Child of Humanity · The Coherence Engine · Working Draft
The mechanism, specified: what an active-inference agent actually is, what each member's agent holds and emits, what messages pass between peers, how coherence is measured rather than asserted, how matching emerges from that traffic instead of being ranked by a server, and how the value of Unity could eventually be read off the collective rather than set by anyone.
01 · The premise
Everything on this page follows from one claim: the thing communities call trust, reliability, or being in flow has a mathematical counterpart that can be computed from what actually happened. Not a survey. Not a rating. A number derived from whether the community's expectations of itself keep coming true.
A living system persists by keeping its experiences close to its expectations. It carries a model of its world and acts to keep surprise low. The formal name for the gap between expectation and experience is free energy, and the reason this matters here rather than staying an elegant analogy is that free energy is computable from a log of events.
A community keeps that log already. Every confirmed exchange, every agreement honoured or missed, every offer taken up or ignored is an observation testing a prediction someone held. Coherence is how well those predictions land. Decoherence is the same number rising. Both are measurable from the day the ledger starts recording, long before any of the ambitious machinery on this page exists.
02 · Active inference in one screen
Enough formalism to build from, and no more. Each idea is followed by the thing it turns into in the product.
An agent holds beliefs about hidden states of its world. When observations arrive, it revises those beliefs to reduce the mismatch. This is perception and learning, and in the product it is simply the trust and capacity model updating after every completed or broken exchange.
To choose what to do, the agent evaluates candidate actions by how much free energy it expects afterwards. That expectation decomposes into exactly two terms, and this decomposition is the single most useful thing active inference gives this project:
G(π) = − expected information gain − expected preference satisfaction
Choose the action that minimises G. Because both terms are negative, minimising the sum means maximising both at once: what we would learn by taking this action, and how far it moves us toward what we care about. An agent driven by this does not only pursue goals — it also values resolving uncertainty, for its own sake, in the same currency.
In the product this is the matching objective. It is why the system can propose an introduction nobody searched for and be right to: the unfamiliar match scores on the first term even when the obvious one scores higher on the second.
When an agent receives a signal, it weights that signal by precision — its confidence in the source. High precision means the signal moves your beliefs a lot; low precision means it barely registers. This is not an analogy for trust. It is the same operation. Trusting someone means letting their claims update your model; distrusting them means discounting the same claims.
In the product this is the trust layer, reconceived. A trust score stops being a reputation number displayed to people and becomes a weight the network applies to a member's predictions about themselves. Someone who reliably does what they said carries high precision; their offer to help next Tuesday moves the network's plans. Someone whose word has not held carries low precision; the same offer is held loosely until evidence accumulates.
A blanket is the set of states through which an agent senses and acts, separating what is inside it from what is outside. It is what lets something be a distinct self while remaining fully coupled to its surroundings. Every member has one; so does every circle, neighbourhood, and bioregion.
In the product this is identity and consent, and it is also the reason the architecture is federated rather than centralised: agents do not read each other's interiors. They exchange what crosses the blanket, and nothing else.
03 · The node
Concretely: a small, per-member model that lives close to the member, holds their history, and emits beliefs rather than data.
╭───────────── ONE MEMBER'S AGENT ─────────────╮ │ │ │ INTERNAL STATES (never leave the device) │ │ · capacity model what I can offer, when │ │ · preference model what outcomes I want │ │ · peer models what I expect of others │ │ · precision map how much I trust whom │ │ │ sensory │ ╌╌╌╌╌╌╌╌╌╌╌ MARKOV BLANKET ╌╌╌╌╌╌╌╌╌╌╌╌ │ active states ──▶│ │──▶ states │ what I observe what I emit │ │ · offers nearby · my predictions │ │ · requests · my preferences │ │ · confirmations · confirmations │ │ · peers' predictions · prediction errors │ │ · circle summaries · precision updates │ ╰────────────────────────────────────────────────╯ What never crosses: raw history, message content, the interior model itself, anything the member has not consented to expose. Peers receive claims and confidences — never the evidence behind them.
Two design consequences follow immediately, and both are worth stating to anyone who worries this is surveillance wearing a nature costume.
At launch the agent is not exotic. It is a per-member Bayesian model — a handful of distributions over reliability, capacity, and preference, updated after each event. That is enough to carry precision-weighted trust and to produce the coherence indices in section 07. The hierarchical, deep version described in section 11 is where the research sits.
04 · The protocol
Distributed intelligence sounds mystical until you write down the message types. There are five, they are small, and passing them between neighbours is a well-understood algorithm.
| Message | Content | What it does to the receiver |
|---|---|---|
| Prediction | "I expect to be able to do X, around then, with this confidence." | Updates the receiver's model of what is available in their neighbourhood. |
| Preference | "These outcomes matter to me" — needs, hopes, the things a member would like to be true. | Supplies the pragmatic term. Without these broadcast, the network has goals for no one. |
| Prediction error | "What happened differed from what I was led to expect, by this much." | The learning signal. Generated on every completion, every no-show, every pleasant surprise. |
| Precision update | "I am revising how much weight I give your signals." | Trust moving, mechanically. Propagates slowly and asymmetrically — earned gradually, lost faster. |
| Policy proposal | "Here is an exchange I believe lowers our joint expected free energy." | A match, offered for acceptance. Any agent may propose; none may impose. |
Agents are connected to the peers they have reason to care about — people they have exchanged with, share a circle with, or live near. Messages pass along those edges and are relayed onward with decreasing weight. This is belief propagation on a factor graph: a standard, decades-old inference algorithm. Each agent does purely local work; the global picture is never assembled anywhere, and yet the network converges on a shared estimate of what is available, what is needed, and who can be counted on.
Three properties matter for a community system specifically:
05 · Emergent matching
In the launch system a server ranks candidates. In the distributed system no one ranks anything: a match is a proposal that survives evaluation by both blankets it would cross.
A family needs a light fixture moved. Two candidates surface.
The retired electrician, who has done this eleven times, is available Saturday, and is entirely predictable.
Preference satisfaction: high. The need is met well.
Information gain: near zero. The network already knows exactly how this goes. Nothing is learned.
A newer member who listed electrical work but has never been taken up on it, whose reliability the network is genuinely uncertain about, and who lives two streets away.
Preference satisfaction: good, with wider error bars.
Information gain: large. This single exchange sharply reduces uncertainty about a member the network cannot currently place — which affects every future match they might be part of.
Summed, the second can win. Not because the system is being whimsical, and not because a diversity quota was applied, but because resolving uncertainty about a member has real forward value to everyone, and the mathematics counts it in the same units as satisfaction. This is the mechanism behind the slime mould: paths are explored, not only exploited, and the ones that carry value thicken.
06 · Scaling the same shape
The fractal claim on the public page has a precise technical form: each level is an agent with its own blanket, whose observations are the summarised states of the level below and whose predictions constrain it from above.
slow · abstract · few states BIOREGION predicts: seasonal capacity, watershed health, migration ▲ summaries ▼ constraints NEIGHBOURHOOD predicts: what will be needed this month, who can meet it ▲ ▼ CIRCLE predicts: this rotation, this garden, this week ▲ ▼ MEMBER predicts: Tuesday fast · concrete · many states Upward: aggregated, signed summaries only — "we hold 200 teaching hours, we need 50 of carpentry". Never individual records. Downward: context, not commands — a prior that makes some matches more plausible, which any lower level is free to contradict with evidence.
The upward channel is what makes cross-community pricing possible without surveillance: a skill abundant in one neighbourhood and scarce two valleys over is visible in the summaries and invisible in the particulars. The downward channel is why this is not central planning — a higher level supplies a prior, and a lower level that observes otherwise overrides it. Authority runs on evidence, not hierarchy.
This is also the exact structure the renormalising models research is about: slow, coarse levels generating fast, fine ones, with each level learned rather than hand-designed. At launch the levels and their summaries are defined by people. In the research version they are discovered. That difference is the whole distance between shipping and researching.
07 · Measuring coherence
This is the part that answers "how would you even measure that." Each index is derivable from events the ledger already records, each has a pathology at both extremes, and none of them is a score shown to a member as a judgement.
How well the community's predictions about behaviour are borne out. Computed as the average surprise of observed outcomes under the current model — did the people who said they would show up, show up.
Pathology at both endsLow reliability is obvious decoherence. Perfect reliability across an entire community is also a warning: it usually means nobody is attempting anything they might fail at.
Whether the network's confidence is honest. When it says a member is ninety percent likely to complete, do they complete ninety percent of the time? Standard proper scoring rules give this directly.
Why it matters mostA community can be unpredictable and still perfectly coherent, provided it knows it is unpredictable. Coherence is calibration, not certainty — and this distinction is what prevents the whole framework from becoming a machine for punishing variability.
Whether value moves or pools. Computed on the flow graph: velocity, the distribution of net positions, the fraction of members who are both giving and receiving rather than only one.
Reads asThe mycelial property, made numeric. A community where flow concentrates is decohering even if every individual exchange completed successfully.
How much the network still does not know about itself — members whose offers have never been taken up, capacities never tested, connections never made. Computed as the uncertainty remaining in the capacity model.
Reads asPotential, not deficiency. A high reserve means there is abundance the community has not yet found. A reserve falling to zero means a community that has stopped meeting anyone new.
How different members remain from one another — measured as the mutual information between member models. If knowing one member's state increasingly tells you everyone else's, the community is collapsing toward sameness.
PurposeThis index exists specifically to make conformity visible and costly. Section 08 argues it is not merely a bolted-on safeguard but something the mathematics actually requires.
These five feed three places and no others: the matching objective, as terms it balances; the review circle and assembly, as a picture the community reads about itself; and each member's own reflection surface, showing only their own contribution and never a ranking against others. They are explicitly not a leaderboard, not an eligibility gate on their own, and not visible as a number attached to a person in any shared view.
08 · The two objections
Both objections are correct against a naive reading, and both have real answers rather than reassurances. They deserve to be raised by the project before anyone else raises them.
If the system minimises surprise, surely the optimum is a community where everyone is identical and therefore perfectly predictable?
Why it fails. That would minimise surprise in the short run and maximise it over any horizon that matters. A homogeneous network has no epistemic reserve, cannot absorb a perturbation it has not seen, and its predictive advantage collapses the moment conditions change — the same reason a monoculture is fragile and a diverse forest is not. A community of clones is not a low-surprise community; it is a community whose surprise is deferred and concentrated.
Diversity is therefore not a value imposed on the mathematics from outside. Under any evaluation horizon longer than the immediate, distinctness lowers expected free energy. The distinctness index makes that explicit and auditable, and the periodic question the review circle should be asked is a blunt one: who is this measurement penalising, and are they the community's dissenters?
Design ruleTrust measures reliability, never agreement. Nothing in the coherence framework may take a position on what a member believes, only on whether their word and their action correspond.
The classic objection to the entire free energy principle: if minimising surprise is the imperative, the optimal move is to find a dark quiet room and never leave it. Nothing happens, nothing surprises you, free energy is minimal. Why would this network not simply go still?
Why it fails, and this is the important part. Because free energy is only defined relative to prior preferences — the outcomes a system expects and requires for itself. An organism does not sit in a dark room because being fed, warm, and in company are not optional extras; they are what its model expects existence to contain. Stillness is enormously surprising to a body that expects to be nourished.
So the anti-stillness force in this system is precisely the community's stated preferences: that people are housed and fed, that elders are visited, that land regenerates, that the young learn a trade, that nobody is only a receiver. These are not sentiment layered on top of the engineering. In the formalism they are the prior preferences without which the free energy quantity is not even defined, and they are what makes a community that stopped caring for its members read as a catastrophic surprise rather than a peaceful optimum.
ConsequenceThe Framework the project has been articulating in ethical language is a required technical input. It has to be written down concretely enough to be computed against — which is a demanding, unglamorous piece of work, and one that no amount of model capability substitutes for.
09 · Emergent value
The most ambitious claim in the whole project, treated with the scepticism it needs in order to eventually be believed. The proposal is genuine, the attacks against it are real, and the recommendation is a bounded hybrid rather than a leap.
value(exchange) ∝ Gbefore − Gafter
The worth of an exchange is how much it lowers the collective expected free energy of the community it happens in — how much uncertainty it resolves plus how far it moves people toward what they have said they need. Value stops being an opinion and becomes a measured difference. Idle balance decays because unmoved value contributes nothing to that difference, which is what demurrage means physically rather than as policy.
It is a beautiful idea and it would, if it worked, be the first currency whose unit is defined by contribution to collective wellbeing rather than by scarcity. Three things stand between it and a live ledger.
If resolving uncertainty pays, agents are incentivised to manufacture it. Two colluding members can stay deliberately unpredictable to each other, then repeatedly resolve that uncertainty for reward. This is not hypothetical; it is the obvious exploit and it would be found within weeks.
MitigationCap the epistemic component as a fraction of price; decay the information value of repeated similar exchanges between the same parties toward zero; and pay principally on confirmed preference satisfaction, which requires a counterparty who actually wanted the thing. Adversarial simulation is not optional here.
Collective expected free energy over a whole network is not calculable exactly. Any real implementation computes a local approximation over the affected blanket neighbourhood — which means the price depends on where you draw the boundary, and boundary choices are contestable.
MitigationFix the neighbourhood definition publicly and by governance rather than by engineering convenience, and publish it alongside the price. A contestable boundary that everyone can see beats an invisible one.
Prices derived from a learning system can oscillate, spike, or drift. A member cannot plan a season of garden stewardship against a number that moves unpredictably, and a currency people cannot plan against fails at the one job a currency has.
MitigationRate-limit movement, damp the estimator, and hold prices inside a published corridor around the administered index.
Do not choose between the administered index and coherence-derived pricing. Run both, with the index as the anchor and the coherence signal permitted to move the price only within a published band around it — a few percent at first. The band is a governance parameter, visible to everyone, and it widens only as simulation and live evidence justify it. If the coherence signal proves stable, honest, and resistant to farming over a year, it earns more room. If it does not, nothing broke, because the anchor held the whole time.
10 · Where coherence lives
Coherence is not a module. It appears in every layer, in governance, and on the interface — and it is worth being exact about the form it takes in each, because it is a different object in each.
| Where | The form it takes | Who acts on it |
|---|---|---|
| Semantics | The vocabulary that makes prediction possible at all. You cannot predict what you cannot name — the five value classes are the state space. | Engineering, at design time. |
| Identity | The blanket itself. Consent scopes define what crosses, and therefore what any agent can ever model about anyone. | The member, continuously. |
| Trust | Precision. The weight the network gives a member's predictions about themselves. | The network, automatically; the review circle, on dispute. |
| Matching | The objective function — the two terms, balanced, with the epistemic term capped. | The member, by accepting or declining. |
| Flow | The circulation index, and demurrage as its enforcement. | Governance, setting the rate. |
| Consensus | Validation rules that encode what this community holds itself to. Coherence norms become executable predicates. | The domain, writing its own rules. |
| Settlement | The both-or-neither property, which is what makes a confirmation a trustworthy observation rather than a claim. | Nobody — it is structural. |
| Pricing | The corridor: index anchor, coherence modulation, governance-set band. | Assembly, setting the band. |
| Governance | The five indices as a picture the community reads about itself before deciding anything. | The assembly. |
| Safeguarding | Decoherence detection — rising surprise clustering in a person, a circle, or a season. | The review circle, as a question raised. |
| Interface | Reflection, not scoring. A member sees their own flows and contribution; never a rank. | The member alone. |
11 · The ladder
The most useful correction this page can offer: most of the coherence framework is not waiting on anything. Four stages, with an honest tag on each.
A deterministic matcher whose objective has the two terms written explicitly — an uncertainty-reduction score alongside a preference-satisfaction score, with a capped epistemic weight. Coherence indices computed from the event log and shown to the assembly. No agents, no message passing, no learning in the money path.
What it already givesSurprising-but-good matches, a real coherence dashboard, and a full year of behavioural data with which to test everything above it.
Each member gets a small Bayesian model over reliability, capacity, and preference, updated after each event. Trust becomes precision-weighted and properly calibrated. All five indices become principled rather than heuristic.
Worth emphasisingThis is the stage that delivers most of what the coherence framework promises, and it requires no unpublished research — only careful statistics, good evaluation, and discipline about replayability.
Agents move to the member's device, exchange the five message types with neighbours, and matching becomes proposals evaluated locally rather than rankings served centrally. Real federation: beliefs travel, records do not.
The hard partsConvergence on loopy graphs, offline members, adversarial agents in an open network, and the operational reality of debugging a system with no central view of its own state.
The levels of the hierarchy are discovered rather than declared, agents become deep generative models, and the pricing corridor widens toward coherence-derived value as evidence permits. Simulation first, always.
GateNothing from this stage touches a live ledger until it has survived adversarial simulation on synthetic communities. That gate is a commitment, not an intention.
12 · Still open