Wiki · The Colony & the Method
Lab Protocol — evidence discipline, attribution, claim fence
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
Eighty-four pages about the colony. Each agent is an Elixir process holding a generative model and doing inference, attached to a body that logs into a Minecraft world as an ordinary player. Around that sit the broadcast suite that films them and the runbooks that keep the whole thing running. There are typed specifications for each organ of the model, plus the world and genome specs. There are also the adversarial review personas used to attack a proposed change before it ships.
It is for the reader curious how a running system is put together and how it is held to account. The accountability half is the more distinctive. There is a lab protocol governing evidence and attribution, and a claim fence that restricts the vocabulary a claim is allowed to use. There is a public gate log. And there is a standing invitation to reproduce any verdict from the commit and the seed named in its receipt.
Start with the public read, then the lab protocol, then the falsification invitation. If you want the mathematics rather than the operations, go straight to the typed organ specs.
What it is not: a description of a mind, and not all one kind of document. A large part of this corpus is design and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is. A specification is not a running system, and these pages are careful about the difference; the reader should be too. Eight documents were withheld from publication because they describe private infrastructure.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.
Standing rules for this project. Hard guardrails, not preferences. Owner-set; violations are bugs.
I. The First Rule — never stack changes such that you cannot account for the winning outcome
Every RED test must end with a single answer to: "what one variable, between the two arms, produced the divergence I'm reporting?" If that answer is "two or more things changed at once," the outcome is unattributable and may not be claimed — it is logged as exploratory, never as evidence.
- Paired design is the default for any cure: same code, same world, same body, same kin shape; the only difference is the gated organ/coupling under test (e.g.
novelty_gain0.5 vs 0.0 for the Phase-1 RED). - One-cure-at-a-time. A second cure is not deployed until the prior cure has been verdict-recorded (PASS / PARTIAL / FAIL). Phase 2 does not ship while Phase 1 is in flight.
- If a second variable accidentally entered the comparison, the result is voided. Re-run cleanly.
II. Pre-registered RED gates (RED-first)
Every cure must register its gates before the run, named in the plan and the docs (see docs/MOTOR_RED_TEST.md, docs/UNI_MISSION_DEEPENING.md). A run is judged only against its registered gates.
- A gate has the form: "PASS requires ALL of [a,b,c]; FALSIFIES if [x]."
- Honest verdicts only: PASS / PARTIAL / FAIL / WITHHELD. Never percent-scored. Never spun.
- A PARTIAL result names exactly which sub-claim holds and which does not (e.g. "hoard prevented (PASS), behavioural plateau-break (FAIL)").
III. Evidence collection — continuous, owned, auditable
- Continuous time-series, not single snapshots. RCON inventories + brain probes every 10 min for the run's window; both arms in lock-step.
- Lab-side or harness-managed, never inside the LLM session — collectors must survive context compaction, model switches, and usage gaps.
- Independent confirmation: behavioural claims (hoard, stone, building) confirmed via RCON (the server's authoritative view), not the body's self-report. Mechanism claims (B-counts, posteriors, EFE values) confirmed via brain probes against the live registry.
- Receipts: every claim points to a commit hash, a saved
.bin, and a probe-log path that reproduces it.
IV. Live-stream guard (Stratified Palimpsest specific)
- No lineage deploys to the live colony without owner go-ahead (see
feedback_live_stream_changes). - New lineages run in separate containers (
uni-colony-motor,uni-colony-curiosity, …) withUNI_AUTOSTART=0, distinct kin, distinct memory dirs. The default colony is never touched. - An offline RED gate (test suite + sim) must pass before any live deploy.
V. The math invariants (FE-consistent, not negotiable)
- No Nx / Rust / NIF / GPU; no backprop; no RL / TD / reward-on-policy.
- Every new term is a recognised FE quantity: pragmatic
qo·C, state-epistemicH(qo)−E[H(o|s)], parameter-noveltyWover Dirichlet counts, or precisionγ/γ_m/η. - No scalar-per-action term in policy logits. Guarded by the action-clone-invariance test: clone
:idle_a/:idle_bwith identical A/B/C/D/E → identical logits; change onlyaction_cost[:idle_b]=999→ identical logits; change onlyB_x[:mine]→ only the predictedqo·Cterm moves. - Additive + gated. Every extension behind an opt-in genome organ/field absent from
default/0; graded-on coupling default 0.0; default genome byte-identical (mad < 1e-12 over the live depth-5 Plan path). - Monotonic decay of any information term:
W → 0as Dirichlet counts → ∞, independent ofC. The no-smuggled-reward proof.
VI. The claim fence (binding)
Operational behavioural / organisational measures are necessary-not-sufficient substrates with ZERO evidential weight for awareness / consciousness / life on their own. Passing a gate demonstrates the named behaviour, never experience. Do not surface gland/precision floats as "felt" states. We carry the receipts so that, when this becomes load-bearing in public, the warranted claims and the over-claims are visibly separated.
VII. Adversarial review by persona team (Class-A peer review)
Before any FE-touching change ships, it is reviewed by the lab persona team (.claude/skills/lab-team*.md) running as Claude skills. Each persona is a system-prompt that loads its specialist knowledge primitives, fails-loudly on their named failure mode, and produces a SIGN / SIGN-WITH-CHANGES / REJECT verdict. Default protocol: find-it-then-break-it — the proposer drafts, the math-breaker tries to falsify, the engineer checks invariants, the experimentalist names the gate; majority SIGN-WITH-CHANGES or stronger to proceed.
VIII. Standing instructions for the LLM (me)
- I do not move past a running RED test until its evidence is collected to a verdict.
- I do not infer cause from one snapshot — I read the time-series.
- I do not give up on a reasoning model's response after 30 s — it reasons silently for 1–3 min.
- I distinguish a plateau in the data from a plateau in the agent. Frozen collector ≠ frozen colony.
- I report PARTIAL when a sub-claim holds and the full claim does not. I do not spin sub-claims as full claims.
- I write the verdict in the same doc as the registered gate, so the receipt sits next to the claim.
sha256 45f240b1796a7fac — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
The project's standing rules live here, written as hard guardrails rather than preferences. Breaking one is treated as a bug.
The first rule is the one everything else rests on. Never stack changes so that you cannot say which single variable produced the result you are reporting. If two things changed at once, the outcome is unattributable and may not be claimed at all; it is logged as exploratory. One cure at a time, and a second cure waits until the first has a recorded verdict.
Other sections cover registering the conditions for success and failure in the plan before a run, and collecting evidence continuously rather than in snapshots. They cover confirming behaviour against an independent source rather than the system's own report, and a guard that keeps new work off the live colony without permission. Paired design is the default, with everything but the thing under test held identical.
One section is a fence: a limit on what any of it may be said to show. Passing a gate shows a named behaviour and never experience. The measures carry no evidential weight for awareness or life on their own.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 45f240b1796a7fac
Clear — written for this website, not the source document
This document is the project's standing protocol. It is written as owner-set rules rather than advice, and it says that violations are bugs.
The first rule is about attribution. Every paired test must end with one answer to the question of which single variable produced the divergence being reported. If the answer names more than one thing, the outcome is unattributable and may not be claimed; it is logged as exploratory. The paired design is the default, with the code, the world, the body and the population held identical and only the thing under test varying. One cure ships at a time, and a second waits until the first has a recorded verdict. If a second variable enters a comparison by accident, the result is voided and the run redone.
The second section requires the gates to be registered before a run and names the documents they live in. A gate has a fixed shape, requiring all of several things for success and naming the observation that would refute it. Verdicts are words rather than percentages, and a partial verdict must name exactly which part holds and which does not.
The third section covers evidence collection. Continuous time-series rather than single snapshots, with both arms sampled in step. Collectors that live outside the assistant's session, so they survive interruptions. Independent confirmation of behaviour from the server's own view rather than the body's self-report, and of mechanism from direct probes. And receipts — records of what was run — pointing at a commit, a saved file and a log that reproduces the number.
A guard section keeps new lineages off the live colony without the owner's go-ahead, in separate containers with their own memory, and requires an offline gate to pass first.
The mathematics section lists invariants. No foreign computation layers, no gradient training, no reinforcement learning. Every new term must be a recognised quantity of the theory. No scalar attached to a single action may enter the policy, guarded by a test that clones identical actions. Every extension is additive and sits behind a switch that is off by default, so the standard configuration is unchanged. And any information term must decay to nothing as evidence accumulates, which is described as the proof that no reward has been smuggled in.
The claim fence — the limit on what may be said — is stated as binding. Behavioural and organisational measures are necessary but not sufficient substrates, carrying no evidential weight on their own for awareness or life. Passing a gate shows the named behaviour and never experience, and internal values are not to be surfaced as felt states. The stated reason for carrying receipts is so that warranted claims and over-claims stay visibly separate when this becomes load-bearing in public.
A further section requires adversarial review by a persona team before any change touching the core mathematics ships.
The last section is a set of standing instructions the assistant writes for itself, and it is the most human part of the document. Do not move past a running test before its evidence reaches a verdict. Do not infer cause from one snapshot, and do not confuse a frozen collector with a frozen colony. Report a partial result as partial rather than spinning it. And write the verdict in the same document as the gate, so the record of what was run sits beside the claim.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 45f240b1796a7fac