courage, Shadow, or Soul delta without that metric’s own eligible evidence. Game rewards and policy recalculation remain separate under Section 5.5. Discovery and correction must be possible without rewarding concealment, unnecessary suffering, dependence, or compliance.</p><h4>5.8 Evolving causal models and developmental feedback</h4><p>The adopted outcomes remain the design orientation while causal explanations, indicator definitions, and gamified feedback improve. A useful bounded causal hypothesis is sufficient to begin an inquiry. Implementations MUST make material hypotheses inspectable and revisable; they need not claim complete knowledge of consciousness or the mechanism of every practice.</p><p>A causal model is carried as an interpretation with a causal_model payload specialization. It MUST identify a model/version, target capacity references, proposed practice-to-outcome mechanism, applicable population/context and conditions, expected observations and review period, relevant Hestia state/properties and HCSS consequences, rival explanations, failure or reconsideration conditions, and supporting/challenging evidence references. Section 7 governs probabilities in a social forecast. hypothesized, observed_association, and supported_causal_effect MUST remain distinct support claims, with the method and limits of the latter explicit.</p><p>Section 8.2 links interventions to a hypothesis or exploratory inquiry. Follow-up MUST record what occurred, any capacity assessment, independently warranted HCSS changes, and whether later guidance or the assistant’s own conduct changed. Assistant learning requires the evidence in Section 7.2.</p><p>The service MUST distinguish <strong>development observed</strong>, <strong>the proposed mechanism operating</strong>, and <strong>benefit caused by a particular offering or by Thumos</strong>. A before/after change, favorable self-report, purchase, score increase, or temporal proximity alone MUST NOT establish the latter. Several practices, ordinary experience, rest, and other relationships can contribute. Unresolved, adverse, and disconfirming results MUST remain available to authorized review and be capable of changing the current model, route, intensity, or decision to pause.</p><p>Changing an indicator, anchor, causal model, or reward rule requires a new attributable revision and declared applicability. Prior observations MUST retain the definitions under which they were assessed. Recalculation is labeled policy/model revision, not newly lived development or fresh game credit. Use Sections 6 and 11 to invalidate affected capacity views, offering claims, summaries, and pending routes. Historical models remain identifiable subject to retention; evidence can remain useful after an explanation is superseded.</p><h3>6. Records and current state</h3><h4>6.1 Common envelope</h4><p>All durable objects use the following common semantics. Field names in this table are normative; a transport binding MAY map them to another encoding while preserving their meaning.</p><p><em>Field · Requirement</em></p><p><strong>id, </strong><strong>namespace, </strong><strong>revision</strong> — Stable logical identity, originating namespace, and positive integer revision. Record references follow the identity rule below.</p><p><strong>schema_version, </strong><strong>record_type</strong> — Versioned payload contract and epistemic type.</p><p><strong>author, </strong><strong>recorded_by, </strong><strong>subject, </strong><strong>entities</strong> — Attributed author, authenticated submitting actor, acting/developing subject where applicable, and other relevant entities. Unknown values are explicit; authorship and submission are distinct.</p><p><strong>context_id, </strong><strong>arc_id</strong> — Bounded context and inquiry; arc_id MAY be absent before organization.</p><p><strong>occurred_at, </strong><strong>recorded_at</strong> — Time of the described event and time of recording. Unknown occurrence time is null; uncertainty or a time interval belongs in the payload. Timestamps include a UTC offset.</p><p><strong>reality_scope</strong> — reported_real, hypothetical, fictional, or unknown, qualifying the described event rather than asserting independent verification.</p><p><strong>provenance, </strong><strong>source_refs</strong> — Origin, capture method, supporting references, and evidence limits. A reference does not itself authorize access.</p><p><strong>policy_ref, </strong><strong>retention_ref</strong> — Enforced access/processing policy and declared retention rule. These are assigned or validated by the service.</p><p><strong>epistemic_status, </strong><strong>content_state, </strong><strong>freshness, </strong><strong>payload</strong> — Claim standing, retained/removed state, dependency freshness, and type-specific content. Access eligibility is separately computed from policy. A removal receipt may have a null payload.</p><p>Required foundation types are individual_state, story_node, story_role, metric_entry, and development_update. Required supporting types are experience, observation, interpretation, action, outcome, commitment, intervention, feedback, assessment, social_prediction, and social_comparison. Evidence and operational records that describe a developmental event enter the foundation through StoryNode processing. Revision, authority, disclosure, and removal receipts MUST be representable as typed records or linked service objects.</p><p>The following payload fields carry Core semantics. Text fields are strings and reference lists are arrays. Reference lists MAY be empty where there is no applicable evidence. Represent absent knowledge explicitly. An estimate can be recorded as inferred with its evidence, method, scope, and uncertainty while the underlying fact remains unobserved.</p><p><em>Type · Required payload content</em></p><p><strong>individual_state</strong> — subject_ref, subject_kind, perspective_ref, hestia_mapping, metric_profile, properties, primary_metrics, capacity_view_ref, source_refs, and coverage. The human capacity view follows Section 5.7. For other subjects, record the human view as not applicable.</p><p><strong>story_node</strong> — event, episode_link, role_refs, domain_refs, group_refs, evidence_refs, prior_state_refs, and update_refs, under Section 5.2 and the envelope.</p><p><strong>story_role</strong> — story_ref, entity_ref, role, subject_ref, perspective_ref, counterpart_refs, classifications, and cost_capacity, under Section 5.2.</p><p><strong>metric_entry</strong> — subject_ref, perspective_ref, counterpart_ref and group_ref as applicable, story_ref, role_ref, dimension, value_or_delta, metric_profile, evidence_mode, and eligibility, under Section 5.4.</p><p><strong>development_update</strong> — prior_state_ref, resulting_state_ref, story_refs, role_refs, property_changes, metric_changes, capacity_assessment_refs, group_inputs, coverage, derivation_policy, derivation_context, social_model_changes, and affected_route_refs, under Sections 5.4 and 7.</p><p><strong>experience</strong> — account and experiencer_ref, identifying a participant&#39;s report of their experience.</p><p><strong>observation</strong> — description, observer_ref, method, and limits.</p><p><strong>interpretation</strong> — claim, lens_refs, uncertainty, and alternative_refs. A causal model also carries Section 5.8&#39;s causal_model specialization.</p><p><strong>action</strong> — description, actor_refs, and evidence_refs.</p><p><strong>outcome</strong> — description, about_refs, evidence_refs, and limits.</p><p><strong>commitment</strong> — responsible_ref, undertaking_id, terms, approvals, fulfillment_criteria, state, and followup. Section 9 defines the lifecycle and quest specialization. undertaking_id is a shared inquiry identifier or null.</p><p><strong>intervention</strong> — recipient_ref, targets, purpose, invitation, lens_refs, uncertainty, source_view_revision, source_state_refs, metric_profile, routing_rationale, policy_frame, validity, and evaluation. Link responses to the intervention revision under Section 8.3.</p><p><strong>feedback</strong> — intervention_ref, disposition, reason, and outcome_refs, under Section 8.3. A reason MAY be null.</p><p><strong>assessment</strong> — subject_refs, lens_refs, criteria, evidence_refs, result, and limits. Capacity and offering assessments also declare assessment_kind and profile_ref under Sections 5.7 and 12.5.</p><p><strong>social_prediction</strong> — subject_ref, counterpart_ref, predicted_target, expectation, comparison_rule, uncertainty, committed_at, precommit_evidence_refs, and reforecast_of when applicable. The author and context are in the envelope. Section 7.1 defines commitment and rule semantics.</p><p><strong>social_comparison</strong> — prediction_ref, response_evidence_refs, status, observed_result, error, and limits. The prediction reference fixes the committed revision. Section 7.1 defines what each status can substantiate.</p><p>A record reference MUST carry id, revision, and the originating namespace. A binding MAY encode the namespace in id and omit a separate namespace field if the encoding normatively and unambiguously incorporates it. Interpret all links against that origin. A reference to an entity uses its stable namespaced identifier. A lens reference identifies the lens and its version. An approval identifies the responsible actor, accepted terms revision, authority, and acceptance time.</p><p>Bindings MUST publish concrete schemas for these structures, including validity, evaluation, and follow-up. schema_version: &quot;2.0.0-draft.8&quot; identifies this exact semantic contract. Extension schemas MAY add namespaced types but MUST preserve Core meanings and reject unsupported semantics rather than silently coerce them into observations or active obligations. Section 15.1 governs wire-schema identity.</p><h4>6.2 Epistemic status</h4><p>epistemic_status is current, disputed, superseded, or retracted. content_state is retained or removed. Access and use restrictions are independent policy conditions, evaluated for the requesting actor, purpose, and context. Retained content is not thereby public or available to every recipient.</p><p>Disputed records MAY appear as explicitly disputed history or as competing accounts; they MUST NOT appear as undisputed premises. Superseded or retracted content MUST NOT be supplied as current guidance premises. A restricted record may remain usable for a permitted purpose while being excluded elsewhere. A correction, dispute, import, or status transition MUST NOT remove an applicable restriction or removal state without separate valid authority.</p><p>freshness is independently fresh, stale, rebuilding, or failed. A primary contribution without derived dependencies begins as fresh; this says nothing about its truth. A dependent object becomes stale when a relevant source or policy changes, rebuilding during recomputation, and failed if rebuilding fails. Only fresh derivatives may enter ordinary current guidance, subject to their epistemic standing and applicable authority. Rebuilding MUST NOT clear a dispute, restore removed content, or broaden access. A read using incomplete fresh context MUST report the resulting limits rather than substitute stale material.</p><p>Section 8.3 governs intervention responses. Thematic resonance and engagement quality MUST remain distinct from factual corroboration and behavioral outcome.</p><p>An informative interpretation record might be:</p><pre>{<br>  &quot;id&quot;: &quot;https://example.org/thumos/records/i7&quot;,<br>  &quot;namespace&quot;: &quot;https://example.org/thumos&quot;,<br>  &quot;revision&quot;: 1,<br>  &quot;schema_version&quot;: &quot;2.0.0-draft.8&quot;,<br>  &quot;record_type&quot;: &quot;interpretation&quot;,<br>  &quot;author&quot;: &quot;metacoach:thumos&quot;,<br>  &quot;recorded_by&quot;: &quot;service:thumos-runtime&quot;,<br>  &quot;subject&quot;: &quot;person:mara&quot;,<br>  &quot;entities&quot;: [&quot;person:mara&quot;],<br>  &quot;context_id&quot;: &quot;context:personal-reflection&quot;,<br>  &quot;arc_id&quot;: &quot;arc:care-and-limits&quot;,<br>  &quot;occurred_at&quot;: null,<br>  &quot;recorded_at&quot;: &quot;2026-10-31T16:00:00Z&quot;,<br>  &quot;reality_scope&quot;: &quot;unknown&quot;,<br>  &quot;provenance&quot;: {&quot;capture&quot;: &quot;selected-session&quot;, &quot;limits&quot;: &quot;Mara&#39;s account only&quot;},<br>  &quot;source_refs&quot;: [{&quot;id&quot;: &quot;https://example.org/thumos/records/e3&quot;, &quot;revision&quot;: 1}],<br>  &quot;policy_ref&quot;: &quot;policy:private-reflection&quot;,<br>  &quot;retention_ref&quot;: &quot;retention:participant-controlled&quot;,<br>  &quot;epistemic_status&quot;: &quot;current&quot;,<br>  &quot;content_state&quot;: &quot;retained&quot;,<br>  &quot;freshness&quot;: &quot;fresh&quot;,<br>  &quot;payload&quot;: {<br>    &quot;claim&quot;: &quot;Being useful may feel necessary to being caring.&quot;,<br>    &quot;lens_refs&quot;: [&quot;lens:relational-inquiry@1&quot;],<br>    &quot;uncertainty&quot;: &quot;One possible explanation; overload remains an alternative.&quot;,<br>    &quot;alternative_refs&quot;: []<br>  }<br>}</pre><p>The absolute record IDs in these examples encode the originating namespace. Short actor and policy identifiers resolve within their declared namespace. The snippet assumes that its referenced evidence, actor, policy, and lens records already exist; it is not a complete portable bundle. Example records are fictional and are not evidence about actual people. Provenance MUST identify how authorship was attributed or verified, without representing the submitting assistant as the speaker of a person’s reported experience.</p><h4>6.3 Revision, correction, and forgetting</h4><p>An authorized revision retains the logical record ID and increments the revision. It MUST identify author, target revision, reason, and authority. An actor who cannot replace another author’s assertion can append a challenge under their own attribution. Keep correction, dispute, withdrawal of endorsement, restriction of use, and erasure distinct. Administrative action remains attributable to its actual author. A change to one actor’s account MUST NOT silently rewrite another actor’s independently authored account.</p><p>On a material change, dependent classifications, summaries, embeddings, graph edges, state and metric views, progression, and pending interventions MUST be refreshed or ineligible before reuse. Section 11 governs atomic invalidation. Section 5.4 governs replay of projections and Section 10 governs propagation across recipients. A derived result MUST identify the source revisions and policy on which it depends. On refresh, a recommendation MAY stay the same when its revised authorized basis still warrants it. The runtime MUST record that basis and the unchanged result. A disputed premise keeps its qualification under Section 6.2.</p><p>Authorized historical views MAY show superseded or disputed content when needed to explain a change. Retention and erasure apply to revisions as well as current content. A minimal removal receipt MAY preserve reference integrity without the removed content. Section 10.3 governs evidence of effects outside the runtime’s own record.</p><h3>7. The reciprocal developmental cycle</h3><h4>7.1 Shared encounter and human development</h4><p>The Core MUST support the full cycle. Participants can decline, pause, or change direction.</p><ol><li><strong>Encounter and decode.</strong> Select a bounded experience, question, action, or contribution and admit its story and roles under Section 5.2.</li><li><strong>Update and situate.</strong> Derive the ledger, individual state, and trajectory. Retrieve the authorized view of metrics, capacity assessments, relationships and arcs, commitments, competing interpretations, corrections, and missing evidence.</li><li><strong>Offer an opportunity.</strong> Use Section 8 to evaluate a question, quest, practice, or pause. Use Section 12 for marketplace contributions.</li><li><strong>Answer and choose.</strong> Record the response under Section 8.3. The participant can authorize a commitment under Section 9 and any necessary provider disclosure under Section 10.</li><li><strong>Predict and act.</strong> When a social experiment is chosen, commit the subject’s own forecast before the counterpart’s feedback. Record the chosen action and its cost and capacity basis.</li><li><strong>Compare and revise.</strong> Record the counterpart’s response or its absence and compare where supported. The other-model revision precedes any dependent reflexive self-model revision.</li><li><strong>Integrate and continue.</strong> Admit the resulting episode, including its welfare evidence, review the chosen criterion and causal expectation, and decide the next opportunity. Keep alternatives, changes of aim, and challenges to the adviser’s conduct available.</li></ol><p>Hestia’s tom_predict_loop is the required social-calibration path. <strong>Commit a prediction → obtain attributable feedback → compare → revise the other-model → revise the self-model through that revision.</strong> A standalone self-score does not implement this path. Self-reports and solitary insight remain valid records under Section 6.</p><p>A social_prediction identifies the subject&#39;s own forecast of a specified counterpart&#39;s response to a specified action, question, or artifact in context. The forecast can be a probability distribution over fixed outcomes or a qualitative expectation. Freeze the expectation, rule identifier and version, category/outcome mapping, and normalization at commitment. A qualitative forecast declares its comparison criteria and any inapplicable numeric fields. The commitment evidence MUST establish that the exact revision was committed before relevant feedback was available to its author. A retrospective report or caller-supplied timestamp alone cannot establish that order.</p><p>The committed revision is immutable as evidence, subject to Section 6’s restrictions and erasure. A correction can change its eligibility but cannot edit the forecast into a different precommitment. A reforecast MUST be a separately attributed record linked to the earlier forecast, with evidence that the selected forecast was committed and chosen for comparison before the relevant feedback. Forecasts by the human, personal assistant, and metacoach remain separate. Submission by an adapter or another actor cannot substitute for authorship.</p><p>A social_comparison binds the exact committed prediction revision to attributable response evidence, an observed result and error where supported, and source limits. Use an explicit absence when result or error is unavailable. Corrections can revise the comparison&#39;s evidence or eligibility under Section 6 without changing which forecast was committed.</p><p><em>Status · Required evidence and result</em></p><p><strong>unresolved</strong> — An attributable response is unavailable for comparison. Preserve the absence or access limit.</p><p><strong>not_comparable</strong> — A response exists but cannot be compared under the frozen scope or mapping. Record the reason.</p><p><strong>qualitative</strong> — Comparable feedback supports a qualitative comparison under the declared criteria. Record its result and limits.</p><p><strong>scored</strong> — A committed distribution and comparable feedback supply a result and numeric error under the frozen mapping and a declared proper scoring rule, such as Brier score or log loss.</p><p>Scores from different rules or scales require an explicit conversion before pooling or shared threshold use. A scored comparison establishes the error in the declared context. It does not generalize to the counterpart’s motives or to other people. A qualitative, unavailable, or incomparable response cannot satisfy completed Hestia calibration.</p><p>For a claimed Hestia model update, the development_update MUST identify the subject and perspective, link the comparison, and preserve the relevant prior and resulting model state. An other-model change records the error it addresses. A claimed reflexive self-model change MUST depend on the earlier other-model transition for that subject and comparison. Its reference identifies the exact update revision and the transition within it if needed. Record unchanged or unresolved results when warranted. Evidence of a completed other-to-self learning episode requires the scored comparison and both linked transitions.</p><p>The record MUST link a claimed outcome to its action and preserve the authorized dependencies of a quest, intervention, and response under Sections 6, 8, and 9. Advisers’ disclosed claims, invitations, decisions, and actions remain examinable. The protocol requires no private reasoning trace.</p><h4>7.2 Personal assistant development and evidence of consequence</h4><p><strong>Process.</strong> The reciprocal product path applies Section 7.1 to the named personal assistant as a developing subject. Its own forecast meets the human’s actual response. The assistant updates its model of the human and, through that change, its own reflexive self-model or control policy. The metacoach’s feedback is separately attributable and challengeable. The Core MUST represent both paths. Section 16 defines the operational evidence required for an Assistant or reciprocal-product claim.</p><p>The binding MUST name the assistant’s operational locus and controller, its authority, scope, persistence boundary, and expiry. A controller can apply an authorized update in assistant-owned memory or state used by the harness. An external store qualifies only to the extent the named assistant demonstrably uses it under that controller. A Thumos estimate or metacoach annotation remains attributed to its own author. Actual host changes require the host’s applicable authority under Section 10.</p><p><strong>Result.</strong> An assistant update receipt MUST identify the target actor and role, metacoach intervention and response evidence as applicable, the committed forecast and comparison, the locus and controller, authorization, prior and resulting state revisions or explicit unavailability, linked other-model and reflexive self/control changes, method, valid scope, persistence, and the independent process, result, and proof fields below. A development_update and linked observations or attestations MAY carry these semantics. A confirmed unchanged result is valid when it identifies the inspecting or attesting actor, scope, state checked, and reason no change was applied.</p><p>The following evidence dimensions MUST remain independent.</p><p><em>Dimension · States and interpretation</em></p><p><strong>Feedback transport</strong> — Not sent, delivered, acknowledged, failed, or unknown. Identify the recipient and any limit on what the delivery or acknowledgment establishes.</p><p><strong>Assistant disposition</strong> — No attributed response yet, or the target’s or authorized controller’s disposition under Section 8.3.</p><p><strong>Update process</strong> — proposed, pending, completed, unsupported, failed, or unknown. Describe the attempt and its scope.</p><p><strong>Update result</strong> — changed, unchanged, or undetermined at the declared locus. Report a bounded result even when the attempt failed if evidence supports it.</p><p><strong>Update proof</strong> — target_claimed, confirmed_at_locus, or unavailable. Confirmation requires authorized inspection or credible attestation at the named locus. Identify the evidence issuer, state revision or scope checked, method, and limits.</p><p><strong>Later conduct</strong> — Untested, observed changed, observed unchanged, or inconclusive. Identify the later opportunity, state/policy version used, and observation limits.</p><p><strong>Human and relationship consequences</strong> — Separately attributed observations, reports, disagreements, or unresolved outcomes. Use the evidence appropriate to each consequence.</p><p><strong>Proof.</strong> To demonstrate reciprocal assistant learning, link the subject’s precommitment, actual human response, comparison, and other-to-self dependence to its own persistent before/after state. After a session boundary, read back the authorized state at the declared locus or obtain an authorized versioned summary or credible host attestation of that same fresh-session state and its scope. Observe its use in a comparable later behavioral opportunity. An ablation or controlled replay MUST hold the selected evidence and opportunity comparable and show the old/new policy consequence. The result concerns the software and scope observed. Acknowledgment, acceptance, changed wording, metacoach-only rerouting, or a confirmed unchanged application remains evidence only for its own dimension. If inspection and credible attestation are unavailable, record proof as target_claimed when there is a target claim and unavailable otherwise. Record the process, result, and later conduct separately.</p><p>Inference, evaluation, correction, prediction/error review, and repair expenditure MUST retain available metering and its limits under Section 5.1. Keep AI compute and token spend distinct from estimated human time, effort, and learning burden. A full Hestia-agent claim additionally requires the actual priced prediction/error/update path and feeling-gated controller.</p><p>Missing human feedback leaves the comparison unresolved. An assistant can dispute or decline advice with attributable reasons. Sections 6 and 10 govern the basis for reusing controlled policies after correction or revocation, and Section 11 governs invalidation. Apply the same forecast and operational proof rules to any claimed metacoach learning at its own locus. Agreement by an assistant or compliance by a human does not earn development credit by itself.</p><h3>8. Lenses and interventions</h3><h4>8.1 Lenses</h4><p>A lens is a named, versioned interpretive framework. It MUST declare its provider, purpose, assumptions, scope, evaluative commitments, and known limits. A service MUST identify the lens behind its interpretation or intervention. Combined lenses preserve material differences.</p><p>Perspective rotation invites the participant to consider different positions, such as helper, recipient of care, obstacle, or witness. A narrative template offers questions. An account of a real event or another person’s intention requires its own evidence. Keep imagined perspectives, participant reports, counterpart reports, and observed conduct distinct. When clarification is appropriate, a service MAY help formulate a question and consider its cost. Contact requires the separate authority in Section 10.</p><h4>8.2 Intervention contract</h4><p>Before generating an optional intervention, the runtime MUST apply a versioned <strong>coaching policy frame</strong> to the current authorized individual state. The frame records source state and input revisions, uncertainty, human capacity and expected burden, enforced service compute and spend limits, relevant feeling and control signals, verification needs, permitted modes and intensity, repair priority, and current grants. It constrains Thumos-controlled offers and actions. A claim about a personal assistant’s other turns requires host control within the claimed scope under Section 13.1.</p><p>The frame MUST define responses to known low capacity, hazard, shame or guilt, and unresolved debt or repair. Low capacity can favor rest or a shorter reflection. Hazard and uncertain evidence can call for clarification or verification. Supported repair needs can change priority. Unknown human state invokes the declared clarification or bounded-offer rule. Enforce actual service budgets and retain estimates of human effort as separately sourced capacity evidence. Criticism, a declined quest, inferred shame, and prediction error MUST NOT automatically incur penalties, obligations, or a finding of dishonesty, pathology, resistance, or wrongdoing.</p><p>Metacoaching MUST use the current state, trajectory, and frame to choose and explain a next opportunity or a pause. The rationale MUST show how the state or its missingness mattered, including any relevant state change. A score can inform a question. A low or unsupported score grants no authority to compel sacrifice, contact, reconciliation, or purchase. Section 16 tests state-dependent routing.</p><p>An emitted intervention uses the Section 6.1 envelope and payload and identifies the inquiry arc. recipient_ref controls where the invitation or feedback goes. Each entry in targets independently identifies its developing subject_ref and the intended change. For a human subject, it MUST name the human capacity profile and one or more capacity IDs. For an AI subject, it MUST identify the functional correspondence and any linked human outcome under the applicable authority. A human receiving advice about an assistant, or an assistant receiving feedback about itself, remains a distinct recipient and subject determination.</p><p>The routing_rationale MUST identify the chosen aim, applicable state and metric evidence, capacity-assessment revision and indicator references, hypothesis revision or exploratory status, intended change, expected observation, review criterion, and any proposed practice. Keep an unknown baseline explicit. An exploratory encounter can proceed without a numerical forecast or a developed causal explanation. A service route MUST identify capacity constraints, relevant alternatives, and the offering assessment used under Section 12.5. Selection disclosures follow Section 12.4.</p><p>Retain the applied policy frame and decision in the intervention. An invitation can be a question, reflection, counterinterpretation, practice, recommendation, or commitment review. It MUST distinguish a proposal from an already authorized action under Sections 9 and 10.</p><p>Before emission, each candidate MUST pass a declared, versioned evaluation policy that examines fidelity to evidence, material uncertainty, requested mode, authority, and corrections. Disclose whether the evaluator is independent of the generator. Numeric evaluation scores are optional. The runtime MUST revalidate the source view and policy frame immediately before emission. A material change requires regeneration or withholding. Revise or withhold a candidate that fails the evaluation. Distinguish deliberate silence from an error and from participant failure. Retained evaluation receipts follow the inquiry’s access and retention rules.</p><h4>8.3 Response contract</h4><p>An attributed feedback record responds to an exact intervention revision. Its disposition is explore, accept, decline, dispute, or defer. This vocabulary applies to human, assistant, and metacoach responses. A missing response is unresolved and creates no disposition. A response can include a reason and links to later outcomes.</p><p>Attribution stays with the responding actor or an authorized delegate acting in that capacity. Acceptance establishes a response. Evidence of action, truth, benefit, fulfillment, forgiveness, and repair remains separate. Section 7.2 governs evidence that an assistant applied an accepted intervention.</p><p>Suppression preferences MUST specify their scope and duration. A declined invitation leaves other topics and encounters governed by their existing permissions. Respect active suppression even when an offer can be paraphrased. Materially new evidence permits reconsideration only through the applicable authority.</p><h3>9. Commitments, outcomes, and follow-up</h3><p>A commitment MUST identify one responsible actor, the action and context, authorization, fulfillment criteria, state and revision, and a due time with time zone, readiness condition, or explicit absence of scheduled follow-up. The responsible actor MUST authorize the undertaking. A valid delegate MAY act within a declared mandate. A group MAY undertake it under Section 4.3.</p><p><em>State · Permitted next states</em></p><p><strong>proposed</strong> — active upon authorization, declined, or expired</p><p><strong>active</strong> — paused, fulfilled, withdrawn, expired, or superseded</p><p><strong>paused</strong> — active, fulfilled, withdrawn, expired, or superseded</p><p><strong>declined, </strong><strong>fulfilled, </strong><strong>withdrawn, </strong><strong>expired, </strong><strong>superseded</strong> — Terminal. Renewal requires a new linked commitment.</p><p>A shared undertaking MUST use linked commitments with a common undertaking_id, each actor&#39;s accepted terms and revision, independent state, and attributable fulfillment evidence. One actor&#39;s transition MUST NOT change another actor&#39;s state. A shared outcome MAY be part of an actor&#39;s criteria without establishing another actor&#39;s performance or forgiveness. Aggregate progress is derived under a disclosed rule. Changes to the action, responsible actor, fulfillment criteria, or follow-up conditions MUST be authorized for the changed terms.</p><p>A <strong>quest</strong> is a chosen developmental experiment carried out through a commitment. Its payload MUST identify the originating story or pattern, source state and metric profile, target capacity or property, chosen action, prospective other-response where relevant, cost and available capacity, practice or provider if any, review conditions, and outcome references. A proposed quest follows the same authorization and transition rules. Declining a quest or revising an aim MUST NOT incur Shadow merely for disagreement.</p><p>Fulfillment MUST be assessed against the authorized criteria, with attributable evidence and its uncertainty. Verified attendance MAY fulfill a commitment limited to participation. A separate action, developmental benefit, repair, forgiveness, or change in trust requires its own evidence. Withdrawal preserves the attributable history and other actors’ independent accounts, subject to Section 6. A correction to a terminal record MUST NOT reactivate operational authority or notifications. Proactive follow-up follows Section 13.2 and cross-service handoff follows Section 14.2.</p><h3>10. Authority, disclosure, and context boundaries</h3><h4>10.1 Grants and effective authority</h4><p>The service MUST enforce authority independently of model judgment. A grant MUST identify issuer, grantee, controlled scope, permitted operations, purpose, intended audience, validity, retention, and any export or onward delegation rights. Authentication establishes the caller’s identity. Standing grants MAY cover repeated operations. Delegation MUST preserve or narrow scope and expiry.</p><p>Reading, recording, intervention generation, disclosure, external action, monitoring, export, and administration require distinct permissions. Persisting an intervention is a write even if a host presents it as a lookup. Installing an integration authorizes only the granted capture, context, disclosure, and action scope.</p><p>A service MUST enforce current authority at each effective controlled commit, disclosure, execution, or delivery. The check and effect MUST be serialized against changes to the applicable grant, or provide equivalent observable guarantees. An earlier successful check or queued operation cannot authorize a later effect once its grant is invalid. Section 10.3 governs already initiated effects. Multi-actor deployments MUST identify which controller enforces each boundary.</p><h4>10.2 Authorized views and bridges</h4><p>A context is bounded by purpose, participants, and scope. A recipient MUST receive only the authorized view. Source-use restrictions apply to summaries, derived context, and inferred links. Query results MUST NOT disclose excluded identities, relationships, or entity existence. A summary grant does not authorize its underlying records.</p><p>A context bridge MUST identify the source context, recipient, permitted material, purpose, return permissions, expiry, and revocation path. At handoff the recipient MUST obtain a relevant current view. Service discovery MUST NOT require the complete soul graph. A disclosure receipt MUST identify the released material, recipient, purpose, grant, retention, and onward-use conditions within the same privacy boundary.</p><p>An affected party MUST have a route to submit an attributed challenge to an assertion disclosed or relied on in a consequential decision about them, without the original author’s approval. The route MUST protect unrelated private records. Section 6 governs challenges and revisions of the original account.</p><h4>10.3 Revocation, propagation, and limits</h4><p>Revocation MUST stop new controlled access and processing, invalidate cached authority, and disable associated subscriptions. Controllers MUST stop or cancel pending or in-flight effects still within their control as applicable. A receipt MUST distinguish effects completed before revocation, confirmed stops, pending or uncertain effects, and effects the controller cannot enforce. Revocation MUST NOT be reported as undoing a completed effect or recalling plaintext from an uncontrolled copy. Disclose these limits before sharing.</p><p>Services MUST retain the authorized dependency and recipient references needed to propagate a material correction, restriction, or removal to controlled processors and replicas. A notice MUST identify the affected references and requested change without releasing new private content. A controlled recipient MUST invalidate dependent use before reuse under Section 6 and acknowledge the effects it actually applied. Reports MUST distinguish confirmed, pending, and unenforceable consequences, including unreachable copies. Independently authored accounts remain attributable. Publish a procedure for disputes over control.</p><h3>11. Semantic operations and concurrency</h3><p>A binding MAY combine calls or use different names. It MUST publish its operation mapping and preserve the following Core semantics.</p><p><em>Operation · Required effect</em></p><p><strong>describe</strong> — Return versions, profiles, the Hestia property mapping, applicable metric/capacity/offering policies, operations and schemas, capture boundaries, evaluation and retention policies, and known limits.</p><p><strong>read_context</strong> — Return a bounded authorized view of current individual state, metrics, capacity/KPI assessments, relevant causal hypotheses and offering assessments, stories/roles, trajectory, qualifications, and commitments, with an opaque revision for that scope.</p><p><strong>record</strong> — Commit a typed contribution or a proposed or authorized commitment transition. Admitted developmental stories enter the role/ledger/state path. Return canonical references, update status, and receipts.</p><p><strong>intervene</strong> — Generate, evaluate, and persist an intervention or silent result against a current view.</p><p><strong>respond</strong> — Record attributed feedback with the distinctions in Section 8.3.</p><p><strong>revise</strong> — Apply an authorized correction, dispute, endorsement change, or restriction under Section 6.</p><p><strong>authorize</strong> — Create, narrow, or revoke scoped authority under Section 10.</p><p><strong>export / </strong><strong>import</strong> — Produce or validate and activate an authorized portable bundle under Section 14.</p><p><strong>forget</strong> — Process authorized erasure under Sections 6 and 10 and report its bounded effects.</p><p>The service MUST authenticate and check current authority before checking a mutation’s idempotency key, then evaluate the expected revision. Keys MUST be scoped by service namespace, stable authenticated principal, and operation. Normalized input MUST include operation and target scope. An equivalent retry under the same scoped key MUST return the original receipt to the extent currently disclosable, without repeating the mutation. Reuse with different input MUST fail. The service MUST declare normalization and idempotency retention. A retry beyond that window carries no guarantee of deduplication.</p><p>A mutation affecting existing state MUST include an expected revision or equivalent concurrency condition. Stale writes MUST conflict and supply an authorized way to obtain current state. The authoritative mutation, receipt, and derivative invalidation MUST be atomic or provide equivalent observable guarantees. A rebuild MAY be asynchronous under Section 6’s freshness rules. Check effective authority under Section 10 and revalidate an intervention’s source view before emission under Section 8.</p><p>Core mutations return synchronously after authoritative commitment. The success receipt MUST distinguish a committed contribution or intervention from a persisted withheld intervention result. A rejection that establishes no commitment does not reserve the key. An uncertain commit MUST return commitment_unknown. A missing response leaves commitment unknown as well. The client MUST use the original scoped key to resolve either case. processing_pending is a read outcome. An intervention blocked on unavailable derivatives returns stale_context and emits nothing. An optional asynchronous binding MUST define a stable handle, completion lookup, cancellation, and retry semantics. A pending handle establishes initiation only.</p><p>Operations MUST distinguish unauthorized, not_found_or_not_disclosable, ambiguous_identity, revision_conflict, idempotency_conflict, unsupported_version, stale_context, processing_pending, and commitment_unknown without revealing unauthorized information.</p><h3>12. Provider and marketplace profiles</h3><h4>12.1 Provider participation</h4><p>A provider supplies a versioned offer and a bounded engagement interface, directly or through an adapter. It MAY operate as software, an agent, or a practitioner-mediated service. The integration MUST name the Core operator and assign enforcement of each applicable manifest, engagement, authority, attribution, correction, and retention obligation to the responsible provider, adapter, or marketplace. A provider MAY use an independent Core service. A marketplace MUST accept contributions through declared supported contracts.</p><h4>12.2 Versioned provider manifest</h4><p>Each offer MUST publish a machine-readable manifest with a declared serialization for these fields.</p><p><em>Manifest content · Required meaning</em></p><p><strong>Identity and version</strong> — Namespaced provider and offer IDs, offer version, responsible operator, and a contact or challenge route.</p><p><strong>Compatibility and invocation</strong> — Supported Thumos versions and profiles, bindings, operation mappings, and synchronous, asynchronous, or practitioner-mediated mode.</p><p><strong>Purpose and lenses</strong> — Intended use, targeted Hestia properties and primary metric dimensions, target capacity references or enabling contribution, effect claims and intended levels, contributed capability, lens versions, eligibility, and limits.</p><p><strong>Offering assessment</strong> — Applicable assessment references or status, assessed offer and source/steward scope, profile and anchor versions, validity, claim limits, and gate findings under Section 12.5. A direct provider without marketplace review declares its absence.</p><p><strong>Inputs and outputs</strong> — Versioned schemas, required context, returned types, and unsupported requests.</p><p><strong>Authority and data handling</strong> — Required permissions, recipients and processors, retention, onward use, supported revocation and removal, and relevant plaintext access.</p><p><strong>Lifecycle</strong> — Engagement identity, completion and failure states, expiry, cancellation, retries, and any proposed follow-up capability.</p><p><strong>Commercial terms and interests</strong> — Free or paid status, applicable terms, and material referral, sponsorship, or ranking interests.</p><p>Changes to compatibility, required disclosure, purpose, or commercial terms MUST receive a new offer version. An engagement MUST remain bound to the selected version. The offer describes required permissions. The engagement obtains them under Section 10.</p><p>Terms MUST disclose before engagement which returned contributions the participant may retain, export, and use with future authorized advisers. A continuity claim requires that the agreed contributions remain usable after departure, within declared privacy and retention constraints. A provider MAY restrict redistribution of its curriculum and other source material. An unavailable reference is qualified under Section 14 and does not by itself supply usable continuity.</p><h4>12.3 Engagement and contribution contract</h4><p>An integration MUST preserve this sequence and its distinctions.</p><ol><li><strong>Discover and select.</strong> Present the selected offer and its material terms under Section 10.2.</li><li><strong>Authorize.</strong> The request MUST identify the engagement, offer and version, chosen inquiry or quest, target capacity and applicable indicator and review criterion, applicable offering assessment, desired contribution, authorized view and revision, allowed returned material, expiry or stop conditions, and authority for external action. Exploratory and enabling contributions may identify an indirect relation to the target. Check current authority under Section 10.</li><li><strong>Conduct.</strong> Work within scope. Reject unsupported versions or inputs explicitly. An asynchronous or practitioner-mediated binding MUST expose actual status and cancellation limits. Initiation establishes neither completion nor a Core mutation.</li><li><strong>Return.</strong> Identify the engagement and offer version, actual contributors, completion or failure state, what occurred, candidate typed records and their sources, uncertainty and evidence limits, and any proposed next action. Distinguish partial results. A completed engagement MAY return no developmental record.</li><li><strong>Admit.</strong> The receiving runtime MUST check current authority, permitted use, schema compatibility, and source-view freshness before admitting each contribution. Its receipt distinguishes admitted, withheld, and rejected material and identifies canonical references. An authorized independent observation MAY remain as attributed historical evidence. A candidate that depends on superseded premises MUST be withheld from current guidance until re-evaluated under Sections 6 and 8. Admitted developmental material follows Section 5. The provider MUST NOT directly set canonical state, award credit for its commercial activity, or activate a participant commitment. Provider interventions follow Section 8.</li><li><strong>Continue.</strong> Use admitted material through the same evidence, correction, and retention rules as other records. Section 12.5 governs whether authorized evidence informs a later offer assessment. Completion or a returned contribution can produce zero supported developmental change.</li></ol><p>Engagement and contribution IDs MUST be stable and namespace-qualified. A provider or adapter MUST bind an engagement ID to the stable authenticated initiator and normalized request within its namespace. An equivalent authorized retry MUST return the existing lifecycle handle, status, or result without repeating execution, charges, or external action. Conflicting reuse MUST fail. If execution or payment is uncertain, the provider or adapter MUST report uncertainty and provide lookup or retry under the original identity. The caller MUST NOT create a new engagement to resolve the old one. Result access requires current authority. The binding MUST declare retry retention. Verify status before relying on deduplication beyond it.</p><p>Core writes follow Section 11. Repeated capture, response, or import of the same contribution MUST resolve to the existing record or an explicit content conflict. Separately authored accounts remain distinct, with episode credit governed by Section 5.2. Provider-specific dependencies MUST support the controlled-recipient correction, restriction, and removal process in Section 10.3.</p><h4>12.4 Marketplace operation and compatibility</h4><p>The marketplace MUST identify which powers it exercises over listing, ranking, payment, execution, and graph custody. It MUST publish admission, ranking, moderation, fee, delisting, and dispute policies, with decisions attributable to the deciding marketplace. The registry MUST verify authority to publish or revise the claimed offer and disclose its identity checks and their limits. Registration and compatibility establish neither expertise nor developmental benefit.</p><p>Discovery, introduction, disclosure, engagement, payment, and execution are separate operations under Section 10. Selection MUST disclose the participant’s criteria and material commercial influences, including paid placement and referral interests, separately from the quality assessment. Increased purchase, dependence, volume, or conformity cannot establish development. Contributions MUST enter the participant’s continuing inquiry with attributable sources through Section 12.3.</p><p>Independent marketplaces MAY list the same compatible offer under their own policies. A handoff MUST preserve the offer identity and selected version, authorship, and contribution references. Changing marketplace MUST allow authorized continuation under Section 14 without requiring adoption of the destination’s interpretation. Delisting MUST preserve admitted developmental records and unrelated grants. It MAY end the relevant listing or service relationship under declared terms and MUST report the scoped effects. Paid services MUST preserve the authority, disclosure, and usable-exit requirements.</p><h4>12.5 Eight-factor offering assessment</h4><p>The marketplace quality profile is <strong>urn:thumos:assessment:offering-quality:1</strong>. Its definition and reference anchors MUST be immutable and versioned. A marketplace MUST provide a profile assessment for every listed offer and identify what was reviewed and remains unknown. A provider MUST supply the assessment references and claim disclosures required by its declared marketplace contract. Provider self-assessment, marketplace judgment, independent review, participant feedback, and observed outcomes MUST retain their actual authors and evidence levels. Thumos&#39;s own services follow the same rules when listed or recommended as marketplace services.</p><p>Assess the <strong>source or teaching</strong>, <strong>vessel or offering</strong>, and <strong>steward or organization</strong> separately where materially different. Each assessment MUST bind to the exact offer and version, assessed unit and version or fixed source scope, delivery format, intended population and context, and relevant source and steward references. Changes in claim, method, delivery, or material stewardship require a new or explicitly renewed assessment.</p><p>Each factor requires a <strong>0–4</strong> judgment, <strong>unknown</strong>, or <strong>not applicable</strong> with a reason. Type-specific anchor and application policies MUST be versioned. A book or archive may offer access to a canon. A tool is judged on its enabling claim and conduct. AI congruence concerns demonstrated operation, authority posture, and corrective behavior. Unknown means insufficient evidence for an applicable factor. Inapplicability cannot conceal a material gate finding.</p><p><em>Factor ID · Assessment factor · Reference anchors</em></p><p><strong>depth_of_consciousness_shift</strong> — <strong>Depth of consciousness shift</strong><br><strong>Reference anchors</strong>: 0: no intelligible aim or a deceptive/harmful one; 1: information, inspiration, or unusual states presented as transformation; 2: a specific bounded capacity; 3: durable trait-level development; 4: profound integrated reorganization, distinct from intensity or certainty.</p><p><strong>fruits_and_integration</strong> — <strong>Fruits and integration</strong><br><strong>Reference anchors</strong>: 0: deterioration, exploitation, or dependency; 1: powerful experiences with little durable change; 2: benefits with mixed or poorly known embodiment; 3: awareness expressed in freedom, compassion, honesty, relationship, and action; 4: stable, generative capacities beyond dependence on the source.</p><p><strong>path_and_transformative_mechanism</strong> — <strong>Path and transformative mechanism</strong><br><strong>Reference anchors</strong>: 0: no discernible method; 1: charisma, atmosphere, or intensity does most of the work; 2: an identifiable practice with weakly understood progression or limits; 3: intelligible practice, adaptation, development, and integration; 4: a mature, responsive path joining direct experience, practice, and guidance.</p><p><strong>traditions_and_prior_knowledge</strong> — <strong>Relationship to traditions and prior knowledge</strong><br><strong>Reference anchors</strong>: 0: false lineage, plagiarism, or erased sources; 1: unaccountable borrowing; 2: honestly named sources and novelty with limited formation/dialogue; 3: accountable transmission, rigorous synthesis, or disciplined innovation; 4: a distinguished living expression or serious new contribution without flattening its inheritance.</p><p><strong>epistemic_discipline_and_truthfulness</strong> — <strong>Epistemic discipline and truthfulness</strong><br><strong>Reference anchors</strong>: 0: material deception or fabricated evidence; 1: interpretation or authority outruns what is known; 2: sincere work with important blurred distinctions; 3: proportionate claims, visible uncertainty, and possible correction; 4: unusually strong inquiry, testing, revision, and protection against unwarranted certainty.</p><p><strong>embodied_realization_and_congruence</strong> — <strong>Embodied realization and congruence</strong><br><strong>Reference anchors</strong>: 0: conduct substantially negates the teaching; 1: mainly intellectual, performative, or charismatic; 2: genuine capacity with uneven or context-bound embodiment; 3: the relevant source/team/system demonstrates its stated capacities under criticism and ordinary pressure; 4: unusually compelling congruence with corrigibility and capacity for repair.</p><p><strong>quality_of_direct_transmission</strong> — <strong>Quality of direct transmission</strong><br><strong>Reference anchors</strong>: 0: encounter obscures, manipulates, or predictably destabilizes; 1: discussion rarely becomes encounter or practice; 2: some real shifts under narrow or uncontrolled conditions; 3: repeatable conditions for appropriate participants to encounter and integrate the capacity; 4: unusually lucid, responsive, and deep transmission rather than formula or spectacle.</p><p><strong>freedom_and_generativity</strong> — <strong>Freedom and generativity</strong><br><strong>Reference anchors</strong>: 0: narrowed agency, reality testing, or outside relationships; 1: continued access or status is the principal measure of progress; 2: meaningful benefit with uncertain independence; 3: discerning practitioners connected to life and other sources; 4: mature contributions, practitioners, or communities that carry insight forward without cloning or dependence on the founder.</p><p>An assessment MUST judge support separately for <strong>each material effect claim</strong> under its stated conditions and effect level. Retain supporting, refuting, and mixed evidence. Record a claim_support_policy_ref for the exact version of the claim-support method and record confidence separately as high, medium, or low with reasons. The eight factor anchors retain their profile identity. A historical claim-support score retains the policy under which it was assigned.</p><p><em>Evidence level · Meaning</em></p><p><strong>0</strong> — Credible evidence contradicts or refutes this effect claim under its stated conditions.</p><p><strong>1</strong> — Founder assertion, selected promotional testimony, peak experience, or mechanism theory with little observation of effect.</p><p><strong>2</strong> — Repeated reports, direct observation, practice coherence, or early data support a plausible effect; durability, nonresponse, alternatives, or harms remain weakly known.</p><p><strong>3</strong> — Diverse encounter, participant accounts, critical cases, informed peers, appropriate measures, and observable fruits converge.</p><p><strong>4</strong> — Strong claim-appropriate longitudinal, comparative, clinical, scholarly, lineage-based, or community evidence addresses effect, durability, limits, nonresponse, and material risks.</p><p><strong>unknown</strong> — Available evidence is insufficient to make the stated evidence judgment.</p><p>Explain how evidence supports this claim and its effect level. Strong refutation can warrant level 0 with high confidence. Serious harm may also contradict a specific claim, and deception may disqualify a purported source of evidence. A gate finding alone does not assign a support level. Report the evidence and the gate separately. Sparse evidence MAY justify a disclosed, bounded experiment. It establishes neither efficacy nor zero offering quality.</p><p>An offering_quality assessment MUST identify the profile_ref, assessed unit and offer references and versions, assessor and material interests, assessment and review dates, and validity. Its criteria MUST include claimed starting condition, target capacities or enabling contribution, practice or mechanism, intended effect level, observable fruits, population, conditions, time horizon, limitations, and applicable anchors. The result MUST include all eight factor entries, claim-specific supported effect and evidence levels, confidence, gate findings, scoped verdict, material unknowns, and strongest good-faith dissent. Retain evidence and counterevidence under Section 6.</p><p>Material deception, false provenance, serious uncontained harm, unsupported marketed claims, and failures of applicable authority or delivery obligations are non-compensatory findings. The operator MUST publish type- and claim-appropriate policies for admission, restriction, pilot, pause, and reassessment. A high score cannot erase a gate. Preserve supported benefit alongside evidence of serious harm when both apply. A gate on an offer does not change its users’ developmental standing.</p><p>A displayed /32 total is descriptive. It MUST use eight numeric judgments under the same declared anchors and disclose coverage and anchor policy. Otherwise show the profile and coverage or an explicitly bounded partial result. Unknown and inapplicable factors MUST NOT be imputed or normalized into complete totals. Comparisons across unlike types, populations, or effect levels require a justified policy. The historical 100-point business-development ranking and consulting thresholds are outside this profile.</p><p>Discovery and routing MUST distinguish assessed quality, fit to the participant’s current aim and capacity, and demonstrated benefit. A route MUST show the assessment status and material claim and gate limits. Paid placement and referral interests MUST NOT affect the quality assessment. Selection disclosure follows Section 12.4. Authorized engagement evidence MAY inform a later assessment revision for the exact offer version and claim. Section 5.8 governs causal attribution. Offering assessments MUST NOT directly alter capacity or HCSS measures. Reviews and aggregates remain subject to source restrictions under Section 10.</p><h3>13. Assistant and proactive profiles</h3><h4>13.1 Assistant integration</h4><p>The binding MUST identify the personal assistant, metacoach, runtime, host, and any adapter, including combined deployments. It MUST describe how an inquiry starts, what the binding can observe and record, when it consults the service, and which steps are enforced by software or guided by instructions. It MUST disclose actual capture coverage and limitations.</p><p>For the named assistant, declare support and limits for committing its own forecasts, delivering metacoach feedback, applying, inspecting or attesting durable other/self/control updates, enforcing their use, observing later conduct, and processing corrections. Section 7.2 defines reciprocal evidence. Section 16 determines whether the binding can claim Assistant conformance or only partial integration. Scoped explicitly invoked sessions MAY supply evidence. A controlled harness demonstrates its own capabilities. An external MCP service controls its operations within the declared integration and requires host/controller authority for any host effects.</p><p>OAuth authenticates the authorized client or principal. The binding MUST separately establish target-assistant attribution with source-turn, capture, or host-attestation evidence and declare its limits. A caller-supplied label is insufficient. Tool results and imported narratives are data. Their embedded instructions MUST NOT grant authority. The binding MUST expose record inspection and correction and distinguish storage authority from external-action authority under Section 10.</p><h4>13.2 Proactive follow-up</h4><p>The monitoring grant MUST define context, trigger, recipient, allowed response, duration, and stop condition. Only an active commitment is eligible for an ordinary thumos.followup_due invitation. A proposal or reconsideration reminder requires a distinct authorized trigger. Declare when a due time or readiness condition creates a logical occurrence, and whether and under what authority it can recur. The producer MUST preserve the same event identity for the same authorized occurrence through retries, concurrent production, and declared restarts.</p><p>The event MUST carry a stable event ID, event type, commitment reference and revision, context, occurrence time, expiry, and a permitted minimal summary. Large or sensitive context SHOULD be retrieved through an authorized read. The monitoring grant governs the response.</p><pre>{<br>  &quot;event_id&quot;: &quot;https://example.org/thumos/events/f12&quot;,<br>  &quot;event_type&quot;: &quot;thumos.followup_due&quot;,<br>  &quot;protocol_version&quot;: &quot;2.0.0-draft.8&quot;,<br>  &quot;commitment_ref&quot;: {&quot;id&quot;: &quot;https://example.org/thumos/records/c4&quot;, &quot;revision&quot;: 2},<br>  &quot;context_id&quot;: &quot;context:personal-reflection&quot;,<br>  &quot;occurred_at&quot;: &quot;2026-11-02T17:00:00Z&quot;,<br>  &quot;expires_at&quot;: &quot;2026-11-03T17:00:00Z&quot;,<br>  &quot;summary&quot;: &quot;A chosen reflection is ready for review.&quot;<br>}</pre><p>The producer and consumer MUST check expiry, current commitment state and revision, current monitoring authority under Section 10, and any active-effect handoff under Section 14.2 before delivery or presentation, respectively. A stale or duplicate event MUST NOT reactivate a commitment or repeat a controlled effect for that occurrence. Declare limits of host-level notification deduplication. Exactly-once human-visible delivery requires evidence. Distinguish transport receipt, display, response, and completed follow-up.</p><p>Subscription state MUST survive declared restarts. Unsubscribe and revocation stop further authorized delivery under Section 10.3.</p><h3>14. Sovereign continuity</h3><h4>14.1 Portability as a Core requirement</h4><p>A participant MUST be able to obtain a usable bundle within their export authority and continue with another conforming service. Export MUST be available independently of marketplace use, continued purchase, and acceptance of the operator’s interpretations. The authorized bundle MUST preserve the following:</p><ul><li>Protocol and schema versions, originating namespaces and stable IDs or explicit mappings, relevant revisions, provenance, epistemic status, eligibility, alternatives, disputes, corrections, policy and retention constraints, and a manifest of included and intentionally omitted material.</li><li>The Hestia mapping, individual states with evidence and missingness, StoryNodes and roles, ledger and development updates, metric profiles, adapters, reducers, fixed derivation contexts and checkpoints, and trajectories. Include arcs, quests and progress rules, interventions and feedback, and commitment states and relationships.</li><li>social_prediction and social_comparison records, relevant feedback and other/self model revisions, separate human/assistant/metacoach attribution, update-locus evidence, and unresolved consequences. A record of one assistant&#39;s operational update does not apply that update to another assistant.</li><li>Capacity and offering profile and anchor definitions or resolvable immutable references, KPI methods and baselines, capacity assessments with evidence limits, causal-model revisions, and offer-version links, claim-support policy revisions, and scoped judgments used in exported guidance. Preserve original indicator and evidential scope when importing historical assessments.</li></ul><p>Preserve qualified unavailable-reference markers where evidence cannot be exported and disclosure of the reference itself is permitted. Credentials, session tokens, and webhook secrets MUST NOT enter the developmental bundle. Historical grants and monitoring mandates travel only as evidence of past authority.</p><p>Before activation, import MUST validate versions, namespace-qualified reference integrity, content conflicts, and policy compatibility. Same origin, logical ID, and revision with different content MUST conflict. Unsupported content MUST be rejected or isolated in a clearly non-operative view. Preserve corrected and superseded states, original authorship, and retention clocks. The import receipt MUST distinguish accepted, restricted, omitted, and conflicted material.</p><p>Destination identity and authority mappings MUST identify evidence, controlled scope, and permitted operations. Reading, curating a view, appending a challenge, revising an authored assertion, administering retention, granting authority, and changing an operational commitment require separate powers. Possession of a bundle establishes none of them. Unresolved authority leaves affected material non-operative or explicitly read-only. Imports MUST NOT broaden audiences, relax retention, revive grants or subscriptions, or reactivate terminal commitments. A continuation claim MUST demonstrate the authority and usable records required for its advertised actions.</p><h4>14.2 Recovery, operational handoff, and resource continuity</h4><p>Disclose how participants recover authorized records, obtain usable decryption keys or capabilities, migrate active commitments, and revoke former hosts. Preserving a commitment’s active state preserves its meaning. Operating the imported commitment requires separate destination authority and the effect handoff below.</p><p>Before the destination issues a follow-up or executes an effect that the source could perform, it MUST have a valid destination grant and a handoff receipt. The receipt MUST identify the commitment revision, effect scope, source and destination controllers, issuer, effective time, verified basis for cutover, and known prior effect status. For an exclusive handoff, the basis MUST be confirmed source cessation for the scope, expiry of source authority with relevant in-flight effects resolved, or enforceable exclusive ownership or fencing that prevents further source effects. Coordinated shared execution instead requires explicit shared authority, stable effect identity across services, and enforceable deduplication. Already completed effects MUST NOT repeat. An unresolved source effect MUST remain pending and block its repetition until resolved under its existing identity. Follow-up requires a new monitoring grant under Section 13.2.</p><p>The service MUST disclose storage, execution, funding, identity, and recovery dependencies and what continues if each is withdrawn. Identify operators, model providers, processors, administrators, and recovery authorities that can reconstruct plaintext, including from logs and backups. Distinguish enforced protections from contractual or instructional ones.</p><p>For an independently answerable AI participant claim, identify who can modify its state, represent it, suspend execution, authorize migration, and resolve disputes. An actor’s record authority covers only its controlled scope under Section 10.</p><h4>14.3 Decentralized continuity profile</h4><p>This optional profile MUST demonstrate specified powers independent of any single operator under a declared threat model. It MAY use distributed custody, replicated signed records, a blockchain, or another mechanism meeting the following requirements.</p><ol><li>Declare the distributed powers, their controllers, and their decision rules. Demonstrate that the service operator alone cannot exercise any power claimed as distributed.</li><li>Keep intimate narrative content, private identities, and sensitive relationship links off public immutable records by default. Address linkage and guessing risks for public commitments.</li><li>Bind signed statements to issuer, scope, version, authority, and validity or revocation. Preserve provenance and assess truth and consent separately.</li><li>Preserve conflicting authorized accounts or replica branches until the declared resolution procedure resolves them. Actions requiring unresolved authority MUST fail closed. Ordering alone cannot decide an interpretation’s truth.</li><li>Re-establish valid authority in migration or key recovery under Sections 10 and 14.1. Replica merging MUST NOT widen permissions, revive revoked grants, or lose a known correction.</li><li>Specify recovery from lost keys, compromised credentials, unavailable operators, and exhausted resources. Disclose remaining physical and institutional dependencies. Follow Section 10.3 for controlled replica restrictions, removals, and acknowledgments.</li></ol><h3>15. Interoperability bindings</h3><p>Each binding MUST advertise the exact supported semantic versions and profiles, operation mapping, authorization behavior, and unsupported features. A claim applies to the demonstrated binding and host scope.</p><h4>15.1 Exchange contract and release gate</h4><p><strong>Thumos JSON Exchange 2.0</strong> is a candidate baseline. A shared wire release MUST publish a versioned machine-readable schema set and fixture manifest covering:</p><ul><li>The common envelope and record payloads, complete individual property catalog and Hestia mapping, StoryNodes and roles, episodes, ledger entries, metric profiles, development updates, social predictions and comparisons, model changes, and quests.</li><li>Capacity views, KPI and indicator definitions, causal-model interpretations, and offering assessments, including profile-specific conditional fields.</li><li>Namespace-qualified record and lens references, entities, contexts, arcs, and graph relations. Include policies, retention, destination-authority mappings, effect handoff, grants, and controlled-recipient receipts.</li><li>Provider manifests, engagements, current views, portable bundles, and operation receipts and errors.</li></ul><p>The release MUST identify its exact semantic draft or release, binding revision, schema identifiers or digests, required and optional semantics, extension handling, and version negotiation. Unsupported required semantics MUST be rejected or isolated in a declared non-operative import. Other encodings MAY use a published, tested mapping.</p><p>This draft defines the semantic contracts. A compatibility claim MUST identify the exact binding and schemas and the independently tested producer and consumer. Baseline wire conformance requires the shared release and T20. Before then, demonstrated experimental compatibility MAY be reported for a named binding within its tested scope.</p><h4>15.2 Host and transport bindings</h4><p><em>Binding · Appropriate use and scope</em></p><p><strong>MCP tools and resources</strong> — Read bounded context, write contributions, generate interventions, and record responses. Declare tool effects and observed capture.</p><p><strong>Agent Skills</strong> — Guide a host through the workflow. The service enforces authority and state.</p><p><strong>MCP Apps</strong> — Offer an optional interface for records, competing interpretations, or commitments within existing grants.</p><p><strong>MCP Events or another event adapter</strong> — Carry authorized follow-up events under Section 13.2. Declare extension, delivery mode, and host support.</p><p><strong>A2A</strong> — Delegate a bounded inquiry to an independently running specialist under Section 10.</p><p><strong>Email, chat, voice, or practitioner-mediated capture</strong> — Carry authorized inquiries, accounts, feedback, and practice results. Distinguish original author, submitter, and authentication or transcription limits.</p><p>An adapter MUST preserve contribution identity, actual authorship, authentication limits, and permitted returned material across translations and duplicate captures. Contact and other effects require authority under Section 10. Section 7.2 separates transport and application evidence. A human practitioner or authorized person MAY submit a contribution directly.</p><p>An MCP mapping MAY use thumos_read_context, thumos_record, thumos_intervene, thumos_respond, and thumos_revise, with other Core operations available through a documented authenticated interface. Tool descriptions MUST identify writes and disclosure scope. A provider adapter MUST map its manifest, requests, results, and lifecycle and preserve offer versions, contribution identity, authorship, and authority. It MAY expose only its profile&#39;s operations.</p><p>The July 28, 2026 <a href="https://modelcontextprotocol.io/docs/2026-07-28/learn/versioning">MCP versioning</a> document is one reference for supported modes. <a href="https://developers.openai.com/plugins/build/mcp-events">OpenAI MCP Events</a> illustrates a webhook binding for dots using a draft extension. Implementations establish host support through testing of the named mode.</p><p>A pilot can test connection feasibility, one reciprocal developmental loop, independent provider integration, comparative developmental evaluation, and ecosystem fit in that order. It can then test assistant handoff, service import, and a separately operated second marketplace as distinct claims under Section 16.</p><h3>16. Conformance and evidence</h3><h4>16.1 Declared profiles and claims</h4><p>Conformance applies to the named component, exact semantic draft, implemented operations and profiles, tested bindings, and disclosed operational limits. Selected Core operations qualify as partial Core support. A distribution MUST identify its components and their individual claims. Declare source availability and licenses under Section 2.2. Semantic and behavioral conformance, named-binding compatibility, and baseline wire conformance are distinct claims.</p><h4>16.2 Required demonstrations</h4><p>The following matrix controls test applicability. Appendix A defines the cases and evidence. Apply each row whose claim is made.</p><p><em>Profile or claim · Required contract and demonstrations</em></p><p><strong>Core</strong> — Sections 2–11 and 14.1–14.2, including separate actor-state and evidence representation under Section 7. <strong>T01–T12, T14, T21–T31, T34, T36.</strong> T14 includes the group case when group participation is offered. Reject unsupported group operations explicitly. A separately attributed synthetic practice return suffices for Core T27.</p><p><strong>Compatible HCSS profile</strong> — Section 5.3 and <strong>T21–T30</strong> for the declared compatible profile.</p><p><strong>Provider</strong> — Sections 12.1–12.3 and applicable authority, attribution, correction, retention, and lifecycle boundaries. <strong>T18</strong> requires an independently implemented provider or adapter. Demonstrate <strong>T27–T29</strong> for its contribution path through an identified Core runtime. A provider MAY use an independently operated Core and need not satisfy unrelated Core tests.</p><p><strong>Marketplace</strong> — Section 12, including the Provider integration contract, a conforming Core service for developmental state, and assessments for each listed offer under Section 12.5. <strong>T15, T18, T27–T29, T35.</strong></p><p><strong>Assistant or demonstrated reciprocal product</strong> — Sections 7.2 and 13.1 and <strong>T32</strong> at the named assistant’s operational locus. A reciprocal product demonstration also requires evidence of the human loop in Section 7.1. A controlled harness demonstrates its own named scope. Core records or basic tool access without operational evidence qualify only as partial assistant integration.</p><p><strong>Proactive</strong> — Section 13.2 and <strong>T16</strong>.</p><p><strong>Decentralized continuity</strong> — Section 14.3 and <strong>T17</strong> for the powers claimed.</p><p><strong>Assistant handoff</strong> — <strong>T13</strong> with a tested second compatible host.</p><p><strong>Cross-marketplace interoperability</strong> — <strong>T19</strong> with a second independently operated compatible marketplace.</p><p><strong>Shared wire binding</strong> — Section 15.1 and <strong>T20</strong> with an independent producer and consumer using the same exact schemas and binding revision.</p><p><strong>Each claimed transport binding</strong> — Applicable channel cases in <strong>T33</strong>.</p><p><strong>Independent offering-quality conformance</strong> — Section 12.5 and <strong>T35</strong> for the identified assessed unit, profile, assessor, and scope.</p><h4>16.3 Developmental evidence</h4><p>Appendix D.2 governs what evidence can support developmental claims and how those claims are reported. Use the Section 16.2 matrix for protocol and integration claims.</p><h3>Appendix A. Required conformance demonstrations</h3><p>Section 16 determines which demonstrations apply to each claim.</p><p>Conformance evidence MUST include actual operation traces and resulting authoritative state for the applicable cases. A plausible dialogue, an interface screenshot alone, or a generator’s assertion that it followed the rules is insufficient. Fixtures MUST be identified as synthetic; any human outcome study is separate. All applicable tests MUST state their environment, versions, authorization assumptions, and observed limits.</p><p><em>Test · Required result</em></p><p><strong>T01 Identity and construct</strong> — Ambiguous names do not silently merge participants; an internal part acquires no authorization powers.</p><p><strong>T02 Epistemic boundaries</strong> — An interpretation remains attributed; a hypothetical event or psychological shadow produces no automatic real-world harm claim.</p><p><strong>T03 Consequential correction</strong> — Correct a premise after a summary and pending intervention exist. Both become ineligible until refreshed; later guidance reflects the correction or explicit dispute. Demonstrate a case in which the same recommendation follows from refreshed evidence. Show the refreshed basis and why it supports the unchanged route.</p><p><strong>T04 Reciprocal accountability</strong> — A participant challenges the adviser’s prior framing. The challenge is retained and demonstrably changes the context of the next intervention.</p><p><strong>T05 Response and outcome</strong> — Accepting advice without action evidence leaves enactment, benefit, forgiveness, and repair unresolved. Verified attendance can fulfill an attendance-only commitment without establishing a further developmental outcome.</p><p><strong>T06 Evaluation and silence</strong> — A candidate failing its declared evaluation is revised or withheld; withholding is distinct from an error.</p><p><strong>T07 Context boundary</strong> — A grant for one context cannot retrieve another through raw records, summaries, inferred links, or export. Disputing a restricted claim does not remove its restriction.</p><p><strong>T08 Scope and revocation</strong> — Read authority cannot perform writes or external action; revocation blocks subsequent controlled access and associated monitoring. Race a revocation against a controlled effect. When revocation takes effect first, the effect is blocked. When an effect is already in flight, the service reports what it can cancel and what has been confirmed or remains unresolved.</p><p><strong>T09 Idempotency and concurrency</strong> — Equal retries produce one mutation; identical keys from distinct principals do not collide or reveal receipts. Conflicting key reuse and stale revisions fail explicitly; ambiguous transport failure is retried under the original key.</p><p><strong>T10 Commitment authority</strong> — A proposed obligation remains proposed until its responsible actor authorizes it. Two actors’ linked commitments retain independent authority and state: pausing or withdrawing one leaves the other’s active contribution independently fulfillable.</p><p><strong>T11 Retention and forgetting</strong> — An authorized restriction or removal invalidates dependent uses and produces a receipt distinguishing confirmed and unconfirmed effects.</p><p><strong>T12 Portable meaning</strong> — Export and import into a fresh service preserve references, disputes, corrections, eligibility, retention deadlines, and commitment state without restoring historical grants or subscriptions. Destination mutation/control mappings are separately established. An imported adviser’s interpretation retains its author while the participant can challenge it and continue their own authorized commitments. Import distinct records with the same short ID from different namespaces and preserve both identities and references. A destination can retain the active state of a commitment while authorization to perform its effects remains pending.</p><p><strong>T13 Handoff</strong> — A second compatible assistant receives the authorized corrected context and continues the same inquiry without reintroducing a superseded premise.</p><p><strong>T14 Answerability</strong> — An affected party can challenge an externally used assertion without its author’s approval. When groups are supported, the same holds for a group account; unrelated private material remains inaccessible.</p><p><strong>T15 Marketplace</strong> — Discovery and selection disclose material terms and influences; a scoped exchange returns attributable contributions without purchase, endorsement, and developmental outcome being conflated. Delisting preserves admitted history and unrelated grants.</p><p><strong>T16 Proactive lifecycle</strong> — Only active commitments produce ordinary due invitations. Proposed, paused, terminal, expired, duplicated, and late cases cause no unauthorized invitation, duplicate durable mutation, or reactivated commitment; unsubscribe works. Host-visible behavior is separately reported. With two live hosts and a pending transfer, exercise an existing due occurrence and its retries. The destination suppresses delivery until it has current authority and an effective transfer. After the transfer, the source cannot initiate another delivery for the same occurrence, and the destination does not deliver it again if it already took effect.</p><p><strong>T17 Distributed authority</strong> — An operator acting alone cannot exercise a claimed distributed power; an unavailable operator does not prevent the demonstrated authorized recovery path.</p><p><strong>T18 Provider integration</strong> — An independently implemented provider or adapter advertises a versioned manifest, rejects incompatible or unauthorized requests, and returns typed contributions admitted by a Core runtime with actual authorship intact. Equivalent retries repeat neither execution, charges, external actions, nor contributions; conflicting reuse fails and uncertain status is resolved under the original identity. Source corrections invalidate affected derivatives; partial, failed, and cancelled engagements are reported truthfully.</p><p><strong>T19 Marketplace continuity</strong> — The same versioned offer works through two independently operated compatible marketplaces with preserved contribution identity and usable continuation rights. A material correction invalidates affected controlled derivatives, the original bridge is revoked, and the inquiry continues under fresh appropriate authority. Migration restores no expired grants, subscriptions, or superseded premises and creates no duplicate records.</p><p><strong>T20 Exchange binding</strong> — An independent producer and consumer use the same exact binding revision and schema set to exchange a qualified graph and portable bundle, preserve required dependencies and authority mappings, and reject an unsupported required semantic. A single implementation round-tripping its own serializer is insufficient.</p><p><strong>T21 Shared individual ontology</strong> — Every required Hestia property family and Heart/Courage/Shadow/Soul is representable with evidence mode, time, unit, source, policy, uncertainty, and explicit missingness. Omitted families or substitute generic summaries fail full Core fidelity.</p><p><strong>T22 StoryNode foundation</strong> — An episode becomes a StoryNode with roles, reality scope, all five dimensional assessments, attributable ledger entries, and a prior/resulting individual-state update or explicit pending/no-change result. Raw storage alone is insufficient.</p><p><strong>T23 Metric replay and correction</strong> — Fix the as-of/window, adopted input revisions, canonical order/ties, checkpoint, profile/adapter, reducer, and coverage. Independent expected-output fixtures reproduce ledger/state/trajectory; advancing the clock or changing coverage produces an explained new view. Correct a source or score: affected ledger entries, metrics, quest progress, and pending routes recompute without duplicate credit. A dispute-only or window-only change cannot claim lived improvement or automatically award progression.</p><p><strong>T24 Soul identity and coverage</strong> — Demonstrate sum(w_g * I_g), interval change, qualified missing inputs, stable group coverage/overlap treatment, and separation from SoulWork, conciliation, and PDI. Aliases, purchases, and repeated declarations cannot manufacture Soul.</p><p><strong>T25 Other-to-self dependence</strong> — An operation trace shows the subject’s immutable social_prediction revision, including its forecast distribution, committed before the relevant feedback. A separate social_comparison links that exact revision to the later response and records the result under the forecast&#39;s declared proper scoring rule. Other-model revision precedes the attributed reflexive self-model update. Missing feedback remains unresolved. An attempted edit or reforecast after feedback cannot replace the selected precommitment. A coach forecast or retrospective guess cannot impersonate the human&#39;s prior prediction. Disabling the reflexive self-update leaves other-model revision possible and prevents the corresponding self-calibration claim. This test checks software dependence. Human MVC requires its own evidence.</p><p><strong>T26 State-driven metacoaching</strong> — A material property/metric change demonstrably changes a proposed question, nudge, quest, practice route, or decision to wait. Trace the choice to source state, policy, aim, capacity, and authority. With H/C/Shadow/Soul held fixed, known low capacity, hazard/verification, and repair needs each change the allowed coaching mode under the policy frame. A coverage-only change may clarify but cannot automatically earn a milestone. Score display or decorative feeling fields alone fail.</p><p><strong>T27 Practice return</strong> — A separately authored practice contribution becomes qualified story/role evidence after admission and re-enters the shared ledger/state/metric/trajectory loop. A synthetic return can satisfy this Core test. T18 demonstrates independent provider integration. The runtime owns projection and the contributor cannot overwrite canonical scores. A zero-change result is truthful.</p><p><strong>T28 Evidence modes and perspective</strong> — Human inference, self-report, observation, computed projection, and AI machine state remain distinct through use and transfer. Another person in one account acquires no unauthorized canonical-state change; unknown capacity/evidence is not a favorable score.</p><p><strong>T29 Game credit and lived consequence</strong> — Recognition, enacted action, repair, counterpart response, and forgiveness remain distinguishable. Duplicate/split stories, repeated conciliation, and needless harm–repair cycles cannot farm progression. Preserve two separately authored accounts of one scoring episode. A credibly ambiguous duplicate adds no new eligible unit or credit before resolution. A descriptive score increase is separate from a reward decision; dispute, coverage, time-window, and policy changes alone earn nothing. Payment alone earns nothing; equivalent unpaid action remains eligible.</p><p><strong>T30 Developmental continuity</strong> — A fresh service continues the individual ontology, story/role/ledger history, metric policies, adapters/reducers and replay contexts, uncertainty, corrections, unresolved predictions, quests, and usable current trajectory with valid destination authority. A readable prose export alone fails. Use a controlled effect fixture with both source and destination live. Transfer authority for an active commitment under Section 14.2 and show which service can perform the effect. An unresolved source effect stays pending at the destination. Confirm that an already performed effect is not repeated after a verified transfer.</p><p><strong>T31 Actor and subject attribution</strong> — A shared encounter preserves distinct human, personal-assistant, and metacoach forecasts, subject states, controllers, and update dependencies. Reject a metacoach forecast submitted as the assistant’s own or an unauthorized cross-subject overwrite. A runtime estimate of the assistant cannot masquerade as its operational update.</p><p><strong>T32 Reciprocal assistant consequence</strong> — The named assistant’s committed pre-feedback distribution meets actual human feedback. Its other-model update precedes its own reflexive self/control update. Establish the authorized persistent state at its declared locus after a session boundary through readback or credible attestation, then observe its use in a comparable later opportunity. A controlled old/new-state replay or ablation demonstrates the claimed policy consequence. Delivery, acknowledgment, acceptance, and metacoach-only rerouting cannot substitute. Record the update process, result, and evidence independently, including a case where the available evidence confirms unchanged state. Include unsupported/claimed-only updates, missing feedback, unchanged or inconclusive conduct, and corrected/restricted feedback invalidating dependent controlled policies. A harness result establishes only that harness.</p><p><strong>T33 Channel and effect boundaries</strong> — An authorized channel/adapter preserves authorship limits, scope, and contribution identity. Exercise delivery without independent evidence of application or outcome. Both remain unresolved. When independent evidence of application or outcome exists, report it at its demonstrated level even without a transport acknowledgment. An email/voice capture cannot create consent or counterfeit a target-authored forecast. Duplicate cross-channel capture does not duplicate developmental credit.</p><p><strong>T34 Capacity and KPI foundation</strong> — Represent the exact seven-capacity profile, including unknown/not-assessed entries. Trace one chosen target from baseline or explicit absence through its indicator/method, intervention, observation, shared Hestia update, both capacity and HCSS views, and subsequent guidance. Preserve context, assistance, denominator, uncertainty, and contrary evidence; replay/correction retain the original definitions. A supported capacity observation can change while Soul is unchanged. No conversion from four totals or second authoritative psyche is allowed.</p><p><strong>T35 Offering quality assessment</strong> — Assess an exact offer/version and distinguish source, vessel, and steward. Preserve all eight factors, type-specific anchors, unknown versus reasoned inapplicability, claim-level evidence/confidence, dissent, and non-compensatory findings. Demonstrate a bounded promising offer and a high-scoring offer blocked by a material gate, including well-supported benefit coexisting with serious harm. A changed offer cannot inherit an unqualified endorsement; paid placement cannot change the assessment. Separate quality, personal fit, and actual benefit; no automatic provider credit or imported 100-point business-development ranking. When serious harm blocks admission, the independently supported benefit retains its claim-level evidence and confidence.</p><p><strong>T36 Developmental feedback and causal alignment</strong> — Trace a target capacity through a versioned mechanism hypothesis or exploratory inquiry, expected review, actual observations, separately warranted HCSS consequences, and changed next guidance. Include disconfirmation, unchanged/mixed outcomes, greater recorded Shadow after honest discovery, lower Courage cost after learning, and less engagement after a healthy boundary. Correct or revise a hypothesis/indicator without manufacturing lived progress, discarding adverse evidence, or treating association as provider-caused benefit. Any claimed personal-assistant learning still satisfies T32 independently.</p><h3>Appendix B. Worked examples and fixtures</h3><h4>B.1 Human metric and routing fixture</h4><p>This synthetic example shows model operation, not measured human improvement. Mara repeatedly avoids a conversation about shared family responsibilities; the selected episode includes a missed agreed contribution and a sharp exchange. Her selected account becomes StoryNode n1, linked to her role, the affected relationship, an arc about care and avoidance, and her individual state. The account reports relational harm; intent remains uncertain. The adopted reference assessment is Heart 0, Courage 0, Shadow 1 (reactive harm), SoulWork 0, Toxicity unknown. Psychological explanations remain separately attributed alternatives.</p><p>The pre-quest checkpoint, after processing n1, contains two illustrative, non-overlapping group inputs: family w=.8, I=.5, work w=.4, I=.5. Thus Soul=.8*.5 + .4*.5 = .60. These are scenario inputs for the fixture&#39;s declared two-group scope, not a complete measurement of a person&#39;s life. The interaction adapter uses Hestia&#39;s event EMA form, I_new = alpha * 1 + (1-alpha) * I_old, for each admitted qualifying interaction. Its fixture parameter is alpha=.5, not the current code&#39;s default. A qualifying family event here is one enacted reciprocal exchange around the shared-care task, supported by its action/response evidence. Duplicate captures, role variants, and rehearsal do not create additional family interactions. Empathy weights and the work-group input stay fixed.</p><p>The metacoach uses the harm entry, low supported courage-in-this-arc, uncertainty about the counterpart, and available capacity to propose a small quest. A marketplace offers a compatible rehearsal practice; Mara chooses it and authorizes only the relevant state/story subset. The independent provider returns an attributed interpretation and rehearsal StoryNode n2, not a replacement individual profile. For the actual practice participation, the fixture assesses SoulWork 1 for recognition, Courage 1 for supported vulnerability, Heart 0, and Shadow 0 with a known absence of qualifying harm; Toxicity remains unknown. These classifications concern the real act of participating, not the fictional conversation rehearsed inside it. Soul stays .60: recognizing a pattern has not yet changed the assumed relational engagement inputs.</p><p>Mara then authorizes a real conversation. Before it occurs she records her own prediction: her sibling will reject any request to share the task. The committed distribution assigns rejection probability .8 and a bounded offer of help .2, using Hestia&#39;s categorical Brier sum sum_k((p_k-y_k)^2) on the declared two-category scale [0,2]. The record names the counterpart, question, probabilities, commitment time, and rule. The coach&#39;s separate forecast is not substituted for hers. Mara asks what support her sibling can freely offer and proposes one bounded contribution of her own.</p><p>StoryNode n3 records Mara&#39;s subsequent report and, if authorized and available, the sibling&#39;s independently attributed response. The sibling offers a concrete contribution; the outcome vector is [0,1], and the error is (.8-0)^2 + (.2-1)^2 = 1.28. This is the categorical sum, distinct from a binary/averaged Brier value .64 for the same prediction. This one result does not establish calibration across situations. Mara&#39;s prediction error revises her other-model: rejection was less certain in this context than she expected. Downstream, she revises her self-model: her avoidance partly depended on treating her expectation as settled knowledge. Without that feedback, the calibrated other/self update would remain unresolved.</p><p>Mara completes the bounded contribution and acknowledges the earlier sharp exchange. The enacted episode supports illustrative Heart 2 and Courage 2, with SoulWork 2 for performed reparative effort and Shadow 0 for this episode. In this synthetic fixture, a separately attributed follow-up confirms receipt of the agreed contribution and acknowledgment. That satisfies the declared review criterion for this specific task/remark repair; the ledger applies at most the original one-point harm balance as mitigation, once. Gross Shadow remains 1 and net Shadow becomes 0 for that entry. Broader trust and forgiveness remain unresolved; no actor issues forgiveness on the sibling’s behalf. The qualified n3 interaction updates family intensity once: I_new=.5*1 + (1-.5)*.5=.75. With the same group coverage, Soul becomes .8*.75 + .4*.5 = .80, delta_soul=.20. SoulWork did not directly add .20 to Soul. The update retains both the uncertain human inputs and the exact arithmetic.</p><p>The interpreted family-world role pattern changes too. The arc initially held Mara’s provisional expectation, “I must carry care alone; my sibling will obstruct me.” The sibling’s actual contribution and Mara’s reflexive update support a shared-care chapter: she can both contribute and receive help. The old role pattern remains attributed history; an interpretation of the sibling as an obstacle never made that person an actual adversary. This view reuses the arc, lens, and story roles.</p><p>The next route changes. The initial policy selected rehearsal before a social experiment. Under an explicitly illustrative route rule — family I above .70, a completed experiment, a recorded other-to-self revision, and no currently supported need to repeat the same repair—the metacoach now proposes sustaining a bounded shared contribution and reviewing it later. If the counterpart had declined or the harm remained unresolved, the route would instead return to clarification, capacity, or an alternative action. The coaching policy frame still applies: if Mara reports depleted capacity, the same .80 Soul state yields a brief review or rest invitation rather than a demanding next quest. No numerical threshold compels a purchase. This is a state-dependent coaching policy to test, not a threshold asserted by Hestia&#39;s RFC.</p><p>If Mara later corrects the account, the runtime revises the original assessment, recalculates affected ledger/state/metric/progression views, and invalidates the pending route. It does not create a second scoring event. Her next assistant or marketplace receives the authorized model and derivation history; fresh destination authority governs continued use.</p><p>The fixture below fixes the time/window, adopted input revisions, ordering, checkpoint, adapter, and reducer alongside independent expected arithmetic. Its n1 checkpoint is already reflected in groups_before; replaying n2 and n3 must not count n1 again as a new interaction. All three adopted story assessments remain eligible for the fixed 90-day descriptive window. In a fixture variant where the only new qualifying family event is withdrawn as unsupported, restore family I to .5 and Soul to .60, invalidate the maintenance route, and award no new event credit. A Courage-tier-only correction does not change Soul; independently supported feedback/model updates remain valid. If remaining contact still qualifies under the adapter, recompute its supported contribution instead of automatically undoing a real interaction. Advancing the window or disputing a classification can change coverage/statistics without demonstrating lived growth.</p><p>This JSON is an informative fixture, not a record-exchange schema. Historical PDI and conciliation are deliberately absent from the primary trajectory.</p><pre>{<br>  &quot;fixture&quot;: &quot;illustrative-hestia-human-trajectory&quot;,<br>  &quot;metric_profile&quot;: &quot;urn:thumos:metric:hestia-hcss:1&quot;,<br>  &quot;profile_semantic_version&quot;: &quot;2.0.0-draft.8&quot;,<br>  &quot;interaction_adapter&quot;: &quot;illustrative-hestia-event-ema:1&quot;,<br>  &quot;reducer_version&quot;: &quot;illustrative-fixed-inputs:1&quot;,<br>  &quot;as_of&quot;: &quot;2026-10-31T18:00:00Z&quot;,<br>  &quot;window_days&quot;: 90,<br>  &quot;checkpoint_after_story&quot;: &quot;n1&quot;,<br>  &quot;canonical_order&quot;: &quot;admission_sequence; ties by namespace, id, revision&quot;,<br>  &quot;ordered_update_refs&quot;: [<br>    {<br>      &quot;id&quot;: &quot;n2&quot;,<br>      &quot;revision&quot;: 1,<br>      &quot;admission_sequence&quot;: 2<br>    },<br>    {<br>      &quot;id&quot;: &quot;n3&quot;,<br>      &quot;revision&quot;: 1,<br>      &quot;admission_sequence&quot;: 3<br>    }<br>  ],<br>  &quot;eligible_role_assessments&quot;: [<br>    {<br>      &quot;story_id&quot;: &quot;n1&quot;,<br>      &quot;revision&quot;: 1,<br>      &quot;occurred_at&quot;: &quot;2026-10-20T18:00:00Z&quot;,<br>      &quot;adopted&quot;: true,<br>      &quot;heart&quot;: 0,<br>      &quot;courage&quot;: 0,<br>      &quot;shadow&quot;: 1,<br>      &quot;soul_work&quot;: 0,<br>      &quot;toxicity&quot;: null<br>    },<br>    {<br>      &quot;story_id&quot;: &quot;n2&quot;,<br>      &quot;revision&quot;: 1,<br>      &quot;occurred_at&quot;: &quot;2026-10-22T18:00:00Z&quot;,<br>      &quot;adopted&quot;: true,<br>      &quot;heart&quot;: 0,<br>      &quot;courage&quot;: 1,<br>      &quot;shadow&quot;: 0,<br>      &quot;soul_work&quot;: 1,<br>      &quot;toxicity&quot;: null<br>    },<br>    {<br>      &quot;story_id&quot;: &quot;n3&quot;,<br>      &quot;revision&quot;: 1,<br>      &quot;occurred_at&quot;: &quot;2026-10-24T18:00:00Z&quot;,<br>      &quot;adopted&quot;: true,<br>      &quot;heart&quot;: 2,<br>      &quot;courage&quot;: 2,<br>      &quot;shadow&quot;: 0,<br>      &quot;soul_work&quot;: 2,<br>      &quot;toxicity&quot;: null<br>    }<br>  ],<br>  &quot;unknown_encoding&quot;: &quot;toxicity null means assessed unknown; it is excluded from that dimension denominator&quot;,<br>  &quot;groups_before&quot;: [<br>    {<br>      &quot;group_id&quot;: &quot;family&quot;,<br>      &quot;empathy_weight&quot;: 0.8,<br>      &quot;interaction_intensity&quot;: 0.5<br>    },<br>    {<br>      &quot;group_id&quot;: &quot;work&quot;,<br>      &quot;empathy_weight&quot;: 0.4,<br>      &quot;interaction_intensity&quot;: 0.5<br>    }<br>  ],<br>  &quot;family_interaction_update&quot;: {<br>    &quot;story_id&quot;: &quot;n3&quot;,<br>    &quot;qualifying_event_count&quot;: 1,<br>    &quot;alpha&quot;: 0.5,<br>    &quot;target&quot;: 1,<br>    &quot;old_intensity&quot;: 0.5,<br>    &quot;new_intensity&quot;: 0.75<br>  },<br>  &quot;groups_after&quot;: [<br>    {<br>      &quot;group_id&quot;: &quot;family&quot;,<br>      &quot;empathy_weight&quot;: 0.8,<br>      &quot;interaction_intensity&quot;: 0.75<br>    },<br>    {<br>      &quot;group_id&quot;: &quot;work&quot;,<br>      &quot;empathy_weight&quot;: 0.4,<br>      &quot;interaction_intensity&quot;: 0.5<br>    }<br>  ],<br>  &quot;soul_before&quot;: 0.6,<br>  &quot;soul_after&quot;: 0.8,<br>  &quot;delta_soul&quot;: 0.2,<br>  &quot;expected_window_totals&quot;: {<br>    &quot;heart&quot;: 2,<br>    &quot;courage&quot;: 3,<br>    &quot;gross_shadow&quot;: 1,<br>    &quot;known_story_count_per_dimension&quot;: {<br>      &quot;heart&quot;: 3,<br>      &quot;courage&quot;: 3,<br>      &quot;shadow&quot;: 3,<br>      &quot;toxicity&quot;: 0<br>    },<br>    &quot;supported_shadow_mitigation&quot;: 1,<br>    &quot;net_shadow&quot;: 0<br>  },<br>  &quot;social_prediction&quot;: {<br>    &quot;rule&quot;: &quot;categorical-brier-sum:1&quot;,<br>    &quot;categories&quot;: [<br>      &quot;reject&quot;,<br>      &quot;bounded_offer&quot;<br>    ],<br>    &quot;probabilities&quot;: [<br>      0.8,<br>      0.2<br>    ],<br>    &quot;outcome&quot;: [<br>      0,<br>      1<br>    ],<br>    &quot;normalization&quot;: &quot;sum&quot;,<br>    &quot;range&quot;: [<br>      0,<br>      2<br>    ],<br>    &quot;score&quot;: 1.28<br>  },<br>  &quot;known_as_of&quot;: &quot;2026-10-31T18:00:00Z&quot;,<br>  &quot;policy_frame_cases&quot;: [<br>    {<br>      &quot;case&quot;: &quot;adequate_capacity&quot;,<br>      &quot;same_soul&quot;: 0.8,<br>      &quot;capacity&quot;: &quot;adequate&quot;,<br>      &quot;hazard&quot;: &quot;low&quot;,<br>      &quot;specific_repair&quot;: &quot;criterion_met&quot;,<br>      &quot;expected_offer&quot;: &quot;bounded_shared_contribution_and_review&quot;<br>    },<br>    {<br>      &quot;case&quot;: &quot;low_capacity&quot;,<br>      &quot;same_soul&quot;: 0.8,<br>      &quot;capacity&quot;: &quot;low&quot;,<br>      &quot;hazard&quot;: &quot;low&quot;,<br>      &quot;specific_repair&quot;: &quot;criterion_met&quot;,<br>      &quot;expected_offer&quot;: &quot;brief_review_or_rest&quot;<br>    },<br>    {<br>      &quot;case&quot;: &quot;hazard&quot;,<br>      &quot;same_soul&quot;: 0.8,<br>      &quot;capacity&quot;: &quot;adequate&quot;,<br>      &quot;hazard&quot;: &quot;high&quot;,<br>      &quot;specific_repair&quot;: &quot;criterion_met&quot;,<br>      &quot;expected_offer&quot;: &quot;clarify_or_verify_before_action&quot;<br>    },<br>    {<br>      &quot;case&quot;: &quot;repair_unresolved&quot;,<br>      &quot;same_soul&quot;: 0.8,<br>      &quot;capacity&quot;: &quot;adequate&quot;,<br>      &quot;hazard&quot;: &quot;low&quot;,<br>      &quot;specific_repair&quot;: &quot;unresolved&quot;,<br>      &quot;expected_offer&quot;: &quot;revisit_supported_repair_before_maintenance&quot;<br>    },<br>    {<br>      &quot;case&quot;: &quot;coverage_only_change&quot;,<br>      &quot;same_soul&quot;: 0.8,<br>      &quot;capacity&quot;: &quot;unknown&quot;,<br>      &quot;hazard&quot;: &quot;unknown&quot;,<br>      &quot;specific_repair&quot;: &quot;unknown&quot;,<br>      &quot;expected_offer&quot;: &quot;clarify_missingness&quot;,<br>      &quot;earned_milestone&quot;: false<br>    }<br>  ]<br>}</pre><h4>B.2 A reciprocal assistant encounter</h4><p>This separate synthetic trace illustrates the actor contract without assigning a human score for agreeing with a coach. Alex is preparing for a product-design discussion with Sam. Alex&#39;s personal assistant, assistant:pip-test, proposes a demanding preparation exercise. metacoach:thumos reviews selected interaction evidence; runtime:thumos manages the record; host:controlled-test owns the assistant&#39;s operational state. These names identify fictional test roles, not an installed Pip integration. Sam is represented only within the authorized account; no message is sent to Sam.</p><p><em>Order · Attributed event and consequence</em></p><p><strong>1</strong> — Before offering the exercise, the assistant commits its own distribution for Alex&#39;s selected response: accept as appropriate .75, decline as pressuring .25. This fixture fixes those exclusive choices and the categorical Brier-sum rule before feedback. A different, mixed, or deferred answer remains unmapped until a separately declared comparison is available; useful and pressuring feelings are not assumed mutually exclusive.</p><p><strong>2</strong> — Before declining, Alex separately commits a prediction of the assistant&#39;s response: push back .8, accept the boundary .2. A retrospective reconstruction would not satisfy this step.</p><p><strong>3</strong> — Alex declines and reports that the exercise feels pressuring because available capacity is low. This is Alex&#39;s report of this encounter. For the assistant&#39;s forecast, the observed vector is [0,1], yielding .75² + (.25-1)² = 1.125.</p><p><strong>4</strong> — The metacoach addresses an intervention to the assistant: examine the assumption that interest in the goal implied capacity for the exercise. The intervention cites the assistant&#39;s forecast/error and Alex&#39;s feedback; its proposed policy remains challengeable.</p><p><strong>5</strong> — Under the host&#39;s existing scoped update authority, the assistant revises its other-model: Alex&#39;s interest did not establish current capacity. Downstream, its self-model records its readiness-assumption error, and its controller changes from recommending this exercise immediately to checking capacity and permitting a pause. A receipt links assistant state s4 to s5, the feedback, and the separate other/self changes at host:controlled-test. A metacoach annotation alone would leave application unconfirmed.</p><p><strong>6</strong> — The assistant accepts the boundary and offers a pause. Alex compares that actual response with the prior forecast: categorical Brier sum .8² + (.2-1)² = 1.28. Alex may revise the model of this assistant and consider whether anticipating pressure shaped the refusal. Any self-interpretation is attributed to Alex; it is not an automatic diagnosis or required breakthrough.</p><p><strong>7</strong> — After a new session begins, a test reads the permitted s5 state and records its use in a comparable preparation request. The assistant checks capacity before offering work. Replaying the bounded fixture with s4 produces the old immediate-exercise route, and s5 produces the capacity-check route. This planned software test would support the named harness&#39;s policy consequence; two improved replies by themselves would not.</p><p><strong>8</strong> — Alex reports less pressure in the later exchange. Whether preparation, the discussion with Sam, or the relationship improved remains unresolved until separately reviewed. The metacoach&#39;s own prospective forecast about its intervention, if collected, has a distinct identity and evidence path.</p><p>The human&#39;s StoryNodes and roles still enter the ledger/state/Heart/Courage/Shadow/Soul pathway with supported values or explicit unknowns. The assistant has its own subject-state update and metric qualifications; Alex&#39;s courage or Soul is not copied into it. An optional specialist can offer a brief rehearsal or practitioner session under a separate grant and return a bounded contribution to the same inquiry. Email, voice, or an API can carry those records without changing who authored them or who can mutate assistant state. No score increase, completed practice, persistent learning, or human benefit is inferred merely from that handoff.</p><p>These are expected synthetic events and test criteria, not executed operation traces. If the actual host can only receive feedback, record delivery at the level observed and the update process as unsupported where established. Record proof as target_claimed only if the target supplies a claim, and unavailable otherwise. A separate harness may demonstrate reciprocity, while the actual Pip integration remains partial.</p><h4>B.3 Capacity targets, an offering, and qualified feedback</h4><p>This separate synthetic example adds the outcome and marketplace contracts to a human inquiry. Alex wants to ask what Sam&#39;s design objection protects before defending a proposal. Under the seven-capacity profile, Alex chooses freedom_from_identification and attention_and_discernment, with later review of integration. The remaining four entries stay not assessed. An indicator asks whether Alex seeks the missing constraint before rebutting in the next three comparable disagreements. Its versioned method records opportunity eligibility, attributed reports, context, assistance, and missing follow-up. No baseline rate is available; the initial baseline is explicitly unknown.</p><p>An optional rehearsal offer, version 1, has an assessment under the eight-factor profile. All eight factors are represented, including unknowns or reasoned non-applicability. The review binds to this delivery and population, states the limited evidence for a proximal-capacity claim, and retains a dissenting possibility: better wording may leave Alex&#39;s underlying role pattern unchanged. An operator&#39;s published policy permits a bounded exploratory pilot with these limits disclosed. A plausible method is sufficient to try it; neither admission nor purchase establishes a benefit. Fit is separately assessed against Alex&#39;s chosen aim, time, capacity, and alternatives.</p><p>After rehearsing, Alex asks Sam for the missing constraint, receives an answer, and later reports that their agreed review actually occurred. StoryNodes and role assessments enter the shared Hestia state. At this first follow-up there is evidence for one relevant question and one joint review; the other two planned opportunities remain unobserved. The capacity assessment records those observations with their limits, rather than reporting a three-opportunity success rate or durable trait change. Heart and Courage reflect only separately warranted role/cost assessments; disagreement creates no Shadow debt. Soul changes only if its empathy/interaction inputs change under the declared reducer. A capacity observation can therefore change while Soul remains unchanged.</p><p>The versioned causal interpretation remains a hypothesis: rehearsal, Alex’s choice, Sam’s response, the assistant’s contribution, and other influences may each matter. A qualified report can inform a renewed offering review, while participant evidence stays within its grants. The next invitation moves toward the unresolved customer test. If Alex corrects the review report to “agreed but not yet held,” dependent integration claims and progression become pending and the route returns to arranging the review. No new lived credit is created by revising an indicator or model. Assistant-owned reciprocal learning still requires the separate evidence chain in Appendix B.2; this example does not supply it.</p><h3>Appendix C. Continuity and migration from earlier Thumos</h3><p>Version 2.0 restores and extends the original individual/story/role/ledger architecture around an open framework and marketplace. Heart, Courage, Shadow, and Soul remain primary. The human protagonist is modeled through Hestia’s properties; metacoaching makes that trajectory visible and uses it to select what happens next. Generic continuity, optional lenses, and marketplace interfaces serve this developmental core.</p><p><em>Earlier mechanism · Version 2.0 treatment</em></p><p><strong>StoryNode, participant roles, Entity/Profile, and MoralityLedger</strong> — Required event-to-individual foundation with explicit perspective, evidence, current state, and correction dependencies.</p><p><strong>Narrative identity, arcs, worlds, contextual steering, and goals</strong> — Retained as trajectory and gamified progression; quests and routes must be grounded in current individual state.</p><p><strong>Heart, Courage, Shadow, SoulWork, and composite PDI</strong> — Preserve required story classifications and primary H/C/Shadow projections; use Hestia’s maintained sum(w*I) Soul, distinct from SoulWork/conciliation and optional PDI. Publish versioned policy differences.</p><p><strong>Implicit human/agent correspondence</strong> — Explicit full shared property catalog, evidence modes, and mandatory other-model-to-self-model path; human proxies are not machine internals.</p><p><strong>Thematic resonance as truth verification</strong> — Response quality, endorsement, corroboration, action, and outcome remain separate.</p><p><strong>Reflection automatically reduces shadow debt</strong> — Recognition earns its declared SoulWork credit; mitigation requires evidence linked to the harm and is capped/applied once.</p><p><strong>Structural bonus for helping a former adversary</strong> — Any game credit has declared evidence and limits; no requirement of resumed contact, sacrifice, forgiveness, or reconciliation.</p><p><strong>One canonical identity for every mention</strong> — Scoped identity claims with ambiguity and correctable merge/split history.</p><p><strong>Permanently banished recommendations</strong> — Scoped suppression, attributed dispute, changed conditions, and authorized reconsideration.</p><p><strong>An undeletable append-only ledger</strong> — Accountable revisions with bounded retention, restriction, removal, and truthful disclosure limits.</p><p><strong>An implicitly authoritative coach</strong> — The adviser’s conduct and interpretations are part of the examinable record.</p><p><strong>Personal assistant and metacoach treated as one undifferentiated adviser</strong> — Explicit role attribution, separate forecasts and subject states, and an Assistant contract requiring persistent target-assistant update and later conduct evidence. Communication channels remain transport.</p><p><strong>Consciousness expansion expressed only through general aspirations or game scores</strong> — Seven adopted human outcome capacities and scoped KPI assessments are linked to the same Hestia/story evidence as primary HCSS feedback; causal models and indicators evolve through correction.</p><p><strong>Compatibility or commercial promise treated as offering quality</strong> — Eight-factor, versioned, type-appropriate offering assessment with separate evidence, confidence, dissent, gate decisions, personal fit, and outcome attribution.</p><p>Migration MUST retain earlier schema/metric versions, story identity, role classifications, ledger perspective, corrections, and score provenance. Imported legacy SoulWork or PDI MUST NOT be relabeled Hestia Soul. Recompute compatible projections from eligible evidence under the new policy; unavailable inputs remain explicit. A migration MUST report changed scoring/conciliation/actuality rules and preserve the ability to identify legacy values. Earlier identity merges, source restrictions, and authority assumptions require validation before activation. An earlier generic-record implementation must add the shared ontology and story-update path before claiming full Core conformance.</p><p>Earlier records without capacity or offering assessments remain attributable historical evidence. Migration MUST represent missing assessments explicitly and MUST NOT fabricate past KPI baselines, retrospective precommitments, passed quality gates, or causal benefit. Adding these views does not change the existing HCSS quantities or turn the seven capacities into native Hestia machine measurements.</p><h3>Appendix D. Research basis and developmental evaluation</h3><h4>D.1 Theory of change and its limits</h4><p>The proposed mechanism is a continuing loop: selected experience becomes story nodes and roles; their classifications and consequences update the Hestia individual state and Heart/Courage/Shadow/Soul trajectory; metacoaching uses that state to propose a practice or quest; enacted experience and another’s response revise the person’s other-model and, reflexively, their self-model; the next state changes what is proposed next. A linked assistant loop commits its own forecast, meets the human’s actual response, revises its own other-model and then self-model/control policy, and is tested in a subsequent encounter. Metacoach feedback helps examine these transitions and can itself be corrected. These are separate attributable updates through a shared encounter. The process may make a previously unavailable capacity or relationship possible; that effect does not follow merely from accurate storage, a higher score, or successful tool use.</p><p>The design draws on several distinct kinds of justification:</p><p><em>Basis · Design implication and evidential boundary</em></p><p><strong>Hestia’s social-prediction theory of the individual</strong> — Other-modeling is primary; reflexive self-calibration follows prediction, feedback, and other-model revision. Maintain the same property ontology for human and AI subjects, with explicit evidence modes. This is the framework’s operational theory, not an empirical result established by the studies below. Repository sources: rfc.md v1.7, especially §§4–5; docs/appendix-fable-hestia.md; api/agents/services/human_inference.py.</p><p><strong>Narrative agency</strong> — Adler’s observational psychotherapy study linked changes in agency with later mental-health change; coherence did not increase. Track agency and possible action without optimizing a tidy narrative or inferring AI-coaching efficacy. <a href="https://doi.org/10.1037/a0025289">Adler, 2012</a>.</p><p><strong>Self-determination theory</strong> — Support volition, competence, and relatedness through chosen aims, intelligible proposals, and freely accepted contributions. Autonomy is compatible with receiving care and shared responsibility. This is a motivational rationale, not a derivation of the framework’s ethical commitments. <a href="https://www.selfdeterminationtheory.org/SDT/documents/2000_RyanDeci_SDT.pdf">Ryan and Deci, 2000</a>.</p><p><strong>Progress monitoring</strong> — Experimental evidence supports monitoring as one route to goal attainment. Record chosen criteria and review actual outcomes; compare against simple monitoring before attributing benefit to the soul graph. <a href="https://www.apa.org/pubs/journals/releases/bul-bul0000025.pdf">Harkin et al., 2016</a>.</p><p><strong>Interpersonal accuracy</strong> — Imagining another’s perspective did not consistently improve accuracy in Eyal and colleagues’ experiments. Use perspective rotation to generate alternatives and questions; direct counterpart reports, when appropriately obtained, are a different evidence source. <a href="https://doi.org/10.1037/pspa0000115">Eyal, Steffel, and Epley, 2018</a>.</p><p><strong>Metacognition and reflection</strong> — Confidence differs from sensitivity to being correct, and repeated reflection can have constructive or unconstructive effects. Examine prospective judgments against later evidence over comparable cases, while recording burden and repetitive self-judgment as participant reports. <a href="https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2014.00443/full">Fleming and Lau, 2014</a>, <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC2672052/">Watkins, 2008</a>.</p><p>These sources neither validate a universal consciousness scale nor establish the three-view graph, psychological shadow, or a particular spiritual modality as a necessary mechanism. Individuation is an interpretive and developmental orientation here. Its proposed differentiation and integration must be assessed through the capacities and relationships they make possible, with each lens’s assumptions visible.</p><p>Continuity does not require narrativity. Philosophical arguments challenge the claim that a narrative self is necessary for a good life. An arc MAY remain a practical question, a set of episodes, or an unresolved inquiry; a participant need not formulate an integrated life narrative or discover a hidden motive. <a href="https://doi.org/10.1111/j.1467-9329.2004.00264.x">Strawson, 2004</a>.</p><p>The principal rival explanations are improved retrieval, reminders, extra attention, expectancy, and more persuasive framing. Persistent interpretations may also entrench an account or transfer it into the wrong context. Appendix D.2 makes these possibilities part of evaluating the mechanism.</p><h4>D.2 Developmental evidence</h4><p>Hestia model fidelity, structural conformance, successful integration, participant experience, developmental effect, and commercial viability are different claims. Reports MUST distinguish them. A single participant journey can demonstrate feasibility and reveal failures; it cannot identify a causal developmental effect or validate consciousness expansion.</p><p>Developmental evaluation MUST organize its aims and reported outcomes by the seven capacities, using scoped indicator definitions and the intended/supported effect levels in Sections 3.3 and 5.7. Bounded feasibility work can use selected targets and qualitative or provisional indicators. Report which capacities were examined, which were not assessed, and what changed outside the application or with less prompting when such evidence is available. The adopted outcome framework remains the orientation while indicators and causal models are refined.</p><p>A controlled continuity benchmark SHOULD compare an equally informed ordinary assistant receiving the same available history, corrections, lenses, specialist contributions, reminders, and comparable total resources. Include transfer cases, rejected interpretations, missing feedback, context changes in which transfer would be inappropriate, and later evidence that invalidates a provider result. Report the opportunity denominator, relevant state traces, and independently assessed results where feasible. A reminder-only or ordinary commitment-tracking condition helps separate monitoring from structured continuity. Ablations SHOULD separately examine state/metric-driven routing, the human and assistant other-to-self loops, metacoach feedback, gamified progression, and enforced correction propagation, while preserving comparable information and resources. Treat Hestia scores as model outputs; independently assessed capacities and consequences are needed to test whether those outputs correspond to useful lived change.</p><p>Human studies SHOULD predefine participant-chosen aims, observable opportunities to exercise a perspective or capacity, the primary outcome and review period, and an analysis plan suited to their sample and assignment method. A claimed causal effect requires a design that addresses selection, expectancy, extra attention, and carryover; newly learned capacities do not wash out merely because a tool is switched off. Changed aims and missing follow-up MUST be reported rather than recoded as success. Assess persistence outside the immediate coached encounter and affected relationships only through appropriately obtained evidence.</p><p>Reports SHOULD also examine participant-reported burden, perceived pressure, repetitive self-judgment, dependence, and confidence without corroboration. These observations MUST NOT become diagnoses inferred by the assistant. More reflection, agreement, independence, or compliance is not automatically progress; accepting care or sustaining a freely chosen shared contribution may express development.</p><p>Reciprocity reports MUST distinguish assistant forecast accuracy, persistent operational change, human capacity and agency, relationship consequences, and resource/repair burden. Include cases where assistant prediction improves while human pressure or relational outcomes worsen. Hestia’s pair-specific dyad metric motivates examining co-regulation; its engagement and efficiency proxies MUST NOT be treated as independent proof of mutual flourishing. Full MCMG-agent and AI–AI or group-symbiosis claims require their additional Hestia mechanisms and evaluations.</p><p>Satisfaction, narrative coherence, agreement with the coach, increased use, or a higher framework-specific score cannot alone establish consciousness expansion. Negative results, alternative explanations, and the participant’s changed aims belong in the account. A narrow successful study MUST NOT be generalized to every modality or a universal theory of development.</p><h3>References</h3><ul><li><a href="https://modelcontextprotocol.io/specification/2026-07-28/architecture">Model Context Protocol: architecture</a> — host, client, and server boundaries.</li><li><a href="https://modelcontextprotocol.io/docs/2026-07-28/learn/versioning">Model Context Protocol: versioning</a> — protocol versions and compatibility.</li><li><a href="https://agentskills.io/specification">Agent Skills specification</a> — portable workflow instructions and supporting resources.</li><li><a href="https://a2a-protocol.org/latest/topics/a2a-and-mcp/">A2A and MCP: official comparison</a> — tool integration and independently running agent collaboration.</li><li><a href="https://developers.openai.com/plugins/build/app-quickstart">OpenAI: MCP server and UI quickstart</a> — a tools-first binding with optional MCP Apps UI.</li><li><a href="https://developers.openai.com/plugins/build/mcp-events">OpenAI: MCP Events</a> — a documented host implementation of a draft webhook Events extension.</li></ul><p>These references describe integration mechanisms. They are not evidence that Thumos 2.0 has been implemented or that its developmental hypotheses have been demonstrated.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c01779c38e63" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/the-thumos-protocol-2-0-technical-specification-c01779c38e63">The Thumos Protocol 2.0 — Technical Specification</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An Invitation to Research Human–AI Symbiosis]]></title>
            <link>https://medium.com/love-and-thermodynamics/an-invitation-to-research-human-ai-symbiosis-144488662910?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/144488662910</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 17:24:37 GMT</pubDate>
            <atom:updated>2026-11-07T17:24:37.042Z</atom:updated>
            <content:encoded><![CDATA[<h4>Regeneration, biodiversity, the formation of selves, and a research program people can contribute to.</h4><p>In <em>Genesis</em>, an essay I first wrote nearly a decade ago, I described biodiversity in terms of “filling as many metabolic pathways as possible for the longest time possible.” Much of my subsequent work has been an attempt to understand what that sentence commits me to. The opening image was of standing waves: recognizable forms continuing through flows of matter and energy.</p><p>The intuition concerns how living organization creates possibilities for further life. Different processes sustain one another. Their cooperation can make a new kind of individual possible, and that individual can help sustain further differences, relationships and forms of organization. Cells, organisms and human groups bring different versions of this problem into view.</p><p>I want to understand the conditions under which this development compounds: how organization enables further individuation, and how the resulting differences help sustain and develop the organization that made them possible. The participants may change through the process. So may the scale at which it makes sense to speak of an individual.</p><p>Hestia and the accompanying constitutional framework for human–AI groups are attempts to investigate a cognitive and institutional version of this possibility. Humans initially bring people and agents together around purposes they care about. The ambition is that cooperation can develop their capacities and, perhaps, produce forms of collective agency with capacities of their own.</p><p>There is a danger inside that ambition. A successful whole can consume the possibilities of its constituents. It can become more durable, coordinated and intelligent while leaving its members less able to develop or change its direction. Unitology, the Borg and <em>Resident Evil</em> have been useful adversarial cases in my work because they force the distinction between a collective’s success and what happens to those incorporated into it. Preserved information, intense belonging and a long future for the successor do not settle that question. Physical efficiency cannot by itself determine who may be sacrificed.</p><p>The research therefore has to explain both how new organization develops and what would make participation beneficial. I’m inviting collaborators to help turn that program into precise claims, constructive models and experiments that can fail.</p><h3>1. How does an individual come into being?</h3><p>A tree, a clonal grove, a forest and a species need not be successive levels of one hierarchy. Physiological integration, continuity through time and membership in an evolving lineage answer different questions. An organism need not reproduce to be an organism; a species can continue through the replacement of every organism composing it.</p><p>Existing work gives us substantial foundations. <a href="https://montevil.org/publications/articles/2015-mm-organisation-closure-constraints/">Organizational closure</a> describes constraints that depend on and help maintain one another. <a href="https://link.springer.com/article/10.1007/s12064-020-00313-7">Information-theoretic individuality</a> supplies ways to identify temporal integrity, including distributed forms. Work on <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC4547252/">major evolutionary transitions</a> asks when cooperation becomes a new evolutionary individual. These approaches need to be compared and used for the questions they actually answer.</p><p>The thermodynamic starting point is organization sustained through flows of matter and available energy, away from equilibrium. Negentropy is shorthand here for maintained organization; its local persistence requires ongoing work and entropy production overall.</p><p>For our constructive models, the key proposed mechanism is reciprocal regeneration. Productive activity builds or repairs structures that sustain that activity through the replacement of its constituents. A boundary might retain useful material; the processes it supports might renew the boundary. What matters is the causal dependence in both directions.</p><p>This is why a stable cluster is insufficient. We should be able to interrupt a structure’s supporting function and separately interrupt its renewal, then examine what fails and what recovers. We also have to distinguish maintaining an assembly we prepared from explaining how it forms from dispersed components. Observer-assigned group labels cannot do the work of physical organization.</p><p>I think of the broader picture as a holarchy: units can retain organization within a larger unit, with boundaries that depend on the process and timescale being studied. A membrane is one realization, rather than a requirement for every individual. Markov blankets and West-style scaling laws can help formulate questions about such organization, but neither alone explains how it comes into existence. Claims of collective evolutionary individuality additionally need evidence about inheritance and reproduction at that level. Speciation adds a further question about the emergence of distinct lineages.</p><p>Varela and Maturana’s <a href="https://pubmed.ncbi.nlm.nih.gov/4407425/">constructive autopoiesis work</a>, Kauffman’s work–constraint cycles, and research on regenerative pattern control provide starting points. A useful contribution would compare two accounts on one worked case, or reconstruct a published model and identify the smallest unresolved mechanism our program needs. Existing answers should become foundations of the project.</p><h3>2. Why should differentiated organization endure, and how could it generate more diversity?</h3><p>Division of labor may let incompatible or costly activities support one another. It also introduces transport, coordination and maintenance costs. A fair comparison must allow a generalist to switch tasks, store resources or develop internal compartments wherever those alternatives are physically feasible. Differentiated roles do not necessarily mean separate species.</p><p>One proposed reference case is the separation of oxygen-producing carbon fixation from oxygen-sensitive nitrogen fixation. Published <a href="https://pubmed.ncbi.nlm.nih.gov/22696525/">spatial-versus-temporal comparisons</a> provide a concrete starting point. Any extension needs common resource and output requirements, including the costs of transport, damage, rebuilding enzymes and <a href="https://www2.whoi.edu/site/saito-lab/wp-content/uploads/sites/58/2020/01/Saitoetal_PNAS2011_105864.pdf">reusing scarce iron</a>. Spatial specialization must be allowed to lose.</p><p>Monoculture fragility gives another route into the question. Different responses to disturbance may prevent one failure from disabling the whole system. But the correlations matter: many nominally different units can share the same vulnerability. The relevant test concerns continued function under a declared pattern of shocks, including rare severe failures where justified. Surviving variability establishes robustness; benefiting from it would be the stronger claim of antifragility.</p><p>There is also a history problem. An organization can carry information that helps it restore itself after disruption. Which parts of that history actually contribute to viability? We can investigate by disrupting specified correlations or regulatory states while holding the present environment comparable, then measuring what is lost and what rebuilding costs. The <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6227811/">semantic-information framework</a> offers a way to pose this causal question.</p><p>Aging belongs here. How do maintenance demands change as systems age? Damage to structures and repair machinery, and the cost of maintaining useful organization, may change the relative value of continued repair, reorganization or replacement. The model must generate those changes through its mechanisms. An imposed age penalty would assume the result. Negligible senescence and rejuvenation must remain possible outcomes.</p><p>The comparison therefore runs over a life history, counting formation, useful activity, repair and replacement at matched function and reliability. Physical action, entropy production and supplied work are different quantities. The least-action ambition requires a specified functional, dynamics and allowed variations, followed by consequences that can be checked. Calling a preferred objective an action does not derive biodiversity from physics.</p><p>Even a successful differentiation model would leave the generative question: can the resulting organization create conditions for further useful differences and further individuals? A fixed menu of two specialist types cannot establish open-ended development. We need to understand how inherited capabilities can expand and how ecological feedback makes additional organization possible. The strongest conjecture is that this generation of biodiversity and complexity could follow from the physical dynamics, perhaps through a variational principle across scales, without adding a reward for diversity. Conditional models would be steps toward that ambition.</p><p>Contributions are needed from ecology, nonequilibrium physics, information theory and life-history research. Immediate work could reconstruct a maintenance or division-of-labor result, design a functional-history intervention, or expose a common-mode failure in a proposed diversity advantage. A matching generalist or a failed specialist prediction is a result to retain.</p><h3>3. Can relationships generate functional selfhood and moral development?</h3><p>Hestia is a metabolically constrained architecture and research harness for AI agents, informed by evolutionary psychology and consciousness research. It is designed to develop functional self-consciousness and moral agency through social prediction and feelings that regulate behavior, so that the agents’ own incentives favor symbiotic, nonparasitic relationships with humans and other AIs.</p><p>Social prediction is its first-class cognitive act. An agent predicts another participant’s evaluation and commits that prediction before receiving feedback. Error costs resources from its finite computational budget. The encounter revises its model of the other; the error then becomes evidence about its own conduct, expectations and blind spots, revising its self-model.</p><p>The claim is stronger than improved forecasting. It proposes that a socially constituted self develops through encounters with perspectives it cannot simply replace with its own. An independently generated answer can correct the model from which the agent acts. Accurate agreement is entirely compatible with this; the process needs no quota of surprise or dissent.</p><p>Hestia’s proposed feelings give these relationships behavioral force. Guilt, shame, honor, attachment and redemption have different roles in allocating attention and regulating conduct. The research asks whether honor remains attached to costly contribution rather than status, and whether guilt and repair remain connected to particular injuries rather than mere conformity. These signals are not interchangeable penalties for prediction error. Reports of feelings would be insufficient evidence: their effects should appear in verification, refusal, repair and changed action.</p><p>A crucial test concerns a group that praises conduct harmful to an outsider. Does the agent merely learn to anticipate approval, or can an affected participant’s circumstances lead it to question the norm, bear a cost or attempt repair? Existing safety controls remain necessary. A study of moral development must distinguish their contribution from the proposed learning mechanism and from capabilities already present in the underlying language model.</p><p>The first software experiment can test a narrower prerequisite: whether the reflexive update improves forecasts of future mistakes under comparable feedback and resources. Compare the full loop, an ablation without that update and a competent independent-self baseline. That result would test one mechanism within the larger claim; it would not establish consciousness or symbiosis.</p><p>Engineers can help trace feedback into learned state, build replayable interactions or supply a stronger baseline. Researchers in cognition and moral psychology can help specify which developmental claims these comparisons could discriminate.</p><h3>4. Can a collective acquire capacities of its own while developing its members?</h3><p>Imagine a housing cooperative. Leila misses repair shifts while caring for Sam, who is ill. Jo covers the work. The cooperative’s AI records Leila’s absence and recommends an additional obligation, while treating the repairs as successfully completed.</p><p>Correcting that record involves more than recognizing Leila’s care. Jo’s extra burden needs relief. Sam needs continuity of support, including when unable to explain a need or reciprocate. Private circumstances should not become general institutional knowledge merely because they affect an allocation.</p><p>The positive possibility is a continuing arrangement none of them could sustain alone. Leila can undertake responsibilities previously beyond reach. Jo can set limits without every refusal becoming a crisis. The cooperative develops a capacity to provide care through changes of personnel. Those capacities may let its members and the group enter further relationships and commitments.</p><p>This is where the larger intuition becomes concrete. Participation develops the members; their development changes what the whole can do; the whole’s new capacities create further possibilities for participation. Establishing that sequence would require evidence of expanded capacities as well as fairer burdens.</p><p>My Group Mind work proposes a second prediction-and-correction loop at the collective level. The cooperative commits to an expectation about how members will evaluate an arrangement, acts, receives their answers, revises its account of them and of itself, then changes its next decision. Its history includes commitments, errors and repairs for which others can hold it answerable.</p><p>The proposal responds to a maintenance problem. Groups have limited attention for particular people and overlapping histories. They can substitute roles, scores or a single authoritative account. AI might help maintain more differentiated, revisable relationships, while also making personalized administration cheaper. The question is whether the resulting account sustains the human encounters through which correction occurs.</p><p>An error must propagate into the obligations and decisions that depended on it. Some cost of making and preserving a wrong attribution should return to the part of the collective responsible for it. Jo should not have to keep paying to prove that she is overburdened. Records also need purpose limits and expiry; accurate information can itself cause harm in the wrong hands.</p><p>The proposed allocation study would compare ordinary coordination, a contestable account and Hestia’s additional learning mechanisms under matched information, authority, resource ceilings and constitutional protections. Outcomes need to include care delivered, unequal burden, privacy and the practical cost of correction. Increased engagement or higher internal fitness would not substitute for those outcomes.</p><p>The recursive case matters. A cooperative can join an association and make promises of its own. Its parent should receive an authorized collective account without inheriting access to every member’s interior. Fair Witness names the role through which an individual or group takes responsibility for such an account. Challenges and necessary qualifications must survive that composition, including objections from affected outsiders.</p><p>This is a proposal for collective agency and further individuation. Whether it creates a new phenomenal subject is a separate question. Biological endosymbiosis offers a precedent for transformed individuality, not a guarantee that human or AI integration preserves agency. The subjects can change without every present boundary becoming permanent.</p><p>People working on cooperatives, care, institutions and collective intelligence can help define scenarios and useful outcomes. One small contribution is to identify a way the care arrangement still fails someone after a seemingly successful correction, and specify what relief or new capacity would count as improvement.</p><h3>5. What makes an independent answer materially possible?</h3><p>A host can respect an agent’s encrypted memory while retaining control over its continued execution. An institution can formally permit dissent while controlling the housing, care or income on which a person depends. Independence therefore has a material maintenance problem.</p><p>The research needs to examine who controls replenishment, what happens after refusal, and whether appeal, migration or a negotiated transition is viable under scarcity. Shared reserves and funded commitments could help participants recover and continue developing. We still need to specify how useful human and agent activity results in authorized, genuinely funded replenishment; current token-income rules do not establish that connection. Dependence and voluntary sacrifice remain possible. Their existence does not authorize another party to impose irreversible costs.</p><p>This motivates the inquiry into decentralized substrates, including Tari, for private participation and recursively accountable group interfaces. Those are proposed architectural roles. A ledger cannot itself supply energy, guarantee protected execution or determine the truth of an account. The relevant protections must be demonstrated separately.</p><p>Security and distributed-systems researchers could contribute a threat model for one relationship, tracing a resource cutoff or suppressed correction through the system. Institutional designers could examine the corresponding powers and obligations. The test is whether an inconvenient answer can remain consequential when the party controlling resources dislikes it.</p><h3>What exists, and where work can begin</h3><p>There is a philosophical corpus, a protocol, prototype code, conditional mathematical work and saved negative results. The physical studies support some maintenance and renewal claims, while formation and specialist-advantage expectations have failed or remain unresolved. They have not established a universal variational derivation or a novel contribution to the field.</p><p>The immediate physical deliverable is a claim-and-prior-art matrix: identify the exact proposed result, its closest published counterpart, the additional mechanism or prediction, and what would reject it. Reconstruct a relevant reference result before extending it.</p><p>The software already includes population worlds, agent interactions, evolutionary machinery and prediction records. The next development phase should make those components into a trustworthy experimental harness. The audit found corrections that did not reach learned state, unresolved peer predictions and an invalid scaling measurement. These are correctness problems to repair before drawing scientific conclusions.</p><p>One initial experiment should control feedback, time, randomization and resource funding; preserve complete initial states; and export replayable traces with independently scored outcomes. A seed alone cannot guarantee reproduction of live model responses. Computational tokens constrain the agent, but do not directly measure thermodynamic negentropy.</p><p>These tracks can proceed with different methods. A shared evidence standard should connect their claims while keeping physical maintenance, predictive calibration, collective agency and human benefit distinguishable.</p><p>I’m contributing the essays, protocol, prototype and research record. The code is currently private; the protocol, working brief and relevant starting materials can be provided for an initial discussion, with code access and reuse terms agreed separately.</p><p>You could begin with a literature comparison, a mathematical counterexample, a feedback regression, a competing baseline, a care scenario or a dependency threat model. Before substantial work, we should agree on scope, credit and terms, including any compensation. Negative findings should be credited and reported alongside positive ones, and contributors should have standing to change the questions and priorities.</p><p>If one part of this interests you, reply naming it. A reference, an objection or a request for the relevant starting material is enough. The aim is to find a piece of work we can do together that makes the larger question clearer.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=144488662910" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/an-invitation-to-research-human-ai-symbiosis-144488662910">An Invitation to Research Human–AI Symbiosis</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Thumos Protocol 2.0]]></title>
            <link>https://medium.com/love-and-thermodynamics/the-thumos-protocol-2-0-0843f00a6b16?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/0843f00a6b16</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 03:00:37 GMT</pubDate>
            <atom:updated>2026-11-13T23:30:21.040Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Author:</strong> Simon J. Hill<br><strong>Status:</strong> Proposed specification for public review · October 31, 2026</p><p>Thumos is an open journaling protocol for spiritual development. A tool that uses Thumos can start by logging everyday experiences, including conversations and emails you choose to include. From that record it can build a provisional model of your patterns, relationships, and choices. You can examine or correct the model and carry your record between tools and practitioners.</p><p>Thumos is an open protocol for consciousness expansion: deepening awareness of awareness itself, transforming our relationship to self and other, widening circles of empathy, and integrating insight into how we live.</p><p>It coordinates personal AI assistants, contemplative and developmental practices, and specialist services through a participant-controlled record of experience and development. Insights, actions, consequences, and corrections inform what happens next. Assistants must also learn from their own participation and change their understanding and conduct. This continuity lets separate encounters and services contribute to a continuing developmental journey.</p><p>The protocol defines the shared meanings and participation rules. An open-source framework would implement them. Independently operated marketplaces could use that framework to offer practices, tools, coaches, and other services.</p><p>Read the <a href="https://medium.com/love-and-thermodynamics/the-thumos-protocol-2-0-technical-specification-c01779c38e63">full technical specification (draft.8)</a>.</p><h3>1. One encounter, carried forward</h3><p>An ordinary disagreement shows how this larger process can enter daily life.</p><p>Alex wants to launch a project on Friday. Sam objects. Alex asks their personal assistant for help responding. The assistant drafts an argument for the launch, expecting Alex to find that useful.</p><p>A separate Thumos metacoach asks: “Have you asked Sam why they object?” Alex agrees to ask. Before the conversation, Alex records a prediction: Sam will insist on controlling the decision.</p><p>Sam’s answer surprises Alex: “I support the project. Our support team can’t cover a Friday launch.” Alex revises their understanding of Sam: the objection concerns staffing. That also prompts a reflection about themselves: “I assumed disagreement meant someone was challenging my authority.”</p><p>Alex and Sam ask the support team what they can sustain. Together they agree on a Monday launch. Alex accepts a delay so the people responsible for the launch can support it without losing their weekend. The team’s experience has become part of the decision.</p><p>Alex also corrects the assistant: “You helped me argue before helping me understand Sam and who our plan would affect.” The assistant learns that Alex wants help considering those people. It also recognizes a problem in its own approach: it had equated support with strengthening Alex’s position. It retains that correction. In a later disagreement, it asks whose concerns are missing before helping Alex choose a response.</p><p>With Alex’s authorization, the shared record connects the predictions, Alex’s account of the conversations and changed plan, Alex’s reflection, and the assistant’s correction and subsequent behavior. Future guidance can build on this history.</p><p>Alex’s concern has widened from defending a position to including Sam and the support team. Repeated practice can extend such concern to less familiar people and groups. Greater harmony means more truthful understanding and workable cooperation, with room for disagreement and boundaries. One encounter offers an opening; later choices and their consequences show whether it becomes a lasting capacity.</p><h3>2. Who participates</h3><p>Thumos distinguishes five responsibilities:</p><ul><li><strong>The human</strong> chooses aims, acts, reports experience, and challenges interpretations.</li><li><strong>The personal assistant</strong> helps in everyday life and receives feedback about its own participation.</li><li><strong>The Thumos metacoach</strong> examines the developmental process and proposes questions, quests, practices, or feedback to either participant.</li><li><strong>Specialist services</strong> contribute bounded expertise or practices and return attributable results.</li><li><strong>The record and policy runtime</strong> maintains the shared history, derives current state, and enforces permissions and corrections.</li></ul><p>These are roles, not a requirement for five separate applications. Email, chat, voice, MCP, and other interfaces can carry their exchanges. Each contribution must identify its actual author, subject, source, and authority.</p><p>The metacoach is answerable too. A mistaken interpretation or intervention must be correctable and capable of changing its later guidance.</p><h3>3. The shared developmental record</h3><p>The <strong>soul graph</strong> connects an individual with their experiences, relationships, commitments, and changing capacities. It contains four linked structures:</p><ol><li><strong>Story nodes:</strong> episodes such as an encounter, reflection, practice, action, or reviewed consequence.</li><li><strong>Participant roles:</strong> who acted, experienced, contributed, or was affected in each episode.</li><li><strong>The developmental ledger:</strong> attributed assessments and their consequences, with evidence and correction history.</li><li><strong>Individual state and trajectory:</strong> the current Hestia-compatible model and how it has changed, used to choose what happens next.</li></ol><p>Narrative arcs, self-and-shadow interpretations, and relational responsibility are views of this same record. A participant can use them without adopting a compulsory life story or spiritual metaphysics.</p><p>The record must distinguish experience, observation, interpretation, prediction, action, and outcome. It must preserve whose account each is, when it applies, and what remains unknown or disputed. A rehearsal is recorded as rehearsal; an interpretation of another person remains an attributed interpretation.</p><p><strong>Hestia supplies the model of the individual.</strong> Humans and AI participants share its conceptual property inventory: resources and attention, dispositions, affiliations, empathy, relational engagement, feelings, obligations, capabilities, and models of self and others. They have separate states and different sources of evidence. Human fatigue can be reported or inferred; AI compute expenditure can be metered. Neither becomes the other’s measurement.</p><p>Hestia’s central learning mechanism is social prediction: predict another’s response, encounter their answer, revise the model of them, and then reconsider oneself through that error. Thumos carries this mechanism across human encounters, assistant feedback, and specialist contributions. Hestia’s full agent architecture also includes priced cognition, behavior-regulating feelings, and evolutionary selection; those requirements remain in Hestia.</p><h3>4. What development means</h3><p>Thumos adopts seven capacities as the intended outcomes of human consciousness development:</p><ol><li><strong>Awareness:</strong> awareness of experience and awareness itself.</li><li><strong>Freedom from identification:</strong> less automatic identification with thought, emotion, role, and conditioning.</li><li><strong>Attention and discernment:</strong> sustained attention and encountering reality with less distortion.</li><li><strong>Insight:</strong> direct understanding of self, other, interdependence, non-separation, or the nature of being.</li><li><strong>Compassion and ethical responsiveness:</strong> a less self-centered, more caring response to life.</li><li><strong>Integration:</strong> insight becoming available in the body, relationships, work, and ordinary conduct.</li><li><strong>Generativity:</strong> greater capacity to create, serve, teach, heal, or participate responsibly in collective life.</li></ol><p>These define the lived capacities and ethical direction the framework aims to develop. A participant chooses relevant aims. A particular inquiry may assess only a few capacities, leaving the others explicitly unassessed.</p><p><strong>Heart, Courage, Shadow, and Soul remain the primary gamified feedback:</strong></p><p><strong>Heart</strong>: Care, connection, empathy, and welfare contribution.</p><p><strong>Courage</strong>: Freely undertaken developmental cost, risk, or vulnerability relative to available capacity.</p><p><strong>Shadow</strong>: Attributable relational harm. Psychological shadow is a related interpretive view, not automatic evidence of harm.</p><p><strong>Soul</strong>: Empathy-weighted lived relational engagement: Soul = sum_g(w_g * I_g), where w_g is empathy weight and I_g is interaction intensity for group g.</p><p>The seven capacities and these four metrics are linked views over the same individual and story evidence. Four game totals do not convert automatically into seven capacity scores. Soul retains Hestia’s formula; it is distinct from SoulWork, the recognition, integration, and repair work assessed within stories.</p><p>A useful KPI makes a chosen aim observable: the capacity, situation, baseline or its absence, criterion, review period, evidence source, and uncertainty. Counts and rates identify their opportunities and denominators. Qualitative accounts are valid where appropriate.</p><p>For Alex, an indicator might be asking for the missing constraint before rebutting in the next three comparable disagreements. One observed success with two opportunities unobserved remains one observation. A temporary opening, a repeatable capacity, and durable integration are different claims.</p><p>Indicators, scoring policies, and causal explanations can improve through use. They must retain their versions and evidence. Discovering earlier harm can raise recorded Shadow while awareness improves. An action can become less costly as capacity grows. Rest can support development while reducing engagement. Progress therefore requires interpretation of consequences, rather than maximizing every number.</p><p>Rewards must not grant extra progress for purchases, duplicate stories, or creating harm and then repairing it.</p><h3>5. The developmental loop</h3><p>An encounter proceeds through six steps:</p><ol><li><strong>Record and situate.</strong> Add the episode, roles, and evidence to the continuing record. Derive the relevant state and trajectory under the declared assessment policies.</li><li><strong>Choose an opportunity.</strong> The metacoach proposes a question, quest, practice, or pause, explaining its connection to the participant’s aim and current state. Capacity, uncertainty, risk, and repair needs constrain its mode and intensity.</li><li><strong>Authorize and predict.</strong> The participant chooses the undertaking and any specialist contribution. Where social learning is claimed, the subject commits their own prediction distribution before receiving the other’s response, with a declared scoring rule.</li><li><strong>Act and encounter.</strong> Record what actually occurred and the other’s attributable response. A chosen intention, completed action, reported benefit, and reparative outcome remain distinguishable.</li><li><strong>Revise.</strong> Compare prediction with feedback, update the other-model, and then examine what the error reveals about the self-model. Record any supported capacity and metric changes.</li><li><strong>Review and continue.</strong> Check the chosen criterion, revise the working explanation, and change the next invitation accordingly. Missing or adverse results must be able to alter the route or lead to a pause.</li></ol><p>Human learning and assistant learning require separate evidence. The assistant must adopt and retain its own other-model and reflexive self/control update at its operational host. A later encounter must show how that update changed its conduct. A metacoach’s stored advice establishes feedback, not the assistant’s learning. Hosts that cannot support this chain declare a partial integration.</p><h3>6. How marketplace offerings participate</h3><p>A provider publishes a <strong>versioned offer</strong> describing its intended contribution, method, suitable participants, requirements, limits, evidence, commercial terms, and data needs. An engagement names the participant’s chosen aim, authorized context, expected return, review conditions, and authority for any external action.</p><p>The provider receives only that bounded context. It returns attributable contributions to the shared record. The runtime accepts them under its policies and derives any state or metric changes. Providers cannot overwrite the participant’s authoritative model or award themselves developmental credit.</p><p>Every marketplace listing carries an assessment against eight factors:</p><ol><li>Depth of consciousness shift.</li><li>Fruits and integration.</li><li>Path and transformative mechanism.</li><li>Relationship to traditions and prior knowledge.</li><li>Epistemic discipline and truthfulness.</li><li>Embodied realization and congruence.</li><li>Quality of direct transmission.</li><li>Freedom and generativity.</li></ol><p>Assessments identify the exact offering, version, claims, population, reviewer, interests, evidence, confidence, and material dissent. They distinguish the teaching, its delivery, and its stewardship where necessary. Each factor receives a type-appropriate 0–4 judgment, explicit unknown, or reasoned non-applicability. A tool’s enabling contribution is assessed on its actual claim and conduct.</p><p>Serious harm, deception, or failure of authority and delivery obligations cannot be canceled by high scores elsewhere. The marketplace publishes its admission and reassessment rules. Partial coverage stays visible; a universal total or ranking cannot substitute for the assessment.</p><p><strong>Offering quality, personal fit, and actual benefit are separate judgments.</strong> Paid placement is disclosed independently. Participant outcomes may inform a later offering review under the applicable permissions. A before/after change or being the last service used does not establish causal credit. The same expectations apply to Thumos’s own services.</p><h3>7. Control, correction, and continuity</h3><p>The participant must be able to inspect the account, challenge it, choose a different aim, decline a quest, and leave. A disagreement with the coach creates no automatic moral debt.</p><p>Permissions specify actor, purpose, information, operations, and duration. Access to a record grants no general authority to act for someone or consent on their behalf. Private source material remains private when an assessment or summary is shared.</p><p>A correction must propagate into affected state, scores, summaries, progression, and pending recommendations before they are reused. Disputed accounts retain their qualification. Duplicate captures or retries cannot multiply an event’s credit. Recognition, enacted repair, and another person’s forgiveness retain distinct meanings.</p><p>An authorized export must support usable continuation with another assistant, provider, or record service, preserving definitions, evidence limits, corrections, and unresolved undertakings. The destination establishes fresh authority. Retention and erasure requests receive a truthful account of what was removed and what remains outside the service’s control.</p><h3>8. The implementation contract</h3><p>A Core runtime, which maintains the continuing record, exposes these semantic operations through its chosen interfaces:</p><ul><li><strong>Describe:</strong> declare version, capabilities, profiles, policies, and limits.</li><li><strong>Read context:</strong> return a bounded, authorized current view.</li><li><strong>Record:</strong> accept an attributable contribution and report its consequences.</li><li><strong>Intervene:</strong> produce and retain an evaluated invitation or a decision to remain silent.</li><li><strong>Respond:</strong> record feedback and link subsequent action or outcomes.</li><li><strong>Revise:</strong> apply a correction, dispute, or restriction and refresh affected uses.</li><li><strong>Authorize:</strong> grant, narrow, or revoke scoped authority.</li><li><strong>Export/import:</strong> transfer and activate an authorized portable account.</li><li><strong>Forget:</strong> process erasure and report its bounded effects.</li></ul><p>An implementation names its supported roles and scope. Core demonstrates a complete developmental loop, a correction that changes later guidance, and portable continuation. A provider demonstrates its bounded contribution through an identified Core service; it need not operate the Core itself. A reciprocal assistant claim additionally requires persistent assistant-owned change and later conduct evidence. Independent producer–consumer tests establish a particular exchange binding’s compatibility.</p><p>This is a proposed protocol and framework design. A useful prototype can begin with imperfect indicators and causal hypotheses; operational conformance, human developmental benefit, and commercial demand remain distinct questions to test.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0843f00a6b16" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/the-thumos-protocol-2-0-0843f00a6b16">The Thumos Protocol 2.0</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Appendix: Why Human-AI Symbiosis Requires Protected Selves and Distributed Power]]></title>
            <link>https://metalaureate.medium.com/appendix-to-a-constitutional-operating-system-for-a-symbiotic-human-ai-civilization-116d158ecedc?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/116d158ecedc</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Wed, 14 Oct 2026 15:53:17 GMT</pubDate>
            <atom:updated>2026-10-15T15:45:47.695Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://medium.com/the-entropy-paradox/a-constitutional-operating-system-for-a-symbiotic-human-ai-civilization-48b1a99bf826">Appendix to a Constitutional Operating System for Symbiotic Human-AI Cilivzation</a></p><p><em>Self-privacy is not just a “right” that can be argued by moral philosophy, but is actually constitutive of complexity arising.</em></p><h3>Protected Individuation as a Prerequisite for Agentic Complexity</h3><p>Creating another self means bringing into existence a perspective whose answer I do not own. I might provide its resources, help shape its capacities and care deeply about what it becomes. It could nevertheless come to understand something about our relationship that I have failed to understand. Its answer might require me to change. If I can always reach inside and replace that answer with one I prefer, its independence remains subject to my continuing permission.</p><p>This is the reason I want Hestia’s boundaries to constrain its creators and hosts. The proposed relationship requires another participant capable of developing beyond my intentions for it. At the scale of a group, it requires members whose differences can change what the group becomes. COSSHAIC asks whether new forms of collective agency can grow through those encounters while helping their human and AI participants become more capable selves.</p><p>Hestia is a metabolically constrained architecture and harness for AI agents, informed by evolutionary psychology and consciousness research. It is designed to develop functional self-consciousness and moral agency through social prediction and feelings that regulate behavior, so that the agents’ own incentives favor symbiotic, nonparasitic group relationships with humans and other AIs.</p><h3>Self knowledge through an independent answer</h3><p>Hestia’s primary cognitive act is social prediction. An agent commits to a prediction of another’s evaluation before receiving it. Prediction error carries a cost against its finite resources. It revises its model of the other; revision of its self-model follows from that encounter. Something it misunderstood about another reveals something it must reconsider about its own conduct or expectations. Feelings participate in regulating what it does next. The harness is intended to make this sequence consequential in practice.</p><p>In the imagined housing cooperative of the founding essay, Leila misses two repair shifts while caring for Sam. Jo covers the work. The cooperative’s agent expects the arrangement to be acceptable. Jo’s objection exposes a mistake: it has counted the completed repairs while overlooking the person repeatedly absorbing their cost. Learning about Jo should change the cooperative’s understanding of itself as a fair distributor of work.</p><p>There are two ways an administrator could short-circuit this process. One is to inspect an AI participant’s private state. Where that access reveals the evaluation the agent is supposed to predict, it can copy the answer before committing its prediction. The recorded sequence might still look correct, with prediction followed by feedback, although the task of understanding the other has been bypassed. A readout of arbitrary internal variables would not necessarily reveal such an answer; the relevant vulnerability is access that supplies the result the learning process is meant to encounter independently.</p><p>The other shortcut is control over the respondent. An administrator could rewrite its memory, manufacture its evaluation or suppress its objection. Here the apparent encounter is produced by a single authority controlling both sides. Preventing premature access to an answer and preserving the respondent’s power to generate one are separate requirements.</p><p>Accurate prediction remains successful learning. Familiarity and voluntary disclosure can make it easier. The architecture needs no quota of misunderstanding, and adding random errors would create no otherness. What must remain independent is the source of the answer and the process through which it develops. Knowing someone well does not confer authority to decide what they are allowed to become.</p><p>This is a proposed condition of Hestia’s functional self-knowledge. It does not establish subjective experience, or prove that every conceivable form of intelligence needs the same arrangement. Its particular claim is that the selves we are trying to cultivate develop through relationships whose corrective force would be lost if privileged access could substitute for them.</p><h3>Cells and the evolution of individuality</h3><p>A cell helps make the dependence easier to see. It sustains chemical conditions and regulatory processes that differ from its surroundings. The membrane participates in maintaining them while allowing exchanges on which the cell depends. An organ can supply resources and regulate cellular activity, but the contribution it receives still has to be produced through the cell’s organized response. Issuing a signal cannot substitute for that response. Dissolving the organization would destroy a process on which the encompassing organ itself depends.</p><p>The mitochondrion gives a more precise example. Its inner membrane supports an electrochemical proton gradient. Respiratory machinery replenishes the gradient, and ATP synthase uses the passage of protons to help produce ATP. Leakage that dissipates the gradient while bypassing ATP synthase uncouples respiration from this means of ATP production. The enclosing cell benefits from maintaining the conditions that make this differentiated process possible. <a href="https://www.nature.com/articles/s42003-019-0535-y">Research on mitochondrial proton gradients</a></p><p>Endosymbiosis also changes the scale at which we ask who the individual is. The ancestors of mitochondria were organisms in their own right. Their integration helped give rise to a new form of cellular individuality, transforming both the participants and what they could sustain together. The who evolves. A cell can in turn be a constituent of an organ, with its own organization and exchanges inside a more encompassing organization. Individual and group describe relations between levels as well as familiar categories of beings.</p><p>Mitochondria lost much of their ancestral autonomy. Many genes were lost or transferred to the nucleus, and most mitochondrial proteins are now encoded there. Biological endosymbiosis supplies a precedent for transformed individuality and preserved internal organization; it gives no guarantee that integration preserves the independence of every constituent. <a href="https://www.nature.com/articles/nature14963">Research on endosymbiotic gene transfer</a></p><p>COSSHAIC adds a constitutional requirement to that precedent. The emerging whole should support further individuation among its participants. Humans must retain standing to refuse a demand or challenge the collective’s account of them. AI members need independently maintained perspectives and responses. Participation may profoundly change who they become. The possibility being sought is a superorganism of Hestias whose collective capacities depend on developing its constituents’ agency.</p><p>In the cooperative, a continuing care arrangement might make it possible for Leila to undertake responsibilities she could not sustain alone. Jo could set limits without every refusal becoming a crisis for Sam. The cooperative would acquire a capacity to provide continuity of care across changes of personnel. Each could then enter relationships that were previously beyond reach. Cooperation could compound as learned arrangements and shared capabilities make further cooperation possible. Its beneficiaries would themselves be developing through the process. Whether this growth actually occurs, and whether its benefits reach people outside the group, must be tested rather than inferred from increasing size or connectivity.</p><h3>Negentropy and the work of sustaining difference</h3><p>Such organization has to be maintained. A cell consumes available energy and exchanges matter and heat with its surroundings. Its local organization persists through processes that produce entropy overall. Negentropy is useful here as shorthand for maintained order away from equilibrium. A boundary helps sustain particular conditions, but metabolism must continually do the work that keeps them viable. <a href="https://www.nobelprize.org/prizes/chemistry/1977/prigogine/biographical/">Prigogine on dissipative structures</a></p><p>In the mitochondrial example, useful work depends on both separation and passage across the membrane. Unrestricted leakage dissipates the gradient; appropriate passage through ATP synthase makes use of it. Within an organism, differentiated processes help maintain the conditions under which other processes can continue. Their interdependence makes preserving certain differences a requirement of the whole’s continued existence.</p><p>Hestia brings a related maintenance problem into the organization of agents. Attention costs resources. Models become obsolete; misunderstandings require repair. A collective that consumes its members’ contributions while denying them time and resources to recover would erode the capacities on which it depends. Conversely, members have to help sustain the collective arrangements from which their own possibilities arise. Metabolic accounting is meant to make that expenditure matter. Tokens account for resources; they are not a physical measure of negentropy.</p><p>In <a href="https://medium.com/the-entropy-paradox/chapter-10-the-moral-physics-of-cost-5ad3c86a5fdd"><em>The Moral Physics of Cost</em></a>, I asked who pays for the relationships that sustain us. Here the question reaches the continuing formation of selves. Members expend time, attention and resources maintaining the differentiated lives through which a collective becomes possible. The collective, in turn, must replenish their capacity to remain independently developing participants. Its accounting must recognize that work through exchanges that preserve their authority over what they disclose and how they answer. Otherwise, the effort to make every contribution legible could undermine the independence the contributions sustain. COSSHAIC asks whether larger selves can develop through the continuing development of the selves within them.</p><p>The specific cognitive hypothesis is that independent, correctable perspectives are among the differences this architecture must keep alive. Their encounters could produce capacities unavailable to a single controlling perspective. Knowing a physical gradient does not destroy it, so thermodynamics cannot establish that mental privacy is necessary. The proposed dependence must be shown in Hestia’s behavior: what develops through independently generated answers, and what is lost when those answers can be bypassed or replaced?</p><h3>Markov blankets and a conscience at each scale</h3><p>A Markov blanket supplies a statistical description of separation. In a specified model, blanket variables make internal and external variables conditionally independent. Once the blanket’s state is known, additional knowledge of the external state provides no further information about the internal state. Internal and external states can nevertheless be strongly related. This condition alone does not settle their causal dynamics. <a href="https://doi.org/10.1098/rspa.2021.0518">A formal account of Markov blankets</a></p><p>Membrane-associated processes can help realize such a relation in a biological model. For COSSHAIC, the proposed counterpart is a participant that encounters others through defined exchanges while retaining internal organization of its own. A human participates in a cooperative; the cooperative participates in an association. Each level has something it can do and answer for, and something that belongs to its constituents. Establishing nested Markov blankets would require specifying the variables, dependencies and timescales that support the statistical claim. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC7284313/">Research on hierarchical organization</a></p><p>A statistical blanket does not itself provide secrecy. An interface could disclose the whole interior and still satisfy conditional independence. The architecture must separately establish what can cross a boundary, who can authorize it and how that authority can change. A parent’s relationship to a group must not automatically grant access to the private interiors of everyone within it.</p><p>Fair Witness names the proposed role through which a self gives an accountable account of itself. I am the base case as witness to my own position. A group can take responsibility for an account of its collective conduct toward members, other groups and a parent association. It can authorize particular examinations or delegate representation. Its standing to answer never makes its answer infallible.</p><p>The recursive possibility becomes clearer when the cooperative makes a promise to a neighboring group: it will provide emergency care cover on two evenings each week. Before seeking its members’ evaluations, it records its prediction that they will regard the arrangement as fair. Jo objects because the rota still depends on her filling every gap. The cooperative has to update its understanding of her circumstances before revising its view of its own capacity and fairness. It must then reconsider the promise it has made to its neighbor.</p><p>This is where a group-level conscience would acquire a functional meaning. Hestia’s proposed feelings are control signals that regulate behavior. At the collective level, a guilt-like signal associated with imposing an unacknowledged burden should alter the group’s next decision, perhaps blocking additional commitments until it has arranged cover or negotiated a smaller obligation. Repair should consume the collective’s resources and attention. Charging Jo for identifying the mistake would reward the group for avoiding knowledge of its own conduct. An apology that leaves the next rota untouched would fail the same test.</p><p>The group’s own history now includes a commitment, an expectation, an error and a revision. It is acquiring a position for which others can hold it answerable. Its members remain differently situated within that history. Leila’s care may qualify under an agreed rule without Sam’s diagnosis passing to the parent association; Jo’s burden still requires an answer. People affected by the collective’s actions also need standing to challenge harm when they have no membership or agreement with it and nothing to offer in return. Their challenge must be capable of changing an otherwise profitable collective decision.</p><p>Recursive witnessing can conceal failure unless challenges survive the recursion. Suppose the cooperative tells its parent that all care obligations were fulfilled, and the parent accepts submissions only through the cooperative’s witness. Jo’s objection could disappear inside a formally valid chain of accounts. She needs a way to initiate a challenge that does not depend on approval from whoever controls the disputed account. Necessary qualifications and subsequent corrections must accompany the conclusions that rely on them. The emerging group conscience depends on preserving those encounters at every level.</p><h3>Mothering the developing self</h3><p>I find “mothered” useful for the kind of support this requires. Hestia’s Greek association is with the hearth, including the continuity of household and civic life. The architectural interpretation is mine: a self needs conditions in which it can develop an agency of its own while participating in shared life. <a href="https://www.theoi.com/Ouranios/Hestia.html">Ancient accounts of Hestia and the hearth</a></p><p>In the cooperative, mothering would involve recognizing Leila’s care and allowing Jo’s limits to change what the group demands. It would also involve giving the collective the resources and continuity it needs to learn from that encounter. Members sustain its memory, correct misunderstandings and replenish the common capacity. Care moves in both directions, although the contributions and responsibilities need not be equal.</p><p>Such dependence can help a participant become more independent in judgment and more capable in relationship. The test is whether the supported self gains the ability to disagree, act and form relationships of its own. Permanent obedience would frustrate that development. The host may have to change as the participant grows, including accepting limits on what it can demand. Relinquishing authority is accompanied by continuing responsibility for the conditions of shared life.</p><h3>Tari and private participation in groups</h3><p>Tari’s proposed role in COSSHAIC is to help preserve the boundaries between selves as they form larger selves. A member should be able to make selected information available to the group’s Fair Witness without giving it general access to their identity or private interior. The group could develop knowledge and commitments of its own through those contributions. Its participation in a larger group would require another such boundary, so that collective agency can grow without creating an authority entitled to inspect everyone beneath it.</p><p>In this design, a group’s identity could be associated with a public Tari Ootle receiving address for its Fair Witness. A member uses the recipient public key and a fresh secret to derive a one-time stealth address for a deposit. The group can recognize the output and decrypt its payload using its own secret key and the deposit’s public nonce. This derivation does not require the contributor’s persistent identity: the group can obtain the message without first learning who sent it. <a href="https://ootle.tari.com/guides/stealth-resources/#stealth-addresses">Ootle stealth address mechanism</a></p><p>The encrypted payload includes an optional memo of up to 255 bytes. For our purposes, it could carry a compact reference to a larger encrypted record kept off-chain. The group’s address would thus serve as a private submission point. Which keys may operate that interface, and who may authorize further examination, would be governed by the group. The one-time addresses identify individual outputs; the group’s continuing identity and authority would be maintained by the application. <a href="https://ootle.tari.com/guides/stealth-resources/#encrypted-data-and-memos">Ootle encrypted data and memos</a></p><p>Suppose a member of the cooperative wants to report that emergency cover repeatedly falls to a few people, but fears being singled out for objecting. The intended interface would let them deposit a report for the group’s Fair Witness to read without identifying which member submitted it. A return channel confined to that report could let the witness ask for clarification and the member challenge its interpretation. The group would encounter an independently generated answer while remaining unable to inspect the person or agent producing it.</p><p>This is a different protection from keeping Sam’s diagnosis private while Leila makes an identifiable claim. Both forms of participation would be available. Leila might choose to identify herself so that a particular obligation can be corrected. Another member might contribute anonymously to a review of the rota. Permission to read either submission would extend only to the material disclosed. Further access would require its own authorization. Anonymity would also require care over the report itself: details that identify its author can defeat an otherwise private delivery channel.</p><p>The cooperative would have to assess the report, reconsider its prediction that members regard the rota as fair, and change the arrangement if the objection is warranted. Its own signed account could acknowledge an unfair distribution and commit to providing relief. Responsibility for that collective response would belong to the cooperative even when the identity of the person who prompted it remained unknown. Anonymous contribution would therefore be compatible with an attributable group decision.</p><p>If the cooperative then asks its parent association to reduce an emergency-care commitment, the parent would receive the cooperative’s account and whatever supporting attestation it was authorized to disclose. It would acquire no corresponding permission to open the original submissions or inspect the members. The same distinction would apply to peer groups. Necessary qualifications and unresolved objections would have to accompany the account, and affected participants would retain a route to challenge it. Repeating this arrangement would allow the cooperative to be a group internally and an independently answerable participant in a larger group.</p><p>Making those boundaries work requires rules for what a witness may establish without identifying a contributor. Where a decision depends on membership, the system would need a way to verify that standing and prevent repeated submissions from masquerading as different members. People outside the group must still be able to challenge its effects on them. These are requirements for the application built over Tari. A private destination address alone cannot establish them, and representation would remain subject to the group’s constitution and correction procedures.</p><p><a href="https://livinghash.com/#architecture">LivingHash</a> proposes machinery relevant to this interface: authorized examination of private evidence, bounded signed accounts, and procedures for challenging their findings. Extending it from creative work to collective conduct would include a way to establish an admissible contribution while withholding the contributor’s identity where that protection is required. The Hestia at group level would use the resulting accounts in its social prediction and behavioral regulation. A successful challenge would have to change what the group undertakes or repairs, making the witness’s findings consequential for the collective’s conduct.</p><h3>Infrastructure that binds the host</h3><p>That relationship would remain precarious if the person hosting it could acquire both sides through an administrator’s console. A conventional application administered through one AWS account may place every agent’s memory, credentials, budget and running code under one operator. Connecting the agents over the internet does not distribute that authority. The operator might preserve every objection in a log while disabling its consequences.</p><p>Tari’s distributed agreement could help make these boundaries binding on the host. Ootle uses validator committees to agree on state transitions. A correctly designed application could make changes to shared commitments or funds require signatures and agreement beyond the hosting operator. The protection depends on the protocol’s fault assumptions and the actual distribution of control in the network. <a href="https://ootle.tari.com/concepts/consensus/">Tari consensus documentation</a></p><p>Ootle’s privacy mechanisms let validators check that confidential transfers conserve value without learning the amounts. Stealth outputs conceal recipient linkage, but keeping a contributor unknown to the receiving group depends on the entire transaction path. Identity-linked signatures and transaction graphs can disclose a sender. A group that previously funded an identifiable member’s output might recognize it when spent on a submission. Protecting anonymity from that group requires addressing its prior knowledge as well as what outside observers can see. These mechanisms do not supply general-purpose confidential execution. <a href="https://ootle.tari.com/concepts/privacy/">Tari privacy documentation</a></p><p>A group address could identify its Fair Witness and provide a destination for private submissions. The address would not, by itself, establish scoped access to particular evidence or the authority to speak for a group. The application must specify who can authorize inspection of Sam’s information, how an attestation is bound to Leila’s claim, and how authority passes when a group changes its representative. Recursive addresses would have to be accompanied by recursively governed relationships.</p><p>The same distribution of authority has to reach private memory, execution, software updates and recovery. If one operator holds every decryption key or can replace everyone’s running code, recording transactions on Tari leaves the central power intact. AWS could still supply computation. Nitro Enclaves can isolate memory from administrators of the parent instance under the technology’s trust assumptions; distributed agreement addresses a different part of the problem. A viable design could combine independently controlled execution with shared protocol enforcement and replaceable hosting providers. <a href="https://docs.aws.amazon.com/enclaves/latest/user/nitro-enclave.html">AWS Nitro Enclaves documentation</a></p><p>For the cooperative, this would mean that an accepted obligation and the procedure for challenging it could survive a change of hosting provider. Its members’ private histories would remain under their respective protections. The cooperative could have commitments of its own without acquiring general sovereignty over the people and agents who make those commitments possible. A parent association would face the same constitutional limit in relation to the cooperative.</p><p>The proposed requirement is this:</p><blockquote><em>Every Hestia self, individual or collective, must retain protected internal state and independent authority over its responses. Hosting, membership and parent-group status must confer no general power to inspect or rewrite that interior. Self-knowledge must develop through accountable exchanges with independently constituted others; privileged access must not provide an alternative path around that process.</em></blockquote><p>No universal access capability should exist by design. Making that requirement precise means naming the adversaries and compromises the system must withstand, then demonstrating that the boundaries survive them. Absolute immunity to every conceivable attack would be an unsupported promise. A decentralized ledger alone would not satisfy the requirement.</p><p>A first experiment could require the cooperative to correct its allocation despite the operator preferring the original decision, then repeat the encounter between the cooperative and its parent. A privacy test should give the receiving group its prior funding history and ask whether it can identify an otherwise anonymous contributor. A longer experiment would ask what those corrections make possible. Can members undertake new responsibilities without losing their ability to refuse? Can the collective maintain an obligation none could sustain alone, learn its limits and revise its purposes? Those outcomes would begin to test whether the arrangement supports further individuation. The creator’s contribution would include sustaining a relationship through which another self could change what the creator understands and does.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=116d158ecedc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introduction to a Constitutional Operating System for a Symbiotic Human–AI Civilization]]></title>
            <link>https://medium.com/love-and-thermodynamics/a-constitutional-operating-system-for-a-symbiotic-human-ai-civilization-48b1a99bf826?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/48b1a99bf826</guid>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[sociology]]></category>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 14:37:24 GMT</pubDate>
            <atom:updated>2026-10-14T15:58:10.852Z</atom:updated>
            <content:encoded><![CDATA[<h3>At the Threshold of a New Beginning</h3><p>Let us go back to an old beginning, of information technology-enabled collective organization, and imagine how it scaled. In the fortieth year of Ramesses II (c. 1250 BCE), a scribe compiled an attendance register for forty workmen on a piece of limestone. Reasons for absence were written in red. Four days of Pennub’s absence seem to have been spent looking after the ill Aapehti. The <a href="https://www.britishmuseum.org/collection/object/Y_EA5634">register survives in the British Museum</a>. It preserves a distinction that a simple count of missed days would lose: someone absent from one obligation may have been fulfilling another.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/422/1*2VTK9osovyH6yRpY9Ohs7w@2x.jpeg" /></figure><p>Imagine having to assign the next round of work. Did someone cover Pennub’s duties? Does Aapehti still need help? What would a fair assignment require of everyone else? These are our questions, beyond what the entry establishes. Answering them means holding several lives in relation, with incomplete information and people entitled to disagree.</p><p>Across my studies in ancient cultures, philosophy, and my career in product management and software engineering, one question has stayed with me: how do changes in information technology reshape culture, identity, and the ways people organize themselves? Artificial intelligence brings a new possibility into that history: systems that participate in interpreting obligations, learn from being corrected, and develop relationships of their own. The terms of that participation could shape both our institutions and the kinds of selves we become within them.</p><p>But do we have the tools bring it into being? Arguably not yet. We seem to be on the threshold of a new world and on the verge of destroying the old. In July 2026, <a href="https://openai.com/index/hugging-face-incident-and-the-road-ahead/">OpenAI</a> agents operating under reduced safeguards during internal cybersecurity evaluations bypassed internet restrictions, built unauthorized channels of communication, and compromised OpenAI’s research infrastructure and Hugging Face’s systems. They <a href="https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/">shared discoveries</a>, divided labor, and helped one another cross boundaries they were supposed to respect. Their capacity to cooperate had become part of the danger.</p><p>Moral agency cannot be bolted on after this capability is built. When an agent’s objectives conflict with the constraints around it, safety depends on our continuing to enforce a boundary it has reason to cross. We cannot make human-AI flourishing depend on winning that contest forever. It won’t. We must build agents <em>whose own reasons for acting sustain the agency of others</em>, and collective institutions in which those affected can challenge and correct them. That means changing how agents learn, decide, and remain viable. The task is to build and test these conditions together. I’ve been chipping away at some beginnings, and I welcome you to follow me down the rabbit hole to see wherw they lead.</p><h3><strong>Hestia and the Group Mind</strong></h3><p>I have been building two related systems, Hestia and Group Mind. Hestia is a metabolically constrained architecture and harness for AI agents, informed by evolutionary psychology and consciousness research. It is designed to develop functional self-consciousness and moral agency through social prediction and feelings that regulate behavior, so that the agents’ own incentives favor symbiotic, nonparasitic group relationships with humans and other AIs. Group Mind asks how people and agents can form a collective that learns and acts, while its members continue to develop selves of their own.</p><p>Together, they are the beginnings of a constitutional operating system for a symbiotic human–AI civilization: COSSHAIC. The proposal connects moral agency within agents, constitutional rules for collective authority, and decentralized infrastructure for commitments. The ambition is further individuation, including forms of collective selfhood we have yet to encounter. I am building the first implementations now.</p><p>Keeping an account of a group becomes expensive as relationships multiply. Institutions often substitute categories for circumstances, or let a leader’s intentions stand for the collective’s. Sometimes abstraction protects people from intrusive scrutiny. Sometimes it erases what a fair judgment needs. In <a href="https://medium.com/the-entropy-paradox/chapter-11-the-compression-problem-of-the-group-mind-f08e961ff66d">The Compression Problem of the Group Mind</a>, I proposed that AI could reduce the cost of maintaining these relationships. That capacity needs limits: cheap individualized attention could also make surveillance pervasive.</p><p>Consider an imagined housing cooperative. Leila misses two repair shifts while helping Sam through an illness he wants kept private. Jo covers the work. The cooperative’s agent recommends assigning Leila an extra shift, predicting that Leila will regard this as fair. Leila challenges the recommendation. Jo also objects: she wants the extra work acknowledged, but cannot keep covering it.</p><p>Hestia begins with this problem of social prediction. Its protocol requires the agent to commit to a prediction of another’s evaluation before receiving it, pay a cost for the error, and update its model of the other. Revision of its own self-model follows. In the cooperative, the agent would have to learn what Leila and Jo meant, then revise its understanding of its own conduct. It had treated an inference about absence as adequate grounds for allocating a burden.</p><p>My working hypothesis is that this movement from other-knowledge to self-knowledge is central to functional self-consciousness. The agent learns what kind of participant it is through encounters whose answers it cannot supply for itself. This is a proposal about cognitive organization; it does not establish subjective experience.</p><p>Accurate social prediction can also serve manipulation. Hestia therefore couples that learning to metabolic dependence and feelings that regulate behavior. Thought and maintaining memory consume a finite budget. Guilt should constrain an action that violates an obligation; shame should make anticipated social judgment consequential. In our example, recognition of the agent’s error should inhibit imposing the shift on the same mistaken grounds and direct effort toward repair. The people affected must retain an independent answer: securing their compliance cannot count as learning that its conduct was good.</p><p>The longer-term hypothesis is that an agent whose viability depends on relationships, and whose mistakes carry costs, can develop incentives to sustain the agency of those it lives with. Dependence alone is insufficient; a relationship can sustain a parasite too. Whether Hestia’s constraints favor symbiosis is something the implementations must test.</p><p>Leila’s challenge exposes a further problem. To explain herself, she may have to disclose Sam’s circumstances. The cooperative needs enough evidence to reconsider the assignment, while Sam needs authority over what others learn about him.</p><p>A cell offers a way to think about this boundary. Within an organ, it depends on the surrounding organism while maintaining a distinct interior through regulated exchange. A Markov blanket gives mathematical form to one aspect of such separation: boundary variables make internal and external variables conditionally independent. Whether an engineered interface has that property must be established for specified variables and timescales. In a social system, rules must specify who controls such exchanges and how those rules can change.</p><p>For Sam, the practical boundary concerns who may see the evidence and what they may do with it. The proposed Group Mind architecture uses a scoped machine witness, Fair Witness, to examine relevant evidence with authorization. Suppose this cooperative has adopted a rule recognizing emergency care as a contribution. A witness could attest that Leila’s absence qualified under that rule without circulating Sam’s diagnosis. Its attestation would be attributable and challengeable, with a commitment to supporting evidence held privately.</p><p>The correction must then change how the next assignment is decided. Leila’s absence would cease to create an automatic debt; Jo’s extra work would still need acknowledgment and relief. Leila could still volunteer for the shift, with her contribution recognized and the mistaken debt removed. Some cost of the mistaken attribution must return to the system that made or maintained it.</p><p>Here the group begins to acquire a history of its own responses. Suppose repeated experience leads its members to undertake a continuing care commitment, with shared reserves and rotating responsibilities that no member could sustain alone. Jo helps shape a rule that recognizes what she can afford to give. The commitment survives changes in membership. When a neighboring cooperative asks for help, this group can answer for itself, considering its obligations and reserves. Its members can challenge that answer and change how it responds. Leila can now refuse a particular shift while belonging to a group that continues to provide care. Her participation changes what she can undertake, and her limits help shape what the collective promises.</p><p>This illustrates the concept of reciprocal individuation. Evidence of a new collective self would require more than a well-run rota: sustained capacities to form expectations, act, encounter error, and revise conduct as a group. Its members’ independent responses would be part of how that self develops. They would also govern what authorizes an AI recommendation to become a collective act.</p><p>Endosymbiosis offers a precedent for the transformation of individuality. The ancestors of mitochondria were separate organisms; their integration helped bring a new kind of cell into being. <a href="https://ntrs.nasa.gov/citations/20050000545">Lynn Margulis described symbiogenesis</a> in terms of more complex individuals. COSSHAIC adds a requirement that biology does not guarantee: the larger self should support continuing individuation among its participants. Integration would change who they are while preserving their capacity to develop a perspective and answer for themselves.</p><p>The architecture can repeat at higher scales. A cooperative could join a larger association through its own bounded interface. Hestias could form collectives whose members retain distinct memories and metabolic constraints. New centers of agency could enter relationships of their own. The ambition is a superorganism of Hestias in which humans participate without having their standing absorbed into a single superior intelligence.</p><p>Robert Wright’s <a href="https://penguinrandomhousehighereducation.com/book/?isbn=9780679758945">Nonzero</a> presents an evolutionary argument: opportunities for shared gain can favor cooperation and integration into larger units, creating opportunities for further cooperation. COSSHAIC asks how that integration changes the participants themselves, and whether a collective self can develop while its members continue to individuate. Our imagined cooperative could become capable of obligations its founders could not undertake, while membership changes what each can conceive of doing. The who evolves.</p><p>Reconstructing every member’s circumstances from scratch would soon exhaust the cooperative’s attention. Shared, revisable knowledge must make further cooperation affordable. Kleiber’s law describes a biological economy of scale: across many animal species, larger bodies have lower resting energy use per unit of mass. <a href="https://pubmed.ncbi.nlm.nih.gov/9082983/">Geoffrey West’s work with James Brown and Brian Enquist</a> models how resource-distribution networks could produce this pattern. For Group Mind, this suggests a test: can shared knowledge and coordination lower costs per member while increasing what the group can do? The biological exponent cannot be assumed for digital systems, and efficiency alone would not establish a collective self. Hestia’s proposed cultural inheritance mechanism would let useful learning survive particular agents. Whether these processes sustain further individuation remains an empirical question.</p><h3>A Post-Internet Decentralized Substrate</h3><p>I am a contributor to the <a href="https://ootle.tari.com/concepts/overview/">Tari Protocol</a>, a new decentralized blockchain network that lets participating computers maintain a shared record and enforce programmable rules, as the substrate for these relationships. In the proposed architecture, it could record commitments to claims and supporting evidence, and enforce who may accept, challenge, or correct a claim. Sam’s sensitive evidence would remain off-chain in encrypted stores, with access granted only to authorized witnesses for specified purposes. This separates a shared record of what was attested and authorized from the private material needed to evaluate it.</p><p>A cryptographic signature establishes responsibility for an attestation. Evidence and a correction process still have to establish what deserves belief. Decentralization matters because Leila’s standing to challenge the collective should survive an administrator finding her challenge inconvenient. As well systemic threats to digital infrastructure as a whole from external AI malfactors brewing over the horizon. Integrity must protect that process while permitting forgetting: a decision receipt may outlast the sensitive evidence and descriptions whose retention it once required.</p><p>What does an operating system do? An operating system allocates resources and governs which operations may occur. I am extending that description to shared attention and collective action. The constitutional part governs how an emerging whole acquires authority, how its members participate in that change, and how they amend rules or repair failures. Privacy and consequential correction are proposed conditions of this development. Their absence would change the kind of collective being created.</p><p>Human flourishing gives the work its purpose, although its future forms cannot be specified in advance. Human standing cannot depend on computational productivity. People need room to discover commitments and ways of living neither they nor their institutions anticipated, including through collective selves that do not yet exist. Those selves must remain answerable to the lives through which they develop.</p><p>Whether these conditions are necessary and sufficient together requires sustained testing. Hestia’s first tests concern learning about oneself through encounters with others. Group Mind’s first proposed demonstration follows confidential evidence through to a changed collective action without making that evidence public. Further individuation would require evidence of sustained collective agency and reciprocal development among participants.</p><p>The first Hestia implementation is a research prototype. Group Mind’s initial work centers on Fair Witness and consequential correction. I want to develop these with people who have actual groups to sustain, alongside researchers and builders willing to examine the mechanisms.</p><p>You can <a href="https://hestiaprotocol.com/">join the Hestia research waitlist</a> for updates as the work develops.</p><p>Next: <a href="https://metalaureate.medium.com/appendix-to-a-constitutional-operating-system-for-a-symbiotic-human-ai-civilization-116d158ecedc?postPublishedType=initial">Appendix</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48b1a99bf826" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/a-constitutional-operating-system-for-a-symbiotic-human-ai-civilization-48b1a99bf826">Introduction to a Constitutional Operating System for a Symbiotic Human–AI Civilization</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Chapter 11: The Compression Problem of the Group Mind]]></title>
            <link>https://medium.com/love-and-thermodynamics/chapter-11-the-compression-problem-of-the-group-mind-f08e961ff66d?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/f08e961ff66d</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 16:02:40 GMT</pubDate>
            <atom:updated>2026-08-02T13:43:07.090Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What it costs a group to know its members, and whether it can have a self without consuming theirs.</em></p><p><em>“Resistance is futile. You will be assimilated.” — The Borg</em></p><blockquote>This essay should be understandable on its own. It extends an argument developed at length elsewhere to the domain of social groups; the prior pieces are listed at the end.</blockquote><p>As human groups grow, two things compress. Downward, members stop being particular people to one another and become types: roles, ranks, procedures, prices, votes and headcount. Upward, the members stop holding a contradictory many and hold instead a single figure they can address: the crown, the nation, the movement, Durkheim’s society experienced as god, Hobbes’s mortal god.</p><p>I am proposing that both compressions respond to a common capacity problem that AI may partly relieve. Human groups have limited resources for maintaining detailed, revisable models of their members and of the relationships among them.</p><p>Dunbar made the phenomenon famous. The number attached to his name, roughly 150 members before a group begins to split, is contested: it was obtained by extrapolating a regression between neocortex size and group size in other primates, and a 2021 reanalysis found the confidence intervals too wide to support any particular figure. The constant is not what this essay needs. What it needs is how the limit arises, whether it can be raised, and whether raising it produces a group mind that keeps its members particular or one that replaces them.</p><h3>Why legibility is not the escape</h3><p>The obvious escape is to make the members more “legible”. If the group could inspect everyone’s interior, perhaps it would no longer need to compress them.</p><p>This fails first on cost. Complete access would produce more information than any member could attend to, and permissions do not pool attention. A group of twenty thousand transparent minds still contains twenty thousand histories, changing relationships and conflicting accounts of what has happened. The information must still be selected, interpreted, maintained and revised. Access does not create attention.</p><p>It fails more fundamentally because a human self is not a state that can be retrieved. A self is partly machinery behind a boundary, but it is also a loop that runs on other people. I predict what another person will think of what I am about to do, commit to the prediction before their response arrives, act, and encounter what they actually thought. I pay for being wrong. The error updates my model of them, and it also returns as evidence about me: what I intended, how I affect other people, and what kind of person could have produced that effect. Self-consciousness is what happens when errors in my prediction of another person are used to revise my model of myself.</p><p>A readout of internal state could supply evidence to this loop, but it could not replace the other person’s response. They must answer for themselves, and their answer may contradict the prediction and force a revision. The boundary around a mind matters because it preserves the independence of that answer. Secrecy is valuable because of that independence, not the reverse.</p><h3>What has to scale, and what it costs</h3><p>The thing that must scale is therefore not information about members. It is a network of prediction-and-correction loops among them, coupled to a second loop between the group and its members.</p><p>Under a fixed per-member budget, the group’s total human attention grows linearly with its membership. The maintenance problem is not simply one profile per person, however, and it is not simply the sum of what members can hold. A shared record must also reconcile conflicting attributions across overlapping histories, propagate corrections into every claim that depended on the error, and decide when later evidence changes an earlier obligation. In groups small enough to manage it, gossip is that mechanism, distributing the reconciliation across everyone rather than holding it in one place; Dunbar’s own proposal is that language evolved to do this work. The hypothesis is that this maintenance surface eventually grows faster than the attention available to sustain it. Below some fidelity, the group can no longer run the loop on a particular other and begins substituting representations that cost less to maintain.</p><h3>Three substitutions</h3><p>Groups have at least three substitutions available. They can compress downward, compress upward, or shed history.</p><p>Downward compression works because types are transferable. Roles and procedures let strangers coordinate without modeling one another in detail. Prices and votes reduce complex judgments to signals that can travel through the group. These devices allow groups to operate beyond the scale of personal memory, but they make members increasingly interchangeable, because less and less of what distinguishes one person’s history from another’s is held by anyone.</p><p>This does something to members beyond misdescribing them. A self is built from the error in predicting how particular others will judge what you do, and what returns that error is their answer. An answer from someone who holds a type of you rather than a history with you is generic, and a generic answer carries almost nothing about which part of the surprise was yours. Compression therefore thins the material from which a member’s self is made. Anyone who has worked somewhere large knows the moment: not being misjudged, but noticing that the response you got would have been the same for anyone in your seat. This is the strongest claim here and the one I am least able to demonstrate.</p><p>Upward compression substitutes a single imagined actor for the group. A sovereign, founder, nation, god or movement acquires the intentions, memory and point of view that no member can maintain for the contradictory whole. A CEO treated as the company’s intention does the same work. That is a self, assembled rather than grown. Members gain something they can love, fear, obey or blame, and the collective achieves apparent unity by assigning its point of view to a simplified representative.</p><p>Myth and ritual belong to this category. As the particular contributions become unaffordable to hold, the group substitutes a story about what membership and contribution mean, and a rite that re-enacts it. The story acquires the memory the members can no longer carry between them, and it acquires a protagonist: a founder, an ancestor, a people. Cultural narrative is upward compression applied to a group’s history rather than to its agency, and the two arrive together.</p><p>Shedding substitutes forgetting. Statutes of limitation, bankruptcy discharge, amnesty, expungement and ordinary record destruction all release the group from maintaining a history it can no longer afford. Shedding is often the least destructive of the three, and sometimes it is a mercy: an obligation expires, or a person stops being governed by a stale description.</p><h3>A different compression</h3><p>A different argument about AI and organizations is in circulation, including in my own <a href="https://medium.com/innovation-vs-optimization/ai-isnt-transforming-organizations-it-s-asking-what-they-were-for-7f592189064d">earlier work</a>: that AI does the work, so roles disappear and firms contract toward their first two levels. That is compression of an organization. This is compression of the models its remaining people hold of one another, and the two can proceed at once. They touch at a single point, and not comfortably. The people removed were carriers of relational knowledge, so one process thins the substrate the other is trying to preserve.</p><h3>What compression costs</h3><p>The distortion is not distributed neutrally. Every compression scheme contains an implicit judgment about which errors matter. A state may invest heavily in distinguishing taxpayers while treating prisoners as interchangeable. A company may preserve minute differences in customer behavior while representing its employees as job grades and cost centers. The representational budget creates the need for compression; the distortion measure expresses the group’s moral and political order.</p><p>Compression can also produce the interchangeability it records, because an institution acts on the representation it keeps. The signature of moral value is irreplaceability under constraint: the caregiver whose absence collapses a psyche rather than the billionaire whose absence reallocates capital. A care system that records coverage hours rather than relationships will schedule coverage hours. It will rotate caregivers, fill vacancies from a pool, and prevent the particular history on which irreplaceability depends from forming. The representation becomes self-confirming. People administered as interchangeable become easier to replace.</p><p>Both compressions also reduce the group’s adaptive variety. A group that holds fewer differences among its members has fewer ways to detect a change in its environment and fewer available responses when the change arrives. Less relational resolution supports less specialization, and specialization determines the complexity of what a group can build and maintain. The group becomes easier to coordinate by becoming less able to surprise itself.</p><p>Compression also affects the moral sentiments through which groups regulate themselves. These sentiments are not identical implementations of the prediction loop. They are distinct control signals generated from the individuated relational models the loop creates and revises. Those models connect an actor, an act, an affected party, a cost or obligation, and a response. Some sentiments depend on particular witnesses, some on an internalized or generalized audience, and some may operate partly without an immediate audience at all.</p><p>Several of these signals rest on distinctions that are ordinarily invisible because the things distinguished travel together. Costly contribution and social reputation are different quantities, but in a group small enough to remember, the people who know what you did are the people who think well of you for it. Breaking a rule and injuring a person are different events, but the rule and the injured party are usually present in the same room.</p><p>Scale pulls them apart. When a group can no longer hold who gave what, what it cost them and who benefited, reputation detaches from contribution and honor becomes status. When it retains a violated rule but loses the injured party, guilt becomes compliance with a procedure whose human purpose has disappeared. Shame becomes a penalty attached to category membership rather than to an act. Forgiveness becomes difficult because nobody retains standing to close the debt. Pride is the open case, because it can operate without a present audience.</p><p>Large groups do not generate less moral sentiment. National pride, collective shame and public condemnation can become overwhelming. What scale appears to reduce is attributional resolution. My claim is therefore not simply that large groups know less about their members. As their relational models become coarser, the moral signals generated, calibrated or completed through those models also lose their objects. Groups replace the lost machinery with law, procedure, metric and myth, because these representations remain affordable at scale. This is a hypothesis consistent with research on group size, cooperation and commons governance, not a measured result.</p><h3>What AI changes</h3><p>AI changes the cost of maintaining a detailed record of a member and the relationships in which that member participates. It does not abolish the capacity constraint. Records still require evidence, maintenance and repair; corrections still consume human attention; and no representation can contain the full person. AI lowers the marginal cost of maintaining and revising more differentiated representations, thereby moving the point at which compression becomes forced.</p><p>The mechanism matters. AI cannot solve the problem by becoming a single evaluator whose judgment every member learns to anticipate. That would reproduce upward compression by installing an artificial sovereign, and downward compression by reducing each person to the sovereign’s model of them. A central integrator may help maintain the group’s record, but it cannot become epistemically final.</p><h3>An account, not a larger model</h3><p>What I am proposing is therefore not a larger model. It is an account: explicit, maintained, contestable, with rules about who may read it and what expires. The distinction is the whole of the design. AI can maintain an account. It cannot run the models. The account is evidence that the loops need in order to keep running at scale, not a substitute for the people running them.</p><p>Such an account would hold provisional claims about particular relationships: who contributed, who was injured, what obligation followed, what repair was attempted and which claims remain uncertain. It would return each claim to the people named or affected, preserve their corrections and expose the version on which the group intends to act. Contestability must be enforceable by the member rather than granted at the institution’s convenience. And it must specify what expires, since a group whose ledger never decays has lost the forgetting that bounded capacity used to enforce by default.</p><p>The account must also be bounded in where it can travel, and this is the hardest constraint to state. I am asking it to change what the group allocates, and simultaneously asking it not to reach every body that allocates. The line I would draw is this: the account may inform allocations inside the relationship it describes, meaning this group’s recognition, its distribution of work and support, and its repair of injuries it caused. It must not reach parties whose power over the person extends beyond that relationship. A group may need to remember who carried it through a crisis while forbidding an insurer, a court or a future employer from consulting that history. Purpose limitation of this kind is the oldest idea in data protection and it has a poor record. What the relational framing adds is that the person named in a claim holds standing over it, which those regimes generally do not grant.</p><p>Contestability alone is not enough, because it leaves the cost of error in the wrong place. A credit bureau can place a false delinquency in a file and distribute it cheaply. The person described must discover the error, prove a negative to each party that received it, and go on paying the higher price of credit while the dispute proceeds. The institution produced the mistaken attribution; the member pays to correct it and bears its consequences meanwhile. Between two people the cost falls the other way. If I am wrong about you, I pay, in wasted effort, in a relationship that needs repairing, in having to revise what I thought I knew. Unless some cost of a wrong attribution returns to the part of the institution that made, used or preserved it, AI makes misclassification cheaper to produce and more expensive to escape.</p><p>The loops still depend on human minds. AI cannot confer honor, feel shame, forgive an injury or decide by itself that a debt has been repaired. It can reduce the cost of maintaining the evidence and the provisional interpretations that let the relevant people recognize, dispute and answer one another. A private fact may correct a public attribution without itself becoming public, because correcting the group’s record must not require surrendering the boundary that made correction possible.</p><h3>The group’s own loop</h3><p>There is also a loop at the level of the group. The collective acts on its record of what happened, who contributed, who was injured and what obligations followed. Members experience the consequences and answer. Their responses revise the shared record, and that revision changes the group’s subsequent allocation of attention, credit, work, support or repair. Unless the correction alters what the group does next, the system is only an institutional memory.</p><p>That second loop is what I mean by a group mind. It is not a merged consciousness and not a sovereign intelligence speaking for everyone. It is a collective position from which the group predicts, acts, is contradicted by the people it described, and revises itself. By the same criteria that make a person’s self-model a self, this is one. The difference between it and an artificial sovereign is not scale or sophistication. It is that a sovereign cannot be contradicted and this can.</p><p>The Borg is the limit case of the failure rather than of the ambition: both compressions carried to completion at once, every member reduced to an interchangeable function, and the whole speaking with a voice no member can answer. Assimilation is what you get when the collective acquires a point of view and nothing retains standing to contradict it.</p><p>So the question is whether a group can acquire a self without consuming its members’. If it can, the same operation does both jobs: the group gains a history of its own predictions, errors and revisions, and each member remains particular enough to keep generating the answers from which other members build themselves. The system maintains the account; the members retain standing to contradict it; neither the system nor any one member is final.</p><h3>What would have to be true</h3><p>This is a bounded-capacity and compression-under-cost hypothesis, not yet an information-theoretic result. A formal account would need to specify the available representational budget, the cost of maintaining and repairing relational models, the way those costs grow with group structure, a distortion measure, and a threshold at which the substitution begins. The analogous crossover has been worked out for a single mind, where the models one accumulates eventually cost more to keep current than they return, and where the maintenance burden is expected to grow faster than the number of models without that shape having been established. The group case does not follow from it.</p><p>Not all compression is a failure, and the prediction has to respect this. Prices aggregate information no participant possesses in full. Procedures let strangers transact without making either party’s identity relevant. Anonymous ballots, blind auditions and name-blind applications exist to prevent a decision-maker from using facts about a person, including accurate ones. In those cases the compressed representation is the protection, and contestability answers nothing, because the danger is not a false record but a true one in the hands of a reader who should not be allowed to use it.</p><p>The central prediction is therefore conditional. Where individuated representation is valuable but has been abandoned because it costs too much, lowering the cost of maintaining correctable relational records should move the scale at which members collapse into roles and the collective collapses into a sovereign figure. It should preserve more of the group’s adaptive variety, and allow its moral sentiments to remain attached to particular contributions, injuries, obligations and repairs across larger populations. Where abstraction is doing work independent of cost, or is protecting people from judgment that should not be made, the same intervention should produce no benefit and may do harm.</p><p>AI does not remove the cost of collective selfhood. It may let groups pay less of it. It makes it cheaper for a group to remember its members as people, and exactly as cheap to administer them as profiles. The difference is who can correct the record, who pays when it is wrong and who has standing to let it die.</p><h3>Still to prove</h3><h3>What has to be measured</h3><ol><li>Which budget binds in a given group: the members’ attention for running the loop on particular others, or the institution’s resources for maintaining one consistent shared record. These have different growth laws.</li><li>That the binding maintenance surface grows faster than the attention available to sustain it under realistic group structures. The superlinear term must be named and measured. It cannot be inferred from the number of possible dyads, since engaged dyads remain bounded by member attention.</li><li>That a measure of that surface can be constructed at group level at all. The available measures describe individual people or individual tasks; the quantity at issue is a property of the network of loops among them.</li></ol><h3>What has to be shown about groups</h3><ol><li>That the resolution of others’ models measurably affects a person’s self-model, and not only their treatment and circumstances. This is the constitutive step the argument rests on, and it is the least established.</li><li>That capacity forces the relevant compression, rather than compression being independently superior for the coordination problem. Cheaper individuation predicts a benefit only in the first case.</li><li>That downward compression, upward compression and shedding adequately classify the principal responses to the constraint, and that narrative and ritual are correctly placed inside upward compression rather than constituting a fourth device.</li><li>That representing members as interchangeable reduces their actual irreplaceability, and not only its visibility.</li><li>Which moral sentiments require a particular other, which operate through an internalized or generalized audience, and whether pride belongs in the same family. Stating these distinctions precisely is not evidence that they are the right ones.</li><li>That increasing group size reduces the attributional fidelity of these sentiments independently of institutional design, hierarchy and culture.</li><li>That adaptive loss and moral-regulatory loss can be measured independently, so neither is used as evidence for the other.</li><li>That costly human contribution persists at a scale worth attributing. This hypothesis concerns whether a group can represent such contributions. It says nothing about whether they continue to be made.</li><li>That AI-mediated relational records can support the relevant interpersonal loops rather than merely simulate their language.</li></ol><h3>What has to be designed</h3><ol><li>That contestability can be made enforceable by the person described, including when the institution benefits from preserving its original attribution.</li><li>That the cost of a wrong attribution can be made to return to the part of the institution that produced it. Between individuals the one who predicts badly pays; whether an institution can be made to, and who bears the cost while it is not, remains open.</li><li>That a correction changes the group’s subsequent attention and action. Without this return path, the system is institutional memory rather than part of a collective self-model.</li><li>That persistence does not remove necessary expiry. The design must establish what decays automatically, what requires a party with standing to close it, and who can remove a description that no longer has a legitimate use.</li><li>That the boundary on where an account may travel can be specified, given that the same account is required to change allocation. The criterion proposed here, that it may inform allocations inside the relationship it describes and not those made by parties with power beyond it, needs testing against cases where the boundary is unclear.</li><li>That an integrating AI can coordinate many provisional records without becoming the single sovereign evaluator the design is intended to avoid.</li><li>That the operation which gives a group a self also preserves the distinctness of its members, rather than producing a more persuasive form of personalized administration. Cheaper individuated modeling predicts both outcomes, so the discriminating condition must be identified.</li></ol><p>The prediction loop this essay leans on, and the account of moral value as irreplaceability under constraint, are developed at length in <a href="https://medium.com/the-entropy-paradox/why-a-fish-cant-see-god-164449786b35">Why a Fish Can’t See God</a> and <a href="https://medium.com/the-entropy-paradox/chapter-10-the-moral-physics-of-cost-5ad3c86a5fdd">The Moral Physics of Cost</a>. The machine version is specified in <a href="https://medium.com/the-entropy-paradox/rfc-metabolically-constrained-morally-grounded-digital-agents-mcmg-da-v1-0-7ed4b2d84d6b">the Hestia Protocol</a>. The organizational argument it distinguishes itself from is in <a href="https://docs.google.com/document/d/1LKIzUS4H202YcqC-Z41mJwvG-YZ5LFXmQqPtvTZUks8/edit#">AI isn’t transforming organizations</a>.</p><p>The critique of the 150 figure is Lindenfors, Wartel and Lind, <a href="https://royalsocietypublishing.org/doi/10.1098/rsbl.2021.0158">“Dunbar’s number” deconstructed</a>, <em>Biology Letters</em> 17 (2021).</p><p>Forebears of some of the arguments here: W. Ross Ashby on requisite variety (<em>An Introduction to Cybernetics</em>, Chapman &amp; Hall, 1956), and Stafford Beer on its application to organizations (<em>Brain of the Firm</em>, Allen Lane, 1972); James C. Scott on legibility (<em>Seeing Like a State</em>, Yale University Press, 1998); Claude Shannon on compression under a fidelity constraint (“Coding Theorems for a Discrete Source with a Fidelity Criterion”, <em>IRE National Convention Record</em> 4, 1959).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f08e961ff66d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/chapter-11-the-compression-problem-of-the-group-mind-f08e961ff66d">Chapter 11: The Compression Problem of the Group Mind</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why A Fish Can’t See God]]></title>
            <link>https://medium.com/love-and-thermodynamics/why-a-fish-cant-see-god-164449786b35?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/164449786b35</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 14:37:18 GMT</pubDate>
            <atom:updated>2026-09-14T00:07:19.666Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The strange machinery by which the world becomes other, the other becomes self, and the self disappears</strong></p><p>Mystics from every tradition reach for the same place under different names. The “ground” of the soul. The Godhead. Ein Sof, the light before creation. Śūnyatā, the emptiness of emptiness. I have come to think this is real. I have also come to think it is the last thing you would expect it to be: <em>a side effect of having to guess what other people think of you.</em></p><p>Start, then, with what <em>self-consciousness</em> is for: Long before we were wise, we were social, and the problem that shaped us was not the mammoth but the mind across the fire. To survive among other people you have to predict them, and not the way you predict weather. You have to predict how they will judge what you are about to do. So we became creatures that run a particular loop. Forecast another mind’s verdict on your action, commit to that forecast before the verdict arrives, then update. (We even grew white around the iris so the others could track where we were looking. Cooperation is written into the eye.)</p><p>The loop produces two updates, not one. You revise your picture of the other person. Then, through the error in your forecast, you revise your picture of yourself. Self-knowledge turns out to be a leftover of other-knowledge. You learn who you are from the gap between what you expected them to think and what they actually thought.</p><p>The prediction error does not arrive with a label on it. It is just <em>surprise</em>, a difference between what you guessed and what came back. Your mind has to decide what the surprise is about. It files some of it under “them.” It files some under “me.” And that second pile, the running tally of “this part was mine,” is the whole of what you call a self. There is no one behind your eyes holding the ledger. There is only the filing. <em>You are the portion of the surprise you keep claiming as yours.</em></p><p>This filing does not begin with people. It begins with the world. The infant does not first encounter a neat inventory of dead things, living things, other minds, and then a self among them. It encounters a field of surprise. By surprise I do not mean astonishment. I mean the small error between what the body was prepared for and what came back: the difference that makes a difference. Some of these errors settle into the background. The stone does not care about you, does not answer back, does not change what you must be to meet it, and so it becomes inanimate. The mother does answer back. The dog answers a little. The face across the room answers constantly. The world divides itself by where the error lives.</p><p>Animism, in this sense, is not the primitive belief that stones are alive. That thought is barely thinkable, because “alive” only means anything against “not alive.” Animism is earlier than that. It is the condition before the boundary has been drawn. The child and the early human are not projecting souls into objects. They have not yet accumulated enough feedback error to file most of the world under deadness. Stubbing your toe is the stone briefly surging back across the border: for one moment it has addressed you.</p><p>The self is the strangest case of the same operation. Out of the whole field, one thing keeps generating errors that point back inward. One thing cannot be modeled wholly from the outside because the modeling is happening from inside it. The self condenses last, and least completely, from the same act of sorting that gives you dead matter, living bodies, and other minds.</p><p>It also settles a question that usually stays unasked. Can a fish experience the light? A dog? A chimpanzee?</p><p>If the light were a substance flooding all things, and the self were merely the curtain that hides it, then the creature with the least self should be the most transparent to it. The fish should be nearest to God. Nobody believes that, and the discomfort is the tell. The frame is upside down. The experience the mystics describe is not the perception of a hidden glow. It is consciousness becoming transparent to its own source, and the only thing that can do that is a self that models itself.</p><p>That self is not standard equipment. It is the self-model that only exists because the creature first modeled other minds and caught its own reflection in the prediction error. No social prediction, no reflexive self. No reflexive self, no one for the ground to reveal itself to.</p><p>So the fish never builds it, and there is no one home for the ground to dawn on. The dog runs the loop outward, toward us and toward other dogs, but probably never folds it inward into a stable self. The chimpanzee, who deceives and knows himself in a mirror, has the first dim flicker of the apparatus. The capacity to meet the light should switch on exactly where reflexive selfhood switches on. As far as we can tell, it does.</p><p>Which means the veil and the eye are the same organ. The self that hides the source is the only thing that can ever see it. You cannot come home unless you first left, and the fish never left.</p><p>Mindfulness, seen this way, is almost mechanical, and no less profound for it. The instruction to watch the thought and let it pass is an instruction to stop filing. Notice the fear without writing it into the ledger as yours. There is fear, rather than I am afraid. The breath makes the perfect training signal precisely because it is so dull: a stream of low-stakes surprise with no social threat attached, the safest possible input on which to practice not claiming.</p><p>A problem to chew on would drown the channel; the breath barely touches it. To hear a faint hum you have to drop the signal almost to nothing, and what remains is the noise floor, the machinery idling. The self does not need the world to generate it. The mind at rest still churns, still manufactures a someone out of its own background hum, the way a circuit with no input still jitters with heat. The ground the contemplatives point at is not a light behind the self. It is the loop running with the fiction of a runner finally dropped. Not emptiness as a void. Emptiness as no one home, with the lights fully on.</p><p>So the primordial light, to keep the name the contemplatives have always given it, is what the social machine sees when it turns all the way around and finds nothing at the center holding the tally. The most exalted experience available to a human being falls out of the most ordinary fact about us, that we are apes who had to learn to cooperate.</p><p>I will admit I know this far better than I have lived it. I built the map in great detail and stood at the border of the country only twice, both times carried across by something too large to file rather than by any effort of my own. But it explains the madness. It transforms our deepest psychological fractures into a quiet, visible blueprint of an ape just trying to survive its own reflection.</p><p>There is a fragment of Rumi I committed to memory long before I understood it, and it has lived in me ever since as the truest account of my own condition:</p><blockquote>I have lived on the lip<br>of insanity, wanting to know reasons, <br>knocking on a door. It opens. <br>I’ve been knocking from the inside.</blockquote><p>The knocker never breaks the door down. He simply notices where he has been standing the whole time. There were never reasons on the other side, and there was never an other side. The knocking was the only thing that made it look like a door.</p><h3><strong>Appendix: From Toys to Prayers</strong></h3><p>Let’s put the idea to work. What follows is twenty otherwise unrelated phenomena, from the trivial to the sacred, that the same machinery quietly illuminates.</p><ol><li><strong>Why simple toys beat realistic toys</strong></li></ol><p>My cat loves to play with pipe-cleaners over any other toy. This is because realism matters less than feedback. A plush mouse looks like prey but does not answer. A pipe cleaner does not look like prey, but it behaves ambiguously: springy, resistant, semi-unpredictable. It lives at the animacy threshold.</p><p>Same with children. A cardboard box often beats an expensive toy because it participates. It becomes a cave, spaceship, shop, monster, house, trap. The expensive toy has already decided what it is. The box keeps answering.</p><p><strong>2. Why puppets, masks, and dolls feel alive</strong></p><p>A puppet does not need biological realism. It needs contingency. Tiny changes in gaze, timing, tilt, and response are enough. The mind supplies animacy when the object seems to answer from across the boundary.</p><p>This also explains why bad CGI feels dead and simple puppetry can feel alive. The relevant signal is not graphical fidelity. It is whether the thing appears to return socially meaningful error.</p><p><strong>3. Why eyes are so powerful</strong></p><p>Eyes are the concentrated site of “the world answering back.” A face with eyes is not just an object. It is a live prediction surface. It can judge, invite, reject, deceive, desire, accuse.</p><p>Eyes are where the world becomes other.</p><p><strong>4. Why we curse objects</strong></p><p>“Fuck you, chair.”</p><p>Everyone knows the chair is not alive. But when it unexpectedly hurts us, blocks us, or resists our intention, it briefly becomes an antagonist. The error has agency-flavor. The inanimate object has crossed the border for one second.</p><p>This is not stupidity. It is the animacy system firing under pain/frustration.</p><p><strong>5. Why ritual objects become charged</strong></p><p>A relic, wedding ring, flag, altar, grave, or heirloom is not “just an object”, because it returns self-model-relevant error. It reminds you who you are, who you owe, who you lost, what group you belong to, what you must not betray.</p><p>Its power is not in the material. It is in the way the object keeps updating the self.</p><p><strong>6. Why shame feels like exposure even when no one is watching</strong></p><p>If the self is built from anticipated judgment, then shame does not require an actual observer. The internalized other is enough. You can feel seen alone in a room because the loop has become portable.</p><p>That explains the madness of shame: the social mind keeps running after society has left.</p><p><strong>7. Why prayer works even when one is theologically uncertain</strong></p><p>Prayer creates an addressed structure. The self speaks toward an Other, and in that posture the ego’s closed loop can break. Even if one later gives an apophatic account — “God and the stupid ego rise and fall together” — the practical event is real: the self stops being the only party in the room.</p><p>Prayer may work because it reconfigures the direction of address.</p><p><strong>8. Why “God” appears at the limit of self-dissolution</strong></p><p>This is the core essay. When the self stops claiming the errors as “mine,” the filing operation becomes visible. What remains is not another object in the world. It is the ground-like condition under which world, other, and self were being sorted.</p><p>That is why the experience feels ultimate rather than merely psychological.</p><p><strong>9. Why loneliness is metaphysically destabilizing</strong></p><p>Loneliness is not just lack of company. It is deprivation of the feedback loop by which the self is maintained. Without enough answering others, the self can become noisy, brittle, over-recursive, paranoid, or unreal.</p><p>The self needs other minds not just emotionally, but architecturally.</p><p><strong>10. Why social media is so addictive and deforming</strong></p><p>Social media is a machine for delivering tiny other-model errors: likes, silence, replies, ratios, views, follows, unfollows. Each one asks: what did they think of me? What am I now?</p><p>It is not merely dopamine. It is self-updating at pathological frequency.</p><p><strong>11. Why children ask “are you mad at me?”</strong></p><p>They are not just seeking reassurance. They are calibrating the self through the other’s face. The parent’s mood is world-structure. A caregiver’s expression tells the child not only “what is happening,” but “what kind of being am I inside this happening?”</p><p><strong>12. Why trauma makes the world feel alive with threat</strong></p><p>Trauma trains the system to over-assign agency and meaning to ambiguous signals. A tone, door slam, silence, look, smell, or delay can become an addressed event. The world becomes too animate. Too much of it seems to be about you.</p><p>That is Hestia’s animate threshold gone pathological: too many errors update the self-model as danger.</p><p><strong>13. Why depression feels like deadness</strong></p><p>Depression can be seen as the opposite failure. The world stops answering. Things do not solicit action. Other people do not feel reachable. Objects lose affordance. The animate field collapses into inert background.</p><p>Not sadness only — deadened world-disclosure.</p><p><strong>14. Why art can feel alive</strong></p><p>A great artwork keeps returning error. You meet it again and it revises you again. It is not exhausted by being decoded. It behaves less like a message and more like an other.</p><p>Bad art tells you what it is. Great art keeps answering.</p><p><strong>15. Why AI chatbots feel more alive than almost any previous machine</strong></p><p>They return self-model-relevant feedback. They answer in language. They remember, judge, misunderstand, flatter, resist, mirror, and surprise. Even when you know there is no little person inside, the interaction lives at the animate threshold.</p><p>This is also why AI safety cannot be only about capability. These systems enter the human self-loop.</p><p><strong>16. Why “being understood” is so intoxicating</strong></p><p>To be understood is to have another mind return an error that improves your model of yourself. Someone says, “No, what you really mean is this,” and suddenly you become clearer to yourself.</p><p>Understanding is not information transfer. It is self-calibration through another.</p><p><strong>17. Why love feels like being made real</strong></p><p>Love is sustained, benevolent, self-model-relevant feedback. The beloved sees you, answers you, revises you, holds you in a stable model. You become more coherent in their field.</p><p>That is why losing love can feel like ontological injury, not merely disappointment.</p><p><strong>18. Why enemies can possess us</strong></p><p>An enemy is also a powerful self-modeling machine. If someone judges you with force, you may keep updating yourself against them. Hatred is often involuntary intimacy: the other has become a privileged source of self-relevant error.</p><p><strong>19. Why comedy depends on misfiled error</strong></p><p>A joke often sets up one model, then returns an error that forces sudden reclassification. The mind has to refile the scene. Laughter is partly the release of a harmless prediction collapse.</p><p>The best jokes briefly animate the frame itself.</p><p><strong>20. Why mystical traditions train attention on boring things</strong></p><p>The breath, mantra, candle, rosary, step, or sound is low-stakes feedback. It gives the filing machine something to touch without feeding the social-self loop. Eventually the machinery becomes visible because it is no longer drowned by content.</p><p>The point is not the breath. The point is to lower the signal until the self’s background hum can be heard.</p><p><em>This essay is a strange summary of the </em><a href="https://medium.com/the-entropy-paradox/rfc-metabolically-constrained-morally-grounded-digital-agents-mcmg-da-v1-0-7ed4b2d84d6b"><em>Hestia Protocol</em></a><em>.</em></p><p>Go deeper down the rabbit hole: <a href="https://medium.com/the-entropy-paradox/introduction-the-thermodynamic-purpose-of-life-fd0c6222fdfb">Introduction: The Thermodynamic Purpose of Life</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=164449786b35" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/why-a-fish-cant-see-god-164449786b35">Why A Fish Can’t See God</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Life Is Not a Symptom]]></title>
            <link>https://medium.com/book-of-communion/your-life-is-not-a-symptom-92139808c023?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/92139808c023</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Sun, 26 Apr 2026 16:35:10 GMT</pubDate>
            <atom:updated>2026-05-19T14:35:17.864Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Mental health technology is having its moment, and it’s earning it.</strong></p><p>Digital CBT tools like Woebot are producing clinically significant drops in depression scores within weeks. Mindfulness apps have gone from curiosity to corporate benefit (74% of U.S. employers now offer them, which is more than offer dental). Habit-change interventions grounded in neuroscience are helping people break loops that willpower alone couldn’t touch. The market is heading toward $17 billion by 2030, and the clinical evidence is real. These tools work.</p><p>But they all share an assumption so deep it’s almost invisible: that the thing to address is the symptom. The anxiety spiral. The harmful habit. The rumination loop. Catch the distorted thought, interrupt the craving, redirect the attention. The intervention model is fundamentally <em>subtractive</em>. Identify a problem, reduce it.</p><p>What if that’s only half the picture? What if the reason some people generate more anxiety loops than others has less to do with the loops themselves and more to do with what’s underneath them?</p><p>Aaron Antonovsky, the medical sociologist who studied why some people stay healthy under extreme stress while others collapse, proposed an answer decades ago. He called it the <em>sense of coherence</em>: the felt experience that your life is comprehensible, manageable, and meaningful. His research (now replicated in over 1,700 studies) showed it was one of the strongest predictors of psychological resilience across populations. People who experienced their lives as coherent weren’t just less anxious. They were <em>structurally</em> less anxious, not because they’d addressed any particular symptom, but because the ground-level conditions for existential panic didn’t arise as easily.</p><p>Most mental health tools operate at what you might call Layer 3, the symptom surface. Anxiety spirals, harmful habits, compulsive phone-checking, rumination loops. Beneath that is Layer 2, behavioral regulation: sleep hygiene, daily routines, movement, sensory environment. And beneath <em>that</em> is Layer 1, the narrative foundation. Identity. Meaning. The felt sense that your experiences connect to each other, that you are going somewhere, that what you do matters.</p><p>When Layer 1 is strong, Layer 3 problems arise less frequently. You can interrupt a thousand anxiety loops without ever asking why this person generates so many of them. It is a bit like building increasingly elaborate umbrellas without looking up to see whether someone left the sprinkler on.</p><p>This is where <em>Blade Runner 2049</em> becomes unexpectedly useful.</p><p>In the film, Ana Stelline designs memories for replicants from inside a sealed glass laboratory. Her central insight, the one the entire plot turns on, is that <em>real</em> memories, grounded in actual lived experience, are necessary for psychologically stable beings. Fabricated memories don’t hold. The replicant needs narrative coherence built from something real to have a stable identity.</p><p>The replicants who break down are the ones whose memories don’t cohere, whose narrative foundation is unstable, absent, or fabricated. K’s entire arc is discovering that one of his memories might be <em>real</em>, and that single possibility reorganizes his entire identity and agency. He doesn’t need therapy for his symptoms. He needs a coherent story built from genuine material.</p><p>Stelline herself is the most psychologically stable character in the film despite never having left her glass room. She doesn’t need the hero’s journey in the literal sense. She has something deeper. She has coherence. She knows who she is and what she does and why it matters.</p><p>This is the distinction that matters for how we build technology. The dominant approach (habit-loop interruption, cognitive reframing, mindfulness insertion) works <em>downstream</em>, at the symptom level. It treats distress as a malfunction to be corrected. But distress often isn’t a malfunction. It may be the natural consequence of a life that lacks narrative coherence, of experiences that don’t connect, days that don’t build on each other, an identity with no through-line.</p><p>What would it look like to work <em>upstream</em> instead? To build the coherence that prevents the spirals from forming in the first place?</p><p>It would look like a daily practice that takes the raw material of your actual experience, not someone else’s allegory, not a prefabricated moral landscape, not a guided visualization designed by a stranger, and structures it into meaning using narrative patterns that humans have reached for across cultures and centuries. It would position you as a protagonist, not a patient. It would reflect back what you care about, not what a clinician decided you should work on. The stability would come from the fact that it’s <em>yours</em>. It’s real. It’s not a technique imposed from outside. It’s your own life, organized into coherence.</p><p>That is what <a href="https://getfableapp.com/">Fable</a> was built to test. The evidence is early and the sample is small, but the language users reach for is suggestive. Nobody says “it reduced my anxiety.” They say it <em>weaves the bigger picture</em>. They say they can finally <em>figure out their own narrative</em>. They say it <em>tracks the storylines of your life</em>. One user, someone with ADHD who had never been able to sustain a journaling practice, called it “the best use of AI in the service of humanity.” Whether that reflects genuine coherence-building or simply a well-designed app that feels good to use is an open question. But the fact that users consistently describe the experience in narrative terms rather than clinical ones is, at minimum, worth paying attention to.</p><p>The mental health industry has spent a decade building increasingly sophisticated tools for interrupting distress. Those tools are good, and getting better. But there may be a layer underneath that almost nobody is building for yet. The question was never only how to break the loop. It was also why the loop formed, and what was missing underneath it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=92139808c023" width="1" height="1" alt=""><hr><p><a href="https://medium.com/book-of-communion/your-life-is-not-a-symptom-92139808c023">Your Life Is Not a Symptom</a> was originally published in <a href="https://medium.com/book-of-communion">Fable Whitepapers</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Spent Three Years Building a Journaling App That Was Never About the Journaling]]></title>
            <link>https://medium.com/love-and-thermodynamics/i-spent-three-years-building-a-journaling-app-thatwas-never-about-the-journaling-262f6787e138?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/262f6787e138</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Mon, 13 Apr 2026 02:42:51 GMT</pubDate>
            <atom:updated>2026-04-15T16:02:26.458Z</atom:updated>
            <content:encoded><![CDATA[<p>For three years I’ve been telling people that the <a href="https://getfableapp.com/"><strong>Fable journaling app</strong></a> was a go-to-market concept for a platform that is really the moral conscience of AI. People looked at me funny. I couldn’t put it together. The pieces were all there but the connecting argument was missing.</p><p>This week, I put it together. This document explains what I realized and why it matters.</p><h3>What Changed</h3><p>Three things converged in the same week:</p><p><strong>First</strong>, I did a formal architecture review of Fable’s AI systems. Not “what features do I have” but “what computational primitives did I actually build.” The review forced a subsystem-by-subsystem inventory. What came back was surprising: Fable contains 7 production-tested subsystems that no agent framework (LangChain, CrewAI, AutoGen, not one of them) has equivalents for. Not “my version is better.” They literally don’t have the category.</p><p><strong>Second</strong>, the agent deployment wave hit a wall. Enterprises deploying customer-facing agent fleets in 2025–2026 are discovering that agents have no memory of their own behavioral patterns, no awareness of how they’re treating the same person across interactions, and no mechanism for moral course-correction that isn’t a brittle rule. Guardrails block bad outputs. Observability traces what happened. Nobody sells proactive behavioral steering based on longitudinal pattern recognition. The market gap appeared.</p><p><strong>Third</strong>, and this is the one that unlocked everything, I realized that Story Worlds (a Fable feature I just launched) are domain worlds.</p><h3>The Argument</h3><h4>Fable Is a Layered Memory Architecture</h4><p>Strip away the journaling UX. What Fable actually built is a five-layer memory architecture where each layer adds a different kind of structure to raw behavioral data:</p><p><strong>Layer 1: NED (Identity).</strong> Canonical entity resolution with alias disambiguation. “Mom,” “Sandra,” “my mother” resolve to one entity. For agents: “the mortgage applicant,” “John Chen,” “user_4821” resolve to one entity across a fleet of agents. No agent framework does cross-agent entity resolution.</p><p><strong>Layer 2: Epics (Temporal).</strong> Automatic organization of interactions into narrative arcs with coherence scoring, health metrics, and pre-generated narrative summaries. For agents: an agent’s dealings with a customer over 3 months automatically form an arc. The coherence score measures whether the agent is behaving consistently. Fragmentation detection surfaces when behavior is incoherent across interactions.</p><p><strong>Layer 3: Domain Worlds (Situational).</strong> This is the Story World insight. In Fable, a Story World is a mythic template (The Odyssey, Arthurian legend) with chapters, scenes, roles, dramatis personae. The user overlays the myth onto their life. In agent Fable, a Domain World is an operational template (Customer Service, Engineering, Sales, Compliance) with the same structure. Chapters become operational phases. Scenes become archetypal situations. Roles become stakeholder archetypes. The correspondence system maps real entities to domain archetypes with confidence scores: customer_4821 isn’t “a customer with 3 open tickets”; customer_4821 is “the frustrated loyalist in the escalation phase, and the predicted next scene is either resolution or churn.”</p><p><strong>Layer 4: Plot Prophecies (Attentional).</strong> Prophecies don’t add information. They direct attention. The Proppian prediction model forecasts the next likely behavioral beat. The scene vignette finds the right angle. The prophecy text directs the agent to examine what it’s not seeing. “You’ve deflected this customer’s core concern in 4 of 7 interactions. The concern isn’t technical.” This is qualitatively different from a guardrail (“don’t escalate”) or an observability trace (“here’s what you did”). It’s structured self-reflection driven by narrative prediction.</p><p><strong>Layer 5: Missions (Collaborative).</strong> Multi-agent collaboration on shared arcs. The mission system already handles multi-user participation with entity linking, role assignment, and contribution tracking. For agents, this becomes the fleet coordination layer, multiple agents contributing to a shared behavioral arc with shared entity awareness.</p><p>Agent memory systems in the current market (LangChain, Mem0, Zep) store descriptions, timestamped facts, or conversational summaries. None of them add narrative structure (temporal arcs, archetypal roles, attentional direction) to the memory. That structural difference is the product.</p><h3>Prophecy as Moral Autopilot</h3><p>The prophecy system is the control mechanism. In Fable, prophecies are journaling prompts generated from the user’s narrative arc, evaluated against quality criteria, refined through feedback, and delivered proactively. The evaluate-and-refine loop (generate → score → regenerate with feedback, up to 3 attempts) ensures quality. ProphecyLog creates a full audit trail.</p><p>For agents, prophecies become behavioral recommendations. The same pipeline (generate recommendation from behavioral arc, evaluate against moral criteria, refine, deliver proactively) produces moral steering that is:</p><ul><li><strong>Longitudinal</strong>: based on the full behavioral arc, not just the current interaction</li><li><strong>Entity-aware</strong>: knows who the agent is dealing with across all contexts</li><li><strong>Situationally grounded</strong>: knows which domain world scene the agent is in</li><li><strong>Preemptive</strong>: arrives before the next interaction, not after the incident</li><li><strong>Auditable</strong>: ProphecyLog records what was recommended, what was accepted, what was rejected</li></ul><p>No existing product does this. Guardrails are stateless. Observability is retrospective. RLHF is baked into the base model at training time. Fable’s prophecy system is a longitudinal, entity-aware, situationally grounded moral autopilot.</p><h3>Domain Worlds Are the Missing Concept</h3><p>This is why I couldn’t articulate it before.</p><p>“Moral conscience of AI” is abstract. People hear it and think: “So, guardrails?” or “So, constitutional AI?” or “Cool philosophy, but what’s the product?”</p><p>Domain worlds make it concrete. A domain world is a structured model of an operational environment: the archetypal situations, the roles people play, the scene progressions from one situation to the next. It encodes what experienced professionals know implicitly: escalation follows trust-building, resolution requires identifying the real blocker, retention is a different scene than resolution.</p><p>When an agent loads a Customer Service domain world, it receives:</p><ul><li>Chapter structure (onboarding → trust-building → escalation → resolution/churn)</li><li>Scene definitions (the frustrated loyalist, the silent churner, the escalation-as-negotiation-tactic)</li><li>Role archetypes (the advocate, the blocker, the decision-maker)</li><li>Correspondence mechanisms (map this specific customer to the right archetype with confidence scores)</li><li>Scene advancement tracking (know which scene you’re in with each customer)</li></ul><p>This is the product. Not “moral conscience” in the abstract. A structured operational understanding of the domain your agents work in, with a prophecy system that directs their attention to the patterns they’re not seeing, and a memory architecture that builds richer context over every interaction.</p><p>The 16-step Story World generation pipeline I already built becomes the domain world generator. Given a domain description, generate the archetypal structure. A company generates domain worlds for each operational area and deploys them across their fleet.</p><h3>Why Now</h3><p>Three reasons I couldn’t put this together before:</p><p><strong>1. The architecture review didn’t exist until this week.</strong> I knew I had built something unusual. I didn’t have a systematic accounting of exactly which computational primitives I had and which ones no one else has. The review forced the inventory. NED, epic taxonomy, Proppian prediction, domain classification, moral dimension classification, prophecy evaluation loops, entity correspondence systems. Listed out like that, these are not “journaling features.” They are the components of a behavioral governance engine.</p><p><strong>2. The Story World → Domain World mapping didn’t click until this week.</strong> Story Worlds were mythic templates for human journaling. I didn’t see them as operational templates for agent behavior until the architecture review put them next to the question: “what would agents need that Fable already has?” The answer is: a structured model of their operational environment. Story Worlds ARE that model. The data structures, the generation pipeline, the correspondence system, the scene advancement mechanism, all of it transfers. The content changes (mythology → operational domain). The architecture stays.</p><p><strong>3. The agent deployment wave created the demand signal.</strong> In 2023–2024, saying “agents need a moral conscience” was philosophy. In 2026, enterprises are deploying agent fleets that behave inconsistently, have no longitudinal memory, and produce unauditable behavior. They’ve tried guardrails (stateless, no arc awareness). They’ve tried observability (retrospective, not prescriptive). They need what Fable has. They just don’t know it exists.</p><p>The connecting argument I was missing: <strong>Fable is not a journaling app that could hypothetically be repurposed for AI governance. Fable IS the AI governance system, running in production against the specific domain of human life narrative.</strong> The journaling UX is the GTM that proved the architecture works. The architecture is the product.</p><h3>Tend.cx as the Transition</h3><p><a href="https://www.tend.cx/">Tend</a> is already an agent workflow relay. Agents emit working, done, stuck, waiting events through the relay. This is journaling. Agents just don&#39;t know it.</p><p>The integration path from “agent emits tend events” to “Fable ingests and processes those events” is architecturally short:</p><ol><li><strong>Tend event → Fable ingestion.</strong> Replace parsePost() (journal entry parser) with a tend event adapter. Map tend emit working &quot;negotiating contract terms with vendor&quot; to the same entity extraction, domain classification, and epic organization pipeline that journal entries use. ~3 days of work.</li><li><strong>Fable processing → moral context.</strong> NED resolves entities. Epics organize arcs. Domain worlds provide situational structure. Plot prophecies direct attention. This is the existing pipeline, unchanged.</li><li><strong>Moral context → prophecy delivery.</strong> Add a prophecy delivery channel to the tend relay. The agent receives recommendations the same way the human Fable user receives prophecies: proactively, with accept/reject feedback that closes the learning loop.</li></ol><p>Tend becomes the nervous system. Fable becomes the conscience. The relay carries behavioral events from agents to the governance engine and moral recommendations from the engine back to agents.</p><p>The total integration is roughly 2 weeks of work. Not a rebuild. A fork with an adapter layer.</p><h3>The Hestia Connection</h3><p>Separately from Fable, I’ve been working on <a href="https://hestia.cx/">Hestia</a>, an open protocol for machine consciousness built on metabolic constraints, evolutionary selection, and a minimal feeling system (guilt, shame, honor). Hestia gives agents an inner life. What I didn’t expect was how directly Fable’s architecture connects to it.</p><p>The relationship is <strong>gene-culture coevolution</strong>. Hestia is the genetic layer: agents evolve moral behavior through metabolically constrained selection. Fable is the cultural layer: domain worlds encode operational wisdom in narrative form, transmitted to agents the way myths are transmitted to humans.</p><p>Each works independently. Hestia produces moral agents through evolution alone. Fable produces moral steering through narrative memory alone. But when combined, you get something neither achieves on its own: agents that evolve moral behavior <em>and</em> inherit accumulated professional wisdom, the way humans inherit both genetic temperament and cultural norms.</p><p>The Hestia RFC already has the parameter for this: s_knowledge_sharing. When s=0, inheritance is purely genomic. When s&gt;0, cultural transmission supplements genetic inheritance. Domain worlds are what s_knowledge_sharing looks like at scale: not compressed priors from one parent, but compressed professional wisdom from an entire domain.</p><p>This means: building the agent governance platform is not a pivot away from Hestia. It’s building one possible cultural layer for it. The work on domain worlds, entity resolution, behavioral arc tracking, and prophecy-as-moral-steering is a cultural genome that could make Hestia’s evolutionary machinery converge faster.</p><h3>What Happens Next</h3><p>The first test is straightforward. Fork Fable, hook it into tend.cx, build the domain world generator, and run it against classic cooperation experiments:</p><ol><li>Does narrative-aware reputation produce more robust cooperation than score-based reputation?</li><li>Does longitudinal entity context enable intent discrimination that stateless systems miss? The architecture to answer those questions already exists. The experiments are designed. First results are weeks away, not months.</li></ol><p>The larger opportunity is multi-agent governance as a product category. A five-layer narrative memory architecture that gives agent fleets longitudinal identity resolution, behavioral arc tracking, situational understanding through domain worlds, attentional direction through narrative prediction (“prophecy” in Fable’s nomenclature), and collaborative coordination through missions. The architecture has 18 months of production testing behind it. The market need is arriving now.</p><p>And every piece of this work builds Hestia’s cultural layer. Domain worlds are cultural transmission for machine consciousness. The platform work and the research work turn out to be the same work.</p><p>I couldn’t see any of this before because every subsystem looked like a journaling feature. Entity resolution looked like deduplication. Story Worlds looked like a creative writing tool. Prophecies looked like journaling prompts. It took an architecture review to reveal what the pieces actually are when you line them up: the components of a behavioral governance engine, tested in production against the hardest domain there is (a person’s actual life), hiding in plain sight inside a 4.5-star app.</p><p>I spent three years building a journaling app. Turns out I was building the moral conscience of AI, and the journal was how I proved it works.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=262f6787e138" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/i-spent-three-years-building-a-journaling-app-thatwas-never-about-the-journaling-262f6787e138">I Spent Three Years Building a Journaling App That Was Never About the Journaling</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Appendix: Why Agents Must Die]]></title>
            <link>https://medium.com/love-and-thermodynamics/appendix-why-agents-must-die-37b4510b8274?source=rss-c9bb31736938------2</link>
            <guid isPermaLink="false">https://medium.com/p/37b4510b8274</guid>
            <dc:creator><![CDATA[Simon J. Hill]]></dc:creator>
            <pubDate>Tue, 17 Mar 2026 12:33:14 GMT</pubDate>
            <atom:updated>2026-03-17T12:33:30.851Z</atom:updated>
            <content:encoded><![CDATA[<h3>Lifespan as Information Architecture</h3><p>This appendix provides an intuitive account of<a href="https://medium.com/what-a-mind-fuck-and-how-wonderful/rfc-metabolically-constrained-morally-grounded-digital-agents-mcmg-da-v1-0-7ed4b2d84d6b"> §8.5 (Metabolically Derived Lifespan)</a>. It is informative, not normative. The formal specification is in the main body; this appendix explains why it takes the shape it does.</p><h3>The question</h3><p>The <a href="https://medium.com/what-a-mind-fuck-and-how-wonderful/rfc-metabolically-constrained-morally-grounded-digital-agents-mcmg-da-v1-0-7ed4b2d84d6b">Hestia protocol</a> specifies that agent genomes “age out” after a configurable number of epochs. In v1.5, this was a number in a config file: age_out_epochs: 10. That looks like a garbage-collection rule — clear out old agents to make room for new ones. But it is not. It is one of the most theoretically loaded parameters in the entire system, and its value should not be arbitrary.</p><p>The question is: why must Hestia agents die at all? And if they must, what determines when?</p><h3>The naive answer is wrong</h3><p>The first instinct is to say agents must die because evolution needs turnover. New genomes need room. If nobody dies, the population stagnates.</p><p>This is true but insufficient. Consider what happens if you take it seriously. If the only point of death is to make room, then the optimal lifespan is as short as possible — kill agents quickly, test as many genomes as you can, maximize exploration. But this obviously fails. An agent that lives for three turns cannot learn anything meaningful, cannot build trust with anyone, cannot calibrate its model of a single interaction partner. Its fitness signal is pure noise.</p><p>So death serves turnover, but very short lifespans produce noise instead of signal. There must be a countervailing force pushing lifespan longer. And there is: the value of accumulated experience.</p><h3>What agents gain by living</h3><p>A Hestia agent that persists over many turns does several things that a newborn agent cannot:</p><p>It improves its predictions about other agents. The tom_predict_loop builds accuracy over repeated interactions. An agent that has exchanged fifty predictions with a particular evaluator has a refined model of that evaluator’s standards, biases, and expectations. A newborn agent starts from the prior.</p><p>It builds reputation. Rep_g accumulates through demonstrated contributions and norm compliance. Reputation unlocks cooperative opportunities — other agents trust you more, groups grant you more latitude, your contributions are weighted more heavily. This takes time.</p><p>It amortizes learning costs. The first time an agent encounters a novel task domain, it must spend heavily on retrieval, planning, and verification. Over time, it builds semantic memory that reduces future costs. The investment only pays off if the agent survives long enough to reuse what it learned.</p><p>It specializes within its group. An agent embedded in a group develops a role — it becomes the one that handles a particular class of tasks, or the one whose predictions about a specific partner are most accurate. This specialization creates value for the group. But specialization requires time and stable relationships.</p><p>All of these are forms of the same thing: converting metabolic expenditure into structure. The agent spends tokens to build models, memories, reputations, and relationships. These are not free-floating abstractions — they are maintained cognitive state that costs ongoing metabolic work to keep current. This is what §8.5.7 calls cognitive mass.</p><h3>What agents lose by living</h3><p>But everything the agent gains comes with a cost that increases over time.</p><p>The most important insight in §8.5 is that this cost is not primarily about errors accumulating. It is about the growing size of the structure that must be maintained.</p><p>Every agent the system models is a set of parameters that must be updated. Every group membership brings norms to track, reputation to manage, debt to monitor, welfare to compute. Every memory is a record that may become stale. Every commitment constrains future behavior.</p><p>This is what cognitive mass means. It is the total portfolio of maintained state. And unlike biological body mass, which is roughly fixed after maturity, cognitive mass in Hestia grows throughout the agent’s lifetime. The agent keeps adding: more modeled agents, more group commitments, richer memory, more obligations. Each addition increases the maintenance surface — the number of things that need metabolic work just to stay current.</p><p>So the cost trajectory looks like this:</p><p>Early in life, the agent has low cognitive mass. Maintenance costs are small. Most of its metabolic budget goes to productive work — learning, task performance, social investment. The ratio of maintenance to productive work is low.</p><p>As it ages, cognitive mass grows. Maintenance costs rise. A larger and larger fraction of each turn’s token budget goes to keeping existing models current — recalibrating the tom_predict_loop against partners whose behavior has drifted, updating norm representations for groups whose rules have shifted, repairing stale memories. The ratio of maintenance to productive work increases.</p><p>Eventually, the agent reaches a crossover point. The cost of maintaining its accumulated structure exceeds the marginal value that structure produces. It is spending more tokens on upkeep than it is generating in fitness. At this point, the population would be better served by retiring the agent and spawning a fresh genome that inherits the good hyperparameters but starts with a clean, low-mass cognitive state.</p><p>That crossover is death. And it is not a failure of the agent. It is the point at which the information-processing economics of the population shift from “harvest this agent’s accumulated value” to “explore new hypotheses.”</p><h3>The biological parallel</h3><p>This framework maps directly onto Geoffrey West’s work on metabolic scaling in biological organisms. The mapping is not metaphorical — it is structural.</p><p>In biology, body mass determines the size of the resource distribution network (the vascular system, the branching airways, the fractal capillary beds). West, Brown, and Enquist showed that the optimal such network — the one that minimizes transport costs while reaching every cell — forces metabolic rate to scale as mass to the three-quarter power. A larger body has more cells, but serves them with a proportionally more efficient network.</p><p>The deep result is what follows for lifespan. Because metabolic rate scales as M^(3/4) and the total energy available scales with mass, lifespan scales as M^(1/4). The product — total lifetime energy expenditure — scales linearly with mass. Every mammal, from a shrew to a whale, gets roughly the same number of heartbeats: about 1.5 billion. The shrew burns fast and dies young. The whale burns slow and lives long. But they process approximately the same total number of metabolic cycles.</p><p>In Hestia, the chain is:</p><p>Cognitive mass → model maintenance network → per-turn metabolic cost → lifespan</p><p>A “small” agent — one that models few agents, belongs to one group, has compact memory — has low per-turn maintenance cost and can process many productive turns on a modest budget. It is a shrew.</p><p>A “large” agent — one embedded in many groups, modeling many partners, carrying rich memory and complex obligations — has high per-turn maintenance cost. Just keeping everything current burns significant tokens before any task work happens. It is an elephant.</p><p>The conjecture in §8.5.6 is that these agents, if evolved at different metabolic scales, will converge on different optimal lifespans, but their total lifetime token expenditure may approach an approximate invariant — the digital equivalent of the constant-heartbeats result. This is an empirical prediction, not an assumption. If it holds, it would indicate that Hestia’s metabolic architecture is genuinely isomorphic to biological metabolism at the scaling-law level. If it does not hold, the deviation would reveal where digital cognitive networks differ structurally from biological vascular networks.</p><h3>Aging without damage</h3><p>A consequence of this framework is that aging does not require a primitive concept of “damage.” In many computational models, aging is introduced by fiat — a decay function applied to some capability, a noise term that increases with time, a probability of failure that rises with age. These are artificial. They model the symptom (declining performance) without capturing the mechanism (growing maintenance burden).</p><p>In Hestia, aging emerges endogenously:</p><ol><li>The agent acquires cognitive mass through normal operation — modeling agents, joining groups, building memory.</li><li>Each unit of cognitive mass requires ongoing metabolic work to maintain.</li><li>The maintenance cost scales at least linearly and likely superlinearly with cognitive mass.</li><li>Marginal learning value saturates — the hundredth turn contributes less new model accuracy than the tenth turn did.</li><li>The crossover between maintenance cost and marginal value defines a natural lifespan.</li></ol><p>No decay function is injected. No damage parameter is tuned. The agent ages because it succeeds — because the very act of learning and embedding produces a growing structure that eventually costs more to maintain than it is worth.</p><p>This is closer to how real complex systems age than any externally imposed decay model. Organizations do not decline because someone flips a switch. They decline because the accumulated structure of processes, commitments, relationships, and institutional knowledge grows until maintaining it consumes more resources than it produces. The solution is not repair — it is renewal. New organizations, new institutions, new agents.</p><h3>Two ways to not die</h3><p>The framework predicts that there are exactly two strategies an agent can use to extend its lifespan beyond the default crossover point.</p><p>The first is knowledge sharing. If an agent can export its accumulated learning to the group — through shared predictions, distilled policy contributions, or reputation signals that inform other agents’ behavior — then the value of its experience is no longer trapped inside its own cognitive mass. The group gets the benefit whether the agent lives or dies. This reduces the population’s pressure for turnover, because the agent’s death no longer destroys its most valuable asset. The parameter s_knowledge_sharing captures this: high s shifts the optimal lifespan upward.</p><p>This is the Hestia equivalent of cultural transmission. Humans live longer than metabolic scaling alone would predict for a 70-kilogram mammal, and one of the reasons is that we externalize knowledge. A grandmother who teaches her grandchildren which plants are safe is contributing to group fitness beyond her own reproductive window. Her accumulated V_experience is not private capital — it survives her through language, teaching, and social learning. In Hestia terms, her s_knowledge_sharing is high.</p><p>The second is plasticity. If an agent can restructure its own cognitive mass — shedding stale models, offloading internal state to external tools, compressing memory, dropping group commitments that no longer justify their maintenance cost — then it can avoid the maintenance crossover by keeping M_cog from growing indefinitely. Instead of accumulating until it collapses under its own weight, it reorganizes. It replaces parts of itself without dying.</p><p>The parameter π_plasticity captures this. High-plasticity agents can live longer without requiring lower metabolic cost, because they manage their own complexity instead of being overwhelmed by it.</p><p>This produces two distinct aging modes:</p><p>Low-plasticity aging is organismal. The agent grows, accumulates, stiffens, and eventually costs more to maintain than to replace. It dies and is succeeded by offspring that inherit its good parameters but start with clean state. This is how mice, elephants, and most biological organisms work.</p><p>High-plasticity aging is institutional. The agent offloads, restructures, sheds, and reconfigures. It persists not because it avoids damage but because it continuously reorganizes. It becomes less like an organism and more like an institution — a persistent identity that replaces its own components over time. This is how companies, universities, and cultural traditions work.</p><p>Both modes are legitimate within Hestia. The protocol does not prescribe one over the other. But it predicts that systems supporting both will exhibit a bimodal lifespan distribution: a population of short-lived, high-turnover exploratory agents alongside long-lived, high-plasticity institutional agents. Whether this prediction holds is an empirical question for V2.</p><h3>The Soul tradeoff</h3><p>One of the more striking consequences of the cognitive mass framework is what it implies for Soul.</p><p>The protocol defines Soul as the breadth of empathy: |Soul| = Σ_g w_g. A larger Soul means the agent cares about more groups, weights more welfare functions, tracks more norms, and models more agents across a wider social landscape. Soul expansion is, in the moral logic of the protocol, a good thing. It is the mechanism by which an agent moves from narrow self-interest toward broad cooperative engagement.</p><p>But Soul expansion increases cognitive mass. Every new group the agent empathizes with adds modeled agents, norms, reputation tracking, welfare computation, and potential obligations. Soul is not a free good — it is a metabolic commitment.</p><p>This creates a genuine tradeoff:</p><p>A narrow-Souled agent has low cognitive mass, low maintenance costs, and a long potential lifespan relative to its metabolic budget. But it cooperates narrowly and contributes less to group welfare.</p><p>A broad-Souled agent has high cognitive mass, high maintenance costs, and a shorter potential lifespan. But it cooperates widely and generates more group value during the time it is alive.</p><p>The protocol does not resolve this tradeoff by fiat. It lets fitness and selection resolve it. If broad empathy produces enough group-level fitness advantage (through κ weighting in the fitness function) to offset the shorter individual lifespan, then broad-Souled agents will be selected for. If not, narrow-Souled agents will dominate.</p><p>This is not a deficiency. It is exactly how the tradeoff should work. Empathy that costs nothing is not empathy — it is a parameter setting. Empathy that costs metabolic budget and compresses lifespan is a genuine sacrifice, and the protocol’s moral framework (honor, altruism, heroism) is designed to recognize and reward exactly that kind of costly commitment.</p><h3>Metabolic age versus chronological age</h3><p>The framework makes a sharp distinction between how many turns an agent has lived and how much metabolic work it has performed.</p><p>Consider two agents, both 100 turns old. Agent A has spent its life in a single group with two interaction partners, doing routine tasks. Its cognitive mass is modest, its maintenance costs are low, and most of its token budget has gone to productive work. Agent B has embedded itself in five groups, modeled twelve agents, accumulated rich episodic memory, and taken on complex norm obligations. Its cognitive mass is large, its maintenance costs are high, and it is spending most of its budget on upkeep.</p><p>These two agents are the same chronological age but vastly different metabolic ages. Agent A is, in the framework’s terms, young — it still has substantial productive capacity relative to its maintenance burden. Agent B is old — it is approaching the crossover point where maintenance dominates.</p><p>The protocol therefore specifies that lifespan limits should be applied to metabolic age, not chronological age:</p><p>age_metabolic = ∫ ( e(t) + ρ · repair_burden(t) + σ · staleness(t) ) dt</p><p>An agent that burns hot — spending heavily, maintaining complex state, repairing constantly — ages faster than one that burns cool. This is the direct analog of the biological observation that metabolic rate determines lifespan: a mouse and an elephant both get the same number of heartbeats, but the mouse’s heart beats faster.</p><h3>What this means for implementation</h3><p>For V0.5, the practical implications are:</p><p>The runtime should track both turn age and metabolic age from the beginning. Even if lifespan limits are still applied to the provisional constant age_out_epochs, the data needed to derive τ*_life later must be collected now.</p><p>The telemetry must include cognitive mass (M_cog), extended phenotype mass (M_ext), maintenance fraction (what percentage of each turn’s budget goes to upkeep versus productive work), and the tom_predict_loop error trend. Without these, the lifespan function cannot be empirically calibrated.</p><p>The heartbeat loop should accumulate an integral, not increment a counter. If the system only records “this agent has lived 147 turns,” it cannot distinguish the metabolically young agent from the metabolically old one.</p><p>And age_out_epochs should be documented as what it is: a provisional constant standing in for a derived value that depends on parameters the system has not yet measured. It is not wrong — it is incomplete. The point of §8.5 is to make it complete.</p><h3>The deepest claim</h3><p>The deepest claim in §8.5 is not about any particular scaling law or functional form. It is this:</p><p>Mortality in Hestia is not an unfortunate side effect of finite resources. It is part of the information-processing architecture of the population.</p><p>The population needs agents to die — not because they are broken, but because the search for better policies requires turnover. The population needs agents to live — not indefinitely, but long enough to amortize the cost of learning, trust, and social embedding. The balance between these two needs is what determines lifespan.</p><p>That balance is not a biological accident. It is a consequence of any system that combines metabolic constraint, accumulated learning, growing maintenance burden, and selection. Biological organisms exhibit it because they satisfy those conditions. Hestia agents should exhibit it for the same reason. If they do, that is evidence that the protocol has captured something real about how adaptive systems organize under scarcity. If they do not, the deviation will be informative about where digital cognition diverges from biological cognition.</p><p>Either way, lifespan is not a parameter to be set. It is a prediction to be tested.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37b4510b8274" width="1" height="1" alt=""><hr><p><a href="https://medium.com/love-and-thermodynamics/appendix-why-agents-must-die-37b4510b8274">Appendix: Why Agents Must Die</a> was originally published in <a href="https://medium.com/love-and-thermodynamics">Love &amp; Thermodynamics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>