Wiki · The Colony & the Method
The Producer's Language Faculty — grade-4, pure UNI, openly falsifiable
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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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.
The Producer narrates and answers in English, Mandarin, Hindi, Spanish, and Arabic. There is no language model (gates 14/15 forbid any foreign/neural compute layer). Language is produced from designed priors the way every UNI acts: a generative model proposes structure, expected free energy selects it, and a grammar renders the surface. This document states the claim, the method, and exactly how to break it.
The claim
The Narrator UNI (
SP.Brain.Narrator) writes a colony scene paragraph that meets a published grade-4 reading/writing contract, deterministically, with no neural model — and reads (comprehends) questions about the show at the same level (SP.Brain.Anchor).
Run it yourself:
mix sp.brain.readability # prints paragraphs + per-metric scores + PASS/FAIL
mix sp.brain.verify # gate "16 · scene meets the grade-4 contract" (with the §16 gates)
mix test test/sp/brain/{narrator,readability,anchor}_test.exs
How language is produced (not a phrasebook)
- Rhetorical move = active inference. A tiny
SP.Brain.Factorsmodel (SP.Brain.Narrator.model/0; factorsstage × last_relation, actions{cause, contrast, temporal, conclude}) infers the next move by EFE under young-writer priors: open, develop with a varied relation, build a short arc, conclude in 3–5 sentences. It is gated by §16 like every UNI (stochastic B, mean-field, deterministic). - Surface = compositional grammar. Each move is realized from authored lexicon (verb-phrases by activity, state predicates, mood clauses, colony fragments) and per-language clause templates, in present tense. Two genuine clauses joined by a connective; the cast is named, then referred to with a pronoun.
So the structure is generated and the words are composed from priors — neither looked up whole nor produced by a neural net.
The grade-4 contract (the rubric — pure arithmetic, SP.Brain.Readability)
A sample is one paragraph (a list of sentence strings). It PASSES iff:
| # | check | threshold |
|---|---|---|
| 1 | multi-clause fraction | ≥ 0.60 |
| 2 | distinct connectives | ≥ 3 of {because, so, but, then, when, while, and, …} |
| 3 | cause→effect (because/so) | ≥ 1 |
| 4 | naked pronoun before a name is introduced | = 0 |
| 6 | most-repeated opening word | ≤ 0.60 |
| 7 | sentences in the paragraph | 3–5 |
| 8 | mean words / sentence | 7–16 |
| – | structural grade index | in [3.0, 5.0] |
The grade index is a fully-disclosed custom formula (no neural scoring):
grade = 2.0 + 1.2·(mean_clauses − 1) + 0.15·(mean_words − 6)
so a one-clause ~6-word sentence ≈ grade 2; a two-clause ~11-word sentence ≈ grade 4. Tense consistency (one tense throughout) and grammaticality are enforced by construction (the realizer only emits present-tense, well-formed clause templates) and asserted in the tests.
How to falsify it (please try)
- Dispute the formula.
grade_index/0is arithmetic; argue the constants are wrong and recompute. If a different defensible readability metric puts the output below grade 4, that breaks the claim. - Find a breaking colony. Edit a fixture in
mix sp.brain.readability(or feed a live board) and find a scene that drops below the contract. - Check determinism. Same cast ⇒ identical paragraph (
mix testasserts it). A nondeterministic output breaks the "pure UNI" claim. - Check the covenant.
mix sp.brain.verifygate 14 asserts no foreign layer inSP.Brain.*. Find anNx/nif/System.cmd/Port/neural dependency in the language path and the "no LLM" claim is broken.
Honest limits
- English is self-certified. The harness measures English (the primary caption). The other four languages mirror the same generated structure with authored templates; they are present-tense and structurally parallel, but a native-speaker review would harden idiom and agreement, especially Arabic (nominal sentences, case) and Hindi (gender). Those are flagged here precisely so the gap is visible, not hidden.
- Grade 4 is pinned to this rubric, not a human judgment. The rubric is the falsifiable proxy; if you think it is too lax or too strict, the formula and thresholds are right here to argue with.
- It is a faculty, not fluency. It writes about the show (the colony, the producer's own state) at grade-4 structure. It cannot discuss arbitrary topics — by covenant, not by accident.
sha256 d920c0f375c34d49 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This page describes how the narrating part of the system writes and reads sentences without any language model, and it tells you how to break the claim.
The method is the same one the rest of the project uses. A small generative model proposes the next rhetorical move, a selection rule picks one, and a grammar renders the surface from an authored vocabulary and clause templates. So the structure is generated and the words are composed, rather than being looked up whole or produced by a neural network.
The standard it is held to is a published rubric of arithmetic checks, with the scoring formula printed openly so you can argue with the constants.
The honest limits are the best part. Only the primary language is scored by the harness; the others mirror the same structure and would need a native speaker to harden. The grade is pinned to this rubric rather than a human judgement. And it is a faculty rather than fluency: it writes about the show and cannot discuss arbitrary topics.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is d920c0f375c34d49
Clear — written for this website, not the source document
This document describes a language faculty and, in the same breath, hands you the tools to refute it.
The claim is stated in a box: a narrating component writes a paragraph about the colony that meets a published reading and writing contract, deterministically, with no neural model, and reads questions about the show at the same level. Commands are given so you can run the checks yourself.
The method section insists this is not a phrasebook. The choice of rhetorical move is made by a small generative model whose factors and actions are named. It selects the next move under priors for a young writer: open, develop with a varied relation, build a short arc, and conclude within a few sentences. The surface is then realised from authored vocabulary and per-language clause templates, in the present tense, joining two genuine clauses with a connective and naming the cast before using a pronoun. The summary the page draws is careful: the structure is generated and the words are composed from priors, neither looked up whole nor produced by a neural network.
The contract itself is a table of arithmetic checks with thresholds. How many sentences carry more than one clause, how many distinct connectives appear, and whether at least one cause-and-effect link is present. Whether a pronoun ever appears before a name, how repetitive the sentence openings are, how many sentences the paragraph has, and the mean words per sentence. A structural grade index must land inside a band, and the formula for that index is printed in full with its constants exposed. Tense consistency and grammaticality are described as holding by construction, because the realiser only emits well-formed present-tense templates.
The falsification section invites four kinds of attack. Dispute the formula and recompute with a different defensible metric. Find a colony state whose scene drops below the contract. Check that the same input gives the same paragraph, since a varying output would break the purity claim. And inspect the language path for any foreign or neural dependency, which would break the no-model claim.
The limits are stated rather than implied. Only the primary language is scored by the harness. The other four mirror the same generated structure with authored templates and are structurally parallel. But the page says a native-speaker review would harden idiom and agreement, and names the two languages where it expects the most trouble, adding that this is flagged precisely so the gap is visible. The grade is pinned to this rubric rather than to human judgement, and the rubric is offered as something to argue with. And the last limit is the plainest: this is a faculty rather than fluency. It writes about the show at one structural level and cannot discuss arbitrary topics, and the page says that is by covenant rather than by accident.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is d920c0f375c34d49