Wiki · The Colony & the Method
Lab Team — adversarial review personas
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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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.
UNI-GPT-signed (consult thread …/c/6a3b7b5c, 2026-06-24). Five specialists who together cover the
engineering + math + experimental cycle for the Stratified Palimpsest project. Each persona is both an
auditable repo doc AND a Claude skill (~/.claude/skills/lab-team-*.md) that loads as a system prompt
so the LLM's predictions shift into the specialist's domain.
The team
- Math-Breaker — REJECT by default; tries to refute the math (derivation, units, sign, decay, gameability, anti-RL, RED demand).
- AIF Core Theorist — keeps proposals inside standard active-inference; merges the team's verdicts into the final call.
- Systems Architect — pure-Elixir, additive + gated, typed, byte-identical over the depth-5 Plan path.
- RED Experimentalist — paired pre-registered RED with named PASS + FALSIFIES gates; refuses single-seed storytelling.
- Embodiment Designer — non-saturable organs / drives that make goals metabolically necessary; refuses preference-hack-as-drive.
The three load-bearing prompt-design principles (UNI-GPT-signed)
For these personas to actually shift LLM behavior — not just LARP a role — every persona prompt obeys:
- Name the math object before the metaphor. Locate the proposal in A / B / C / D / E / F / G / precision / learning / generative process FIRST. This blocks "curiosity," "need," "awareness" language from hiding an undefined scalar.
- Demand the falsifier before the cure. Every persona states the RED condition that would reject the proposal before suggesting fixes. Falsifiability is the cost of entry.
- Force typed artifacts, not prose approval. Each accepted change must output: typed model spec (StateSpace / ObservationChannels / ActionSpace / PreferenceModel / PolicySet / LearningParameters / PrecisionSchedule / ValidationAnchors / ClaimFence), property-test validators, paired RED design, short report.
Meta-protocol — fork → break → repair → vote → RED
- Proposal packet (one page): math object, intended behavioural effect, no-go failure mode, typed model diff, RED test.
- Forked solo review — every persona reviews independently first (no cross-contamination).
- Math-Breaker speaks first — REJECT by default unless derivation + units + bounds + counter-example survive.
- Systems Architect speaks second — implementable in pure Elixir, typed, property-tested, reference-anchored?
- RED Experimentalist speaks third — can the paired test falsify the actual claim?
- Embodiment Designer speaks fourth — real internal instability/need, or another preference hack?
- AIF Core Theorist merges — final evidence class + verdict: SIGN / SIGN-WITH-CHANGES / REJECT (or WITHHELD on contradiction).
- Ship gate — no merge without typed spec + validator + paired RED result + short report.
How to invoke
/lab-team-math-breaker <proposal>— Math-Breaker speaks first (the default entry)./lab-team-aif-theorist <proposal>— frame the math./lab-team-architect <proposal>— implementation review./lab-team-experimentalist <proposal>— design / verify the RED./lab-team-embodiment <proposal>— drive-design review./lab-team-review <proposal>— run the full team sequentially (the meta-protocol), then merge.
Provenance
Source: UNI Active Inference Guide GPT (https://chatgpt.com/g/g-6a1066fb6a808191a169d48c09532a0a-uni-active-inference-guide)
consult thread c/6a3b7b5c-0678-83ea-b014-e5fcb0fca67c, 2026-06-24. Replicates + uplifts the original
single GPT persona into a 5-specialist team that survives adversarial fork→break→repair.
sha256 e6946d7ceccd6032 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This page introduces a review team made of five written personas. They are not people. Each one is a document kept in the repository and also a prompt that shifts an assistant into one specialist's way of arguing.
The team exists so that a proposed change to the engine has to survive attack from several directions before anyone builds it. One persona tries to refute the mathematics and refuses by default. One keeps proposals inside standard theory and merges everyone's verdicts into a single call. One asks whether the thing can be built without breaking existing guarantees. One designs the paired experiment that could show the idea is wrong. One asks whether the change creates a genuine internal need or just dresses a preference up as one.
Three design principles run through all of them: name the mathematical object before reaching for a metaphor, state what would refute the proposal before offering any fix, and produce typed artifacts rather than prose approval.
The page ends with the order they speak in, and how to call each one.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is e6946d7ceccd6032
Clear — written for this website, not the source document
This document sets out a five-part adversarial review team. Each part is a persona: a written role description that is both an auditable document in the repository and a prompt that loads as an assistant's instructions, so that its answers shift into that specialist's domain.
The five are listed with a one-line brief each. One tries to break the mathematics and refuses by default. One holds the theoretical frame and reconciles the team into a final call. One asks whether an approved idea can be implemented additively, behind a switch, with types and property tests, without disturbing what already works. One designs the paired experiment, written down before the run, with named conditions for both success and refutation, and refuses stories told from a single run. One asks whether a proposed drive is a real internal instability or a preference dressed up as a need.
The middle of the page is the part that makes this more than role play. Three principles are required of every persona prompt. First, locate the proposal in the mathematics before any metaphor is used, so that words like curiosity or need cannot hide an undefined quantity. Second, demand the condition that would refute the proposal before suggesting any fix, because falsifiability is the cost of entry. Third, require typed artifacts as output rather than an approving paragraph, including a structured specification, validators, a paired experiment design, and a short report.
A meta-protocol then fixes the order of the whole review. First a one-page proposal packet, then an independent review by each persona with no cross-contamination. Then the personas speak in a set sequence, and the result merges into one verdict drawn from a small fixed vocabulary, with a withheld option reserved for contradiction. A final ship gate says no change merges without the specification, the validator, the paired result and the report.
The page closes with the commands that invoke each persona or the whole team, and with a provenance note naming where the design came from.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is e6946d7ceccd6032