Universal Natural Intelligence

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.

This is a scoping artifact, not narration. No final prose is written here. Each beat carries an intent — one line saying what that beat has to accomplish — plus the visual and the receipt. Prose comes next, per beat, and is synthesised locally by piper per lab/film/welcome/TOOLCHAIN.md.

Scoped LONG first, deliberately. The operator's production ladder: plan the ~36-minute documentary, then derive the 12-minute main cut and the 3-minute short from it. Scoping long is the cheapest act available; any smaller frame authored first silently drops a wing, and the wing it drops is always the same one — the honest state, because it is the least comfortable to keep.

Two corrections to the brief, made before the work rather than after it

1. M0..M9 is ten movements, not nine. The brief says "the nine movements (M0..M9)" and then names ten things: invitation, mission, estate, how to watch, how it works, the truth machine, the honest state, the cookbook, how to engage, the close. Ten names, ten identifiers, one wrong count. This outline uses ten movements, M0 through M9, with those ten names in that order. The discrepancy is recorded here rather than silently resolved, because a count corrected in silence is how CLAUDE.md's own banner came to say 25 gates in one paragraph and 23 in another.

2. lab/film/welcome/SPINE.json does not exist. Checked 2026-08-01. This document is therefore the spine's source, not its rendering. When SPINE.json is authored it must be generated from this file, not typed beside it — the estate has already paid for the other arrangement (viewer/generate_state_blocks.cjs, and the six-place drift it exists to stop).

The house rule this film inherits

lab/film/script/reels_cues.json — reel3_proven:

"We took three big, beautiful, hopeful ideas about the universe, and we tested every one of them, fairly, all the way down. And across the entire film, we never once used the word proven."

That boast is a constraint, and this film matches it exactly. The word is not a stylistic tic to avoid — it is a banned token in this estate's machine-checked vocabulary: production/schemas/claim_fence.json fences prov(e[sd]?|en|ing) and proof, case-insensitive, word-boundary-anchored, alongside the consciousness, life, experience and over-claim families.

Three consequences, all binding on the narration:

The narration never speaks that token, in any inflection. Not in a claim, not in a denial, not in a joke. lab/film/QC.md recorded the same check for TRAVELERS — "Word 'proven' anywhere in 67 min of narration? ✅ never" — and it was a hand-typed check on a hand-typed sheet. Here it is a gate: the narration corpus is run through the fence regex built from claim_fence.json before any audio is synthesised, and a hit fails the build.
The fenced families are fenced too. No "conscious", "sentient", "self-aware", "alive", "living", "experience", "feels", "suffers", "world-first", "breakthrough", "AGI", "human-level". Where a beat below needs to gesture at one of these, it names what a gate demonstrated instead.
One place the token legitimately appears on screen, and never in the voice. The cookbook's UNI fence uses proven as a ledger value (UNI.Architect/UNI-Encyclopedia-Cookbook/encyclopedia/CLAIM-LEDGER.md), while this repository's fence bans the same token. That divergence is real, it is a finding, and it is beat W-M7-03. When that ledger is on screen the word is legible on the card; the narrator does not say it.

This outline was run through the fence, and it failed. Six of its own lines carried fenced tokens on the first pass — living in W-M2-05, felt in W-M4-04, proof in W-M6-03, W-M6-08 and W-M8-01, and re-proved in the production constraints. All six are repaired above. Recording the failure rather than quietly fixing it is the point: a scoping document that declares a fence gate and then does not survive it is exactly the class of self-inconsistency this estate keeps paying for. Two exemptions are declared for the gate rather than discovered later: receipt paths are exempt (a filename such as L12-creativity-awareness.md is a citation, not a claim), and on-screen quotation is exempt while the voice is not (W-M6-03 and W-M7-03 both show a fenced token and paraphrase it aloud).

TRAVELERS is not reused. Different film, different subject — ozone, pressure, bioenergetics. Nothing from lab/film/script/MASTER_SCRIPT_EN.md or the five segment cue files is carried across. What is carried across is the discipline, and one self-implicating fact about TRAVELERS itself (beat W-M6-12).

Running time — how the estimate was derived, and why it is an estimate

input	value	source
TRAVELERS EN narration	~10,165 words	`lab/film/QC.md:18`
TRAVELERS spoken duration	~67 min	`lab/film/QC.md:18`
implied rate	~152 words/min	derived from the two above
Piper measured, this machine	4 words → 1.683 s → ~143 wpm	`lab/film/welcome/TOOLCHAIN.md:44-46`
planning rate used here	150 wpm	midpoint, stated so it can be checked

At 150 wpm, 36:00 ≈ 5,400 narration words ≈ 74 words per beat across 73 beats.

This is an estimate and is labelled as one. The real duration is the sum of the measured WAV durations, which only exist after synth_narration.cjs runs and writes audio_manifest.json (voice model, sha256, exact text, output duration per beat — TOOLCHAIN.md:52-56). The QC sheet for this film is generated from the gate's output, precisely because lab/film/QC.md states 143 scenes in its probe table while its own segment table sums to 157 (TOOLCHAIN.md:70-74). A hand-typed running time is a claim with a half-life. This one is marked PREDICTED until the manifest exists.

Movement map

id	movement	beats	in	out	duration	share
M0	INVITATION	5	00:00	02:30	2:30	6.9%
M1	THE MISSION (intent, not measurement)	6	02:30	05:30	3:00	8.3%
M2	THE ESTATE	8	05:30	09:30	4:00	11.1%
M3	HOW TO WATCH	5	09:30	12:00	2:30	6.9%
M4	HOW IT WORKS	8	12:00	16:00	4:00	11.1%
M5	THE TRUTH MACHINE	8	16:00	19:45	3:45	10.4%
M6	THE HONEST STATE	12	19:45	25:45	6:00	16.7%
M7	THE COOKBOOK	7	25:45	29:15	3:30	9.7%
M8	HOW TO ENGAGE	9	29:15	33:30	4:15	11.8%
M9	THE CLOSE	5	33:30	36:00	2:30	6.9%
	TOTAL	73			36:00	100%

M6 opens at 19:45 of 36:00 — 54.9% of the running time. Not at the end, and this is not a preference. CLAUDE.md's communication contract is explicit: "A FAIL, a retraction, a falsified prediction, a NOT_CLEARED — these are spoken and said first, never appended at the end where they read as a footnote." Ten and a quarter minutes of film run after the honest state — the cookbook, the engagement, the close. Everything that follows has to survive it. An adverse result the audience can still walk out on is a caveat; an adverse result the rest of the film must live with is a finding.

Threads that cross movements

Three ideas are too big for one movement and are deliberately braided:

FALSIFIER THREAD — what a falsifier is (M5), what it looks like when the estate is caught (M6), the drill you personally run (M8). Beats: W-M5-07, W-M5-08, W-M6-05, W-M6-07, W-M8-04, W-M8-05, W-M8-06.
TWIN-TERM THREAD — every technical idea arrives twice, plain then precise, in the same breath. Never one or the other. Register below.
STOPS THREAD — the things an agent, a visitor, or a contributor must not do, and why each prohibition protects evidence rather than protecting the estate. Beats: W-M6-03, W-M8-08, W-M8-09.

Twin-term register (binding on narration)

Every row must appear in the film with both halves audible, in the same sentence, in this order: plain first, precise second. The plain term is never dropped for brevity; the precise term is never dropped for accessibility.

plain (layman)	precise (science)	first appears
a good guess about what is out there	a posterior belief over hidden states	W-M4-01
being surprised	surprisal, `−ln p(o)`	W-M4-02
the running cost of being wrong	variational free energy, `F`	W-M4-02
a hunch you brought with you	a prior	W-M4-03
how much a body trusts its own senses	precision, `γ`	W-M4-04
a plan judged before you act	expected free energy over policies	W-M4-05
curiosity, made countable	the epistemic term, `H(qo) − E[H(o	s)]`
the same start gives the same run	seeded determinism, `SP.Determinism`	W-M4-06
a promise written down before the run	a pre-registered gate	W-M5-02
an honest scoreboard	the four verdicts: PASS / PARTIAL / FAIL / WITHHELD	W-M5-03
a list of words we will not say	the claim fence	W-M5-04
a receipt that can be re-derived	a hash-anchored ledger entry	W-M5-05
a check that can actually bite	a gate with a mutation test	W-M5-06
breaking our claim on purpose	falsification	W-M5-07
we saw it	`OBSERVED` — a source-pinned recorded measurement	W-M3-02
we rebuilt it	reconstruction / simulation / derived field	W-M3-02
a different bug entirely	species boundary (E. coli vs Salmonella vs Bacillus)	W-M3-04
the tail on a bacterium	the flagellar motor	W-M2-07
the one computer everything lives on	the canonical host — "the chip"	W-M2-03
the thing that says no to going live	the go-live guard, `presence_evident`	W-M6-08
nature already ran the experiment	convergent evolution as constraint-optimum evidence	W-M7-05
not everything is there for a reason	Gould & Lewontin 1979 — drift, inertia, spandrels	W-M7-06

Hard production constraints (from the toolchain, measured)

No live IPs, hostnames or keys on screen, ever. The estate ran a remediation that converted 29 live IP literals to names and re-ran the check to confirm it still bites (viewer/ip_fence.cjs, commit 7fdcfc0). A film that puts 10.190.245.x on a title card undoes that in one frame. Boxes are named — the chip, the studio, the relay — never addressed.
No GPU, no WebGL, no Three.js. Chrome is invoked --headless --disable-gpu and only ever screenshots a static SVG (TOOLCHAIN.md:60-62). A film about a laboratory that refuses to ship those things must not be rendered by them.
No container in the build. Piper is native on PATH. This is the specific defect that makes TRAVELERS unbuildable today, and the Welcome film must not inherit it (TOOLCHAIN.md:8-28).
No music bed without Michael's explicit rights attestation, and if one is added, dgst/RIGHTS.md must name the grant (TOOLCHAIN.md:63-66).
Nothing is fetched at build time. Every asset is on this machine or is generated by the build.
The studio stays cold. No OBS, no MediaMTX, no presence token, no go-live path is touched by this pipeline (TOOLCHAIN.md:67-68).

M0 · INVITATION — 00:00 → 02:30 · 5 beats

Job: get a stranger through the door without asking for a single unit of trust, and tell them the worst news is coming at minute twenty so it does not feel like an ambush.

[W-M0-01 | FRAME]

Intent — Open on the door itself. This is a laboratory that publishes what it has measured and publishes what it has failed at, and you are about to be handed the tools to attack both.
Visual — door.html rendered as a still, then the title card resolving out of it. Palette from the repo, no photography.
Receipt — viewer/door.html; viewer/hub.html

[W-M0-02 | FRAME]

Intent — Name who this is for: someone who has never seen the estate, has no account, and owes us nothing. Nothing here needs permission to read.
Visual — The public site's front page as a still, the four top-level sections legible.
Receipt — UNI.Public/README.md; docs/PUBLIC_README.md:5-7

[W-M0-03 | CLAIM]

Intent — State the promise plainly: "you can read every claim we have and every receipt behind it without asking us for anything" — and say immediately that this promise is currently incomplete, because the source repositories are private (M8 carries the detail).
Visual — The promise as a quote card, with a small amber marker on the word "every" pointing forward to beat W-M8-03.
Receipt — docs/PUBLIC_README.md:7; UNI.Public/generators/sources.json (visibility: private)

[W-M0-04 | FRAME]

Intent — Say what this film will not do. It will not tell you the estate has succeeded; it will not use the fenced words; it will not put a number on screen that a person typed rather than a machine generated.
Visual — The fence classes from claim_fence.json as a legible list, animating in — the vocabulary that is off the table, on screen, before any argument is made.
Receipt — production/schemas/claim_fence.json

[W-M0-05 | FRAME]

Intent — Lay the shape of the next thirty-six minutes and flag the turn: at minute twenty this film stops describing the laboratory and starts listing what is wrong with it. That is not the ending; ten minutes run after it.
Visual — The movement map as a timeline strip, M6 highlighted at the 55% mark, with the remaining third of the bar visibly still to come.
Receipt — This document, § "Movement map"; CLAUDE.md § "Never bury an adverse result"

M1 · THE MISSION — 02:30 → 05:30 · 6 beats

Job: state the mission in full, once, and fence it as intent. The fence is the point of the movement, not a footnote to it.

[W-M1-01 | INTENT]

Intent — Read the mission in full, unedited, as one sentence: "worldwide peace through free and open understanding; the public pursuit and ownership of General Universal Natural Intelligence; the end of extract and extortion; making safe all wombs and Gaia — earth, our solar system and beyond."
Visual — Full-frame type, the whole statement at once, no animation, held long enough to read twice.
Receipt — The operator's declaration, 2026-08-01. Not currently held in any repository file. Authoring it to lab/film/welcome/MISSION.md and citing that path is a prerequisite for this beat — see § "Open questions for the operator", Q1.

[W-M1-02 | INTENT]

Intent — Immediately mark the whole statement: this is intent, and intent is a different class of thing from measurement. Nothing in it has been measured. Saying so is the first honest act of the film.
Visual — An INTENT — NOT MEASURED badge lands on the mission card in the estate's own badge style, matching the truth-class badges used everywhere else.
Receipt — CLAUDE.md § "Truth contract"; docs/LAB_PROTOCOL.md § VI (the claim fence)

[W-M1-03 | INTENT]

Intent — Clause one — worldwide peace through free and open understanding. What would even count as evidence here? Say honestly: nothing in this estate measures it, and the only thing downstream of it that is checkable is whether the work is genuinely readable by a stranger.
Visual — The clause isolated; beneath it, the one checkable descendant — the public documentation coverage figure — with its generator named.
Receipt — UNI.Public/safety/verify_coverage.cjs; UNI.Public/content/coverage-manifest.json

[W-M1-04 | INTENT]

Intent — Clause two — the public pursuit and ownership of General Universal Natural Intelligence. Say what "public" is cashed out as here: pre-registered gates, published negatives, a seed that reproduces a verdict. Say what "General" is not cashed out as: no claim of that kind exists in the ledger.
Visual — Split card — left, the four things "public" means operationally; right, an empty column headed by the claim that has never been made.
Receipt — docs/PUBLIC_README.md:11-17; docs/gates/PUBLIC_GATE_LOG.md

[W-M1-05 | INTENT]

Intent — Clause three — the end of extract and extortion. The only operational shadow of this clause in the codebase is a product contract: CPU-only, no LLM inference, no analytics, no accounts, no hidden network calls. That is a small thing measured against a large thing said.
Visual — The runtime prohibitions as a list, each with the file that enforces it; then the clause above them, visibly larger than its enforcement.
Receipt — UNI-FLAGELLUM/CLAUDE.md § Mission (product contract); mix.exs (deps → []); docs/FALSIFICATION.md § 1

[W-M1-06 | INTENT]

Intent — Clause four — making safe all wombs and Gaia — earth, our solar system and beyond. Name the one thing in the estate that carries the name: Gaia is a projection surface with a lint gate, not a planetary claim. The gap between the name and the thing is stated, not smoothed.
Visual — The Gaia surface rendered, its gate name beside it, then the clause above it at scale.
Receipt — viewer/gaia/gaia_server.cjs; viewer/gaia/gaia_lint.cjs; viewer/gate_registry.json (gaia, gaia-lint)

M2 · THE ESTATE — 05:30 → 09:30 · 8 beats

Job: a map. Four repositories, three machines, one colony, one microscope, one public surface. No addresses on screen.

[W-M2-01 | FRAME]

Intent — There are four repositories and they do four different jobs. Show them as a map before any of them is explained.
Visual — Four-node diagram: UNI.Minecraft, UNI-FLAGELLUM, UNI-Encyclopedia-Cookbook, UNI.Public. Each with its one-line job.
Receipt — UNI.Public/generators/sources.json

[W-M2-02 | CLAIM]

Intent — Three machines with three non-overlapping roles, and the plain statement that every failure this project has had came from conflating them.
Visual — Three boxes, named not addressed: the chip, the studio, the relay. Arrows one-way: chip → studio → relay.
Receipt — docs/UNIVERSE.md § 1.1

[W-M2-03 | CLAIM]

Intent — The chip — the one computer everything canonically lives on. The colony runs there, rootless, with no graphics hardware at all.
Visual — The chip alone, its two surfaces labelled, the broadcast surfaces greyed out with "never, here".
Receipt — docs/UNIVERSE.md § 1.1, § 1.2; production/docs/adr/ADR-PROD-013-colony-host-placement.md

[W-M2-04 | CLAIM]

Intent — The studio and the relay, and the three-day failure that taught the split: overlays software-rendered to a black frame on a graphics-less machine. Say the failure, then the rule it produced.
Visual — A black frame, held uncomfortably long, then the corrected topology.
Receipt — docs/UNIVERSE.md § 1.1 (node2 row); production/docs/adr/ADR-PROD-011-native-windows-obs-on-render-host.md:11-18

[W-M2-05 | CLAIM]

Intent — The colony: bodies inhabiting a real world, each running one predict-act tick from start to finish. Plain first — small agents that guess, act, and update; then precise.
Visual — The world view still, then the tick decomposed into its stages.
Receipt — docs/UNIVERSE.md § 0; lib/sp/brain/; docs/SYSTEM_OVERVIEW.md

[W-M2-06 | CLAIM]

Intent — The producer: the thing that watches the colony and puts it into words and pictures — and the hard rule that it may never say a name or a number it cannot see in the live state.
Visual — A spoken line, then the state it was grounded against, then the injected fake name being refused.
Receipt — docs/FALSIFICATION.md § 6b; SP.Brain.Speaker.grounded?/2; mix sp.brain.verify gate 19

[W-M2-07 | CLAIM]

Intent — The other laboratory: a bacterial flagellar motor — the tail on a single-celled organism, the smallest rotary engine anyone has measured — with licensed microscopy beside a deterministic reconstruction. Two different classes of thing, kept apart on purpose.
Visual — Motor schematic; the observation and the reconstruction side by side with different badges.
Receipt — UNI.Public/generators/sources.json (uni-flagellum); UNI-FLAGELLUM/CLAUDE.md § "Truth contract"

[W-M2-08 | CLAIM]

Intent — The public surface: roughly 290 documents from these repositories, rendered as they are written, plus thirteen authored articles. Say the one rule the whole site is built on — every volatile number is generated from the repository it describes and carries the commit it was read from.
Visual — The site's own six-row "what is here" table, live.
Receipt — UNI.Public/README.md; UNI.Public/content/articles/

M3 · HOW TO WATCH — 09:30 → 12:00 · 5 beats

Job: hand the viewer the reading instrument before they need it. Every badge, verdict and boundary they will see for the next twenty-four minutes.

[W-M3-01 | INSTRUCTION]

Intent — Everything on screen from here carries a label saying what kind of thing it is. Teach the label before the content.
Visual — The badge set, isolated, each one held.
Receipt — viewer/honesty_card.json; docs/gates/PUBLIC_GATE_LOG.md § "Evidence classes"

[W-M3-02 | INSTRUCTION]

Intent — The first and hardest boundary: we saw it versus we rebuilt it. Only a source-pinned recorded measurement is OBSERVED. A reconstruction, a simulation, a derived field or a model output may never be relabelled as one — that relabelling has a name in this estate, truth laundering, and there is a mutation test that detects it.
Visual — Two images side by side, one observed, one reconstructed, then the label swap attempted and caught.
Receipt — UNI-FLAGELLUM/CLAUDE.md § "Truth contract"; viewer/lab/verify_shot.cjs --mutate

[W-M3-03 | INSTRUCTION]

Intent — The four verdicts, and the rule that there is no fifth: PASS, PARTIAL, FAIL, WITHHELD. Never percent-scored. A PARTIAL names exactly which sub-claim holds and which does not.
Visual — The four words; then a real PARTIAL row expanded to show its two halves.
Receipt — docs/LAB_PROTOCOL.md § II; docs/gates/PUBLIC_GATE_LOG.md

[W-M3-04 | INSTRUCTION]

Intent — The species boundary, using the flagellar lab: behavioural evidence from one organism is never merged with structural evidence from another, and neither may be implied to have come from one measured specimen.
Visual — Three organism cards, physically separated, with the merge animation refused.
Receipt — UNI-FLAGELLUM/CLAUDE.md § "Truth contract"

[W-M3-05 | INSTRUCTION]

Intent — The last reading rule, and the one that makes the rest usable: a failed, blocked, external or not-run gate stays visible. If you cannot find the failures on a surface, the surface is not finished.
Visual — A gate list with its ci:false rows deliberately left in — listed, never run, never a fabricated pass.
Receipt — viewer/gate_registry.json (colony, hud, overlays); CLAUDE.md § generated block uni.state.gates

M4 · HOW IT WORKS — 12:00 → 16:00 · 8 beats

Job: the science, twice — plain and precise, in the same breath, every time.

[W-M4-01 | CLAIM]

Intent — Start with the guess. A body cannot see the world; it can only see its own sensors. So it keeps a good guess about what is out there — a posterior belief over hidden states — and acts on the guess.
Visual — A body, a wall between it and the world, and the guess drawn on its side of the wall.
Receipt — docs/UNIVERSE.md § 2 (FEP in one page); docs/specs/generative_model.md

[W-M4-02 | CLAIM]

Intent — Being surprised has a number. Surprisal is minus the log of how likely what you saw was. Free energy is the running cost of being wrong — an upper bound on that surprise, and the one quantity everything here minimises.
Visual — The two quantities written out, units carried, with the bound drawn as a bound.
Receipt — docs/UNIVERSE.md § 2; docs/FALSIFICATION.md § 2 (gate 3: F ≥ −ln p(o))

[W-M4-03 | CLAIM]

Intent — A prior is a hunch you brought with you. It is not a flaw — it is how anything survives a world it cannot see all of. It is also how a thing becomes certain too early.
Visual — A flat prior and a peaked prior, same evidence, different conclusions.
Receipt — docs/specs/generative_model.md; docs/UNIVERSE.md § 2

[W-M4-04 | CLAIM]

Intent — Precision is how much a body trusts its own senses right now. It is a real parameter with a name, γ, and it is deliberately never dressed up as an inner state — the fence forbids it.
Visual — The same signal at two precisions; then the fence card refusing the sentimental reading.
Receipt — docs/LAB_PROTOCOL.md § V.2, § VI; production/schemas/claim_fence.json

[W-M4-05 | CLAIM]

Intent — How a plan gets chosen: every candidate plan is scored before it is acted on, and one part of that score is curiosity made countable — how much a plan would reduce what the body does not know.
Visual — Three policies scored side by side, the pragmatic and epistemic terms separated and labelled.
Receipt — docs/LAB_PROTOCOL.md § V.2; lib/sp/brain/

[W-M4-06 | CLAIM]

Intent — The same start gives the same run, byte for byte. Not approximately — identically. There is no random number generator anywhere outside the seeded one.
Visual — Two runs at one seed, byte-identical; then the seed changed by one, diverging.
Receipt — docs/FALSIFICATION.md § 3; SP.Determinism; mix run --no-start -r runs/real_evidence.exs

[W-M4-07 | CLAIM]

Intent — What is not inside: no language model, no foreign mind, no network call. Zero external libraries at all — so there is no library a language model could be hiding in.
Visual — sed -n '/defp deps/,/]/p' mix.exs on screen returning an empty list; then the detector list that fails the build if any of them appears.
Receipt — docs/FALSIFICATION.md § 1; mix.exs; mix sp.brain.verify gates 14/17/18

[W-M4-08 | CLAIM]

Intent — The fence between what a gate demonstrates and what it does not. Passing a behavioural gate demonstrates the named behaviour, and carries zero evidential weight for anything beyond it. This is stated by the estate about its own work, unprompted.
Visual — A passing gate, then the wall, then the empty space on the other side of the wall.
Receipt — docs/LAB_PROTOCOL.md § VI; docs/PUBLIC_README.md § "What we are careful not to say"

M5 · THE TRUTH MACHINE — 16:00 → 19:45 · 8 beats

Job: how a claim is made here, and — the beat this film exists for — how a stranger takes one apart.

[W-M5-01 | CLAIM]

Intent — The first law, stated as the estate states it: never stack changes such that you cannot account for the winning outcome. If two things changed, the result is unattributable and may not be claimed at all.
Visual — A paired design: two arms, one variable different, everything else identical.
Receipt — docs/LAB_PROTOCOL.md § I

[W-M5-02 | CLAIM]

Intent — A gate is a promise written down before the run: PASS requires all of these; FALSIFIES if this. Registered before, judged only against itself.
Visual — A real pre-registration document, its PASS and FALSIFIES clauses highlighted, dated before its run.
Receipt — docs/LAB_PROTOCOL.md § II; docs/receipts/red_preregistration_forage_pureworld_graduation.md

[W-M5-03 | CLAIM]

Intent — The ledger. Every gate the project claims has to be represented as a row, and the row names a real file whose bytes carry the evidence. State the measured size of that ledger now.
Visual — evidence/gates.ndjson scrolling, with the generated tally card beside it.
Receipt — evidence/gates.ndjson; CLAUDE.md generated block uni.state.gate_ledger — 207 rows / 110 unique names; last row per name: 93 PASS · 4 PARTIAL · 12 PENDING · 1 FAIL

[W-M5-04 | CLAIM]

Intent — The claim fence again, now as machinery rather than promise: a versioned JSON file, case-insensitive, word-boundary-anchored, with an agreement guard holding the JavaScript and Elixir copies to the same behaviour. A flagged line is dropped honestly, never reworded into a subtly-different claim.
Visual — The fence file, then the two implementations, then the snapshot test that holds them equal.
Receipt — production/schemas/claim_fence.json; test/sp/brain/fence_snapshot_test.exs

[W-M5-05 | CLAIM]

Intent — The anchor: a hash chain over the control-plane ledger, so a changed history is visible. Say precisely what it gives you — tamper-evidence — and set up M6, where the film says what it does not give you.
Visual — The chain, the tip hash, the anchor's declared length agreeing with it.
Receipt — CLAUDE.md generated block uni.state.control_plane — 32 entries, tip b90b7498… at seq 32, anchor agrees; evidence/control_plane/

[W-M5-06 | CLAIM]

Intent — A check that cannot fail is not a check. So the checks are attacked: a mutation is introduced on purpose, and the gate has to catch it. Show one doing exactly that in greyscale, with no text read.
Visual — Two rendered fixtures, the mutation applied, the gate turning red.
Receipt — viewer/lab/verify_shot.cjs --mutate; viewer/lab/l6.html:52-53; viewer/gate_registry.json (lab-l0…lab-l6)

[W-M5-07 | INSTRUCTION]

Intent — What a falsifier is. Not a critic, not a sceptic — someone who takes a specific written claim, finds the exact condition under which it would be false, and goes and checks that condition. A claim nobody can imagine falsifying is not science.
Visual — One claim on screen; beneath it, the estate's own written falsifier for it, in its own words.
Receipt — docs/FALSIFICATION.md (opening); docs/FALSIFICATION.md §§ 1–6b

[W-M5-08 | INSTRUCTION]

Intent — Being caught is the product working. Say it without hedging: if you break a claim here, the estate owes you a correction, and the correction is published where the claim was. That is not generosity — it is the only arrangement under which any of the rest of this is worth reading.
Visual — A real retraction: a claim, its withdrawal, and the receipt that carries both.
Receipt — docs/PUBLIC_README.md:19 ("that is a legitimate falsification and we owe you a correction"); docs/receipts/phase1_curiosity_red_CORRECTION.md

M6 · THE HONEST STATE — 19:45 → 25:45 · 12 beats · 54.9% in, never last

Job: say what is wrong, in order of severity, with the receipt for each — including the several occasions on which this estate's own instruments caught this estate lying to itself. Ten and a quarter minutes of film run after this movement, and every one of them has to survive it.

[W-M6-01 | FRAME]

Intent — Mark the turn out loud. Everything up to here described how the laboratory is supposed to work. What follows is what is actually true today, measured, and it is not flattering.
Visual — Hard cut. Palette shift. The word HONEST STATE full frame, the clock visible at 19:45 of 36:00.
Receipt — CLAUDE.md § "Never bury an adverse result"

[W-M6-02 | ADVERSE]

Intent — The off-box witness is compromised. The remote machine that was supposed to independently hold a copy of the ledger accepts the writer's own key. The count of independent custodians is zero. So the anchor stands on git alone: tamper-evident, and not unforgeable.
Visual — The witness diagram with the independence arrow drawn and then struck through; the literal independent_custodians: 0.
Receipt — viewer/gaia/collectors.cjs:806-825; viewer/bag.cjs:17,81; CLAUDE.md § "Four things that must not be softened", item 1

[W-M6-03 | ADVERSE / STOP]

Intent — And the obvious fix is forbidden. Removing that key is the one repair an agent must not perform — using write access to erase the evidence of write access is not restoring a witness, it is destroying the last evidence that it happened. The defect stays visible until a human fixes it properly.
Visual — The stop card, S1, quoted verbatim on screen. (The plan's own wording uses a fenced token; it is legible on the card and paraphrased in the voice — same rule as W-M7-03.)
Receipt — evidence/remediation/phase9_plan.json → stops[id=S1]

[W-M6-04 | ADVERSE]

Intent — The gate registry and the canonical ledger do not agree. Of 33 registered gates, 1 appears in evidence/gates.ndjson and 32 do not — while the row schema says every gate the project claims must be represented there.
Visual — Two columns, 33 and 1, the 32-wide gap drawn as a gap.
Receipt — CLAUDE.md generated block uni.state.registry_ledger_gap; viewer/generate_state_blocks.cjs; viewer/gate_registry.json; evidence/gates.ndjson

[W-M6-05 | ADVERSE / FALSIFIER]

Intent — Worse, and better: four governing documents said that intersection was empty for two weeks after a row had already landed — and they said it inside the very paragraph claiming these numbers were generated. They were hand-written. The instrument that caught it is the same instrument that now writes the number.
Visual — The false sentence, then the row that falsified it, then the generator replacing the sentence.
Receipt — CLAUDE.md generated block uni.state.registry_ledger_gap; viewer/generate_state_blocks.cjs

[W-M6-06 | ADVERSE]

Intent — The governing document was wrong in seven places, corrected, then wrong in six more within six hours — including a next-instruction that told a fresh reader to rebuild something that had shipped six hours earlier. One of those numbers was false 176 seconds after it was written.
Visual — The half-life, on a clock: 176 seconds, ticking.
Receipt — CLAUDE.md §§ "Corrected 2026-07-28", "Corrected AGAIN 2026-07-29"; generated block uni.state.how_to_measure

[W-M6-07 | ADVERSE / FALSIFIER]

Intent — The third correction was of the second correction's own claim — and the reason is the sharpest thing in this film. The gate whose entire purpose is to catch a stale instruction could not see one of the three copies of the document it guards, because that copy was defined as a root and then declared as no document. The audit's own count was therefore an undercount: it reported five, and the true number was six.
Visual — Three copies of the same banner; the instrument's beam illuminating two of them; the third, unlit, carrying the stale line.
Receipt — CLAUDE.md § "Corrected A THIRD TIME 2026-07-30"; viewer/state_blocks.cjs (OUT_OF_TREE); viewer/verify_claims.cjs

[W-M6-08 | ADVERSE]

Intent — The guard that refuses to put this estate on air is real, it covers every path, and it is honest about its own ceiling: its claim level is presence_evident, not unforgeable. A process cannot authenticate a person; it can only refuse whenever the evidence of one is absent, and name which condition failed.
Visual — The six code paths that reached the live button, four of them with no check at all; then the guard refusing each, with the refusal text legible.
Receipt — viewer/golive_guard.cjs:1-30; viewer/gate_registry.json (golive-refuses-agents)

[W-M6-09 | ADVERSE / STOP]

Intent — And on the studio machine, the control socket that drives the broadcast has no authentication at all. That one is the operator's to close, not an agent's, and it is named in the stops so it cannot be quietly done.
Visual — The port, unauthenticated, on the studio box; the S2 stop card beside it.
Receipt — CLAUDE.md § "Four things that must not be softened", item 4; evidence/remediation/phase9_plan.json → stops[id=S2]

[W-M6-10 | ADVERSE]

Intent — The science is not yet run. The gate that would graduate the colony is blocked on a pre-registered test whose runner does not exist — the file raises a scaffold error instead of executing. State it plainly: no verdict has yet been authored about a real scientific claim here.
Visual — The runner file open at the raise, the pre-registration beside it, the contract visible in the exception text.
Receipt — runs/pureworld_qa_gate.exs:38-44; docs/receipts/red_preregistration_forage_pureworld_graduation.md; evidence/remediation/phase9_plan.json → road_to_air.gates[colony_on_program]

[W-M6-11 | ADVERSE]

Intent — The broadcast track record, entire: one go-live, six minutes against a four-hour slot. The emergency stop has never been fired by a human in the platform's life — and a kill switch that has never been pulled is a decoration.
Visual — A four-hour bar with six minutes filled; then the unpressed button.
Receipt — evidence/remediation/phase9_plan.json → road_to_air.track_record, road_to_air.free_moves_today[A2]

[W-M6-12 | ADVERSE / SELF]

Intent — And this film's own predecessor cannot be rebuilt from its own source today. TRAVELERS synthesised its narration by copying audio out of a container that no longer exists — 143 hardcoded identifiers pointing at nothing — and its hand-typed quality sheet contradicts itself about how many scenes it has. That is why this film's build has no container in it and this film's quality sheet is generated.
Visual — The docker cp line; the 143 identifiers; the two contradicting scene counts, 143 and 157, on the same sheet.
Receipt — lab/film/welcome/TOOLCHAIN.md:8-28, :70-74; lab/film/QC.md:17, :30

M7 · THE COOKBOOK — 25:45 → 29:15 · 7 beats

Job: introduce the cookbook as a thing a stranger takes home and uses tonight, and be exact about the one place its vocabulary and this repository's vocabulary disagree.

[W-M7-01 | FRAME]

Intent — There is a second body of work here that is not a codebase at all: two books and a build, written so someone outside this estate can cook the same dishes.
Visual — The repository as a shelf: encyclopedia, cookbook, the pack, the build.
Receipt — UNI.Architect/UNI-Encyclopedia-Cookbook/README.md

[W-M7-02 | CLAIM]

Intent — Two sovereign ledgers, and the rule that organises everything: one vocabulary describes what this programme has built; the other describes what nature has been observed to do. Merging them is the cardinal error. Reading a published law about biology raises nothing here.
Visual — The two ledgers side by side, their value sets legible, the merge arrow refused.
Receipt — UNI-Encyclopedia-Cookbook/encyclopedia/CLAIM-LEDGER.md; UNI-Encyclopedia-Cookbook/encyclopedia/NATURE-LEDGER.md

[W-M7-03 | ADVERSE / VOCABULARY]

Intent — And here the estate disagrees with itself, on camera. One repository's fence bans a certain word outright; the other repository uses that same word as an admissible ledger value. Two books, one estate, opposite treatments of one token. It is on screen; it is not in this narrator's mouth; it is unresolved and it is the operator's to rule.
Visual — The fence file's banned classes and the cookbook ledger's value row, side by side, the shared token highlighted on both. No voice-over of the token itself.
Receipt — production/schemas/claim_fence.json; UNI-Encyclopedia-Cookbook/README.md § "The two sovereign ledgers"; § "Open questions", Q3

[W-M7-04 | CLAIM]

Intent — The ladder: thirteen developmental rungs, L0 to L12, from a genome to the open questions at the top — and the honest position printed in the book's own front matter, that roughly two of eleven-plus rungs are earned.
Visual — The ladder with two rungs lit and the rest dark; the printed position quoted.
Receipt — UNI-Encyclopedia-Cookbook/cookbook/recipes/L0-genome-zygote.md … L12-creativity-awareness.md; UNI-Encyclopedia-Cookbook/README.md

[W-M7-05 | CLAIM]

Intent — The NATURA recipes: twelve build catalogues, from rocks and water and air to stars, DNA, ants, whales, bats, humans — and the open question beyond. State why nature is treated as an authority at all: it is the only system that has already run the experiment, over a very long parallel search, with the failures deleted.
Visual — The twelve recipe cards dealt out.
Receipt — UNI-Encyclopedia-Cookbook/cookbook/recipes-natura/CN-01-rocks.md … CN-12-beyond-human-open-question.md

[W-M7-06 | CLAIM]

Intent — And its counterweight, which travels with it everywhere: not every trait is an adaptation. Drift, inertia, developmental constraint and frozen accidents are real — the vertebrate retina is wired backwards. "Nature does it this way" generates hypotheses; it settles nothing, and the bio-inspired design still has to beat a tuned conventional baseline on a pre-registered metric or be recorded NEGATIVE.
Visual — The backwards retina, drawn honestly; then the pre-registered-metric requirement.
Receipt — UNI-Encyclopedia-Cookbook/README.md § "Nature as the authority" (Gould & Lewontin

[W-M7-07 | INSTRUCTION]

Intent — How to take it home. Clone the repository; the books are plain markdown you can read in any editor. Run the build to produce the pack for your own assistant; run the site generator, which is standard-library Python only — no packages to install, nothing fetched. Start at the two index files. Everything needed is in the clone.
Visual — Three commands on screen, executed, with their real outputs.
Receipt — UNI-Encyclopedia-Cookbook/tools/build_gpt_pack.py; UNI-Encyclopedia-Cookbook/deploy/deploy_cookbook.py; UNI-Encyclopedia-Cookbook/cookbook/00-INDEX.md; UNI-Encyclopedia-Cookbook/encyclopedia/00-INDEX.md; UNI-Encyclopedia-Cookbook/gpt/README-START-HERE.md

M8 · HOW TO ENGAGE — 29:15 → 33:30 · 9 beats

Job: turn a viewer into a falsifier and a contributor, with steps they can execute, and be honest about which of those steps does not work today.

[W-M8-01 | INSTRUCTION]

Intent — Start where a stranger starts: the public site. Thirteen articles, roughly 290 documents from the repositories rendered as written, the gates, the receipts, the computed coverage and — deliberately — a page listing everything that was withheld and why.
Visual — The site's navigation, each section opened briefly, ending on the omissions page.
Receipt — UNI.Public/README.md; UNI.Public/app/omissions/; UNI.Public/app/coverage/

[W-M8-02 | INSTRUCTION]

Intent — Every page can be read at the depth you want — the same document, three ways in. You do not need to be an engineer to start, and you do not get a simplified version that says something different from the technical one.
Visual — One document, three lenses, the claim identical across all three.
Receipt — UNI.Public/content/lenses/; UNI.Public/content/curation.json

[W-M8-03 | ADVERSE]

Intent — And the honest defect in that offer: today, none of those citations can be opened. The source repositories are private, so every citation says so on its face. A citation you cannot follow is an appeal to authority — marking it is the difference between documentation and a brochure. When a repository is published, one field changes and the same citations become links.
Visual — A citation with its "cannot be opened today" marker; then the single field in the source manifest that would change it.
Receipt — UNI.Public/README.md § "Three things worth knowing", item 1; UNI.Public/generators/sources.json (visibility: private)

[W-M8-04 | INSTRUCTION / FALSIFIER]

Intent — The falsifier's drill, step one: pick a claim and read its written falsifier. Do not argue with the summary — go to the claim's own document, where the estate has already written down the exact observation that would break it. Six of them are written out, numbered.
Visual — The falsification document, one numbered claim expanded, its "Falsify:" clause highlighted.
Receipt — docs/FALSIFICATION.md §§ 1–6b

[W-M8-05 | INSTRUCTION / FALSIFIER]

Intent — Step two: reproduce the verdict. Check out the repository at the commit named in the receipt. Run the launcher named in the receipt with the seed named in the receipt. Compare your outcome to the pre-registered PASS and FALSIFIES conditions. The seed alone is enough.
Visual — The three commands, run, with a real verdict landing at the end.
Receipt — docs/PUBLIC_README.md:13-19; runs/*.exs; docs/PUBLIC_REPRODUCIBILITY_BUNDLE.md

[W-M8-06 | INSTRUCTION / FALSIFIER]

Intent — Step three: attack the check, not just the result. Change one thing that should make the gate fail and confirm that it does. A gate that stays green under a mutation is the real finding — bigger than any single wrong number.
Visual — A mutation applied to a passing gate; the gate turning red as it should; then the same move against a gate that does not, drawn as the prize.
Receipt — viewer/lab/verify_shot.cjs --mutate; UNI-FLAGELLUM/CLAUDE.md § "Required validation" (mutation tests)

[W-M8-07 | INSTRUCTION]

Intent — How to contribute, concretely. A finding here is not a bug report — it has a shape: severity and the claim it hits, the file and line, the command that reproduces it, the root cause, the smallest correction, and the test that must fail first. Write it in that shape and it can be acted on the same day.
Visual — The required fields as a form, filled in with a real past finding.
Receipt — UNI-FLAGELLUM/CLAUDE.md § "Communication" (the reported-delta fields); docs/LAB_PROTOCOL.md § III

[W-M8-08 | INSTRUCTION]

Intent — Where to put it. Comments land append-only on the project's own tracking surface and are version-controlled, never edited. Pre-registrations have a desk that prints the exact line each one needs. Nothing you send is quietly deleted; a correction is a new record, never an overwrite.
Visual — A comment posted, appearing in the append-only file; the desk printing a pre-registration line.
Receipt — viewer/track/track_server.cjs; evidence/track_comments.ndjson; viewer/lab/desk.cjs; viewer/lab/l5.html

[W-M8-09 | INSTRUCTION / STOP]

Intent — And what is not yours to do — the same list that binds every agent working here. Do not write to a frozen artifact; a correction is a new record with a new hash. Do not edit the gate ledger. Do not amend the contract. Every one of these protects the evidence, not the estate.
Visual — The stops list, rendered from the plan file, each with its one-line reason.
Receipt — evidence/remediation/phase9_plan.json → stops, not_mine

M9 · THE CLOSE — 33:30 → 36:00 · 5 beats

Job: land it without triumph, without sentimentality, and without a single fenced word.

[W-M9-01 | FRAME]

Intent — What actually stands, said in one breath: an instrument that catches its own operators writing false numbers, a vocabulary that refuses the words that would flatter it, and a published list of everything that is currently broken.
Visual — Three cards, held.
Receipt — CLAUDE.md generated blocks; production/schemas/claim_fence.json; docs/control-plane/LIMITATIONS.md

[W-M9-02 | FRAME]

Intent — What does not stand, said just as plainly: the mission is intent and remains unmeasured; the science gate has not run; the witness is compromised; the citations cannot yet be followed.
Visual — The same three-card frame, inverted, four items.
Receipt — M1, M6 and M8 beats above, restated by reference — no new claim in the close

[W-M9-03 | FRAME]

Intent — The word this film has not used, and why. Not because the estate is timid, but because a machine-checked list of forbidden words is the only version of restraint that survives contact with an ambitious person at two in the morning.
Visual — The fence file, one last time, whole.
Receipt — production/schemas/claim_fence.json; lab/film/script/reels_cues.json (reel3_proven)

[W-M9-04 | FRAME]

Intent — The invitation, restated as an ask rather than an offer: come and break something. Being caught being wrong is the product working, and the estate has now shown you four separate occasions on which it was caught by its own instruments.
Visual — The door again, open, with the falsifier's three steps as a card beside it.
Receipt — docs/PUBLIC_README.md:19; beats W-M6-05, W-M6-07, W-M8-04–W-M8-06

[W-M9-05 | FRAME]

Intent — Credits, and the reproduction line: how this film was built, which voice model spoke it, and where the manifest that lets you re-derive every narration line lives. The film ends by telling you how to check the film.
Visual — Credits over the generated quality sheet; the reproduce block; the audio manifest.
Receipt — lab/film/welcome/TOOLCHAIN.md; lab/film/welcome/render/synth_narration.cjs (to be built); lab/film/welcome/QC_<cut>.md (generated)

Deriving the two shorter cuts (do this later, from this file, never beside it)

The 12-minute main cut and the 3-minute short are subsets of these 73 beats, not new writing. Two rules govern the derivation, and both exist to stop the honest state being the thing that falls out:

M6 keeps its share. In every cut, the honest-state movement holds at least 16% of the running time and opens at 50–58%. In the 12-minute cut that is ≥ 1:55 opening around 6:30; in the 3-minute short, ≥ 29 s opening around 1:35. A cut that cannot afford the honest state cannot afford to exist.
Every retained claim keeps its receipt. A beat may be dropped; a beat may not be kept with its citation removed for time.

Provisional 12-minute skeleton (33 beats): M0 ×2, M1 ×3, M2 ×3, M3 ×3, M4 ×4, M5 ×4, M6 ×6, M7 ×3, M8 ×4, M9 ×1. Provisional 3-minute skeleton (9 beats): M0 ×1, M1 ×1, M2 ×1, M4 ×1, M5 ×1, M6 ×2, M7 ×1, M8 ×1. The short drops the close, not the honest state — a three-minute film has room for a finding or a farewell, and the finding is the product.

Gates this outline owes before any audio is synthesised

gate	what it checks	status
`welcome-fence`	narration corpus vs `production/schemas/claim_fence.json` — zero hits, all classes	NOT BUILT
`welcome-receipts`	every `CLAIM` / `ADVERSE` beat names a receipt path that resolves on disk	NOT BUILT
`welcome-twin-terms`	every row of the twin-term register appears with both halves in one beat	NOT BUILT
`welcome-honest-state-position`	M6 onset ∈ [50%, 58%] and M6 share ≥ 16% in every cut	NOT BUILT
`welcome-audio-manifest`	every beat has a WAV with recorded voice model, sha256, text, duration	NOT BUILT
`welcome-qc-generated`	`QC_<cut>.md` is byte-identical to the gate's own output	NOT BUILT
`welcome-no-network`	build performs zero network calls; no container invoked	NOT BUILT

None of these are registered in viewer/gate_registry.json yet. Registering them is agent work; appending their rows to evidence/gates.ndjson is S4 — the operator's.

Open questions for the operator (his calls, not an agent's)

Q1 — the mission's home. The mission statement is not currently held in any repository file. Beat W-M1-01 reads it verbatim and cites nothing. Where should it live — lab/film/welcome/MISSION.md, or somewhere more central? Until it has a file, that beat's receipt is a person, and a person is not a receipt.
Q2 — narrator voice. en_US-lessac-medium (TRAVELERS' voice, continuity) or en_GB-jenny_dioco-medium (the estate's spoken-channel default). Both are on disk (TOOLCHAIN.md:48-50).
Q3 — the vocabulary divergence. One repository bans a token the other uses as a ledger value (beat W-M7-03). The film shows the disagreement rather than resolving it. Resolving it is a contract amendment — S5.
Q4 — music bed. Still absent, still requires an explicit rights attestation and a named grant in dgst/RIGHTS.md (TOOLCHAIN.md:63-66).
Q5 — captions. TRAVELERS carried machine-authored Marathi and recorded native-speaker review as its largest editorial risk (QC.md:66). Same language, a different one, or English only?

What this document does NOT claim

It does not claim a running time. 36:00 is predicted from a words-per-minute figure derived from another film. The measurement does not exist until audio_manifest.json does.
It does not claim the beats are the right beats. No beat has been storyboarded, timed, or read aloud. 73 is a plan, not a result.
It does not claim every cited receipt says what the beat needs it to say. Every path in this document was checked to exist on disk on 2026-08-01; the sufficiency of each receipt for its beat is checked by the welcome-receipts gate, which is not built.
It does not claim the honest state is complete. M6 carries twelve findings. They are the ones the estate's own governing documents and generated blocks currently surface. There is no basis in this document for believing that is all of them.

sha256 d8c9ad3d60553d9b — of the original file, so what was ingested stays checkable.

WELCOME TO UNI LABS — narrative outline (the DOCUMENTARY cut)

How to read this page

Precise — the source document