02 / SOCIETY OF MINDS
Colony is the living city. MoME is how it chooses experts. Society Protocol is how those experts talk — without one of them becoming observer, judge, and motor at once.
A society earns a coherent answer when no single voice owns the loop.
claim ceiling · architecture illustration, not phenomenal consciousness
THREE NAMES / KEEP THEM DISTINCT
The public name Society of Minds is the architecture. Colony is the city that lives it. MoME is the router inside that city. Confusing the three turns a dispatch ticket into a claim of mind.
An inspectable multi-agent system: bounded mentats, a substrate of programs and receipts, and a human sponsor who keeps final authority.
A two-level router: which mentat should speak, then which visual profile if the task is seen rather than only said. The route is disclosed. It is never motor permission.
Classical multi-agent law. Mentats talk only through Steward. Contradiction is kept. One persuasive component cannot be observer, planner, judge, and historian at once.
01 / COLONY
Colony is Valentin Kalinin's living system for evidence-bound agents. YAML is the contract. The CLI is the truth path. Chat executes a turn; it does not remember. A diagram on this page is architecture, not a runtime receipt.
The human keeps the constitutional share of the city. Agents propose; they do not inherit the right to close a claim.
Named crafts: orchestration, proof, software, embodiment, language, image, story. Each has a seat. None owns the whole loop.
Cognition programs, memory, and sealed afterstates. Work that cannot show a receipt is not yet city work.
Projects what was expected, why an action was taken, what changed, and which hypothesis suffered. It has no outgoing motor edge.
02 / MoME
MoME expands mixture-of-experts from “which model” to “which mentat, then which visual profile.” Callers attach the route to the turn so a later audit can see who was actually asked.
A bounded job with a named goal. Prestige of a model is not a routing signal.
Assistant-level expert: Nova to orchestrate, Inquisitor to prove, SWE to patch, Archi to embody a puzzle, Tim to teach.
When the task is seen, MoME also selects a visual expert profile. Language-only work skips this rung.
The trace names the mentat, optional pet pairing, and why that coalition was offered. Hidden routing is a lie.
A MoME ticket is advisory. Motor permission is a later deterministic gate, or abstention.
Some routes pair a mentat with a totemic companion. That is routing texture, not a second personality with motor rights.
When several experts are offered together, the city records coalition structure. That is not a vote that the answer is true.
Synergy weights begin as disclosed priors. A learned coupling is a later claim, and only with evidence.
03 / SOCIETY OF MINDS
The name honors Minsky: intelligence as a society of bounded agents. Colony implements that as classical protocol and receipts — not as a recreation of the 1986 book, and not as a physics channel.
A faux council is a lie. If a mentat was not invoked, the city does not speak in that voice.
Cross-mentat state moves through Steward. Direct assistant-to-assistant bypass is forbidden.
When no evidence-bound continuation survives, the motor gate stays shut. Uncertainty is not laundered into action.
LOAD-BEARING SEATS
HOW A TURN RUNS / MAS OPERATIONAL AWARENESS
A Society of Minds earns a coherent answer through bounded specialists, deterministic action authority, sealed afterstates and a durable memory of contradiction. The reflective layer watches this loop; it never becomes the motor.
Ht → Ât+1 → at → At+1 → δt → Mt+1
hypothesispredictionactionsealed afterstatedifferenceupdated self-model
OBSERVABLE MANIFESTATION
A useful self-model must alter future behaviour in ways that can be falsified across embodied turns.
A contradiction changes the next candidate set instead of becoming a footnote to the same behaviour.
Evidence-backed progress survives several embodied transitions without collapsing into fresh guesswork.
The system closes the forecast against the sealed world response and records what failed to match.
The revised move points back to observations and causal receipts, not to mood, confidence or model prestige.
An empty glass is a valid terminal state. Uncertainty cannot be laundered into motor authority.
DIVISION OF EPISTEMIC RESPONSIBILITY
The Society is not a cast of characters. It is an architecture that prevents one persuasive component from becoming observer, planner, judge and historian at once.
Supplies current frames and typed world state. It does not explain them.
Produces competing hypotheses from language, spatial, causal, memory and domain-specific views.
Projects evidence and continuity. It has no outgoing motor-control edge.
Freezes the expected afterstate before action, so surprise cannot be rewritten later.
Admits one scoped action. An LLM label or critic verdict is never execution permission.
The environment, not the Society, decides what actually changed.
Closes prediction against evidence and injures the hypothesis that failed.
Keeps supported phases, retractions and provenance available to the next turn.
Stops the loop when no evidence-bound continuation survives the gate.
CLAIM CEILING / FIXED
This contour can support a falsifiable operational claim: the system models its own causal history and changes future behaviour accordingly. It cannot establish, detect or simulate phenomenal consciousness.
ENTER ONE EXPERIMENT