Wiki · The Colony & the Method
Producer Language & Speaking Uplift — a genuinely-speaking UNI that manages the show, live
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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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.
Companion to docs/UNI_SIGHT_PLAN.md and docs/FALSIFICATION.md. Status: APPROVED, in progress.
Context
The producer UNI already DIRECTS the show by active inference (EFE over telemetry → camera cuts,
narration beats, colony/health) and now READS free language (rung 1–3: SP.Brain.Reader — learned
word→meaning, online learning, compose, surprise/free-energy). It SPEAKS today via:
SP.Brain.Narrator— a FEP move-selector + AUTHORED, GROUNDED, grade-4, multilingual clause templates (fact slots bound from live state ⇒ no hallucination), certified bySP.Brain.Readability.SP.Brain.Anchor— Q&A (learned reading + grounded handlers).SP.Brain.Director.narration_line/add_line+ the producer's{:line, u}beat — the live caption seam shown on/stream.
The gap. The producer's SPEECH is still largely AUTHORED templates, not LEARNED. To be a genuinely-speaking UNI that manages the show in its OWN voice, its narration + answers + announcements must be COMPOSED from a language model it LEARNS (from a corpus + its own show) while staying GROUNDED (facts from live state, never invented). No LLM (gate 18), falsifiable (grade-4 + surprise-drop + a NEW grounding gate), deterministic, pure FEP.
Honest ceiling. Pure-FEP language is n-gram/HMM/topic-grade — coherent, on-topic, grade-4, growing with the corpus — NOT LLM-fluent. Rung-4 grammar/morphology is the open frontier and may hit a wall; we measure where, honestly. The grounding constraint deliberately bounds generation so the producer never states a fact it cannot see.
Approach (phased; each independently shippable + falsifiable)
Phase A — Corpus + the learning loop
Seed corpus = language priors (show-relevant authored phrases + optional small public-domain text
under priv/corpus/). Online learning from the SHOW (its own narrations, the questions, event
descriptions) — self-supervised, no drift. Persist (runs/producer_reader.bin) so language
survives restarts. Reuse Reader.learn/learn_corpus + Anchor online-learning.
Phase B — Richer learned generation (rung 4: word order + morphology)
Extend SP.Brain.Reader: higher-order transitions (trigram w/ backoff to bigram/unigram) for word
ORDER; light morphology (deterministic inflectional stemmer: strip -s/-ed/-ing/-est/-ly) so
inflections generalise. compose/2 → beam search (length/fluency); surprise/3 stays the
falsifiable learning metric. Honest: short coherent clauses, not fluent paragraphs.
Phase C — Grounded learned speech + a GROUNDING GATE
New SP.Brain.Speaker: (1) choose STRUCTURE (Narrator move-selector / Anchor intent), (2) bind
FACT SLOTS from live state (names/counts/health — never invented), (3) REALIZE the surface from the
LEARNED model (Reader.compose) inside the slot frame. NEW §16 grounding gate
(SP.Brain.ValidationEngine): every fact-token in an utterance ⊆ the state it came from. Certify
generated speech at grade-4 via SP.Brain.Readability.
Phase D — Live: the speaking producer manages the show
Route the producer's voice through the Speaker: Director.narration_line (captions), the producer
{:line, u} beat, Anchor answers, and a new cut-announcement — all composed, learned, grounded.
EFE decisions unchanged; only the VOICE is uplifted. Apply live with NO blackout (hot-load on
uni@Thinker; no model-shape change ⇒ no colony reset). Observe on /stream.
Phase E — Observable proof
Extend mix sp.uni.prove (language): corpus size, surprise DROP over a training pass, readability ≥
grade-4 on generated speech, the grounding gate (no hallucinated facts), live samples. Update
docs/FALSIFICATION.md.
Critical files
lib/sp/brain/reader.ex · NEW lib/sp/brain/speaker.ex · lib/sp/brain/anchor.ex ·
lib/sp/brain/narrator.ex + narration.ex · lib/sp/brain/director.ex · lib/sp/producer.ex ·
lib/sp/brain/readability.ex · lib/sp/brain/validation_engine.ex (+ grounding gate) ·
lib/mix/tasks/sp.uni.prove.ex · NEW priv/corpus/ · docs/FALSIFICATION.md.
Verification
mix test (Reader/Speaker: stemmer generalises, grounded compose binds facts, no hallucination) ·
mix sp.brain.verify (grade-4 gate 16 + NEW grounding gate + gate 18 no-LLM green) ·
mix sp.uni.prove (surprise drops with corpus; readability ≥ grade-4; grounding holds). Falsifiable:
inject a fact not in state → grounding gate FAILS. Live: hot-load, no blackout, observe on /stream.
Honest risks / ceiling
Pure-FEP fluency ceiling (n-gram/HMM/topic-grade, not LLM). Grounding bounds generation (no hallucinated facts — a feature). Corpus size gates fluency; learning from own output risks echo (mitigated by the authored seed + grounding/keyword teachers). No LLM, no RL, deterministic, gates green; EFE decisions untouched — only the VOICE is uplifted.
sha256 095185a0f7e4574e — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This is an approved plan, in progress rather than finished. It describes making the narrating part of the system speak in a voice it has learned, instead of filling in authored templates.
The gap it names is honest. The narrating component already selects what to say by the same method the rest of the system uses, but the words themselves are largely written by hand. To speak in its own voice, the sentences would have to be composed from a model it learns, from a starting corpus and from the show itself, while staying tied to what it can actually see.
The ceiling is named in the same breath. Language built this way is coherent and on-topic but not fluent, and the plan says the next rung may hit a wall, which they intend to measure rather than hide.
Five phases follow, each meant to ship and be testable on its own, ending with a check that surprise falls as the corpus grows.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 095185a0f7e4574e
Clear — written for this website, not the source document
This is a plan, marked approved and in progress, for changing how the narrating component speaks.
The context section describes what already exists: a component that directs the show by the same inference method the rest of the system uses, and that already reads free language, learning word meanings online and measuring its own surprise. It speaks today through a move-selector paired with authored, grounded clause templates in several languages whose fact slots are bound from live state, plus a question-and-answer path and a caption seam.
The gap is stated plainly. The speech is still largely authored templates rather than learned. For the component to speak in its own voice, its narration, answers and announcements would have to be composed from a language model it learns, from a seed corpus and from the show itself. They would also have to stay grounded, so that facts come from live state and are never invented. The constraints are repeated: no foreign language model, falsifiable through named checks, deterministic.
The honest ceiling comes immediately after, before any of the plan. Language built this way is described as coherent, on-topic and growing with the corpus, but not fluent in the way a large model is. The plan says the next rung of grammar and morphology is an open frontier that may hit a wall, which they intend to measure rather than conceal. It adds that the grounding constraint deliberately bounds what can be generated so that a fact is never stated that cannot be seen.
Five phases follow, each said to be independently shippable and testable. The first assembles a seed corpus and the learning loop, learning from the show's own output and persisting so language survives a restart. The second extends the model to higher-order word ordering with fall-back, and adds a simple deterministic handling of word endings so inflections generalise, with a search over candidate surfaces. The honest note repeats: short coherent clauses, not fluent paragraphs. The third introduces a component that chooses structure, binds fact slots from live state, and only then realises the surface from the learned model inside that frame. It comes with a new gate requiring every fact-token in an utterance to come from the state it was generated from. The fourth routes the live voice through that component with no interruption to the broadcast, changing only the voice and leaving the decision-making untouched. The fifth extends an existing proving command to show corpus size, a drop in surprise over a training pass, a readability threshold on generated speech, and the grounding check.
The verification section names what would refute it: inject a fact that is not in the state and the grounding gate should fail. A closing section on risks repeats the fluency ceiling, notes that corpus size limits fluency, and flags that learning from its own output risks an echo, with the mitigation named.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 095185a0f7e4574e