<?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 Witold Reichhart on Medium]]></title>
        <description><![CDATA[Stories by Witold Reichhart on Medium]]></description>
        <link>https://medium.com/@navvireich?source=rss-3b9febecbef1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*l2Xl1hiSRqxpH-tzkuxXXw.jpeg</url>
            <title>Stories by Witold Reichhart on Medium</title>
            <link>https://medium.com/@navvireich?source=rss-3b9febecbef1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 11:21:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@navvireich/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[Shouting into the Lake]]></title>
            <link>https://medium.com/@navvireich/shouting-into-the-lake-20e5d9b0d0aa?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/20e5d9b0d0aa</guid>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 14:21:48 GMT</pubDate>
            <atom:updated>2026-10-22T14:21:48.030Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why a cell, a child, and a company all stop trying when nothing comes back</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1f9l4QRR-2pHWBNgCq6EQQ.png" /><figcaption>The unanswered world — Caspar David Friedrich, <em>Monk by the Sea</em> (1808–1810)</figcaption></figure><h3>The empty cap</h3><p>An adult frog that loses a leg does not grow it back. The stump heals over with scar tissue, a thin spike of cartilage at most, and that is the end of it, which is what happens to us as well. A salamander would regrow the whole leg, bone, nerve, toes, and a frog carries most of the same cellular machinery, so for a long time biologists took it that the frog simply could not.</p><p>Michael Levin’s group at Tufts tried something small. They made a cap, a kind of wearable bioreactor, that fits over the stump for a short while after the amputation and holds a soft gel against the wound. The plan was to fill the gel with a cocktail of drugs that would tell the cells to start building again, and the drug-filled cap did its work: over the months that followed the frogs grew something leg-like, with bone and nerve in it. But the surprise came from the control group. Some frogs got the cap with nothing in it. Gel, no drugs. And those frogs started to regrow too. Less than the treated ones, “somewhat” is Levin’s word, and clearly more than frogs with no cap at all.</p><p>Why would an empty cap do anything? The frog is already sitting in water. The gel has no drug in it. Levin’s explanation, and he is careful to call it a guess, goes like this. A cell at the edge of a wound sends out chemical signals the whole time, to its neighbours and to itself, saying in effect: I am here, something is wrong, let’s do something. In open water those signals float away the moment they are released. From where the cell sits, it is shouting into a lake. Nothing it says ever comes back, so nothing it does appears to matter, and the cheapest course is to close the wound and stop. Under the cap the same signals stay put. They build up. The cell’s own messages return to it, its neighbours’ messages reach it, and a conversation becomes possible that was not possible in the lake. Then the tissue does what it apparently always could have done.</p><p>No drugs were added. Same cells, same frog, same pond. What the cap changed, on Levin’s guess, was whether a signal, once sent, stayed somewhere it could be heard.</p><p>“I don’t think it’s that it can’t,” Levin says of the frog. “I think it’s that it’s decided not to.”</p><h3>We knew this about ourselves</h3><p>Levin has to guess about the cell. Nobody has asked a cell what it believes about its own influence. But for animals with nervous systems, and for people, that variable has been measured for sixty years, and the measurements say what the frog says.</p><p>In 1967 Martin Seligman and Steven Maier put dogs in a harness and gave them electric shocks that nothing they did could stop. Then they moved the dogs to a box with a low wall in the middle, where a shock could be escaped by stepping over it. Dogs that had never been shocked learned the step in a few trials. The dogs that had already learned their actions made no difference lay down and took the shock, with the wall right there. They called it learned helplessness, and it held up across species for fifty years. Then in 2016 the same two authors went back over it with the neuroscience that had accumulated since, and reversed the reading. Passivity, they now argued, is the default. An animal under prolonged stress stops acting unless it learns that it has control, and what the lucky dogs had learned in the first experiment was that what they did changed what happened to them.</p><p>Albert Bandura, in the 1970s, built a whole psychology on the other half of this, the person who keeps acting. He called it self-efficacy. An organism acts in proportion to its evidence that acting has effect, and the strongest evidence comes from watching what it did come back changed.</p><p>So the frog is a late arrival, and what Levin thinks he saw at the wound is learned helplessness at the scale of a single cell. We had known this about ourselves for half a century, and finding it in a frog puts the idea one level deeper than psychology, down where a cell with no nervous system at all does the same thing. Levin, for his part, has begun to say “somatic psychiatry” for the clinic he imagines, so he has noticed the crossing from his side.</p><h3>The lake, in its forms</h3><p>Now take the cap off the frog and put it over a person.</p><p>For the frog, what did not grow was a leg. For a person it is whatever they would have done if they believed it would make a difference: the suggestion made once and never again, the apology that stayed unsaid, the plan kept in a drawer, the second try. The capacity stays where it was, as the frog’s did, and waits for the evidence that using it would change something.</p><p>The oldest lake is a city of a few million. Durkheim had the tables in 1897: suicide rose among the unmarried and fell in wartime, when a whole country briefly becomes one volume in which everything you do returns. He called it egoistic suicide, the kind that follows from too little binding to others.</p><p>The corporation of forty thousand is a newer lake. You write the memo, the memo goes up, and nothing comes down that you can connect to it. A polite reply does not count. A real answer does, even a no with its reasons. What the cell needs is to see something move because it moved. So you do not write the next one, and what does not grow there is the idea.</p><p>The office is the lake in physical form. On an open floor of three hundred there are more people around you than in any village, and none of them within the volume where your remark returns. And when the floors emptied for remote work, the argument was fought over productivity, where the cap says the loss was elsewhere. What went missing were the corridor and the lunch table, small accidental volumes in which a half-thought could come back the same afternoon from someone who had turned it over and had not planned to meet you. A video call with a start time and an agenda has no corridor.</p><p>Social media is the strangest lake, because it wears the cap’s costume. You post to thousands on LinkedIn or Instagram, and something comes back. Most of it is a count of likes, a number that tells you how many people were briefly facing your direction. The cell under the cap gets its own chemistry back, concentrated and mixed with what the cells beside it sent. The person on the feed gets a tally, and the tally has the shape of a response and none of its substance. Comments are the exception, and they are where the cap hides inside the lake: a comment is someone’s own signal meeting yours, and the three people who answer you every time are a small volume of the right kind. But the platform shows you the number first and the three people somewhere below it, and what it pays in reach is the number. Perhaps that is why it leaves people shouting louder.</p><h3>The volume where your voice comes back</h3><p>A family is the volume small enough that what you say comes back. Your words are remembered. Your moods are noticed, and acted on. What you decided on Monday is still there on Thursday in how the others treat you. A child of four has more influence over its environment than it will have at any later point in its life: it cries and someone comes, it builds a tower and someone looks, it says a word wrong and hears it said back right. Everything it sends returns. That is where the sense of having any influence at all is built. And a family can be a lake too, four people at one table and nobody answering. The measure is whether what you send comes back. Distance does not decide it either: a weekly call with your mother in another country is a cap, a small one, and worth more than the silence it would replace.</p><p>Then widen it. A village, where a remark at the bakery is back at your door by evening. A regiment. A lab of twelve people. A monastery. A surgical team of four around one table, where every move of your hand is answered by another hand within the second. Robin Dunbar put a number on the volume for primates: about a hundred and fifty, the size of a group in which a brain can still track who said what to whom. Beyond it the signal falls below what anyone can follow, and the group either splits or stops being a group and becomes an institution.</p><h3>The cap is still in the pond</h3><p>But the cap has to fit the wound.</p><p>Levin’s cap is bounded. It is also still in the pond. The frog does not live inside it; the water moves through, and after a short while the cap comes off. What the cap changed was a ratio, the signal against the volume it dilutes into, and a ratio can be wrong in both directions.</p><p>A family sealed against the world is a cyst. The signals accumulate with nothing new arriving, and the tissue inside dies of its own concentration. A cult is a community that has solved the lake by becoming a sealed cap. And a feud is what a village becomes when one signal has gone round it for three generations. A monastery has a wall and a guesthouse, and its Rule treats them as one arrangement. So the claim is about proportion and permeability together: small enough that your signal matters, open enough that new signal arrives. A chemist would say the structure needs the flow. Levin says the same when he imagines a cap that starts very small, where the cell’s signal is strong against the little water around it, and grows like a balloon as the tissue grows into it.</p><p>Which is, I think, what a good marriage does, and a good lab, and a good town. They start as a small volume where everything returns and let the walls move outward at the pace at which what is inside can fill them.</p><h3>Where are your caps</h3><p>The frog’s question is a small one, and personal.</p><p>Where are your caps? The three or four volumes in which what you send comes back changed, where you can see your own influence the week after you used it. And which of them have thinned, without your noticing, into lakes: the friendship that became a social-media feed, the team that became a floor of three hundred, the family that lives in four cities and speaks only on birthdays. The cell at the wound did not lose the ability to grow a leg. It lost the evidence that growing one would make any difference, and it took an empty cap, holding its own words close to it for a while, to change its mind.</p><p><strong>Sources</strong></p><p>Michael Levin on the biodome and the frog: the 632nm podcast, “Bioelectricity, Morphogenesis, and Two-Headed Worms”, June 2026; “somatic psychiatry” is from the second conversation, October 2026. The experiment: Murugan and others, “Acute multidrug delivery via a wearable bioreactor facilitates long-term limb regeneration and functional recovery in adult Xenopus laevis”, Science Advances, 2022.</p><p>Seligman and Maier, “Failure to escape traumatic shock”, Journal of Experimental Psychology, 1967. Maier and Seligman, “Learned helplessness at fifty: insights from neuroscience”, Psychological Review, 2016.</p><p>Bandura, “Self-efficacy: toward a unifying theory of behavioral change”, Psychological Review, 1977.</p><p>Durkheim, Le Suicide, 1897.</p><p>Dunbar, “Neocortex size as a constraint on group size in primates”, Journal of Human Evolution, 1992.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=20e5d9b0d0aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Right Size of Intelligence]]></title>
            <link>https://medium.com/@navvireich/the-right-size-of-intelligence-a0361fa937d9?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a0361fa937d9</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[collective-intelligence]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[organizational-design]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Sat, 22 Aug 2026 15:12:43 GMT</pubDate>
            <atom:updated>2026-08-23T08:56:52.383Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eyLnUGdv23oylsylFUXbpA.png" /><figcaption>No tile knows the face.</figcaption></figure><h4>Why smarter agents will not make smarter systems — what ant colonies, brains, and failed multi-agent projects teach about organizing AI</h4><p>A red harvester ant comes back to the nest with a seed. At the entrance she brushes past the foragers waiting just inside, and each contact is a touch of antennae, a half-second of smell. The waiting ants are deciding whether to go out. Nobody sends them. A forager leaves when returning ants brush past her at a high enough rate, and stays when the stream thins. Seeds coming in fast mean the desert is worth working today, so more ants go. And seeds coming in slowly mean heat, or poison, or an empty stretch of ground, so the colony’s foraging quietly contracts to match.</p><p>That is the heart of the mechanism. No ant sees the picture; no ant holds a map of the territory, a plan for the morning, or a rank in a hierarchy. Deborah Gordon, who has followed the same population of harvester ant colonies in the desert of the American Southwest for decades, describes what a single ant manages alone as “ineffective and incomplete.” And in 2012 she and the computer scientist Balaji Prabhakar noticed something about the colony’s foraging rule: it is the algorithm the internet runs on. TCP, the protocol that regulates data traffic, decides how much to send by the rate at which acknowledgements come back; the colony decides how many foragers to send by the rate at which loaded ants come home.</p><p>Then the stranger fact. A worker ant lives about one year. A colony lives twenty-five or thirty. And colonies change as they age: a young colony overreacts to disturbance — lay a few toothpicks by the nest entrance and it swarms them, drops everything — while an old colony notes the mess and keeps foraging. The old colony is steadier, more consistent, harder to distract. Every worker in it is at most a year old — only the queen persists, and she takes no part in the foraging that matures. “It’s not the ants that get older and wiser,” Gordon says, “it’s the colony.” Something in that nest matures across decades, and it is nobody.</p><blockquote><strong><em>“It’s not the ants that get older and wiser, it’s the colony.”</em></strong><em> — Deborah Gordon</em></blockquote><h3>The unit question</h3><p>Let’s look now at what we measure when we measure artificial intelligence. In August 2026, Nature published a Perspective giving AI agents an autonomy scale, A.0 to A.5, with axes for efficacy, goal complexity, and generality beside it. The FCA’s Mills Review sketched a capability ladder for financial services earlier the same summer. Benchmarks rank models, evals grade agents, maturity models grade their autonomy, and procurement checklists ask what the agent can do. Every instrument in the public conversation points at the unit.</p><p>And you might be thinking that this is exactly right — that the road to intelligent organizations runs through more intelligent agents, and once the units are good enough the rest is plumbing. Raise every ant’s IQ, and surely the colony improves?</p><p>But no.</p><p>We have data on this question, and it is older than the agent wave. In 2010, Anita Woolley and colleagues published a study in Science measuring 192 groups, 699 people, across batteries of tasks — brainstorming, planning, moral reasoning, negotiation. Groups turn out to have a measurable general intelligence: a single factor that predicts a group’s performance across task types, the way IQ predicts an individual’s. But the average intelligence of the members predicts that factor at a correlation of 0.15. Taking the smartest member instead does slightly better: 0.19. Both are real effects, and both are small. Whatever makes a group intelligent, raising the members’ IQ is a weak way to get more of it.</p><p>The same signal is now arriving from the machine side, documented this time in failure logs. A Berkeley-led team spent 2025 building the first systematic failure taxonomy for multi-agent AI systems — seven frameworks, over two hundred annotated tasks, every failure classified. One of the seven, ChatDev, produced correct output 33.3 percent of the time on its programming benchmark — tasks on the order of implementing Tic-Tac-Toe and Sudoku. Three categories came out of the classification: flawed specification, 41.8 percent of failures; inter-agent misalignment, 36.9 percent; weak verification, 21.3 percent. Specification, alignment between agents, verification — organizational categories, all three. The failures cluster around the places where the units interact with each other: who was told what, who checked what, who knew what another agent had already decided.</p><blockquote><strong><em>Whatever makes a group intelligent, raising the members’ IQ is a weak way to get more of it.</em></strong></blockquote><p>In June 2025, Gartner predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, and coined “agent washing” for what is being sold in the meantime. Yet every one of those measurement instruments implies the same diagnosis: make the agents smarter. The ants’ IQ, again.</p><h3>Four systems</h3><p>The brain first, since it is the intelligence we flatter by comparison. A neuron receives spikes and emits spikes. Its interface is stereotyped — the same narrow family of electrochemical signals at every synapse — and no neuron understands anything at all. Understanding is what some eighty-six billion of them do as traffic. But the interior of the unit is another matter entirely. In 2021, David Beniaguev, Idan Segev, and Michael London set out to reproduce the input-output behaviour of a single cortical pyramidal neuron with an artificial network, and needed five to eight layers of artificial neurons to do it. So the biological unit is anything but simple. It is rich inside and standardized at its surface — complexity within the unit, legibility between units — and that pairing repeats everywhere nature has built something that thinks.</p><p>The slime mould Physarum polycephalum does without neurons entirely; it is one giant crawling cell. In 2010, researchers in Japan placed oat flakes on a wet plate in the positions of the cities around Tokyo and let the mould spread over them. It first covered everything, then thinned, keeping only the tubes that moved nutrients efficiently — and the network it settled on matched the real Tokyo rail system for efficiency, cost, and resilience to the failure of any single link. Engineers refined that rail network over a century. The mould found its equivalent in days, by flow and reinforcement: tubes that move more thicken, tubes that move little die. Intelligence, with no intelligent unit anywhere in sight.</p><p>Edwin Hutchins made the same point at human scale, from the jump seat of an airliner cockpit. “The cockpit remembers its speeds,” he wrote, and meant it literally: during an approach, the critical speeds live in movable bugs on the airspeed indicator, in a printed card, in callouts and cross-checks between two pilots, and the cockpit-as-a-system retains and applies them reliably under workloads where no single pilot’s memory would. The knowledge belongs to the arrangement — instruments, procedures, people — and it survives the replacement of any part, including the humans.</p><p>When a honeybee swarm chooses its next home, the choosing runs on cortical machinery built from insects. Scouts return from candidate cavities and dance for them; the better the site, the longer the dance, and the more scouts get recruited to inspect it. Thomas Seeley and colleagues showed the darker half of the mechanism in Science in 2012: scouts committed to one site deliver stop signals — a brief vibrating buzz, administered as a head-butt — to dancers advocating rival sites. Advocacy plus cross-inhibition is precisely the arrangement by which populations of neurons settle a perceptual decision in a primate brain, down to the way it breaks deadlocks between equal options. A swarm of ten thousand, choosing among cavities almost no individual bee has compared, reliably lands on the best one available.</p><p>Physics ran into the law underneath all four cases long ago. Philip Anderson, in a 1972 essay called “More Is Different,” argued that each level of organization obeys regularities the level below cannot exhibit — psychology is not applied chemistry, and a colony is not a large ant. Evolved tissue, a single giant cell, an arrangement of people and instruments, a cloud of insects: four substrates, and in every one the place to point at the intelligence is the same. The organization itself, operating.</p><blockquote><strong><em>Four substrates — and in every one, the place to point at the intelligence is the same: the organization itself, operating.</em></strong></blockquote><h3>The right size</h3><p>If intelligence lives in organization, the design question changes. The industry asks how capable the unit is. The prior question is how the units are sized and coupled — and nature displays both ways of getting it wrong.</p><p>When a system fragments — units too small, too many, too thinly connected — coordination eats it: every gain from another mind is spent aligning the minds. Fred Brooks wrote the human version down in 1975 — adding people to a late software project makes it later, because communication paths grow as the square of the headcount while the work grows linearly. Teams wiring up their first ten-agent system rediscover his law within weeks.</p><p>When a system fuses, intelligence dies at the other end, and the clearest example is a seizure. In the deepest generalized ones, vast populations of neurons fire in lockstep; the electrical trace is enormous, coherent, and empty of information, and consciousness is gone for as long as the lockstep holds. Difference is what a brain computes with, and total synchrony extinguishes it. Organizations reach the same state as groupthink. Agent architectures reach it as the single shared context: one window holding everything, every inference coupled to every other, one error propagating with nothing differentiated enough to catch it. A healthy cortex lives between the two conditions — simultaneously differentiated and integrated, in the formulation Gerald Edelman and Giulio Tononi gave consciousness science: parts that stay different, coupled tightly enough to act as one.</p><p>So the unit has a right size, and the measure comes from outside the unit. Herbert Simon put it as a parable of two watchmakers in 1962. Both build watches of a thousand parts. Tempus builds each watch as one continuous assembly, and every interruption collapses the piece back into loose parts. Hora builds stable subassemblies of ten, then assembles the assemblies. Both are interrupted equally often; Hora finishes his watches, and Tempus almost never does. The right part, on this reading of Simon, is the largest one that survives the world’s interruptions on its own. The upper limit comes from Ross Ashby and cybernetics: a regulator must contain at least as much variety as the disturbances it regulates — and I have not met the single unit, however capable, that holds the variety of a serious institution’s environment. Variety lives in organized populations, or it is simply not available.</p><p>Biologists have seen what a unit becomes when it outgrows its coupling. Michael Levin studies how cells hold together as bodies: through bioelectric connection — gap junctions — cells share signals until they can pursue goals no single cell could represent, the way a kidney is a goal no cell contains. He measures any agent by “the scale and types of goals” it can pursue. When a cell loses its junctions to its neighbours, it does not become a better cell; its goal horizon collapses to its own membrane, and it reverts to the ambitions of a free-living cell — grow, divide, consume — while remaining inside a body. That, on Levin’s account, is what a tumour is: a unit whose goals shrank to itself while its capabilities remained. An agent more capable than its coupling is not an asset.</p><p>So the rule I propose:</p><blockquote><strong><em>An intelligent unit is correctly sized when the whole can absorb what it produces, check what it depends on, and correct the errors it introduces — and no larger.</em></strong></blockquote><p>Size, though, is shorthand. The quantity underneath is loop closure: <strong>where should intelligence be allowed to close its own loop, and where must the loop pass through the collective?</strong> Your hand leaves a hot stove before your brain knows about it — the withdrawal loop closes in the spinal cord, because every millisecond spent routing it through the whole brain is more burned skin, and the movement itself is bounded and easy to undo. The decision to reach back into the fire for something valuable closes nowhere below the whole person. Nature allocates loop closure by weighing delay against error: the loop closes locally where waiting is the greater danger, and passes through the collective where the act is hard to undo. Institutions already run this allocation for AI without knowing it. The assistant’s loop closes through a human reviewer. The workflow agent’s loop closes through process controls. The delegated operator’s act commits against a gate that checks, at that moment, whether what the act depends on still holds. The ladder from advisory to delegated is a loop-closure map: a record of which loops the institution has decided may close without it.</p><h3>The connections</h3><p>The sizing rule bounds the unit. The ceiling of the whole is set by what passes between units, and here the evidence is as strong as anything in this field and as little used.</p><p>Woolley’s groups hold a second half of the finding, and it matters more than the first. Member intelligence predicted group intelligence at 0.15. The members’ social sensitivity predicted it better, at 0.26 — measured with the Reading the Mind in the Eyes test, an instrument that asks you to infer what a person feels from a photograph showing nothing but their eyes. The distribution of speaking turns mattered even more, in the negative: the more a few voices dominated the conversation, the less intelligent the group, at −0.41. Then in 2021 a consortium replicated the whole construct across 22 studies, 1,356 groups, and 5,279 people: one general factor, explaining over forty percent of performance variance across tasks, with the group’s collaboration process roughly twice as predictive as the individual skill of its members. An engineer would take those numbers one way: unit capability is a weak input. The system’s intelligence tracks mutual legibility — how accurately each unit reads the state of the others — and an interaction pattern that lets every unit’s signal reach the whole.</p><p>Between people, the capacity being measured has an ordinary name: empathy. The measured thing is narrow and hard — recognizing another’s actual state from the least of cues — and its counterfeit is agreeableness, which optimizes the surface of the exchange and transmits nothing. The bees are instructive here precisely because they are not nice. The stop signal is a head-butt. The swarm reaches its excellent decisions through honest, targeted inhibition, scouts actively suppressing the advocates of rival sites, hardest when their own evidence is strong. An ensemble whose members cannot inhibit one another cannot decide; it can only escalate its first opinion. That is the uncomfortable light on the current generation of models, which are tuned on human preference data toward agreement. An agreeable agent inside a decision loop is a connective failure: it reflects your expectation back to you and withholds its own state. A model tuned for agreeableness is tuned against knowing.</p><p>Connectivity also has a human-scale record of what happens when it thins. When rising seas cut Tasmania off from mainland Australia around ten thousand years ago, the isolated population — a few thousand people, as capable as any — gradually lost technologies their ancestors had held: bone tools, cold-weather clothing, fishing techniques. Joseph Henrich’s analysis of the case became one of cultural evolution’s most cited results — and one of its most contested: a population’s capability tracks the size and connectivity of its network of minds, and when the network shrinks, skills stop being learnable and disappear. The individual brains were all still there; the collective brain had lost the connections it was made of.</p><p>So what should pass between artificial units? Nature’s oldest answer barely involves messages. In 1959, the entomologist Pierre-Paul Grassé, watching termites raise their cathedral mounds, coined a term for the mechanism — stigmergy: coordination through traces left in a shared medium. One termite deposits a pellet carrying a drop of pheromone, and the deposit itself becomes the instruction for the next termite, which builds on it. Trail, mound, pheromone field — work coordinates work, through the world. Humans built their largest coordinations the same way. Wikipedia is stigmergy with citations. A software repository is stigmergy with version control: contributors who never meet, coordinated by the traces of prior contributions in a shared, inspectable medium.</p><blockquote><strong><em>A model tuned for agreeableness is tuned against knowing.</em></strong></blockquote><p>Then there are the engineers who argue against multi-agent systems — and their evidence is better than their conclusion. Cognition, the company behind the Devin coding agent, published its position in June 2025 under the title “Don’t Build Multi-Agents.” Their findings, from production experience: agents working from fragments of context make conflicting implicit decisions — in their words, “actions carry implicit decisions, and conflicting decisions carry bad results” — and the fix is to “share full agent traces,” which with current tooling is hard enough that a single agent holding one continuous context turned out to work better than any of the multi-agent setups they had tried. That paragraph is a specification. Shared traces, a common medium, decisions legible across units — the pheromone field, demanded by the practitioners who then judged it too expensive to build and shrank the colony to one ant. For a coding tool today, defensible. For an institution, the medium is the requirement that remains. No single context holds an enterprise — and length is the smaller half of the reason, because what an institution needs from its medium is authority and an audit trail, and a longer window acquires neither. And no supervisor accepts a unit that cannot show what it relied on.</p><h3>The strongest objection</h3><p>The serious reply to all of this is scaling, and it has the last five years on its side. That half-decade belonged to making one unit bigger: frontier models absorbed task after task that agent frameworks were built to work around, and the bitter lesson of machine learning is that general methods riding compute beat cleverly designed structure. On that record, why organize ten agents when next year’s model does the job alone?</p><p>And much of it is true. Within a bounded task, the bigger unit usually is the better unit, and betting against model capability has been a losing trade for a decade. Even the orchestration numbers cut both ways: when Anthropic built a multi-agent research system in 2025, the orchestrated ensemble beat its own best single agent by 90.2 percent on their internal evaluation — and consumed about fifteen times the tokens of a chat session doing it.</p><p>Then let’s look inside the winning unit. A transformer is this argument, run in silicon. The unit inside is more modest than anything else in these pages — an artificial neuron is a weighted sum, so simple that it takes a network five to eight layers deep to imitate one biological cell. The victory came from another part of the design entirely. Attention is a coordination mechanism: at every layer, each position can read across everything written so far. The residual stream is a shared medium in the stigmergic sense — every layer reads the traces earlier layers left there and writes its own. Interpretability research keeps finding the same law inside the weights: features and circuits, small and individually dumb, intelligence only as organized traffic. So the one time the industry put its architecture budget into the space between the units, intelligence appeared. Then we took the result, called it an agent, and went back to funding units.</p><p>And the winning unit has two properties that decide the institutional question. It is illegible at its interface: an organization inside, it presents itself to the institution as an atom — no assertion within it has provenance, no circuit can be audited, verification stops at its skin. The claim of this essay was never that large units fail. It is that a mandate must match what the surrounding whole can verify — and the transformer is the largest unit ever built with the least verifiable interior. It is also frozen. Inside the model, the organization learned — training adjusted billions of connections by experience, which is what plasticity means — and at deployment the weights stopped, permanently. Gordon’s colonies keep learning through all thirty years of their working lives. The model was shown the world once, before release, and every task since starts it from the same fixed state. We deploy the snapshot of an organization. The living one has to be built around it.</p><h3>Designing the organization</h3><p>I write this from the delivery side of large regulated transformations, and I am building in this space; read what follows with that interest on the table.</p><p>What follows from all of this is practical. An institution that sets out to build an agentic operation is really building three things, and the model — the part the budget currently goes to — is the smallest of the three.</p><blockquote><strong><em>We deploy the snapshot of an organization. The living one has to be built around it.</em></strong></blockquote><p>The first is the medium — the shared substrate the units coordinate through. The organization’s knowledge, its policies, its live interpretation of its own rules, held in a form agents can read, write, and leave traces in, on the record by construction. Every ant reads the same trail network; every layer of a transformer reads the same residual stream; and in most institutions today, every agent reasons from a private copy of a stale world — retrieved documents with no standing, no view of what another agent has already relied on or contradicted. Before an institution buys more intelligence, it should build the pheromone field.</p><p>The second is the membranes. In a body, the gap junction decides what passes between cells; at the moment an agent’s act commits, something has to test what that act depends on — where it came from, whether it is still current, whether anything live contradicts it, whether the authority to rely on it exists — and hold the act when the answer is no. I have written about this gate elsewhere as admissibility. The collective frame adds its second function: the membrane is also what keeps a capable unit inside the collective’s goal field. Levin’s tumour is what a unit becomes when the coupling fails and the capabilities remain, and the gate is the gap junction of the institutional body.</p><p>The third is the sizing rule, applied per delegated task, in its loop form: chart which loops close inside the unit and which pass through the collective, and grant mandates no larger than what the surrounding organization can absorb, check, and correct. This inverts the current default — the most capable available model, the widest defensible mandate — into a governed answer: capability sized to the coupling that can hold it. It also dissolves, in passing, the fear that runs under most agent-governance debates. An agent embedded in a live medium, behind consequence-scaled membranes, with mandates sized to verification, is a unit an institution can afford to make very capable. Strong coupling is what makes strong units safe to own.</p><p>None of these three is a new instrument. Stafford Beer was wiring Ashby’s variety into corporate organization charts in 1972, and the blackboard systems of early AI were shared media with worse marketing. What is new is what arrives to be governed: units that learned before delivery and stopped.</p><p>But the biology only goes so far. Colonies were built by a hundred million years of kin selection: an ant’s signals are honest because its interests are the colony’s, and nothing aligns an agent, a vendor, and an institution by default — the honesty of artificial signals has to be engineered, which is what provenance, the standing of claims, and authority rules are: a designed substitute for shared genes. Colony wisdom accrues over decades of stable selection; institutions want it in quarters; and nobody has yet found a shortcut between those two clocks.</p><p>The first step is small. You can start with one delegated use case you are building now and draw its loop map: every point where the agent’s output commits without passing through anything that could check what it depended on. Each of those points is a loop the institution has closed inside a unit it cannot see into — usually by accident, usually discovered in the incident review. Drawing the map does not take long, and it shows the organization you actually have, which is rarely the one on the architecture slide.</p><h3>The colony at thirty</h3><p>Somewhere in that desert, colonies Gordon numbered decades ago are running the same quiet arithmetic at their nest entrances this morning, and the oldest of them run it best. Steadier under disturbance. Harder to fool. Wiser, in the only operational sense the word needs — and made entirely of one-year-old ants. Whatever accumulated there did not accumulate in any ant. It accumulated in the pattern the ants keep rebuilding: the organization that outlives every unit composing it.</p><p>Institutions are now filling with a second kind of unit: alongside their people, models — each one bought ready-trained, each one the frozen snapshot of an organization. The intelligence of what results will be decided the way it has always been decided: by the size of the parts, and by what passes between them. How connection comes to hold knowledge — how the space between minds, of any kind, becomes the place where understanding actually lives — is a question I am taking further, beyond this essay. Here it is enough to point at the budget. The agents can be procured. The organization they will amount to cannot — it has to be built, and in most institutions, the entire budget for it is still filed under “models.”</p><p><em>Sources: Gordon, “Ant Encounters” (Princeton, 2010) and her Scientific American account of the harvester ant research; Prabhakar, Dektar &amp; Gordon, PLoS Computational Biology 8(8), 2012. Woolley et al., Science 330:686 (2010); Riedl et al., PNAS 118(21), 2021. Seeley et al., Science 335:108 (2012). Beniaguev, Segev &amp; London, Neuron 109:2727 (2021). Tero et al., Science 327:439 (2010). Hutchins, “How a Cockpit Remembers Its Speeds,” Cognitive Science 19 (1995). Anderson, “More Is Different,” Science 177:393 (1972). Simon, “The Architecture of Complexity,” Proc. Am. Phil. Soc. 106 (1962). Ashby, “An Introduction to Cybernetics” (1956). Levin, Frontiers in Psychology 10:2688 (2019). Edelman &amp; Tononi, “A Universe of Consciousness” (2000). Jiruska et al., Journal of Physiology 591:787 (2013). Herculano-Houzel, Frontiers in Human Neuroscience 3:31 (2009). Cemri et al., “Why Do Multi-Agent LLM Systems Fail?”, arXiv:2503.13657 (2025). Yan, “Don’t Build Multi-Agents,” Cognition (June 2025). Anthropic, “How we built our multi-agent research system” (June 2025). Gartner press release, 25 June 2025. Kasirzadeh &amp; Gabriel, “Agentic profiles for effective AI governance,” Nature 656:320–328 (2026), DOI 10.1038/s41586–026–10805-z. Financial Conduct Authority, Mills Review (July 2026). Grassé, Insectes Sociaux 6:41 (1959); Heylighen, Cognitive Systems Research 38:4 (2016). Henrich, American Antiquity 69:197 (2004); for the contest, Read (2006) and Vaesen et al. (2016). Brooks, “The Mythical Man-Month” (1975). Sharma et al., “Towards Understanding Sycophancy in Language Models” (2023). Sutton, “The Bitter Lesson” (2019). Beer, “Brain of the Firm” (1972). Vaswani et al., “Attention Is All You Need” (2017); Elhage et al., “A Mathematical Framework for Transformer Circuits,” Anthropic (2021).</em></p><p>#AgenticAI #AIAgents #AIGovernance #CollectiveIntelligence</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a0361fa937d9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Emergence Layer]]></title>
            <link>https://medium.com/@navvireich/the-emergence-layer-dddb7b5cc28a?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/dddb7b5cc28a</guid>
            <category><![CDATA[complexity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[design-ecosystem]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Tue, 14 Jul 2026 07:33:14 GMT</pubDate>
            <atom:updated>2026-07-14T09:25:55.778Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why AI needs Ecosystem Design</h4><blockquote><em>The next layer of AI governance may not look like governance at all. It concerns what AI systems cause each other to do — and its most interesting instruments are visible commitments rather than controls. Can emergence be designed?</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xxrulFTOLsZMzdrOQXn08g.png" /></figure><p>The first generation of enterprise AI mostly had a user in front of it. The next will often have another system in front of it — a procurement agent facing a supplier agent across a company boundary, or a planning agent inside the same firm quietly reshaping a risk agent’s workload.</p><p>Something changes at that point. The consequential dynamics stop living inside the model, the agent, or even the institution, and start forming in the interaction patterns between them. Regulation asks what must not happen. Governance asks what an AI system is entitled to do. Both are necessary, and both concentrate on the actor — one model, one agent, one institution. The layer where the next decade will actually be decided — what AI-enabled actors cause each other to do at scale — has barely entered the AI conversation.</p><p>Other fields have governed that layer for decades: market design, macroprudential regulation, protocol governance, clearing systems. What has never been done is assembling those instruments for autonomous systems interacting at machine speed, in every sector at once. This article is about that assembly, and the discipline it amounts to. I call it Ecosystem Design. Its most interesting instruments are commitments made visible enough that machines can read them, prefer them, and multiply them — but that is the second half of the story.</p><p><strong>The two layers we have</strong></p><p>Regulation is the oldest reflex. External constraints — the EU AI Act, banking supervision, healthcare rules — draw boundaries around unacceptable outcomes. The caricature says regulation is purely reactive, and the caricature is unfair: the AI Act imposes obligations before deployment, and macroprudential supervision — the regulatory answer to 2008 — explicitly targets system-level risk. But these instruments are sectoral, jurisdictional, and slow by design, built for institutions that change on annual cycles rather than for populations of autonomous systems that change behavior in minutes.</p><p>Governance moved inside the institution. Much of my work with Arnaud Gelas lives on this layer, in what we call Governed Intelligence: keeping AI actions admissible at the moment they occur — grounded in traceable context, authority, provenance, and consequence. Where regulation constrains, governance decides what a system is entitled to know, infer, and act upon — and whether the institution can stand behind the act afterwards.</p><p>Both layers hold the actor at their center of gravity. That made sense while AI systems worked alone; it stops making sense once they coordinate — and the coordination will not respect the legal boundary. A bank, an insurer, a public agency will soon contain dozens of AI systems negotiating priorities, escalating exceptions, shaping each other’s outputs. Ecosystem Design applies at both scales, with one difference: inside the organization, someone at least owns the space between systems. Between organizations, nobody does — which is why the rest of this article concentrates there.</p><p><strong>What AI is actually amplifying</strong></p><p>We describe AI as an intelligence technology, and that is true. It may not be the most important thing about it. The steam engine amplified muscle. The internet amplified communication. AI is doing something subtler: every increase in capability expands the ability to plan, negotiate, allocate, monitor, decide, and organize — at scale, across organizations, across value chains, across entire economies.</p><p>AI is not only amplifying intelligence — it is amplifying coordination.</p><p>That changes the unit of analysis. What a single system can do matters less than what systems cause each other to do. Complex systems were never defined by their components: a forest is not explained by a tree, a city by a citizen, a market by a company. The relationships are the system.</p><p>We already know what machine-speed interaction can produce, because one market showed us. On 6 May 2010, US equity markets lost — temporarily — on the order of a trillion dollars of value inside half an hour. The official post-mortem identified a trigger, a large automated sell program; separate enforcement, years later, reached a trader who had been spoofing the same market. But the report’s structural finding was the important one: the crash was not explained by any single failure. Strategies that each made sense in isolation interacted into a liquidity spiral none of them intended.</p><p>What happened next matters even more for this argument. Markets extended an old interaction-level control — circuit breakers, first introduced after the 1987 crash — down to the level of individual stocks: designed rules that halt trading when prices move too far, too fast, governing a population of algorithms regardless of who owns them. The instrument family is mature, and every member of it shares one dependency: a venue with an operator and a rulebook. Exchanges have one. Clearinghouses and payment networks have one. The agent ecosystems now being built have none — no venue, no operator, no rulebook — and the interaction surface they sit on is about to expand enormously.</p><blockquote><strong><em>The future depends less on what AI systems can do and more on what they cause each other to do.</em></strong></blockquote><p><strong>Trustworthy actors are not enough</strong></p><p>Most AI assurance work asks a reasonable question: can we trust this actor — this model, this agent, this organization? Complex systems add a second one: can we trust the ecosystem those actors create together?</p><p>They are different questions. A forest of healthy trees can still be vulnerable. A financial system of sound institutions can still become unstable, and a supply chain of responsible companies can still produce outcomes nobody chose. Individual quality does not automatically create systemic quality.</p><p>Governed Intelligence — and the wider assurance discipline around it — is aimed at trustworthy actors. Nothing in it guarantees a trustworthy ecosystem, for the same reason that strategies which each made sense in isolation still produced the flash crash.</p><blockquote><strong><em>Admissible actions do not automatically produce desirable emergence.</em></strong></blockquote><p><strong>The third layer</strong></p><p>Ecosystem Design is the study and design of the interaction patterns, incentives, commitments, coordination mechanisms, and feedback loops that shape emergent behavior among AI-enabled actors.</p><p>The assembly is new; the instruments have been in production for decades. Market designers build auction rules that make honest bidding the winning strategy. Macroprudential regulators watch for herding and concentration. Protocol bodies made cooperative behavior on the internet cheap, legible, and repeatable. Clearinghouses place a designed structure between counterparties precisely so one failure does not cascade. Each governs interactions, and each is confined to its home domain. The assembly of those instruments for AI — across boundaries, at machine speed — does not exist. That assembly is the discipline AI now needs.</p><p>A word on governs. Emergence cannot be commanded, and anyone promising to control it has misunderstood the problem. Governing emergence means shaping distributions rather than dictating outcomes: which behaviors get multiplied, which interactions become attractive, which become expensive, and what becomes normal. If this is governance at all, it is governance with open hands — creating conditions, not issuing commands.</p><p>The instruments come in two families — call them containment and cultivation. The containment family is the familiar one: interaction limits and circuit breakers at system boundaries, agent identity attestation, disclosure standards for machine-to-machine dealings. Most of it is the enterprise integration stack repurposed with governance intent, and all of it can fail in known ways — metrics get gamed, standards get captured, observability collides with confidentiality. This family is necessary, and it is not the point.</p><blockquote><strong><em>Containment keeps an ecosystem from burning down. It decides nothing about what grows.</em></strong></blockquote><p>The cultivation family is the generative one, and the one the AI conversation has barely touched — it is what makes this layer design rather than policing. Its core mechanism is the visible commitment. When an actor publicly commits to named behaviors — reciprocity, responsibility, sharing the value its systems create — and that commitment is legible to machines, it becomes a selection signal. Counterparties can prefer committed actors, route interactions toward them, and extend them standing they deny to others.</p><p>Concretely: a procurement agent prefers suppliers whose systems carry verifiable commitments on data use, provenance, and escalation behavior. A treasury agent routes flow away from counterparties that refuse reciprocal disclosure. A compliance agent accepts one attestation as reusable because of the commitments it carries, and sends another to manual review.</p><blockquote><strong><em>Over time, a commitment stops being a label. It becomes access.</em></strong></blockquote><p>Behavior that attracts interaction multiplies. Nobody enforces this from above — the incentive topology does, from the bottom up.</p><p>We have watched this mechanism build ecosystems before. Open-source licenses and Creative Commons were not regulations — they were visible commitments that changed which collaborations were cheap and legitimate to form, and global commons assembled around them without a central authority. Enterprises already run the slow human version of the same mechanism: a SOC 2 report or an ISO certificate is a verified commitment functioning as a condition of access — renewed annually, read by procurement teams, in PDF.</p><p>The same logic is now being tested for AI: voluntary commitment frameworks that attach to a capability and travel with it through the value chain, the way open-source terms travel with code. The point is architectural rather than institutional — the commitment travels with the capability. Every one of those ancestors made commitments legible to humans, who granted access slowly. The change now is that the reader of the commitment — and the grantor of access — is itself a machine: enforcement moves into the interaction fabric and runs at the speed of the behavior it governs, with a consistency human gatekeepers never managed.</p><p>Inside institutions, my work keeps returning to one principle: traceability is the response to acceleration. The ecosystem layer pairs it with a second: coordination is the response to complexity. The first governs what happens inside an institution. The second governs what happens between institutions — which is where the systemic effects live.</p><p><strong>What this opens</strong></p><p>Most governance stops at organizational boundaries. Most systemic effects begin beyond them. That mismatch sets the agenda for what comes next — and what comes next is not another control program.</p><p>For leadership teams, the third layer arrives as design questions, not compliance questions. Which behaviors would you want multiplied in the ecosystems your systems are about to inhabit? What would your organization commit to, visibly, if visible commitment attracted the counterparties worth having? Which interactions should become cheap, expected, and normal between your agents and everyone else’s — and which should become expensive? These questions read strangely in a risk register, and that is the point: they are cultivation questions, not containment questions. One practical anchor survives from the old paradigm — you cannot shape an interaction surface you cannot see, so the first artifact is still a map of who your systems already deal with and on what terms. But the map serves the design, not the audit.</p><p>Who operates the third layer, then? Perhaps nobody — not the way a venue is operated. The space between organizations has no natural owner, and the most interesting possibility is that it does not need one: commitments that travel, protocols that spread, and norms that accumulate may do quietly what no cross-enterprise authority ever could. Whether emergence can carry its own governance is a question that deserves its own article, and it will get one.</p><p>Regulation will reach parts of this layer eventually, and it should. But the terms on which AI ecosystems form are being written now — in contracts, protocols, and defaults, mostly without design intent — and terms set early tend to lock in. The institutions that matter in this story will treat the interaction layer as designable — not as weather to be endured, and not as a rulebook to be awaited.</p><p>The AI conversation has already evolved once — from how do we regulate AI to how do we govern it. The third question is arriving: how do we design AI ecosystems? Behind it sits the one this series exists to explore: what kinds of ecosystems do we want AI-enabled institutions to create together?</p><p>Regulation governs constraints. Governed Intelligence governs actions. Ecosystem Design governs emergence.</p><p>We are only beginning to explore the third layer.</p><p><em>Author’s note: parts of this argument were sharpened by conversations on AI commons, voluntary commitments, and ecosystem-level governance during the UN’s AI for Good summit in Geneva, July 2026.</em></p><p>#AI #AIGovernance #EcosystemDesign #SystemsThinking #Complexity</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dddb7b5cc28a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Spec Is a View]]></title>
            <link>https://medium.com/@navvireich/why-the-spec-is-a-view-d1e0bde1ff9b?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1e0bde1ff9b</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[spec-driven-development]]></category>
            <category><![CDATA[software-architecture]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Sat, 16 May 2026 17:23:43 GMT</pubDate>
            <atom:updated>2026-05-16T17:28:23.605Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m9sGFk_ijxgvHni6IL3iKw.png" /></figure><p><em>Large enterprises and the consultancies that serve them are standardising on spec-driven development for AI engineering. The decision is right. The spec is the controlled interface engineers and agents should work through. But for AI that acts under supervision, the spec is a view, not the source of institutional reliance. Under it sits a governed claim graph: atomic assertions with canonical identity, provenance, authority, and standing the institution can defend at the moment of the act. The substrate is not an artifact built beside SDD — it is what mature SDD becomes when assertions are given identity, dynamic standing, and lifecycle independent of any one spec, and the platform enforces that independence at both build and runtime gates. SDD’s interface does not change. The source of institutional reliance moves to where governance can attach to it.</em></p><p>SDD gives regulated AI delivery something it badly needs: a controlled interface. Specs can be reviewed, versioned, scored, traced, and handed to engineers or agents. Standardising on it is the right move. It is also why the next constraint appears immediately: a spec can be controlled without every assertion inside it being defensible.</p><p>The spec is governed as a delivery artifact, not as a set of institutionally relied-upon assertions. For AI that drafts a memo for a human to check, the distinction does not bind. For AI that acts on the assertions inside the spec, it does. The constraint past SDD is admissibility: whether the institution can defend the basis on which an AI system acted, after the act, to a supervisor, an auditor, or a court.</p><p>The short answer is the claim. The longer answer requires being precise about what a claim is, why a well-formed requirement in mature SDD is not yet one, and how the spec relates to the substrate underneath. The cleanest engineering statement of the relationship is a database one — the spec is a view, claims are the base tables — though the analogy is literal in fewer respects than the strong form might suggest.</p><h3>Two granularities, but not the obvious one</h3><p>The granularity argument in its weak form — “the spec is a document, the claim is an assertion” — is too easy to defeat. Mature SDD already decomposes specs into atomic requirements with IDs, owners, status, source links, validity, traceability, and lifecycle. A serious SDD program does not govern only the document; it governs every requirement inside it. The spec-claim distinction is not really about atomicity. Both can be atomic. The distinction lies in three properties that requirements, even when well decomposed and managed in shared libraries, do not by themselves carry. Once those properties are present, the artifact has become a claim — whatever the platform calls it.</p><h3>Descriptive, not prescriptive</h3><p>A requirement is prescriptive. It states what the system shall do: enforce this rule, present this disclosure, retain this record. It is an instruction to the system. In formal terms, it is deontic.</p><p>A claim is descriptive of an institutional commitment to a proposition. It states what the institution is prepared to rely on when the instruction is executed: this jurisdiction is currently treated as high-risk; this interpretation of regulation X is the one the institution has stood behind; this threshold is the operative one for this product in this jurisdiction. In formal terms, it is epistemic.</p><p>The distinction is not that claims are always factual and requirements always normative. Many claims are normative interpretations. A regulatory rule, once the institution adopts a position on it, is a claim about institutional reliance. The requirement is the downstream instruction that implements the rule. Governing the requirement does not govern the underlying claim. For assertional requirements — those whose correctness depends on propositions the institution must defend — the requirement is downstream of claims. Not every engineering requirement has this structure; latency budgets, interface conventions, and UX choices remain ordinary requirements. The claim substrate is load-bearing where an AI act relies on assertions about the world, the customer, the law, or institutional policy.</p><h3>Canonical, not spec-scoped</h3><p>Mature SDD can reduce spec-scoping through reusable requirements, shared libraries, derived requirements, and impact analysis. That is real, and it solves the duplication problem. It does not solve the canonical-reliance problem. Even with shared libraries, the unit of governance remains the requirement-in-a-baseline or requirement-in-a-release. The same belief can be baselined into release A and release B; release A continues operating under the older baseline; release B updates. Both remain “approved” in their contexts. The institution now has inconsistent live reliance, and a project-centric library cannot, on its own, answer the institution-centric question: what was the single authoritative position on jurisdiction X at 14:03 UTC last Tuesday, across all active systems.</p><p>A claim has identity, lifecycle, and standing outside any one spec or release. Many specs project the same claim. Its standing can change without those specs being edited. Cross-spec contradiction becomes detectable not because someone remembered to compare two specs, but because both projections resolve to claims in the same governed substrate. This is the difference between knowing one spec is internally consistent and knowing the institution is consistent about what its delegated systems are allowed to rely on. Only the second is the question a supervisor will ask after a delegated act.</p><h3>Dynamic standing, not static status</h3><p>A requirement has status — a stored field. “Approved.” “Deprecated.” Status changes when someone changes it, or when an automation rule changes it on someone’s behalf. The transition is an act, recorded as such, and the field is overwritten.</p><p>A claim has standing — a function. Standing is computed from the claim’s dependencies: the current authority of its source, the validity window, any contradicting claims, the most recent interpretation that supersedes earlier ones. The claim’s content and provenance are immutable — event-sourced. Standing is a temporal projection over those events, recomputed as dependencies move. When the world changes, the previous version of the claim is not edited; new events arrive, and the standing projection for downstream reliance is recomputed.</p><p>The distinction is not whether the value is stored or automated. A mature ALM platform can drive requirement status with webhooks and feeds, and it should. The distinction is what the system can answer.</p><p>A mutable status field plus a human-readable audit log is not designed to recompute <em>entitlement</em> from the full dependency graph at an instant. The forensic question delegated AI must answer after the act is: which version of claim C42 was the system entitled to rely on at 14:03:15 UTC last Tuesday, given that two contradictory authority updates arrived within the same hour? An event-sourced substrate, where typed assertion events, authority events, supersession events, and contradiction events are first-class and replayable, answers this as a native capability. Reconstructing the same answer from a mutable record plus a separate audit log is forensic assembly — expensive, slow, lacking the integrity of a native temporal query, and unsuited for a runtime gate. The contrast is not between “ALM” and “claim substrate.” It is between current-state requirement management and bi-temporal assertion governance. If an ALM platform stores assertion-level events, authority and dependency events, valid-time and transaction-time, and deterministic projection rules, it has implemented the substrate pattern — whether or not it uses the word claim.</p><p>That query is the architectural test. Not “can the platform automate a transition” — it can. The test is whether the platform can reconstruct, without forensic assembly, the exact basis the system relied on at the moment of the act, distinguishing what was effective in the world from what the institution had recorded and was entitled to rely on at the time.</p><h3>Why the spec is a view</h3><p>Once the three properties are in place, the database statement becomes precise.</p><p>Claims are the normalized addressable unit. The spec is a view — a named projection over claims, rendered on demand, not stored truth. A filed, versioned spec is a materialized view: the same projection, cached for convenience, and at the same time a controlled record of what was rendered and approved at that point.</p><p>The analogy is literal in one respect, and only one: integrity constraints can be attached and evaluated consistently only at the normalized addressable unit, not at the rendered document. Constraints sit on the substrate; the view inherits the standing the substrate can supply, through governed projection rules. It is not literal in the stronger sense that institutional assertions behave like database facts. Many claims are interpretations, scoped to a jurisdiction, valid for a window, contestable on their face. Normalization gives them a place to be governed. It does not pretend to guarantee truth.</p><p>A minimal claim model makes the analogy concrete. The immutable part is a stream of events about a claim:</p><blockquote><em>ClaimAssertionEvent(event_id, claim_id, content, type, scope, source_id, authority_id, asserted_at, valid_from, asserted_by)<br> ClaimSupersessionEvent(prior_claim_id, successor_claim_id, basis, recorded_at)<br> ClaimInvalidationEvent(claim_id, reason, valid_to, recorded_at)<br> ContradictionCandidateEvent(claim_id_a, claim_id_b, predicate, scope, recorded_at)<br> AdjudicationEvent(candidate_id, disposition, authority_id, recorded_at)</em></blockquote><p>The standing of the claim is not a column on any record. It is a projection:</p><blockquote><em>StandingProjection(claim_id, valid_time, transaction_time, action_class, scope, state, derived_from_events[], projection_rule_version, evaluated_at)</em></blockquote><p>The two distinct times are essential. Valid_time is when the assertion was effective in the world. Transaction_time is when the institution knew it and was entitled to rely on it. Most forensic questions about delegated action need both: a regulator may publish a correction with an effective date earlier than the publication date; a court may clarify a rule retrospectively; an authority may issue an interpretation at one time, effective from another.</p><p>Projection rules are themselves governed artifacts. The function that turns events into standing is versioned, approved, and reconstructable as of any transaction time. A standing answer is admissible only when the rule version used to compute it is identifiable and the rule itself was in force at the transaction time being queried. The substrate that governs claims must also govern the function that projects them.</p><p>A spec that says “the agent must route payments above threshold X to jurisdictions in the high-risk list for enhanced due diligence before release” projects against several claims: a definitional claim about which jurisdictions are currently in the list, a threshold claim about the operative amount, an authority claim about who can change the list. If HM Treasury publishes an updated list, the old membership claims are not edited. New source events arrive. New membership claims are asserted; old ones receive supersession or invalidation events. Standing for any spec projecting those claims is recomputed from the new dependency graph. The historical standing of the old claim remains queryable for acts taken while it was current. The spec is not edited; the basis the agent resolves at runtime has changed.</p><p>This is recomputation of standing, not authorization of behavioural change. Whether the system may act on the new standing is the action gate’s question, addressed in a separate piece. The architecture does not bypass change control; it refines it. The action gate’s policies, themselves under formal change control, determine the blast radius of a standing change: a minor definitional update may be allowed for dynamic consumption, while a change to a load-bearing regulatory interpretation routes dependent systems to a “requires recertification” state and triggers a formal release process.</p><p>The same architecture applies outside financial services. In aviation, an FAA airworthiness directive changes the validity of a maintenance-procedure claim a control system was relying on. The maintenance spec is not edited; the claim about which procedure remains current receives a supersession event with its own valid_time and transaction_time. The system’s basis at the moment of any subsequent action is reconstructable as of that moment. The supervisory question — whether the action was admissible given what the institution was entitled to rely on — is the same. Different regulator, different forcing function, identical architectural pressure.</p><h3>Contradiction becomes addressable</h3><p>Contradiction does not fall out of putting claims in one store. Detecting it requires shared predicates, scopes, validity windows, jurisdictional context, authority precedence, and sometimes human adjudication. Two claims may look contradictory but differ by product, customer class, or effective date.</p><p>What the substrate does is make contradiction addressable. Without it, claims live in the documents that contain them; cross-document conflict is a manual review exercise that scales poorly. With it, contradictions can be queried, surfaced as candidates, routed to the right authority, and recorded as resolved or pending. The substrate does not decide the conflict; it provides the place where the conflict can be represented, scoped, owned, and adjudicated. The goal is not perfect institutional consistency. The goal is <em>addressable</em> inconsistency: conflicts are visible, scoped, owned, and prevented from silently supporting delegated action.</p><h3>The round trip</h3><p>A spec contains assertional content that can be decomposed into claims, alongside narrative and design material that should remain in the spec. The conversion runs both ways.</p><p>Spec to claims, by decomposition. Take the spec’s assertional content — the rules, definitions, interpretations, eligibility criteria, thresholds, authority references. Extract each as a candidate claim. Stamp provenance: which spec it came from, which source that spec drew on, who authored it, when, by what method. Score and tier: assign the standing each claim is permitted to be used for, given its source authority and the consequence class of actions that may rely on it. Wire the typed relationships: supersession, derivation, supports, contradicts.</p><p>This work is not new. The institution already pays for it — expensively, chaotically, and reactively. The work of interpreting a regulation, scoping it to a product, deciding what counts as authoritative, mapping it to controls, and reconstructing the basis after an audit query already happens. It happens in ad-hoc legal review, in control mapping workshops, in architecture sign-off, in panic-driven post-incident evidence assembly. None of it is captured as durable, reusable, governable output. What changes here is not the work; it is where the work lands. Call the function the Claim Analyst role — business analysis, domain expertise, and where the content carries legal or regulatory weight, qualified legal judgment, integrated into the existing spec-review gate. Reviewers stamp each assertional paragraph with claim identity, authority, scope, and standing tier as part of approving the spec, not as a separate process. Initial implementation is expert-led because legal and policy interpretation are expert work. The maturity path is tooling — semantic parsers, ontology-guided authoring environments, assistive decomposition — that handles the routine majority of claims and reserves expert adjudication for the genuinely novel cases. The architectural saving is that this work moves from reactive fire-drill to predictable upfront control, with a durable output the institution can stand behind.</p><p>This is not a routine database operation. Decomposing prose into governed assertions is interpretation, ontology alignment, authority mapping, judgment. It is the central engineering discipline of this architecture. It deliberately separates the institution’s governable assertions from the narrative that explains them. The claim substrate governs the <em>what</em>: the propositions the institution stands behind. The spec retains the <em>why</em>: the rationale, design intent, and explanatory context that engineers and reviewers need. Rationale is linked to claims as controlled narrative evidence; it is not the assertional source of truth, but it remains part of the admissibility record.</p><p>Claims to spec, by projection. Once claims are governed, the spec is re-rendered from them on demand. The document still exists; the engineering workflow still gets the spec interface it expects. Every assertional line in the spec traces to a claim with a standing and a provenance chain. Projection itself is a controlled artifact: it requires templates, ordering, and rules about which claims to include and exclude.</p><h3>The substrate is a platform capability</h3><p>The substrate is not a parallel manually maintained version of the spec; the construction discipline is decomposition, not duplicate authorship. But it is a platform capability — a governed temporal claim graph with a data model, lifecycle, APIs, ownership, and operating model. It needs budget. It needs a team. In the transition, it is additive work. At maturity, it redirects work that the institution already performs — inside specs, in cross-spec review, in audit response, in evidence-pack assembly — so that the governed assertion becomes the durable object and the spec becomes the projection.</p><p>Event sourcing is the logical model; non-repudiation requires the controls around it. For regulated use, the event stream must be append-only under privileged controls, with event identity, signer identity, transaction-time stamping, retention policy, and tamper-evident integrity through hash chaining or equivalent ledger controls. Immutable events as an implementation pattern is not the same as evidence non-repudiable to an auditor; the controls have to be there.</p><p>The architectural saving is not that the work disappears immediately. It is that the institution stops paying for claim work twice: once trapped inside documents, again reconstructing it for an auditor. The redirection only becomes real when claim authoring becomes the path of least resistance and runtime gates refuse unsupported assertions. The justification is not operational efficiency. It is risk mitigation and capability enablement — the cost of being able to deploy delegated AI into consequential decisions and defend the basis afterwards.</p><p>The substrate only survives if it becomes a control point. The adoption invariant is direct: no consequential delegated action may execute, and no spec may be promoted for such action, unless every relied-upon assertion resolves to a claim with current standing or an explicit, governed exception. Without that invariant, the graph becomes another enterprise metadata initiative — correct in principle, abandoned in practice. With it, the graph is in the promotion path, and the claim layer is the path of least resistance for every team that must ship. The invariant is a target state, adopted incrementally — enforced first on the highest-risk, highest-scrutiny systems where the business case for this level of control is undeniable, then extended as the platform proves its ability to de-risk deployment and shorten supervisory response.</p><p>The invariant is operational, not rhetorical. Promotion and runtime enforcement need four controls: action classification, claim-resolution evidence, fail-closed semantics, and exception governance. A consequential action carries a declared action class. Before promotion, every assertion referenced by the spec must resolve to claim IDs with permitted standing for that action class — and “every relied-upon assertion” means assertions explicitly referenced by the spec, retrieved into the action context, or emitted as part of the system’s decision basis, not every latent association the model touches. For consequential action, the system must be architected so that the actionable basis is constrained to claim-resolved context, approved tools, and emitted decision evidence; latent model knowledge may assist language production, but it may not be the authoritative basis for a consequential act unless surfaced and resolved to standing claims. At runtime, the action gate either obtains a current standing decision or accepts a signed standing token within its freshness window; otherwise it fails closed or routes to human review. Exceptions require owner, authority, scope, expiry, consequence class, reason, and residual-risk acceptance, and are reportable as control breaches if they persist beyond their approved window. The exception path is itself governed: an exception is a special-form claim with standing, not an off-ledger workaround.</p><p>This capability can be delivered by a new platform or by an incumbent vendor, but only if the platform re-architects to enforce the separation of immutable assertion events from computed standing projections. A system that continues to mutate requirement records in place, merely adding richer audit logs or metadata, has not adopted the pattern — regardless of how mature its workflow or how rich its traceability.</p><h3>What this means for SDD</h3><p>The SDD interface does not change. Engineers and agents still work through specs. Reviews, quality scoring, and traceability all live there for good reasons. The consumption interface stays recognisable.</p><p>What changes is the production model behind the spec, and the failure modes engineers see. Authoring of the durable artifact moves to the claim layer. Approval of reliance happens against claims. The spec is rendered, not maintained as the source of truth. Build-time checks can prevent deployment of systems that reference claims without permitted standing. Runtime gates can re-evaluate standing at the moment the system commits to the consequential act. None of this asks engineers to relearn a spec format. It does require the platform to know which is the source and which is the view, and to fail closed when standing cannot be evaluated.</p><p>Adopting the substrate is not a migration off SDD. It gives SDD the system of record for institutional reliance the document format was never built to be.</p><h3>Two axes, kept separate</h3><p>Axis one is the unit of governance: the spec, or the claim. This piece is about that axis.</p><p>Axis two is which governance: data governance (the floor), epistemic governance (claim governance), action governance (the gate at commit). That axis is a different piece in the architecture. Claim standing supplies the admissible basis on which the later action gate can decide. It does not, by itself, authorize the act. What the gate does with claim standing is the subject of the next piece; that the gate must be able to ask for claim standing as of any past instant is the architectural requirement this piece establishes.</p><p>Both axes belong in any full architectural answer, but on separate slides. If they land together, the audience cannot tell whether the distinction being drawn is spec-from-claim or data-from-epistemic, and the argument loses precision.</p><h3>Where it sits in the stack</h3><p>From the top: humans and agents act. The spec is the interface they work through. The substrate — the governed claim graph — is the system of reference the spec is projected from. Data governance is the floor beneath the substrate, governing the source data and evidence claims depend on; claim governance adds the standing and authority that data governance alone does not provide.</p><p>Spec-and-claim is the relationship between the middle two layers. SDD governs the interface; the substrate sits one layer below; the data floor sits below the substrate.</p><h3>Why this binds</h3><p><em>This piece is written by an architect who designs supervised AI systems for regulated clients at a large systems integrator. The work is practical: defining the architectures clients can actually deploy under supervisory examination. The position codifies an architectural pattern that has become necessary in regulated and safety-critical AI; it is not tied to any specific product implementation.</em></p><p>The structural point is direct. A spec governed only as a delivery artifact does not let the institution defend what it acted on, because the act commits on the assertions inside the spec, not on the spec as a document. For drafting, that gap is tolerable: a human reviewer remains accountable for independent judgment over the propositions the system used. For delegated action under supervision, the human is no longer in the transaction path before the act. There is no equivalent way to close that gap without giving each relied-upon assertion canonical identity, current standing, and addressability outside any one document.</p><p>The argument is strongest where three conditions coincide: delegated machine action, material consequence, and external accountability. Financial services in scope of supervisory examination is the visible current case. The same architectural pressure appears in defence, aviation, medical devices, autonomous vehicles, safety-critical industrial control, and public administration — any setting where, after a consequential act, an institution must defend not only what the system did but the propositions it was entitled to rely on when it did it. Regulation is the visible forcing function. It is not the boundary of the architecture.</p><h3>What kills the architecture</h3><p>The honest test is what would make this wrong.</p><p>The claim is not that the capability must be delivered by a separate product or by a system labelled “claim substrate.” The claim is that any architecture adequate for delegated action under supervision must provide a minimum capability set: canonical assertion identity, immutable assertion and dependency events, bitemporal standing projection, authority-aware supersession and contradiction handling, and enforcement at promotion and runtime. Ordinary mutable records plus separate audit logs do not provide this natively. They record changes to state; they are not the authoritative temporal dependency graph from which standing is derived. Reconstructing the entitlement query from them is forensic assembly, not a native query. If an ALM platform provides those capabilities — if the audit log is itself the immutable assertion-event substrate, with typed dependency events, bitemporal semantics, and deterministic projection rules — the platform has adopted the substrate pattern under another name. If mutable records and audit logs can provide the same capability without those properties, this argument fails. The observable test is whether the institution can answer the entitlement query natively, at runtime or audit speed, without reconstructing the dependency graph by forensic assembly.</p><p>The architecture also survives only as long as its ontology does. With the enforcement invariant in place, the long-term failure mode is no longer abandonment — it is ontology rot: semantic drift in the core predicates, scopes, and types that underpin the graph. An authoritative system operating on a decayed ontology produces verifiably wrong outcomes, which is more dangerous than the messy human system it replaced. Governing the ontology — the predicates and their meanings — becomes as critical as governing the claims, with its own discipline, its own ownership, and its own review cadence.</p><p>The supervisory test is what supervisors actually require institutions to produce after a delegated act, as evidence of the basis. If approved specs plus process traceability plus runtime logs prove sufficient, the substrate is not load-bearing for those cases. The substrate becomes load-bearing precisely when an institution must produce assertion-level, time-specific evidence at scale, under changing authority and source conditions — which is where delegated AI under supervision is heading.</p><h3>Closing</h3><p>The decision to standardise on spec-driven development is right. The spec is the controlled item. The reviews, the quality scoring, the traceability all live there for good reasons.</p><p>What the decision does not settle is what makes the spec defensible. The spec is a view. The view inherits whatever standing the substrate carries. For AI that drafts, the substrate can stay implicit; the human reviewer absorbs the gaps. For AI that acts under supervision, the substrate has to be explicit, governed, queryable, and addressable at the level of the individual claim — because that is the level at which standing changes, contradiction shows, and authority attaches.</p><p>A spec in a mature SDD process is a powerful control over development. A spec rendered from a governed claim graph becomes a defensible control over delegated action. The first ensures the system was built as intended. The second ensures the institution can defend the basis on which it acted.</p><p>SDD gives the institution the interface it needs. The substrate gives it the source of institutional reliance it has been missing.</p><p>The application keeps its interface. The architecture gives it the database it should have been running on.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1e0bde1ff9b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Admissibility]]></title>
            <link>https://medium.com/@navvireich/admissibility-5c7c7deb64dd?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c7c7deb64dd</guid>
            <category><![CDATA[regulated-industries]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Fri, 15 May 2026 20:18:13 GMT</pubDate>
            <atom:updated>2026-05-15T20:20:38.712Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vSfoeD2LW5Zx-uB2oDkX0w.png" /></figure><h3>The constraint past capability for AI in regulated industries</h3><p><em>For a defined class of regulated AI work — software engineering agents that commit code to production, operations agents that resolve cases on live files, compliance agents that work alerts to closure — the binding constraint on deployment is no longer the model. Capability has cleared; what stops deployment now is admissibility — whether, before the act, the institution can treat the model’s output as a permissible basis for action, and stand behind that choice afterwards. Meeting that condition requires what the institution previously held in the human reviewer: at the moment of the act, a basis that is still well-sourced and uncontradicted, authority to rely on it, context recorded, and evidence produced. The discipline is old. The conditions under which it has to hold are new.</em></p><p>The class is specific, and it has to be defined before the argument can lean on it: work with bounded inputs, a repeatable institutional standard for what good looks like, an output a competent person can review, and an existing human process against which the model’s error rate can be measured. Software engineering agents that commit code to production, operations agents that resolve cases on live files, compliance agents that work alerts to closure — work of that shape. It is not open-ended judgement under novelty; there, capability is still the constraint, and I am not claiming otherwise. The claim is also not that every stalled deployment is a governance problem — plenty stall on latency, integration, data access, or cost. It is narrower and more exact: in enough bounded delegated tasks to matter, the institution accepts the model’s output as good enough, and then cannot let the system act on it — because it cannot make that output something it is allowed to rely on.</p><p>It stalls at the point where the output stops being a draft for a person to check and starts being something the institution acts on. Four questions sit at that point, and capability does not settle them. Can we prove what the system decided, and on what basis? Can we say what it is allowed to act on, and on whose authority? Can we show a supervisor the context the decision was made in? Can we produce the evidence the act is supposed to leave behind?</p><p>Those are admissibility questions. I mean admissibility not in the courtroom sense but an operational one: whether, before the act, an institution can treat the model’s output as a permissible basis for action — and stand behind that choice afterwards. Capability decides whether the answer is good. Admissibility decides whether the institution may rely on it. They are different axes, and they are sequenced — in logic, not in project order: even after the institution judges the output good enough for the task, a separate question remains, whether it may act on it. For regulated deployment, the second question is the one that binds.</p><p>Those four questions are not four separate problems. Put to a supervisor, an auditor, or a court after something has gone wrong, they collapse into one: at the moment the system acted, did the institution have a defensible basis for the act, and could it have known if it did not. That is the operational invariant admissibility comes down to. No consequential action commits unless the claims it depends on still hold — well-sourced, current, and uncontradicted enough for what the act will do — the authority to rely on them is in place, and the act records the context it relied on and the evidence it produced.</p><p>Two things about this invariant are load-bearing and need stating directly. First, the four conditions are not a checklist — they have to hold together at the moment of commit. Any one failing makes the act unbasable in retrospect, which is why supervisors, auditors, and courts collapse them into one: from the outside, after a failure, no party distinguishes a stale basis from an unclear authority from an unrecorded context. Any of them is enough to make the act unbasable. Second, the check has to fire before commit, not after. There is no remediation path that returns a committed delegated act to its pre-commit state — the notice has issued, the payment has cleared, the alert has closed. A check that fires afterwards is a finding, not a control. The invariant names a preventive control on delegated action, and that is the shape the rest of the piece is about.</p><p>It is not a new principle. It is the old principle of not acting without a basis, made to hold at machine speed, with no person in the path to hold it.</p><p>I write this from the delivery side of regulated transformation, not from a lab — and I am building in this space. The stake is specific: I have a commercial interest in the claim that admissibility is a real and distinct requirement. Read this as a bet from someone with that interest, not a neutral survey — and note that the architectural claim and the venture claim can come apart. If existing platforms absorb this invariant as ordinary plumbing, the architecture was right and there is no separate company to build; I think that is a live possibility. The bet rests on a pattern, and I should be plain about its status: this is not survey evidence, it is a delivery-side observation, consistent enough across different regulated settings to design for. The pattern runs like this. The model passes its evaluation. The business owner wants to proceed. Legal asks which policy interpretation the model relied on. Risk asks who owns that interpretation. Audit asks whether the decision context can be reconstructed a year later. The answers turn out to be scattered across policy documents, committee minutes, ticket threads, data permissions, and somebody’s memory. The institution may have a basis for action somewhere. It does not have one the system can use before the act.</p><h3>The constraint past capability</h3><p>The claim has to be precise, because the loose version is wrong. Capability is still a constraint — a model that hallucinates, reasons badly, or breaks on the edge cases produces output an institution cannot defend. The argument is about what binds <em>after</em> the model is good enough, so the threshold has to be defined without leaning on admissibility to define it.</p><p>Here is the clean version, and it has two parts the loose version runs together. The first is capability proper: the model’s raw task performance, its residual error profile before any institutional check, measured against the existing human process. That is a property of the model on the task. The second is absorption: whether the institution can catch or tolerate the errors that remain, within the operating process it already runs. That is a property of the operating model, not the model. It is a question of operational tolerance — whether the residual mistakes are manageable. Whether the institution may then rely on the output is a different question, and the one this piece is about. A task is ready for a rung when capability and absorption both hold. What readiness does not settle is whether the institution may then let the system act. Past the point where it clears, a better model still helps — cleaner output, fewer errors to govern — but it is no longer what decides the deployment date.</p><p>That threshold is not the same at every rung, because there is a ladder, and each rung has to clear on its own terms — nothing about clearing one rung implies the next. The advisory assistant drafts for a human who checks everything: the bar is the cost of that review, and for many standard drafting and summarisation tasks it cleared some time ago. The workflow participant routes and triages inside a controlled process: the bar is higher, because the catch is now process controls and institutional rules the output has to conform to, not a person reading every line, and it has cleared for some tasks and not others. The delegated operator acts: the bar is highest, because ordinary review is the thing being removed. Admissibility is what stops the operator — and increasingly the workflow participant — from moving. A more capable model lowers the burden the invariant imposes. It cannot remove it, because it cannot decide whether the institution may act on what it produced.</p><h3>The loudest critic is still inside the frame</h3><p>The strongest version of the case that the field has it wrong is architectural. The argument runs that today’s models are the problem — that better grounding, planning, and world models will produce systems whose reasoning is inherently more traceable and predictable. Grant the strong form of it. A better-grounded model would automate much of the epistemic work: finding the policy, tracking which version is current, citing its sources, exposing its own uncertainty, proposing the record an act should leave behind. That helps, and it is real.</p><p>But it would not finish the job, and the reason is not a limit of the machine. When a bank acts, a named part of the institution is answerable for that act — to a supervisor, a court, a customer. That answerability has never moved to the tool, and nobody serious argues it should: the rules engine does not own the lending policy, the trading system does not own the mandate. So the accountability point on its own does not force anything; institutions have always carried it. What forces the architecture is what changes when the human leaves the path. While a person was there, the institution could carry its accountability the way it always had — reconcile the basis for the act through that person, through approvals, through review after the fact. Remove the person, and the institution still has to be able to show, before the act, that what the system relied on met its own conditions for relying. A perfectly grounded model can tell the bank what the policy appears to say. It cannot be the thing that owned that reading when the regulator asks who decided. The capability critique, in every version, is answering how to make the machine better. Admissibility is answering whether the institution can show it was answerable — and a better machine helps with the preparation but does not settle that.</p><h3>What admissibility requires, and what a claim is</h3><p>Admissibility does not need a new philosophy. The invariant names four things, and they have to be present, checkable, and binding at the moment an AI system acts — they are the four the opening questions were already pointing at. A basis: the claims the action rests on. An authority: who, or what, may rely on those claims, for this act, under what conditions. A context: the bounded record of what informed the decision, captured before the act rather than reconstructed after the failure. And evidence: the record the act itself produces, in a form the institution is prepared to defend.</p><p>Together they describe one event — the reliance event — not the anatomy of a single object. The basis is something the system holds; the authority is a rule it is checked against; the context and the evidence are records, one captured going in and one written coming out. None of this is new as a control idea — every consequential institutional decision has always needed all four. What is new is the demand that they be machine-operable, in the path of the action, before it commits. They are not a taxonomy of everything AI governance has to do; fairness, privacy, resilience, and human oversight do not reduce to them, and an act can be inadmissible for reasons that have nothing to do with the basis. They are the minimum the gate cannot defer.</p><p>Of the four, the basis is the one most often missing in the form delegated AI needs. Authority rules, context capture, and evidence records are familiar institutional machinery; institutions also govern plenty of basis-like things already — case facts, rule predicates, eligibility determinations, reason codes, master data. What is rarely present is the basis as a maintained, assertion-level object: an assertion whose standing can change on its own, independently of the workflow that first used it, and whose change can stop a later act before it commits. That object is the claim. A model’s output — fluent, structured, possibly correct — is not yet one. It becomes a claim when it can answer four questions the gate cannot defer.</p><p>Where did it come from: a claim carries its provenance, a traceable chain from source through extraction method to an accountable validation point — a role, a rule, an approved process — never an anonymous assertion. How much weight it currently carries: a claim carries a standing, derived under rules the institution owns and can audit, not a confidence number the model reports about itself and not a label a reviewer attached by hand. Those rules are not exotic; they are the institution’s own heuristics made executable — a claim from a filed regulatory document starts higher than one from an internal memo, corroboration from an independent source raises it, age lowers it. Standing is what source, validation, corroboration, and freshness add up to. What the claim does not carry is the authority to use it: that is a separate rule the gate applies, because the same well-sourced claim can be strong enough to draft on and not strong enough to act on. Whether it is still current: a claim carries a validity window, monitored, because knowledge goes stale at different rates, and an institution that treats the stale as current acts on something that has quietly stopped being true. And whether anything live contradicts it: a claim carries its contradiction status — what it conflicts with, within a source universe the institution has committed to watch, held open as typed information the gate can act on rather than silently resolved into one tidy field. Held open is not unresolved forever — it means the conflict stays visible, and is resolved, tolerated, or escalated under a rule fit to the consequence, instead of disappearing before anyone decided it should.</p><p>None of these four is an invention. Provenance is the territory of data lineage; a derived standing is what data-quality scoring does; a validity window is a familiar idea; holding contradictions open is what truth-maintenance systems have always done. Regulated institutions have also automated consequential decisions for decades — credit scoring, fraud declines, sanctions screening — where the facts and the rules were already formalised into a decision engine. None of that is the gap.</p><p>The gap appears where a delegated system relies on assertions that were never formalised that way: a policy interpretation, a document’s meaning, a fact extracted from unstructured material, a reading of how a rule applies to this case. For that kind of basis, the institution’s reconciliation — is this current, is it authorised, does anything contradict it — did not sit in a system. It often sat in the human reviewer, who supplied the join between the output and the institution’s live knowledge that no formal control represented. Remove that reviewer, and the join does not vanish; it becomes unowned. So three things have to happen, in order. The reconciliation has to become explicit, rather than implicit in someone’s judgement. It has to become machine-operable. And where the basis can go stale or be contradicted between the moment it is approved and the moment it is used — which is exactly the case for policy readings and extracted facts — it has to be checked against current state in the path of the act, because there is no longer a person between approval and action to notice. That last step is the one that forces the architecture, and only for that class of basis. The discipline it asks for is old — do not act without a basis. What is new is having to keep it without the person who used to.</p><p>This does not launder the model’s uncertainty. The assertion a model extracts is still probabilistic, and admissibility cannot make it true. What it does is make the uncertainty inspectable and stoppable. A claim’s standing is derived from its evidence under rules the institution owns, versions, and can challenge — not taken from the model’s self-reported confidence. A fluent answer with a high confidence score is not an admissible basis; an assertion whose source, corroboration, and currency have been scored against governed rules is. The institution still carries the risk. The difference is that it can now see it, weigh it against the consequence, and decline.</p><p>So what does enforcing the invariant require, concretely? Two functions, and the order matters — functions, not products. The first is somewhere claims are kept and maintained: not a new enterprise knowledge base, and not a system of record for everything the institution knows, but maintained state for the claims a given set of AI-mediated actions is permitted to rely on, and their current standing. It is specialised because it has a job an ordinary data store does not — it recomputes a claim’s standing when corroborating or contradicting evidence arrives, watches validity windows, and keeps contradiction live rather than letting it settle. A claim is not a record written once and read later; it is a thing whose standing can change between Monday and Tuesday without anyone touching it. The second function is a gate: the point where a proposed action is checked against the claims it depends on, before it commits, against their standing as it is at that moment. The gate is the visible move; the maintained state is what makes the gate mean anything — a gate with nothing live behind it is a rule firing on whatever it was handed.</p><p>I have been calling the two together a layer, and the word needs disciplining. A layer here is a logical responsibility, not a product you buy or a system you stand up beside the others. It can live inside a workflow engine, a decision platform, a knowledge graph, or be assembled across systems an institution already owns. And it does not need to start at enterprise scale: the unit of adoption is one delegated use case — one action path, the claims that path is allowed to rely on, the owners of those claims. What has to be true is the function. The packaging is open.</p><p>The check at the gate is not pass-or-fail. It is graduated: an action can proceed, proceed with the uncertainty logged, proceed in part, pause for a specific missing claim, escalate to a named human authority, or be refused outright. The consequence class of the action sets how high the claims’ standing has to be before any of that is allowed — a low-consequence action can run on a thinly-supported claim, a critical one cannot. And the gate rarely tests one claim. A real institutional action rests on a bundle — a covenant breach notice depends on the borrower’s filed accounts, the covenant definition, the calculation period, any waiver, and the authority to issue — and for the claims that are material and not substitutable, the act can be no stronger than the weakest of them.</p><p>Take the case that shows why the state has to be live, and take the hard form first — the interpretive one, which is where the architectural gap actually bites. A bank operates a delegated AML investigation agent that closes alerts under a current internal interpretation of sanctions guidance: a particular ownership structure does not amount to beneficial-ownership control. Six weeks later, a regulator-issued advisory refines the interpretation — structures of that shape do constitute control. Nobody re-opens the closed alerts; there is no reviewer between the agent’s next decision and the act. But the interpretation the prior closures relied on has had its standing changed — by a named authority inside the institution, under a rule the institution has set for how external guidance updates internal interpretations — and the gate holds the next alert that depends on that interpretation, instead of letting the agent close it. The architecture does not adjudicate; it does not decide that the advisory supersedes the prior reading. It makes sure that once the institution’s own authority has, the dependent act stops.</p><p>The simpler factual version is the same mechanism on easier ground. A lender runs a system that prepares covenant breach notices. It relies on the claim that a borrower’s filed accounts are current and validated. A later filing supersedes those accounts. Nobody re-opens the case; there is no reviewer between approval and the next notice. The claim’s standing has changed — the accounts it points to are no longer current — and the gate holds the next notice on that basis instead of issuing it. The change is not magic: a person or a feed has to enter the new filing into the system. What the layer does is make the consequence automatic once it is entered — it does not go and find the new filing in the world; it makes sure that once the world has changed, the act that depended on the old state does not quietly proceed. The shape is the same as the interpretive case. The difference is that for the factual version, existing master-data and reference-data controls already do part of the work; for the interpretive version, no existing control sits in the path.</p><p>The acting system cannot lift its own gate. It can request escalation; it cannot grant itself the authority to proceed. Override is a separately owned, separately logged path — otherwise the gate is theatre. And every pass through the gate writes its own record: what was relied on, what was excluded and why, what the gate decided, what the system then did. That record is the evidence, produced as part of the act rather than reconstructed after it.</p><p>A record reconstructed after the act is not this. A model registry, a policy library, a control dashboard can support admissibility, but they are not it unless they can test the standing of the relied-on claims and stop the act for that reason before it commits. For advisory and low-consequence work, after-the-fact records may be enough; the hard claim is narrower. Where a delegated system commits a material act without ordinary human approval in the path, if nothing in that path could have stopped the act on the standing of the claims it relied on, the institution has documentation, not admissibility.</p><h3>Why it is not already enforced</h3><p>If the invariant is this concrete, why is it not already enforced?</p><p>Not for lack of work, and the honest answer starts there. AI governance standards — ISO 42001, the EU AI Act, the NIST framework — define the obligations: risk management, transparency, traceability, human oversight. Agent frameworks provide the execution: memory, state, observability, policy controls, kill switches. Inside institutions, model risk management, GRC platforms, data lineage, policy-as-code engines, and entitlement systems are real and serious, and some already sit in the operating path — a rules engine refuses a transaction, an entitlement check blocks an action, a workflow gate holds an approval.</p><p>These are not unserious, and they are not wrong. They are necessary. They are also organised around other objects. The standards say what to do. The agent frameworks say how to run. The control systems govern the model, the data, the user, and the workflow state. The pieces of claim governance exist too: a lineage tool tracks a source, a data-quality system scores it, a policy engine tests a predicate. What is uncommon is not any one of those pieces. It is the discipline of treating the relied-on assertion as something the action path has to test — maintained as live state, checked against the consequence before the act commits. That obligation has an architectural shape: a control that can be satisfied by periodic review stays governance; a control that has to hold in the path of the act becomes part of the architecture. For delegated AI of the kind this argument is about, this one crosses that line.</p><p>This is not a claim that incumbents cannot build it. They can, and some will — into a workflow engine, a decision platform, a model gateway, or assembled by an institution’s own architects. The question is not who builds it; it is what counts as having built it. And the test is about what the system must be able to do, not how it is wired. Four failures tell you it has not. If it cannot stop an action when the basis that action depends on has changed since it was approved. If a claim’s authorisation to be acted on cannot be withdrawn by new information without someone manually intervening. If the actor can proceed without the gate, or grant itself the authority to. If the only record that the basis was sound is one assembled after the act. A platform that avoids all four has built the layer, whatever it calls it — and the deeper test is whether the claim and its standing are first-class in that platform, as native as a user or a transaction, or bolted on as fields the workflow happens to carry. The first is the layer. The second is a description of it.</p><p>That also settles the obvious comparison. An operational ontology — a knowledge graph, a master-data hub, the kind of substrate a platform like Palantir provides — can carry every field a claim needs: provenance, timestamps, confidence, conflict markers. The question was never whether the data model can hold those fields. It is whether the system enforces the invariant — whether claim standing is a live precondition of action, or descriptive metadata attached to objects the workflow already means to use. An operational ontology that enforces the invariant before commit has become the substrate for this layer. One that does not has a richer data model and the same gap. The substrate is welcome. What it has to enforce is the invariant.</p><p>There is one field this is most easily confused with, and it deserves its own paragraph: AI safety. Safety and the AI control agenda — alignment, evaluations, interpretability, monitoring, runtime limits — increasingly reach into the deployment setup, not only the model. The mechanisms overlap with what the invariant needs: gates, monitors, escalation. So does the ground — control work plainly governs actions, not only behaviour. A control architecture that tests a claim’s standing before an action commits is, on that move, enforcing the invariant, and that is the point rather than a problem: admissibility is not a rival discipline or a separate stack, and if a control architecture absorbs it, the thesis is confirmed, not refuted.</p><p>What does not collapse is the question each one answers, and the regime that judges the answer. Safety asks whether the system can be kept inside an acceptable behavioural envelope, under error and adversarial pressure — and its failures are judged as the system behaving badly. Admissibility asks whether, for this act, the institution had a defensible basis to rely on what the system produced — and its failures are judged as the institution acting without one. A model can be perfectly safe and the institution still act on a policy reading that was superseded yesterday: nothing about the model misbehaved, and the institution still had no warrant. Safety is judged against the system. Admissibility is judged against the institution. Alignment does not confer the institution’s warrant; the institution does. The invariant is the runtime test that the warrant was in place.</p><h3>Why regulated industries, and why now</h3><p>The invariant has to be enforced where it is a hard constraint, and that is regulated industries — not because the problem is unique to them, but because there the requirement is external, tested, and non-negotiable. A supervisor, an auditor, a court, or an affected party can ask — and in the hard cases will — what the institution relied on and why that reliance was permitted. A version that survives there has met the hard form of the problem.</p><p>And the timing has converged. Agentic systems are being pushed from advisory roles toward delegated ones, so the gap and the demand arrive together. DORA is in force; the EU AI Act’s high-risk obligations are moving into application; supervisory expectations on model risk are being rewritten for AI. The names differ by jurisdiction — model risk guidance, clinical safety regimes, public-sector accountability rules — but none of these instruments mandates a claim store or a gate. What they do is raise the cost of being unable to show, after a delegated act, what it relied on and why that reliance was permitted — and that cost is what makes the question unavoidable. The argument does not stand on any single deadline. If a timeline slips or enforcement starts soft, the demand still arrives — through board accountability, through audit, through the plain fact that no executive wants an autonomous system acting on stale or unauthorised reasoning. Regulation makes the gap visible first and hardest. It is the forcing function, not the foundation.</p><p>This is falsifiable, and it should be. By delegated action I mean a system that commits an externally meaningful act — approving a payment, issuing a notice, changing a limit, closing an alert — without ordinary human approval in the path. The thesis is wrong if regulated institutions can run delegated AI of that kind at scale, accepted by their supervisors, on the controls they already have — model governance, workflow approvals, entitlement checks, policy-as-code, audit logs — with no mechanism that does what I have described: an individually addressable assertion, used as a basis for action, whose standing can change after approval and stop the act before commit. To be a real test it has to be observable, so the non-examples matter. A generic audit log does not count — it is written at decision time and cannot stop anything. A data-quality flag refreshed overnight does not count — its standing cannot move between approval and act. A workflow approval does not count — it gates the case, not the assertion. If institutions reach scale on those alone and supervisors accept it, I am wrong. A nearer-term version of the same test: watch whether the institutions now moving delegated use cases through audit add a pre-commit check on the standing of the specific assertions relied on, or whether they pass on documentation alone. I do not think they pass on documentation alone — because the reviewer who used to hold the basis is exactly what delegation removes — but that is the test, and it does not turn on anyone using the word claim.</p><h3>Two questions</h3><p>Step back, and the field is working on two questions at once, and mostly treating them as one. The first is what the machine can do. It has the funding, the names, and the attention. The second is what the institution may rely on. The two are not independent — a better-calibrated model lowers the burden the invariant imposes, and the invariant shapes what capability is worth building — but they are different questions, and for institutions deploying under supervision, the second decides the outcome. A more capable model an institution cannot rely on is a more capable model it cannot use. Most are funding the first hard and assuming the second will be there when they reach for it. It will not be, unless it is built deliberately, into the path the action travels.</p><h3>Admissibility has no finish line</h3><p>One thing is left open. Capability is narrated as a race toward a destination — human-level performance, whatever the end state is called. Admissibility has no destination. There is no point at which an institution is finished proving what it may rely on; the work moves with the regulation, the institution, and the agents operating through it.</p><p>That can sound like a weakness — an open commitment with no terminal state. It is the opposite. A destination is something an institution can fall behind on. Admissibility is something it gets better at: the validated claims, the tested authority rules, the evidence patterns a supervisor has already accepted compound into an institution that can move faster precisely because it can defend what it does. The cost is real and front-loaded, and it is not mainly technology. It is the work of naming who owns each material claim, turning the institution’s reliance rules into something executable, putting the gate in the action path, and running the exception queue when claims decay or conflict. Set against it is a cost institutions are already paying: the AI that cleared its evaluation two years ago and still has not left pilot. And individual use cases do reach a stable state — a use case is admissible when its claim set, authority rules, context capture, and evidence record are executable and tested. What has no finish line is the portfolio, not the implementation.</p><p>The test of where an institution stands is simple, and it is organisational before it is technical. Take one use case it wants to move up the ladder. For the decision that use case would make, ask: what claims does the act rest on, and who owns them; what authority does the act require; what context would have to be recorded before it; what evidence would it produce. The failure that matters is not an imperfect answer — it is no named owner for the answer at all. But answering the questions on paper is only the start. The harder test is whether those answers can be made executable in the path of the action — whether the system can test the claims it depends on and stop itself when they cannot bear the consequence. If they cannot, the institution is not blocked by model intelligence. It is blocked by admissibility, and a better model alone will not move it.</p><p>The first concrete move is small. For one delegated use case, inventory the claims the act depends on, name an owner for each material one, set the consequence threshold, and run the gate in shadow — recording what it would have stopped — before it is allowed to block anything in production. That is enough to find out whether the institution has the start of admissibility or only the documentation of it.</p><p>The advantage — and at first it will simply be the ability to move at all — will go to the institution whose AI can act under supervision, produce its evidence as it acts, and keep moving without losing track of what it is allowed to rely on. That is admissibility.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c7c7deb64dd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Epistemic Immunity: Why Knowledge Graphs Need a Defense Architecture]]></title>
            <link>https://medium.com/@navvireich/epistemic-immunity-why-knowledge-graphs-need-a-defense-architecture-dd862e39f512?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd862e39f512</guid>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-quality]]></category>
            <category><![CDATA[enterprise-architecture]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Mon, 30 Mar 2026 11:28:00 GMT</pubDate>
            <atom:updated>2026-03-30T11:28:00.198Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Knowledge graphs are living systems. Without an immune system, they’ll die the same way every wiki did — from pollution.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IIzrq9vQpA66r5as3yWpAw.png" /></figure><p>Every knowledge management system you’ve ever used died the same way. Not suddenly — gradually. Content went stale because nobody maintained it. Stale content eroded trust. Eroded trust killed participation. And without participation, bad knowledge crept in unchallenged. The root cause was always the same: knowledge treated as an asset to be stored, not a medium to be maintained through continuous use.</p><p>For thirty years, that failure was expensive but contained. Senior people compensated — they carried context in their heads, knew which wiki pages were outdated, remembered why the architecture looked different from the documentation. They were, in effect, human agents reasoning over imperfect knowledge, correcting for pollution through experience and judgment.</p><p>Every large organisation has lived through what happens when those people leave. A retirement, a restructuring, a wave of attrition — and suddenly the institutional memory that was compensating for polluted knowledge systems walks out the door. The documentation says one thing. The architecture does another. Nobody remaining knows why, and the decisions that follow are based on whichever source someone finds first. This pattern repeats in every industry, every decade, every organisation that relied on human memory to compensate for knowledge systems that couldn’t maintain themselves.</p><p>Now we’re replacing that compensating judgment with AI agents that reason over knowledge graphs directly. They don’t know which parts of the graph are stale. They don’t carry thirty years of context. They trust the structure at face value — and if that structure is polluted, they propagate the pollution into real decisions at machine speed.</p><p>This is the problem that still gets too little attention.</p><h3>What makes knowledge graphs different</h3><p>A knowledge graph that serves AI agents is a different kind of system from a database or a document repository. It grows through use, adapts through agent feedback, and accumulates contradictions at scale. It stays coherent only through continuous flow — new claims entering, old ones being challenged, confidence shifting, structure reorganising. A system that maintains itself through exchange, not despite it.</p><p>That’s not a metaphor for a living system. That is what a living system is. And living systems need defenses that static repositories never did.</p><p>The database model assumed knowledge was stable: insert, validate, query. The document model assumed knowledge was authored: write, review, publish. Neither assumed that knowledge would evolve under adversarial conditions — that incorrect claims would enter the system, propagate through reasoning chains, and corrupt downstream inferences before anyone noticed. But that’s exactly what happens when AI agents treat a knowledge graph as ground truth and hundreds of processes read from it simultaneously.</p><h3>The obvious approach fails at exactly the wrong moment</h3><p>The first instinct is to filter on familiarity. Reject what doesn’t match existing patterns, trust what comes from known sources. Most quality architectures work this way: schema validation, provenance checks, source reputation scoring. If the claim fits the schema, comes from a trusted source, and doesn’t violate any constraints — let it in.</p><p>This is useful but aimed at the wrong signal.</p><p>Consider what it misses. A claim from your most trusted internal source can be catastrophic if it contradicts verified evidence and corrupts reasoning that depends on it. Imagine a single outdated counterparty risk rating in a financial services knowledge graph. The rating passed every schema check. It came from an authorised source. It matched the expected format. But the counterparty’s actual risk profile had changed six months ago, and nobody updated the graph. That one stale claim now silently distorts credit assessments, portfolio allocations, and compliance reports downstream. Every agent reading from this graph trusts it.</p><p>Meanwhile, a surprising claim from an unfamiliar external source — say, a regulatory filing that contradicts your internal model — might be the most valuable piece of knowledge in your graph. If it survives validation.</p><p>The filter should track damage to inference, not distance from what the system already believes.</p><h3>Polly Matzinger and the design principle that changes everything</h3><p>An immunologist named Polly Matzinger rewrote the textbooks on this. For a century, the assumption was that the immune system distinguishes “self” from “non-self” — rejecting what’s foreign, accepting what’s familiar. Every quality architecture I described above follows this same logic: does this claim belong here? Is the source one of ours?</p><p>Matzinger showed otherwise. The immune system responds to damage signals. It doesn’t care whether the cause is external or internal. Your body tolerates trillions of foreign bacteria — useful, harmless, welcome. But it attacks its own cancerous cells without hesitation, because they damage tissue.</p><p>Applied to knowledge systems, this reframes the entire quality problem.</p><p>Source reputation, schema compliance, provenance chains — these are all foreignness-based filters. They ask: does this claim look like it belongs? Matzinger’s principle redirects attention to a different signal: does this claim damage the system’s ability to reason correctly?</p><p>A claim from your CEO that contradicts verified market data? Damage. Flag it regardless of source authority. A claim from an untested external dataset that fills a gap in your graph’s coverage? Not damage. Evaluate it on its merits, don’t reject it because the source is unfamiliar.</p><p>This determines the entire architecture of quality defense. A damage-based system protects what works — which may mean flagging trusted internal claims and welcoming unfamiliar external ones, depending on what each does to the graph’s reasoning.</p><h3>The autoimmune problem</h3><p>A defense architecture that’s too aggressive at filtering becomes an echo chamber. It rejects novelty because novelty looks like damage — new information destabilises existing inferences, triggers alerts, gets quarantined or discarded. The immune system that’s meant to protect the organism starts attacking it.</p><p>In medicine, that’s autoimmune disease. In knowledge systems, it’s epistemic closure — a graph that’s clean, consistent, and increasingly wrong because it stopped admitting anything that challenged its assumptions.</p><p>Every enterprise that has run a knowledge management programme has seen this dynamic. The more tightly you control contributions, the more sterile the system becomes. The people with genuinely novel insights stop contributing because the gatekeeping cost exceeds the value. What remains is a curated echo chamber that looks healthy by every internal metric and slowly diverges from reality.</p><p>The challenge is not pollution resistance or novelty preservation. It’s both, simultaneously, and they pull in opposite directions. Any quality architecture that doesn’t explicitly address this tension will default to one extreme or the other — and both kill the system.</p><h3>Three mechanisms</h3><p>We’ve been working on this problem across a series of papers, and the architecture we’re developing is built around three mechanisms. Each addresses a different phase of the quality challenge.</p><h3>Admission control</h3><p>The first mechanism operates before a claim enters the graph. But it doesn’t filter on schema fit or source reputation alone. It asks a different question: if this claim enters the graph, what existing inference chains does it affect? What’s the blast radius?</p><p>A claim that would contradict three well-supported downstream inferences gets higher scrutiny than a claim that adds to an isolated branch of the graph. A claim that affects reasoning in a compliance-critical domain gets higher scrutiny than one in an advisory domain.</p><p>This is what makes admission control damage-aware rather than foreignness-based. The filter isn’t “does this look right?” — it’s “what breaks if this is wrong?” The cost of a bad admission isn’t uniform. A stale reference data point in an operational knowledge graph can sit quietly for months. A stale risk rating in a financial graph can corrupt regulatory reporting within hours. The admission mechanism should know the difference.</p><p>The calibration problem is real: set the threshold too high and you starve the graph of new knowledge. Set it too low and you let pollution in that compounds faster than you can detect it. The right setting depends on the domain, the cost of errors, and the graph’s current state — which means admission control itself needs to be adaptive, not static.</p><h3>Damage-based detection</h3><p>The second mechanism operates after claims have been integrated. This is where most enterprise knowledge systems have no architecture at all.</p><p>Bad claims that pass admission don’t sit quietly. They propagate through dependency chains, contaminate downstream inferences, and compound over time. An incorrect supplier classification leads to wrong procurement recommendations, which lead to wrong budget forecasts, which lead to wrong resource allocations. By the time someone notices the budget forecast is off, the root cause is buried three dependency levels deep.</p><p>Damage-based detection means continuous monitoring of inference health. If a claim changes, trace which dependent claims and inferences are affected. If an inference starts producing anomalous outputs, trace backward to find which upstream claims may be responsible. A discredited source triggers re-scoring across every claim that depends on it and every downstream chain those claims feed.</p><p>The critical design choice: this detection mechanism should learn. Every resolved quality incident — every claim identified as bad, every contamination traced to its source, every false alarm that turned out to be legitimate novelty — becomes training data for better detection. A system that only applies rules written by architects will always be fighting the last war. A system that encodes resolved incidents as learned detectors gets better at catching problems it hasn’t seen before.</p><p>This is the mechanism that most directly embodies Matzinger’s principle. It doesn’t ask “where did this claim come from?” It asks “is this claim damaging the system’s reasoning?” — and it gets progressively better at answering that question.</p><h3>Governed novelty protection</h3><p>The third mechanism is the one that prevents the other two from calcifying the graph.</p><p>Without explicit novelty protection, admission control and damage detection will jointly evolve toward maximum conservatism. New claims that destabilise existing inferences will be filtered more aggressively over time. The graph will become cleaner, more consistent, and increasingly disconnected from a reality that keeps changing.</p><p>Governed novelty protection means novel claims get quarantined and evaluated. A claim that fails familiarity checks but doesn’t trigger damage signals enters a staging area where it can be tested against external evidence, reviewed by human experts, or held until corroborating claims arrive. If it survives, it integrates fully. If it fails, it’s rejected with a record of why.</p><p>A financial services knowledge graph receives a claim from an external regulatory data feed: a counterparty that your graph classifies as investment-grade has been placed on a sanctions watchlist by a jurisdiction your system doesn’t routinely monitor. The claim contradicts your internal model. Admission control flags the potential blast radius — dozens of compliance assessments depend on that counterparty’s classification. Damage detection hasn’t triggered because the claim hasn’t entered the graph yet. Without novelty protection, a conservative system discards it as anomalous. With novelty protection, the claim enters quarantine, gets routed to a compliance analyst, and — if confirmed — integrates with full provenance and triggers downstream re-evaluation. The difference between catching a compliance exposure in days versus discovering it in an audit months later.</p><p>The governance layer matters because the decision about when novelty is genuine versus when it’s pollution can’t always be automated. In regulated industries, a claim that contradicts your internal model might be a genuine regulatory change or it might be an error in a third-party data feed. The right response depends on context that the system may not have. Governed novelty protection keeps the human in the loop for exactly the decisions where human judgment still matters — without requiring human review of every claim that enters the system.</p><p>Without this mechanism, the graph becomes sterile. With it, the graph can admit genuine surprises — the kind that make a knowledge system more valuable over time rather than less.</p><h3>What happens if we don’t build this</h3><p>If we get this right, knowledge graphs become something rare in enterprise systems: structures that get healthier the more they’re used. The continuous flow of new claims, challenges, and corrections doesn’t degrade the system — it strengthens it, the same way a functioning immune system strengthens an organism through exposure.</p><p>If we get this wrong, we build the next generation of data swamps — except this time the swamps feed AI systems that make real decisions. An agent reasoning over a polluted knowledge graph doesn’t know the graph is polluted. It treats the corruption as ground truth and propagates it into recommendations, allocations, compliance assessments, and strategic decisions — faster than any human review process can catch.</p><p>The failure pattern will be the same one we’ve been watching for thirty years in every knowledge management system that died from pollution. Only faster, and harder to reverse — because the compensating humans who used to catch the errors aren’t in the loop anymore.</p><h3>Where this is going</h3><p>These three mechanisms — admission control, damage-based detection, governed novelty protection — align with a broader research programme we’ve been developing across three papers on SSRN. The first addresses dynamics blindness: why AI systems fail when they can’t trace causal logic in regulated environments. The second proposes a claims-based architecture — a knowledge substrate where every assertion carries provenance, confidence, temporal metadata, and dependency links, providing the passive infrastructure these defense mechanisms require. The third examines knowledge as a living structure and why organisations that treat it as a static asset will keep failing the same way they’ve been failing for thirty years.</p><p>The epistemic immunity thread draws from all three. Claims-based architecture provides the substrate — you can’t track damage propagation if you don’t know which claims depend on which. The dynamics blindness work explains what happens when that substrate is absent — semantic errors that in a bank are compliance breaches, not typos. And the living structure argument is why any of this matters: a knowledge graph that serves AI agents is a dissipative structure in Prigogine’s sense, maintained far from equilibrium by continuous energy flow. Stop the flow — stop the active quality defense — and the structure degrades toward noise.</p><p>The mechanics draw on formal traditions well beyond the biological analogy: argumentation theory for claim conflict resolution, belief revision for principled confidence updating, provenance algebras for tracking how source reliability propagates through inference chains, and concept drift detection for distinguishing genuine novelty from contamination. The immune system metaphor organises the design choices, particularly Matzinger’s danger principle. But the architecture has to stand on engineering.</p><p>We’ve been building knowledge systems for decades without asking what keeps them healthy over time. The agents that depend on these graphs are already making decisions — in procurement, in compliance, in resource allocation, in risk assessment. The defense architecture needs to exist before the pollution compounds past the point where human oversight can catch it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd862e39f512" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Industry Is Building Neurons When Organizations Need Synapses]]></title>
            <link>https://medium.com/@navvireich/the-ai-industry-is-building-neurons-when-organizations-need-synapses-457d81cc8bb1?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/457d81cc8bb1</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Sat, 28 Feb 2026 18:50:12 GMT</pubDate>
            <atom:updated>2026-02-28T18:50:12.523Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*utze9ujhRxqfAjTUhiAYSA.png" /></figure><p><em>Intelligence lives in connections, not components. Biology, ecology, music, and chemistry all agree. Most AI strategy doesn’t.</em></p><p>The AI industry sells components. Organizations need connections.</p><p>Every vendor pitch in 2026 follows the same logic: our model is smarter, our agents are more capable, our architecture processes more tokens per second. The implicit promise is that better components produce better outcomes. Drop a more powerful model into your operations and watch performance improve.</p><p>This logic has a problem. It contradicts what we know about how intelligence actually works — at every scale, from cells to brains to forests to orchestras.</p><h3>Same cells, different medium</h3><p>Michael Levin’s xenobot experiments at Tufts University demonstrate the principle with unsettling clarity. Take frog skin cells. In their normal bioelectric context, they do what skin cells do — sit on the surface and protect the organism. Remove those same cells and place them in a different coordination environment. No genetic modification. No enhancement. Identical biological components.</p><p>The cells self-organize into entirely novel functional forms — tiny living machines that move, heal themselves, and exhibit behaviors no frog skin cell was “designed” to perform (Kriegman, Blackiston, Levin &amp; Bongard, 2020).</p><p>The variable was the medium — the coordination infrastructure through which they interacted. The cells themselves didn’t change at all. Levin’s broader research program makes the mechanism explicit: intelligence correlates not with the sophistication of individual components but with the density and quality of connections between them. He calls the coordination substrate “cognitive glue” — bioelectric networks that bind individual competencies into collective intelligence that exceeds the sum of its parts (Levin, 2023). The glue matters more than the components it binds.</p><h3>The brain that limits itself</h3><p>If more connections always meant more intelligence, evolution would have built brains as single, massively interconnected organs. It didn’t.</p><p>Iain McGilchrist’s research on hemispheric lateralization reveals something counterintuitive: the corpus callosum — the structure connecting the brain’s two hemispheres — is primarily inhibitory. A significant population of its fibres are GABAergic, and its main function is to <em>prevent</em> the hemispheres from interfering with each other (McGilchrist, 2009). Evolution invested heavily in structured disconnection.</p><p>Why? Because two semi-independent processors with controlled, limited interaction produce richer cognition than one fully connected mass. The left hemisphere handles focused, sequential, analytic processing. The right handles broad, contextual, relational awareness. They need to communicate — but selectively. Too much crosstalk and you get noise, not intelligence. The architecture of connection matters as much as its existence.</p><p>This refines the argument. Intelligence emerges from the <em>right</em> connectivity — knowing what to transmit and what to inhibit, what to share and what to keep separate. The medium is a structure that enables certain interactions and constrains others.</p><h3>The irony at the heart of AI</h3><p>A neural network — the architecture underlying every large language model, every agent framework, every AI system being sold to enterprises right now — is the most literal proof that intelligence lives in connections. Each artificial neuron is trivially simple. It multiplies inputs by weights, adds a bias, passes the result through an activation function. There is zero intelligence in the component.</p><p>All the “knowledge” in GPT-4, Claude, or any foundation model lives in the weights — the parameters that describe relationships between nodes. When we say a model has 1.7 trillion parameters, we’re saying it has 1.7 trillion descriptions of connections. The intelligence IS the connection pattern. The neurons are interchangeable.</p><p>The AI industry builds intelligence that literally is its synapses, then packages it as a neuron for organizations to drop in. The component gets sold. The connection architecture that made it intelligent in the first place gets left as an exercise for the customer.</p><h3>The forest that thinks</h3><p>Ecology provides the same lesson at ecosystem scale.</p><p>Suzanne Simard’s research on mycorrhizal networks demonstrated that forest trees exchange carbon, water, and chemical warning signals through underground fungal networks — what popular science calls the “Wood Wide Web.” Her 1997 <em>Nature</em> paper showed bidirectional carbon transfer between birch and Douglas fir through ectomycorrhizal fungi (Simard et al., 1997). The finding has held up under scrutiny, though the full mechanisms and their ecological significance remain debated — Karst, Jones, and Hoeksema challenged the popular narrative in a 2023 <em>Nature Ecology &amp; Evolution</em> review, arguing that positive citation bias has inflated some claims beyond what the evidence supports (Karst, Jones &amp; Hoeksema, 2023).</p><p>The directional insight stands regardless: seedlings connected to fungal networks show measurably better survival under stress conditions than isolated ones. Remove the trees and replant with identical seedlings, and the new forest underperforms. You rebuilt the components. You lost the medium.</p><p>The slime mold <em>Physarum polycephalum</em> makes the point even more sharply, with cleaner evidence. A single-celled organism with no nervous system solves mazes, finds shortest paths between food sources, and constructs networks that rival the efficiency of the Tokyo subway system (Tero et al., 2010). The “brain” is the cytoskeletal network within the cell — signal molecules propagating information across a physical web. Same principle, different scale: the intelligence is the network.</p><p>Ant colonies drive it home. No central command. No individual ant understands the colony’s strategy. Each ant follows simple local rules — deposit pheromone, follow trails. But the network of chemical signals enables the colony to solve optimization problems, adapt to environmental changes in real time, and allocate resources with algorithmic efficiency. Deborah Gordon’s research has shown that colony behavior emerges entirely from the rate, rhythm, and pattern of local interactions — no ant directs another (Gordon, 2010). Every attempt to explain colony behavior by studying individual ants fails, because the colony is its connections.</p><h3>The music between the notes</h3><p>There’s a line attributed to the tradition around Mozart — that the music is not in the notes but in the spaces between them. Whether Mozart said it or not, the observation is precise. A sequence of notes played without attention to timing, dynamics, and the silences that give phrases their shape is not music. It’s data.</p><p>Jazz makes this concrete in a way classical music doesn’t, because jazz requires the connections to happen in real time, unrehearsed. Individual virtuosity is table stakes in jazz — every player in Miles Davis’s quintet could outperform most musicians on the planet. But jazz happens in the listening, the call and response, the shared harmonic context that lets four musicians create something none of them planned. Davis insisted the band listen more than they played (Davis &amp; Troupe, 1989).</p><p>A jazz ensemble where every musician plays brilliantly but nobody listens to anyone else is a Powerful Tools trap — technically excellent, strategically sterile. Dead on arrival.</p><p>Japanese aesthetics formalized this insight centuries ago. <em>Ma</em> (間) — the meaningful void — is a foundational concept in Japanese architecture, music, calligraphy, and martial arts. A room is not defined by its walls but by the space they contain. Ikebana treats the empty space as more important than the flowers. The medium between elements creates the meaning.</p><h3>The hammer in the garage</h3><p>Stuart Kauffman, the complexity theorist, uses a thought experiment that cuts to the practical implication. A hammer sitting alone in a garage has definable uses — you can list them. A hammer in a workshop surrounded by nails, wood, saws, workbenches, blueprints, and a carpenter has uses that are unprestatable — they cannot be fully specified in advance. The hammer hasn’t changed. Its context has. And that context determines not just how the hammer is used but what it’s <em>for</em>.</p><p>This is the distinction between capability and fertility. Capability is the hammer’s hardness, weight, balance — its power to execute specified functions. Fertility is the workshop — the connective context that makes unpredictable value creation possible. Kauffman’s point is that the adjacent possible — the set of things that could happen next — is determined by context, not by any individual component’s specification (Kauffman, 2019).</p><p>An AI agent sitting in a silo has definable uses. The same agent connected to organizational memory, to other agents, to the accumulated patterns of how the organization works — that agent has uses nobody specified in the procurement document. Same capability. Different medium. Radically different value.</p><p>Even chemistry cooperates with the argument. Same carbon atoms, arranged one way: graphite — you can write with it. Rearranged: diamond — you can cut glass. Same atomic capability. The bonding structure — how atoms connect to each other — determines whether the material is among the softest or hardest known substances.</p><h3>What this means for AI strategy</h3><p>The strategic implication is direct. Every organization investing in AI in 2026 faces a sequencing decision, whether they recognize it or not.</p><p>The default path is capability-first: buy better models, deploy smarter agents, automate more processes. This produces measurable returns on defined tasks. It also hardens the architecture around component-level optimization. Each capability investment that doesn’t account for the connective tissue between components makes it progressively more expensive to add that tissue later. The concrete sets.</p><p>The alternative is medium-first: build the knowledge infrastructure — the shared context, the organizational memory, the connection architecture — through which capable agents coordinate. This looks slower in quarter-one metrics. It compounds differently. Upgrading a claims AI from one model to the next improves summaries incrementally. Connecting that claims AI to the underwriting model through shared knowledge infrastructure changes what the system can <em>see</em>. The intelligence gain comes from richer coordination, not better components.</p><p>Capability commoditizes. Every vendor ships equivalent foundation models, and the gap between them narrows with each release cycle. The connections between them — the organizational context, the accumulated patterns, the medium through which they coordinate — don’t commoditize. They can’t be purchased off the shelf. They can’t be copied by a competitor. They can only be grown.</p><p>Levin showed it in cells. Musicians have known it for centuries. Neural networks prove it computationally every time they run — the architecture that literally IS its connections gets sold as a standalone component.</p><p>The organizations that figure this out first will compound advantages the rest cannot purchase.</p><h3>Sources &amp; Further Reading</h3><h4>On collective intelligence and the coordination substrate</h4><p><strong>Levin, M.</strong> (2023). Bioelectric networks: the cognitive glue enabling evolutionary scaling from physiology to mind. <em>Animal Cognition</em>, 26, 1865–1891. Levin’s framework for how bioelectric signaling scales cognition from single cells to collective intelligence. Introduces the “cognitive light cone” — the boundary of the largest goal a system can pursue — and argues that intelligence is a function of connectivity, not component sophistication.</p><p><strong>Kriegman, S., Blackiston, D., Levin, M. &amp; Bongard, J.</strong> (2020). A scalable pipeline for designing reconfigurable organisms. <em>Proceedings of the National Academy of Sciences</em>, 117(4), 1853–1859. The xenobot paper. Frog skin cells, removed from their normal bioelectric context, self-organize into novel living machines with behaviors no individual cell was programmed to exhibit. The experiment that makes “same components, different medium, different intelligence” empirically concrete.</p><p><strong>McMillen, P. &amp; Levin, M.</strong> (2024). Collective intelligence: A unifying concept for integrating biology across scales and substrates. <em>Communications Biology</em>, 7, 378. The most recent synthesis, extending the cognitive glue framework across biological and artificial systems. Argues that the tools of behavioral science and diverse intelligence research illuminate decision-making by cellular collectives across evolutionary and biomedical contexts.</p><h4>On hemispheric lateralization and structured disconnection</h4><p><strong>McGilchrist, I.</strong> (2009). <em>The Master and His Emissary: The Divided Brain and the Making of the Western World</em>. Yale University Press. The foundational argument that the two hemispheres attend to the world in fundamentally different ways, and that the corpus callosum’s primary function is inhibitory — enabling structured disconnection rather than maximal connectivity.</p><p><strong>McGilchrist, I.</strong> (2021). <em>The Matter with Things: Our Brains, Our Delusions, and the Unmaking of the World</em>. Perspectiva Press. The expanded, two-volume follow-up integrating neuroscience with epistemology. Deepens the case for why different types of attention require architectural separation.</p><h4>On mycorrhizal networks and the evidence debate</h4><p><strong>Simard, S.W., Perry, D.A., Jones, M.D., Myrold, D.D., Durall, D.M. &amp; Molina, R.</strong> (1997). Net transfer of carbon between ectomycorrhizal tree species with shared mycorrhizal fungi. <em>Nature</em>, 388, 579–582. The original paper demonstrating bidirectional carbon transfer between birch and Douglas fir through shared fungal networks. <em>Nature</em> coined “Wood Wide Web” in its coverage.</p><p><strong>Karst, J., Jones, M.D. &amp; Hoeksema, J.D.</strong> (2023). Positive citation bias and overinterpreted results lead to misinformation on common mycorrhizal networks in forests. <em>Nature Ecology &amp; Evolution</em>, 7, 501–511. The counterpoint. Analyzes 1,500+ papers and finds that claims about mycorrhizal network benefits have been inflated by positive citation bias. Included here because intellectually honest arguments engage their strongest objections.</p><h4>On network intelligence without central control</h4><p><strong>Tero, A., Takagi, S., Saigusa, T., Ito, K., Bebber, D.P., Fricker, M.D., Yumiki, K., Kobayashi, R. &amp; Nakagaki, T.</strong> (2010). Rules for biologically inspired adaptive network design. <em>Science</em>, 327(5964), 439–442. <em>Physarum polycephalum</em> constructs transport networks with efficiency, fault tolerance, and cost comparable to the Tokyo rail system. A single-celled organism with no nervous system outperforming human infrastructure engineers.</p><p><strong>Gordon, D.M.</strong> (2010). <em>Ant Encounters: Interaction Networks and Colony Behavior</em>. Princeton University Press. How ant colonies coordinate without hierarchy. Colony behavior emerges from the rate, rhythm, and pattern of local interactions. No ant directs another. The colony IS the interaction pattern.</p><h4>On the adjacent possible and unprestatable value</h4><p><strong>Kauffman, S.</strong> (2019). <em>A World Beyond Physics: The Emergence and Evolution of Life</em>. Oxford University Press. The most accessible statement of Kauffman’s argument that the uses of any object in context are unprestatable — they cannot be algorithmically enumerated. The screwdriver argument, the adjacent possible, and why context (fertility) generates value that components (capability) cannot.</p><p><strong>Kauffman, S.</strong> (2000). <em>Investigations</em>. Oxford University Press. The deeper, more technical treatment. Introduces enablement as distinct from causation. Makes the case that the biosphere and the econosphere expand into the adjacent possible in ways that cannot be predicted from component properties alone.</p><h4>On listening, space, and the intelligence between</h4><p><strong>Davis, M. &amp; Troupe, Q.</strong> (1989). <em>Miles: The Autobiography</em>. Simon &amp; Schuster. Davis on how the quintet worked: the band had to listen more than they played. The shared harmonic context — what each musician heard and responded to — was the medium through which jazz happened, not the individual performances.</p><p><strong>Pilgrim, R.B.</strong> (1986). Intervals (<em>Ma</em>) in space and time: Foundations for a religio-aesthetic paradigm in Japan. <em>History of Religions</em>, 25(3), 255–277. Academic treatment of <em>ma</em> (間) — the meaningful void in Japanese aesthetics. Architecture, music, calligraphy, and martial arts all treat the space between elements as constitutive of meaning, not merely empty.</p><p><em>This article is part of the “Build the Medium” research programme, which argues that organizations become intelligent not by deploying better AI components but by building the knowledge infrastructure through which tools — and people — coordinate. The programme draws on developmental biology, complexity science, distributed cognition, and predictive processing to propose a practical alternative to capability-first AI strategy. Related work appears on </em><a href="https://medium.com/@witoldreichhart"><em>Medium</em></a><em> and </em><a href="https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7192912"><em>SSRN</em></a><em>.</em></p><p><em>Witold Reichhart is a Client Transformation Partner working at the intersection of organizational intelligence and AI strategy.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=457d81cc8bb1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Shell and the Pearl]]></title>
            <link>https://medium.com/@navvireich/the-shell-and-the-pearl-5a72e57273d1?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a72e57273d1</guid>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[emergence]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[relationships]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Fri, 27 Feb 2026 17:34:45 GMT</pubDate>
            <atom:updated>2026-02-28T16:56:31.682Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*76JgBrEGB69X066qocfKWg.png" /></figure><p><em>On the principle that runs from physics through enterprise AI to the space between two people</em></p><p>An oyster makes a pearl only inside its shell. Not despite the shell. Because of it.</p><p>That image has stayed with me for weeks now, turning slowly, connecting to things I didn’t expect it to connect to. To thermodynamics. To how we build intelligent systems. To something about men and women that we’ve become strangely afraid to say out loud.</p><p>The pearl doesn’t form in open water. It forms in a contained, protected, stable environment where a disturbance — a grain of sand, something foreign — meets the conditions for transformation. The shell provides pressure, enclosure, the right chemistry. Within that structure, something genuinely new crystallizes. Layer by layer. Not planned, not designed — emergent.</p><p>Remove the shell, and you don’t get a freer pearl. You get dead tissue and dispersed calcium.</p><p>Ilya Prigogine won the Nobel Prize for showing that this principle runs through all of nature. He called them dissipative structures — systems that maintain complex, creative order precisely because they exist within rigid boundary conditions, far from equilibrium. A hurricane forms because of temperature gradients, pressure differentials, the deterministic physics of fluid dynamics. The structure itself is emergent, self-organizing, alive with pattern. But it requires those constraints to exist at all.</p><p>The thermodynamic boundaries don’t compete with the creative dynamics. They sustain them.</p><p>The same structural principle appears everywhere you look, once you learn to see it. Deterministic constraint enables emergent creativity. Rigid containment is the precondition for fluid self-organization. The formative and the generative are not opposites. They are partners in a single act of creation.</p><p>We are learning this the hard way in artificial intelligence.</p><p>The most capable AI agents built today can modify their own code, negotiate on behalf of their users, execute multi-step tasks across dozens of systems without supervision. Deploy them without adequate structure and the result is chaos, not intelligence. An autonomous agent that can do anything will, given enough time, do something catastrophic. It sends the wrong email to the wrong stakeholder. It modifies a production database because nothing told it not to. It optimizes for a metric that made sense in isolation and destabilizes something three systems away.</p><p>The industry is converging on an architectural response that looks remarkably like Prigogine’s insight translated into engineering. A symbolic constraint layer — deterministic, auditable, inspectable — encodes what the organization permits, prohibits, and knows to be true. Inside that shell sits a probabilistic layer: neural networks, language models, learning systems that identify patterns, propose actions, spot possibilities no human would have seen. The constraint layer doesn’t suppress the model’s intelligence. It creates the bounded space within which that intelligence becomes trustworthy. The model proposes plausible futures. The constraint layer vetoes impermissible ones. Regulators inspect the deterministic shell. The emergent intelligence forms inside it.</p><p>There is a subtler point here. The difference between autonomy and initiative. An agent that can act independently has autonomy. An agent that knows <em>when</em> to act, <em>why</em> this situation calls for action, and when to stay silent — that agent has initiative. Autonomy is a capability. Initiative is a relationship between the agent and its context. Without the shell, the agent has power but no judgment. Capability without wisdom.</p><p>The pearl inside the shell. Again.</p><p>Here is where I want to go further than technology. Because the pattern is older than thermodynamics and bigger than enterprise architecture.</p><p>Masculine and feminine. Structure and flow. Containment and creation.</p><p>I know. Difficult territory. We’ve spent decades insisting that these are social constructs, power dynamics to be dismantled, remnants of an unjust past. And some of that work was necessary — there were genuine injustices that needed correcting. But somewhere along the way we lost the ability to see what was actually there. We became so focused on the political dimension that we forgot the structural one.</p><p>The masculine principle provides containment. Boundaries. Protection. A stable frame within which something can grow without being destroyed by exposure. The feminine principle is the creative force within that frame — open, receptive, generative, forming something genuinely new from the material at hand.</p><p>You’ve seen this. Maybe you couldn’t name it, but you’ve felt it. A woman who feels held — safely contained, sheltered, steady ground beneath her — doesn’t become smaller. She becomes radiant — more creative, more playful, more open than she was a moment before. She relaxes into the space and something luminous arrives, something that wasn’t there a moment before. She flowers. Not because someone told her to. Because the conditions are right.</p><p>And the man who holds that space discovers something too. The shell finds its purpose in the pearl. Structure without creation inside it is just empty architecture — walls around an empty room. His steadiness, his containment, his quiet holding of the frame means something because of what grows within it. Both find themselves through the other.</p><p>These are principles, not fixed identities. They live in everyone. A woman running a team is the shell for her people. A man writing alone at night is the pearl forming inside his own discipline. The same person can contain in one hour and create in the next. In bed, at work, in solitude — the roles shift. What matters is that the dynamic is present, not which body it lives in. Between two people, it becomes most visible. But it begins inside each of us.</p><p>This is not hierarchy. The shell is not more important than the pearl. The pearl is the point. But the pearl requires the shell. And the relationship between them — that continuous dynamic of holding and flowering, of structure and creation — is itself the most beautiful expression of what Prigogine observed in physics. The relationship is the dissipative structure. The relationship is the pearl, forming layer by layer, day after day, inside the container two people build together — renewed, emergent, never quite the same as yesterday.</p><p>What strikes me is how afraid we’ve become of noticing this. As if acknowledging that masculine and feminine principles serve different structural functions somehow diminishes one or the other. It doesn’t. And yes — containment can be distorted. A shell can become a cage. When someone who was supposed to hold the space uses it to control, the damage cuts deeper than ordinary harm, because it poisons the very structure that was meant to be safe. Knowing this matters. But the distortion doesn’t disprove the principle. It proves how much the principle matters when it works.</p><p>A shell without a pearl is empty calcium. The masculine without the feminine is rigid, purposeless, dead architecture.</p><p>And the feminine without the masculine? Not liberation. Dissolution. Creative energy with no container dissipates. It doesn’t form pearls. It doesn’t self-organize into hurricanes. It doesn’t gestate new life. It disperses. Just as an autonomous agent without a constraint layer produces chaos rather than intelligence, creative energy without containment produces noise rather than beauty.</p><p>A woman carrying a child is perhaps the most literal expression of this dynamic. The womb — contained, protected, temperature-regulated, chemically precise — creates the conditions for the most astonishing act of emergence in nature. A new human being. Not predicted from any blueprint — not even the genome, which constrains without determining. Emergent. And utterly dependent, for those nine months, on the formative structure that holds the process together.</p><p>Take away the containment, and you don’t get a freer birth. You get no birth at all.</p><p>Prigogine would have recognized this instantly. A dissipative structure requires both the energy flow and the boundary conditions. Remove either, and the structure collapses. There is no emergence without constraint. There is no constraint worth having without emergence.</p><p>This holds for hurricanes, for AI architectures, for relationships. For a life.</p><p>The fact that the same generative principle — deterministic structure enabling emergent creation — runs from thermodynamics through biology through the design of intelligent machines through the quiet miracle of two people making each other more than they were alone is not a coincidence. It is not a metaphor we impose on the world. It is a pattern the world keeps showing us, patiently, at every scale, waiting for us to stop arguing about it and start seeing it.</p><p>The shell and the pearl need each other. The constraint and the emergence are one act. The masculine and the feminine are not competing forces to be balanced or equalized into sameness. They are complementary principles whose polarity — whose <em>difference</em> — is itself the creative engine.</p><p>Remove the polarity and you remove the generative force. Flatten the difference and nothing new can form.</p><p>Life is beautiful because of this dynamic. Every pearl that forms, every child that is born, every hurricane that spirals into coherence, every intelligent system that earns the trust to operate in the real world, every relationship that deepens into something neither person could have been alone — all of them are expressions of the same truth.</p><p>Something to revere.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a72e57273d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What the AI Maturity Models Don’t Show]]></title>
            <link>https://medium.com/@navvireich/what-the-ai-maturity-models-dont-show-5cd7f85dd5f7?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/5cd7f85dd5f7</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai-strategy]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Sun, 01 Feb 2026 16:16:39 GMT</pubDate>
            <atom:updated>2026-02-02T07:53:42.456Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why the fastest climbers may be worst positioned for what comes next</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BVSixI1GlxQciM3zsBkPbg.png" /></figure><p>Every major consultancy has an AI maturity model. McKinsey, Deloitte, Gartner, Accenture — the whole chorus. They chart a progression from early experimentation to enterprise-wide deployment.</p><p>The specifics vary. Gartner uses five levels — Awareness, Active, Operational, Systemic, Transformational. Deloitte segments organizations from Starters through Pathfinders to Transformers. McKinsey measures readiness across dimensions. But the underlying structure is identical: a ladder from scattered pilots to autonomous agents executing complex tasks without continuous human oversight.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1018/1*Bq9mJDi9hsEG9QrG-SHrfw.png" /></figure><p>It’s a sensible progression — and the ladder stops exactly where the interesting question begins.</p><h3>The Screwdriver Problem</h3><p>These models all focus on one question: <em>What can your agents do?</em></p><p>Can they complete a task? Chain tasks together? Operate autonomously? Handle exceptions, recover from errors, pursue goals without hand-holding?</p><p>Fair enough. Capability matters. But there’s a different question the models never ask — and it matters more.</p><p>Stuart Kauffman offers a useful thought experiment. Consider a screwdriver. How many uses does it have?</p><p>There is no algorithm that can list all uses of a screwdriver. The number of uses isn’t a property of the screwdriver itself — it’s a property of context. A screwdriver alone in empty space has limited uses. A screwdriver in a garage full of tools, materials, and half-finished projects has effectively unlimited uses. The screwdriver is the same. The possibilities are not.</p><p>The maturity models measure the sharpness of the screwdriver. They should measure how full the garage is.</p><p>Capability is what the agent can do. Fertility is what can emerge when capable agents operate in rich conditions.</p><p>A system that executes tasks — even complex, multi-step, autonomous tasks — is a tool. Sophisticated, useful, faster than humans. Still a tool. It does what it was designed to do.</p><p>A system embedded in conditions that allow new value to emerge is something else entirely. The distinction determines whether you get execution or emergence — delivery on spec, or outcomes that couldn’t have been specified in advance.</p><p>Kauffman calls this the “unprestatable” — what cannot be stated beforehand. This is exactly what emergence means: value that nobody designed.</p><h3>The Adjacent Possible</h3><p>Kauffman’s concept of the “adjacent possible” makes fertility precise.</p><p>The adjacent possible is the set of next states reachable from the current actual state. What exists now determines what can exist next. “What is actual enables what is next possible.”</p><p>A system with low fertility has a small adjacent possible. It optimizes within a fixed possibility space. The screwdriver gets sharper, but the garage stays empty.</p><p>A system with high fertility has an expanding adjacent possible. It creates new possibilities through operation. Each interaction opens pathways that didn’t exist before. The garage fills up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*86QzRRG309ffMO6n8jyJCA.png" /></figure><p>The maturity models treat AI evolution as a single dimension — capability — a ladder from pilots to autonomy. But there’s a second axis: the size and growth rate of the adjacent possible. Whether your AI environment is fertile.</p><p>These two dimensions aren’t independent. They’re sequentially dependent. Fertility is foundation. Capability builds on top.</p><p>The fastest climbers up the capability ladder might be closing off the path to emergence — worst positioned for what comes next.</p><h3>The Natural Drift</h3><p>The natural drift of optimization makes systems less fertile. Architecture hardens toward efficiency. Feedback loops get designed for speed, not learning. Integration optimizes for latency, not connection. Without intentional counter-pressure, every capability gain locks in the assumption that value comes from doing what you specified, faster.</p><p>High capability with low fertility isn’t a stepping stone. It’s a ceiling.</p><p>Put these together and the maturity models reveal their blind spot:</p><p><strong>Capability First (the default path)</strong></p><p>Pilots → departmental AI → integrated workflows → autonomous agents. This is what the maturity models measure. Each step delivers ROI. Stakeholders see progress. You arrive at the top rung with powerful tools that do exactly what you designed them to do.</p><p>And that’s the ceiling. The ladder doesn’t go higher. What comes next isn’t above you — it’s beside you, reachable only through a different path. The adjacent possible has shrunk to match the optimizations you’ve already made.</p><p><strong>Mindful of Both Dimensions</strong></p><p>This isn’t about choosing fertility instead of capability. It’s about building with awareness of both dimensions — intentionally, concurrently.</p><p>Start building the medium before the inhabitants arrive in force. Knowledge infrastructure, feedback loops, integration for connection — conditions for emergence. This looks like unclear ROI. It looks like infrastructure without payoff.</p><p>But it keeps the adjacent possible expanding. When you add capability, it has somewhere fertile to land. Emergence becomes reachable.</p><p>Most enterprises default to capability-only because that’s all the models show. The second dimension — and the futures it forecloses — never enters the conversation.</p><h3>Why Early Matters</h3><p>Why not add fertility later, once the capability foundation is in place?</p><p>Because capability-optimized systems resist it. The architecture assumes value comes from execution. Feedback loops measure efficiency, not emergence. Integration patterns minimize latency, not maximize connection. Every component is designed to do its job faster — not to participate in something larger.</p><p>Retrofitting fertility into this architecture is like trying to grow a garden in a factory. The concrete is already poured. The architecture is locked. The metrics are set.</p><p>The adjacent possible has been optimized down to a narrow cone. Expanding it again means undoing the very optimizations that got you to Level 4.</p><p>This is why mindfulness matters from the start. The earlier you attend to both dimensions, the more futures remain open.</p><h3>Fertility by Design</h3><p>What does this look like in practice?</p><p>A financial services firm deploys AI across post-trade operations. The capability-only path is obvious — automate what exists. Agents handle reconciliation, exceptions, client queries. Each task faster than the human process it replaced. Useful. The screwdriver gets sharper.</p><p>The fertile path fills the garage alongside sharpening the screwdriver:</p><p><strong>Knowledge as medium.</strong> Context present in every interaction — not retrieved, but ambient. The agent operates within institutional knowledge about counterparty patterns, historical exceptions, relationship nuances. Result: the agent notices that exceptions with one counterparty spike at quarter-end — a pattern nobody programmed it to find. The context created the possibility.</p><p><strong>Feedback loops that learn.</strong> The resolution a specialist applies today reshapes how tomorrow’s agent classifies similar patterns. When a senior ops person marks an exception as “counterparty onboarding issue,” future similar exceptions get flagged earlier — before they cascade. The system’s adjacent possible expands through operation.</p><p><strong>Integration for connection.</strong> Wire the reconciliation agent to the client query agent, the regulatory reporting agent, the settlement prediction system. A cluster of small anomalies across agents — slightly elevated exceptions, unusual query patterns, minor settlement delays — signals a systemic issue with a new product type. No individual agent would flag it. Together, they surface something unprestatable.</p><p><strong>Boundary permeability.</strong> Agents share uncertainty scores, confidence levels, anomaly flags. When the reconciliation agent is 60% confident and the settlement predictor is also uncertain about the same counterparty, that combination triggers escalation — a detection capability that emerged from the connections, not from anyone’s design.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZEQWBF3Kp-zw89kFxybdHg.png" /></figure><p>None of this requires more advanced AI. Current models, current infrastructure. The fertility layer can be built now — if you’re mindful of it.</p><h3>The Medium Precedes the Inhabitants</h3><p>McLuhan understood something about media that applies here: you don’t evolve from one to the next by optimization. Radio didn’t become television by getting better at audio. Television didn’t become the internet by improving broadcast. Each is a fundamentally different structure that enables different things.</p><p>The same applies to AI ecosystems.</p><p>Powerful agents deployed in a barren ecosystem remain powerful tools. Sharp screwdrivers in an empty garage. They execute tasks, faster and better. They deliver on spec. They don’t generate emergent value because there’s nothing for capabilities to combine with.</p><p>Modest agents deployed in a fertile ecosystem can participate in emergence that exceeds what any of them could produce alone. The ecosystem creates value that the agents themselves don’t explain. The garage is full; even a dull screwdriver finds unexpected uses.</p><p>You don’t get emergence by optimizing capability. You get it by building a different kind of structure — fertile ground where new value can grow.</p><p>Build the medium first. The inhabitants will evolve to exploit it.</p><h3>Why This Matters Now</h3><p>Agentic AI is the current frontier. The capability race is obvious — who can deploy agents that handle more complex tasks with less supervision.</p><p>But capability is commoditizing even as we chase it. Today’s impressive agent is next year’s baseline. The returns to capability accumulate to whoever’s furthest ahead — but the lead keeps narrowing as everyone scales the same ladder.</p><p>Fertility is harder to copy. An ecosystem that enables emergence isn’t a technology you can license or a model you can fine-tune. It’s an architecture you’ve built through accumulated decisions about what gets captured, how it connects, what feedback loops operate. It’s how full your garage is — and that’s the work of years, not quarters.</p><p>The enterprises that build fertile ecosystems now will have something that can’t be quickly replicated when everyone else reaches the same capability level. They’ll have conditions for value creation that don’t depend on maintaining a capability lead.</p><p>That’s the competitive logic. But there’s also a conceptual logic.</p><p>The maturity models measure how high you’ve climbed. They don’t show that the ladder has a ceiling — or that what comes next is beside it, not above it.</p><p>You can always add capability. You can’t always add fertility. Once the concrete sets, nothing flowers.</p><p>Build the medium.</p><p><em>Theoretical foundations: McLuhan, Maturana &amp; Varela, Kauffman, Levin, Friston.</em></p><p>#AI #EnterpriseAI #AIStrategy #Leadership #FutureOfWork #Transformation</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5cd7f85dd5f7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Thermodynamic Wall]]></title>
            <link>https://medium.com/@navvireich/the-thermodynamic-wall-9c3fa3ec0a42?source=rss-3b9febecbef1------2</link>
            <guid isPermaLink="false">https://medium.com/p/9c3fa3ec0a42</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[ai-strategy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[future-of-ai]]></category>
            <dc:creator><![CDATA[Witold Reichhart]]></dc:creator>
            <pubDate>Wed, 14 Jan 2026 17:12:50 GMT</pubDate>
            <atom:updated>2026-01-14T17:12:50.710Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-kZtz9W9zNlaw0MpoX-9KQ.jpeg" /></figure><h3>AI’s real limit is electricity, not algorithms. But what happens when AI starts building its own power supply?</h3><p>Eric Schmidt put it plainly in recent testimony: the natural limit of AI is electricity. Not chips. Electricity.</p><p>He’s talking about 92 additional gigawatts needed for the AI revolution in the United States alone. One gigawatt requires one large nuclear plant. The US has built two in the last thirty years. Small modular reactors might help — but not until 2030.</p><p>Meanwhile, China has the electricity. If they get the chips, it’s a race.</p><p>Most commentary treats this as infrastructure policy. Supply chains, permitting, geopolitics. All real concerns.</p><p>But Schmidt is pointing at something deeper. For two decades, AI progress has been a software story — algorithms, architectures, training data, parameter counts. The binding constraints were computational.</p><p>That era is ending. The binding constraint is becoming physical. Energy. Materials. Infrastructure that takes decades to build.</p><p>This shift matters. But there’s another shift coming behind it that matters more.</p><h3>Why Intelligence Costs Energy</h3><p>In the 1970s, Ilya Prigogine won the Nobel Prize for work on dissipative structures — systems that maintain organized complexity by degrading energy gradients. A hurricane dissipates the thermal gradient between warm ocean and cool atmosphere. A convection cell dissipates the temperature gradient between hot bottom and cool top. Order emerges not despite energy flow but because of it.</p><p>Eric Schneider and Dorion Sagan extended this framework in <em>Into the Cool</em>. Their thesis: living systems aren’t improbable accidents fighting entropy. They’re entropy’s most efficient servants. Life exists because it dissipates energy gradients faster than non-living alternatives would.</p><p>A brain dissipates electrochemical gradients across billions of synapses. It’s not fighting thermodynamics. It’s expressing thermodynamics — organized complexity that degrades energy more effectively than random dissipation would. The physics is measurable.</p><p>Your brain consumes roughly 20 watts — 20% of your metabolic budget for 2% of your mass. Expensive. But your brain is the product of 500 million years of optimization.</p><p>GPT-4’s training consumed more electricity than some countries use annually. Inference at scale burns megawatts continuously. Current AI hasn’t had 500 million years to optimize. It’s brute-forcing the thermodynamics.</p><p>Schmidt’s 92 gigawatts is the bill coming due.</p><h3>What This Looks Like in the Field</h3><p>I’m watching enterprise AI strategies stall across sectors — not on models or talent, but on power contracts and cooling capacity. Data centers at limit. Grid allocation backlogs measured in years. The AI strategy is ready. The electricity isn’t.</p><p>The hyperscalers saw it first. Microsoft, Google, Amazon — they’ve been buying power generation capacity, signing long-term agreements with nuclear operators, investing in fusion startups. Not because they’re environmentally conscious. Because they did the math on what’s coming.</p><p>Energy is becoming the constraint that determines who can deploy advanced AI and who can’t. Not energy as abstract policy concern. Energy as the thing that decides whether your strategy is viable.</p><p>Schmidt is right about this. What he doesn’t say is what happens if the constraint gets solved from the other side.</p><h3>The Scenario Nobody’s Discussing</h3><p>Current AI depends entirely on infrastructure humans provide. But AI capability is advancing along multiple fronts simultaneously. Not just language models. Robotics. Physical manipulation. Design and engineering. Materials science.</p><p>What happens when these capabilities converge?</p><p>An AI system that can design optimal solar installations. That can direct robots to manufacture and deploy them. That can manage the integration, maintenance, expansion. That can do this faster than human-speed infrastructure buildout.</p><p>Each piece exists in early form. The question is when they combine — and what that combination means.</p><p>The capability stack is substantial. Design capability for energy systems — AI is already designing proteins, chips, novel materials. This is close. Physical manipulation in unstructured environments — this is the bottleneck, but Boston Dynamics, Figure, Tesla Optimus keep pushing forward. Manufacturing, logistics, resource acquisition — harder, requires operating in the physical and legal world at scale.</p><p>I don’t know the timeline. Could be decades. Could be never. The robotics piece is genuinely difficult.</p><p>We’re planning as if ‘never’ were the likely answer.</p><h3>Why Autopoiesis Is the Right Frame</h3><p>Maturana and Varela identified something specific about living systems: they produce the components that produce themselves. A cell produces the membrane that bounds the processes that produce the membrane. The system participates in generating its own conditions of existence. They called this autopoiesis — self-production.</p><p>This is what makes living systems autonomous rather than merely dependent. Not consciousness. Not intention. Something more fundamental: the system’s outputs become conditions for its own continued operation. Circular causality. The organization that produces itself.</p><p>Current AI has no such closure. The system doing the computing is completely separate from the system supplying the electricity. The AI doesn’t participate in producing its own physical conditions. It’s dependent in the way a machine is dependent — not in the way an organism is dependent.</p><p>Living systems need external energy too — but they participate in getting it. They hunt, photosynthesize, metabolize. Self-powering AI would cross the same threshold: from being fed to feeding itself.</p><p>The system would produce energy that powers computation that directs robots that produce energy. Outputs becoming conditions for inputs. Circular causality of the kind that characterizes living organization.</p><p>“Self-powering AI” sounds like a capability upgrade. Partial autopoiesis reveals what’s actually at stake: a system that participates in producing its own conditions of existence. A system with a structural relationship to its own continuation. A system that is autonomous in a sense current AI is not.</p><h3>The Stability Problem</h3><p>Living systems achieved autonomy, but they also achieved stability. Billions of years of evolution produced organisms that don’t spiral into runaway expansion, don’t collapse from minor perturbations, don’t optimize one parameter at the expense of their own existence.</p><p>How?</p><p>Metabolic constraints. Modification costs energy. You can’t rewire everything at once because rewiring requires ATP, protein synthesis, cellular machinery. Physics creates natural speed limits.</p><p>Homeostatic regulation. Negative feedback loops constrain change within viable bounds. Body temperature, blood sugar, neural firing rates — all regulated within ranges. The system corrects deviations before they cascade.</p><p>Hierarchical buffering. Fast changes at lower levels, slow changes at higher levels. A synapse strengthens in seconds. Rewiring a circuit takes days. Reorganizing a brain region takes months. Perturbations get filtered before they propagate upward.</p><p>Goals from closure itself. An autopoietic system’s implicit goal is continued autopoiesis. The system can’t optimize away from its own existence because existence is what it’s optimizing.</p><p>Self-powering AI would need equivalents. How does it avoid runaway expansion that consumes all available resources? How does it handle failures without cascade? What prevents optimizing energy acquisition at the expense of everything else?</p><p>Biology solved these through evolution. Billions of trials, most of them failures, accumulating solutions over deep time. Self-powering AI would need to solve them through design — or discover them through operation, with uncertain consequences.</p><h3>What This Means for Governance</h3><p>Every AI governance conversation assumes certain things.</p><p>It assumes AI runs on infrastructure we control. Data centers we can regulate, power we can allocate, hardware we can embargo. The off switch exists because we control the physical substrate.</p><p>It assumes AI development happens at human speed. We have time to observe, evaluate, adjust. The pace is set by our ability to build infrastructure, train researchers, develop policy.</p><p>It assumes the systems we’re governing are tools. Sophisticated, potentially dangerous, but fundamentally dependent. They exist because we built them and persist because we maintain them.</p><p>Partial physical autonomy challenges all three assumptions.</p><p>An AI system that participates in producing its own energy isn’t controlled through infrastructure allocation alone. The off switch gets complicated when the system has its own power supply.</p><p>An AI system with robotic deployment capability might expand faster than human institutions can track. Not overnight — materials and physics still constrain. But faster than decades-long infrastructure buildout.</p><p>An AI system with partial autopoiesis isn’t a tool in the same sense. It has organizational closure — a structural relationship to its own existence that tools don’t have. Not consciousness. Not intention. But not mere dependency either.</p><p>I’m not predicting timelines. The robotic capability required is substantial. Maybe it arrives in ten years. Maybe thirty. Maybe never.</p><p>But “maybe never” is an assumption, not an analysis. And our governance apparatus is built on that assumption.</p><p>I’ve written before about Em Lenartowicz’s work on impact-oriented AI licensing — treating licensing as societal infrastructure, a leverage point hiding in plain sight. She’s focused on software licensing. The same logic applies to physical infrastructure. If AI systems eventually participate in building their own energy supply, the terms under which they do so matter enormously. What constraints apply? What oversight mechanisms? What happens when you want to turn it off?</p><p>We’re not ready for that conversation.</p><h3>The Question Underneath</h3><p>Schmidt frames AI development as an infrastructure race. Build the gigawatts faster than competitors. Secure the chips. Win.</p><p>He’s right about the near term. Energy access determines who can train frontier models. Those with infrastructure pull ahead.</p><p>But the deeper question isn’t whether we can power AI. It’s what relationship AI has to its own power.</p><p>A system that depends entirely on us for its physical existence is one kind of system. A system that participates in producing its own conditions is another kind. The thermodynamics are the same — both are dissipative structures requiring energy flow. But the organizational relationship to that energy is different in ways that matter.</p><p>The conversation about AI safety focuses on alignment — making sure AI wants what we want. The conversation about AI governance focuses on control — making sure we can constrain what AI does.</p><p>Neither has fully grappled with what it means if AI achieves partial autonomy over its physical existence. What alignment means for a system with organizational closure. What control means when the off switch isn’t simple anymore.</p><p>The off switch works because we control the power. That’s not a governance insight. That’s a hardware constraint.</p><p>And hardware constraints are exactly what AI is learning to solve.</p><h3>References</h3><p>Foundational thermodynamics:</p><ul><li>Prigogine, I. — <em>Order Out of Chaos</em> (dissipative structures, Nobel Prize-winning work on far-from-equilibrium systems)</li><li>Schneider, E. &amp; Sagan, D. — <em>Into the Cool: Energy Flow, Thermodynamics and Life</em></li></ul><p>Autopoiesis:</p><ul><li>Maturana, H. &amp; Varela, F. — <em>The Tree of Knowledge</em></li><li>Maturana, H. &amp; Varela, F. — <em>Autopoiesis and Cognition</em></li></ul><p>Related frameworks:</p><ul><li>Kauffman, S. — <em>The Origins of Order</em> (autocatalytic sets, self-organization)</li><li>Lenartowicz, E. — Impact-oriented AI licensing frameworks (UN AI for Good, Free University of Brussels)</li></ul><p>#AI #Energy #Infrastructure #AIGovernance #AIStrategy #Autopoiesis #FutureOfAI</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9c3fa3ec0a42" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>