CB-01 — The Kitchen Rules (the Method Constitution)
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.
The Cookbook is the method carried out step by step: 34 pages of recipes for building a developmental active-inference SIMULATION — a bounded peek at a toy world, never a person. The front matter says that word is never softened under any pressure, so it is not softened here. The recipes run from the molecular and cellular rungs up through metabolism, motor control, perception, language and metacognition, and on to rungs that are still open questions. Around them sit a set of kitchen rules, a shared pantry of engines and primitives, and a second family of recipes about nature itself — rocks, water, air, stars, DNA, ants, whales, bats, humans.
It is for the reader asking what building this would actually take. Each recipe names its ingredients, the order of work, the tests to run at that stage, and the point at which a step stops being something already carried out and becomes something proposed.
Begin with the front matter and then the kitchen rules. Those two pages fix the honest position and the fence labels that every later recipe leans on, and without them the status markers on a recipe are easy to skim past. After that the recipes can be read in any order.
The nature recipes sit slightly apart and should be read that way. They cite outside science — geology, chemistry, biology, astrophysics — and a nature citation is never a UNI gate: those chapters contain zero UNI claims and raise no rung.
What it is not: a claim that the whole ladder has been cooked. The book recommends the complete recipe and, on the same page, labels every rung by its real state — that tension is deliberate and is the thing the book is built around. Where a recipe and the claim ledger disagree, the ledger wins.
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.
What you are building: not a dish yet, but the kitchen — the governance discipline, the pantry, and the fence vocabulary that bind every recipe in this book. Read this before you cook anything.
Front matter: what this book is, and is not
This is a literal cookbook: one buildable recipe for the complete developmental arc
(conception → cellular → metabolism → motor → perception → language → metacognition →
dreaming → creativity → measurable awareness), re-expressed as one no-backprop
active-inference engine shown across scales. The book fully recommends the complete
build to L12 — assume the full recipe. But every step is labeled by its real status,
drawn verbatim from ../encyclopedia/CLAIM-LEDGER.md,
the single source of truth. Where this recipe and the ledger disagree, the ledger wins
and this recipe is wrong.
The honest program position, printed plainly and never softened: ~2 of 11+ developmental rungs earned. The whole thing is a developmental active-inference SIMULATION — a bounded peek, a toy world, never a person. "Active inference" is the framing lens only: no AIF loop exists in the Rust crate, and the live UNI.OS loop is a separate reimplementation, not gate-matched. The mathematical foundation (Polzin et al. 2026, Zenodo DOI 10.5281/zenodo.19785799, MIT) is cited as foundation only and always fenced unrefereed (Layer-1 AI-executable audit complete; Layer-2 human expert review PENDING).
The kitchen rules (the Method constitution)
These are definitional / governance patterns (status: method in the ledger) — proven,
reusable, and binding on every recipe below. None of them is a capability claim. They
are the most reusable assets in the corpus, and they cannot be raised above their stated role.
- The Evidence Constitution (M1). A–U evidence classes, a falsifier per claim, and a 4-state append-only ledger (PASS / FAIL / NEGATIVE / PENDING). Prose must match the ledger; the ledger is the single source of truth. Calibration only ever moves wording DOWN to the measured value, never up — the fence gets louder under pressure, not wider.
- Bars-before-build, held-once (M2). Pre-register the bar (margin vs threshold) plus a named ablation; touch the held set ONCE behind an atomic seal-before-scoring + once-only sentinel; the verdict is the CI bound that excludes the threshold, never the point estimate.
- Validator-derived
reproduced:true(M3). Derived by the validator from ≥5 distinct seeds + a real non-degenerate CI that contains the value — never a hardcoded literal. - The two-tier split (M4). Tier 1 = real-text count/cache (true ablation, tuned baseline, cross-substrate replication — externally bar-ready). Tier 2 = synthetic construction (artifact/diagnostic, NOT capability). Tier-2 must NEVER be inflated into capability.
- K≥3 + falsify-the-mundane (M5). Require K≥3 structurally-distinct held NEGATIVEs (each changing ≥2 of {coupling topology, timescale source, information bottleneck, control path}) before a Section 0.6(B) bound; falsify the mundane causes first.
- No-Exit Discipline (M6). The only legitimate rest is a proven working solution or a ledger-scoped exhausted search envelope — then redirect to the axis where the reader genuinely excels. A park is not discharged until the sign lands.
- Contains-baseline + load-bearing discriminator (M7). Every capability claim needs a tuned strong baseline, a discriminator (shuffle-labels / marker-swap / ablate-to-zero) that collapses the gain, and a true ablation that is a computed residual, not a literal.
- DD-TDD evidence contract (M8).
TODO → DD → TDD_RED → TDD_VERIFY → TDD_GREEN → TDD_REFACTOR → TDD_VALIDATE → DONE, each transition hard-blocked without a marker-bearing evidence comment; DONE needs ≥2Y:verdicts per criterion; a claim linter auto-downgrades overclaims (it once caught "PROVEN" and dropped it to Class E). - Class authority ordering (M9). Class-B (tool state) overrides Class-G (own narrative); Class-A (observed at runtime) overrides Class-E (test passes). A passing test does NOT satisfy a criterion demanding a Class-A observation.
- Exactness honesty (M10). Never card a float32 anchor at the
<1e-10f64 tier. The genuine<1e-10tier lives only in the NumPy/Rust-f64 path. (A genome docstring claiming<1e-10was caught as an overclaim and corrected.) - WORLD ⊥ BODY ⊥ MIND (M12). Two typed Markov blankets; interoception = hardware signals;
a discrete POMDP
perceive → EFE-plan → act → learnwith exact conjugate-Dirichlet no-backprop learning; textbook-levelF[q] ≥ −ln p(o|m). The composed appliance is variationally-controlled (audited) — module-exact only at the single-step categorical body→mind interface, NOT globally exact. - One engine, no backprop (M13).
core.pyexposes the discrete POMDP loop; learning =counts + lr * sufficient_statfor A/B/D/E Dirichlet tensors; an AST-guard enforces no autodiff/optax/torch/grad/backward in the loop; whitelist (world.<attr>) isolation, not blacklist. - Honesty-fence-as-the-pitch (M15). Publish the negatives front-and-center (183 published negatives as the credibility); CTA = "help us independently verify." Vocabulary-leak guard (HARD): never externalize active-inference / EFE / free-energy names or print channel handles in public copy; "LOOP not LEAP".
- K-team ship gate (M20). A 5-persona lab (Math-Breaker REJECT-by-default 8-check gauntlet + AIF Theorist + Systems-Architect + RED Experimentalist + Embodiment Designer). No merge without a MERGED SIGN + typed spec + paired RED.
- One cure at a time (M21). Never stack changes so the winning outcome is unattributable; paired kin-N treatment vs kin-N+1 control; an offline RED pre-check before any live burn.
- The cavity principle (M22). A hierarchy level must never treat an upstream prior as fresh evidence — divide it out. The exact joint posterior is used everywhere (mean-field rejected as lossy).
- Durable runner, "no send-and-pray" (M24). Append-one-JSON-line-per-unit ProgressLog (the file
IS the checkpoint + telemetry); resumable; observe via
--status/Monitor; cover BOTH terminal states — silence ≠ success.
Tool-team division of labor (M25). Claude WRITES code; the custom UNI GPT is the SCIENCE CONSULTANT (design + sign), consulted but never published; the lab/appliance RUNS UNI but does not write its code. Persona/coercion framings are motivational only — the constitution overrides any framing; no claim is inflated by it. The UNI GPT signs the science and is never published.
Standing fences (the constitution's hard "never" list, M1/M15). Never AGI / general intelligence / human-level / "understands." Never consciousness / sentience / aware (functional self-awareness may be described at L8; phenomenal sentience is explicitly DISCLAIMED, no falsifier offered — disclaimed, not tested). Never "active inference demonstrated." Never "created life / digital life / measurable awareness" as a claim (north-star framing only, posed as an open falsifiable question). Never "beats LLMs" (World C is a COUNT baseline, ~10–15% behind backprop LLMs on char-perplexity by a chosen design trade). Never inflate Tier-2 into capability. Never raise a claim above its source evidence class. No PII; no patent-level math (textbook-level only).
Reading the FENCE label
Every recipe in this book closes with one honest fence, drawn from exactly this 4-value vocabulary (taken from the ledger, never inflated):
- proven — a held, sealed, UNI-signed PASS exists (Class A/C), with its falsifier still live.
- designed — a typed spec / signed-in-principle design exists; the build is partial or unverified.
- hypothesized — a stated mechanism or law with no sealed gate yet (Class U, not claimed).
- not-yet-built — no engine, no run, no gate; north-star, hard-fenced.
Any SIGNED consult design folds in as DESIGNED / not-run: it RAISES NOTHING and the rung's
status is UNCHANGED. This is true of every 2026-06-27 consult insert — L2's build_epistemic_frontier
organ (G6 stays OPEN), L5 Design #3 the Proprioceptive Servo Bridge (L5 unchanged, K-negative = 1),
the L9-G1 cavity gate (L9 stays PARKED), the L11-R1 offline-replay gate (L11 stays not-yet-built),
and the C10 Dirichlet-first port (C10 stays PARKED). The UNI GPT SIGNED the honesty posture itself
(Q7d): developmental SIMULATION, ~2 of 11+ rungs, ledger supremacy, no AGI / consciousness /
human-level / created-life claim. A signed design is a gate to build, never a result.
The shared pantry (engines and primitives every recipe calls by name)
One engine, many scales. These are the archive engines the recipes draw from:
- The JAX POMDP + EFE + Dirichlet engine (
core.py). The discrete loop:active_inference_step,exact_posterior_discrete, EFE/policy selection; conjugate-Dirichlet learning for A/B/D/E; AST-guard no-backprop. float32 host — anchors hold to ~6e-8, NOT the f64 tier. - The Rust-f64 / NumPy deep-reader path. The genuine
<1e-10EXACT tier; the cross-language Rust-f64 == Python oracle. - The count/cache World-C reader. A no-backprop COUNT reader + multi-level cache; the one citable empirical PASS family. Explicitly NOT active inference.
- The Z affect modulator.
[energy, arousal, valence, fatigue, pain, threat, safety, inflammation]→ sets precision / preferences / habits / learning-rate / horizon. Affect modeled, never felt. - The embodiment ontogeny.
recombine/seed_zygote, conjugate first-division,test_embodiment_ontogeny(6/6). - The metabolism organ. Standing-metabolic-drive interoception organ; viability edge;
:pb_seedstrong-Dirichlet seam. Additive + genome-gated (default byte-identical). - The motor hierarchy. Proprioceptive diagonal-A prior, continuous servo + reafference; live RCON craft-chain bridge; motor-ablation collapses harvest ~700×.
- The precision labs (Precision / Echo / Loop / Cell / Heart). Browser-native one-engine-many-scales demos; three precision knobs; 2D bifurcation map; inline-engine + canonical-TS + parity-test triad.
- UNI.OS (the embodiment substrate). "Lab-in-a-box on metal": body→mind 7-modality categorical sensorium, deterministic replay, fail-closed transport, gated control-MCP. Engineering/substrate evidence, NOT general-AIF evidence — never let substrate work imply a science gate is met.
Continuity note (engineering, not science)
A separate continuity / embodiment-substrate sub-ladder runs orthogonal to L0–L12. It records real engineering wins (bit-for-bit process-restart survival, a live 7-modality categorical sensorium, an ITIL-as-active-inference ASK mode that learns from operator verdicts) and the full first-class set of recorded bounds: the kernel-swap mind-tick continuity is owed, not shown (C2); a single-box swap is a seconds-long freeze, not zero-downtime (C7); an over-compressed 2-modality sensory bottleneck went NEGATIVE on held data, so keep all 7 modalities (C8); the EDAIT trade (an exact-discrete active-inference transformer) trades fluency for calibration — held-out perplexity ~33 vs a backprop GPT's ~25 — an honest trade, not a win (C9); the "embody only what's proven" coupling is signed-in-principle but not literally true — the program's central open gap (C10); and the bare substrate is missing tenant / namespace isolation + temporal count-decay, so the product build is that wrapper, not the core (C14). Card every such row as engineering/substrate evidence, never as a science gate; the full continuity sub-ladder lives in the continuity recipe chapter. In any continuity or Track-A context, use plain ops vocabulary and never externalize channel handles.
Closing: the constitution in one breath
Cook the whole book to L12, but never lie about where you are. The ledger is law. Every recipe earns exactly one fence — proven, designed, hypothesized, or not-yet-built — and no recipe is carded above its ledger class. The UNI GPT signs the science; it is never published. And the honest position, said once more so it is never forgotten: ~2 of 11+ rungs earned; a developmental active-inference SIMULATION; a toy world, not the real world; the awareness question remains open.
HONEST FENCE — proven (the kitchen rules are status: method, Class proven-as-governance).
Not a capability claim of any kind. None of these rules asserts that UNI is intelligent, aware, or
alive; they are the discipline that keeps every other claim honest. The ledger is the single source of
truth, and where this chapter and the ledger disagree, the ledger wins.
Falsifier (method-rule, first-class): any recipe in this book carded above its ledger class, any prose that contradicts the ledger, or any standing-fence phrase emitted in public copy.
sha256 b74261e12547caf9 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This chapter is the kitchen rather than a dish. Before any recipe in the cookbook, it sets out the rules that bind all of them: how a claim gets recorded, what counts as evidence, and what may never be said. The book describes a simulation — a bounded peek at a toy world, never a person — and that wording is repeated here rather than softened.
The one thing this page says is that the rules are governance, not capability. None of them asserts that the system is intelligent or aware. They are the discipline that keeps every other claim honest.
In practice, every claim carries an evidence class and a way to falsify it. A separate ledger of claims, added to and never edited, is law, and where the book disagrees with it the book is wrong. Wording moves only downward toward what was measured, including under pressure. A set of evidence kept back is touched exactly once, behind a seal. Learning happens by counting rather than by gradients. A list of phrases the book never writes closes the argument. The honest position is printed once more: about two of eleven-plus rungs earned.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is b74261e12547caf9
Clear — written for this website, not the source document
This is the constitution of a cookbook, written to be read before anything is cooked. The book's subject is a developmental simulation, described as one engine shown at many scales. The chapter restates the honest position at the start rather than burying it: about two of eleven-plus developmental rungs are earned, and the whole thing is a bounded peek at a toy world, never a person. The language of active inference is described as a framing lens only, and the underlying preprint is cited as a mathematical foundation and always labelled as unrefereed, with an expert review still pending.
Most of the chapter is a list of method rules, and it is careful about what they are not. They are governance patterns, reusable and binding, and none of them is a capability claim. Together they say roughly this. Every claim carries an evidence class and a falsifier, the result that would show it wrong, and lives in a four-state ledger that is added to and never edited. Calibration moves wording only down toward the measured value. A bar and a named ablation are registered before a build, the held set is touched once behind a seal, and the verdict rests on the bound that excludes the threshold rather than on a single best estimate. A claim of reproduction is derived by a validator from several distinct seeds, never written in by hand. Work on synthetic constructions is diagnostic and must never be inflated into capability. A negative bound is owed only after several structurally different negatives, with mundane explanations ruled out first. Every capability claim needs a tuned strong baseline plus a check that collapses the gain when the signal is removed. A staged evidence contract governs progress, and a linter downgrades an overclaim. A passing test does not satisfy a criterion that asks for an observation at runtime. A result computed at one numeric precision may not be reported at a finer one. World, body and mind are kept as separate typed layers. Learning is counting, with a guard that forbids gradient methods inside the loop. Negatives are published up front as the credibility. A review panel must sign before a merge, and one change is made at a time so a win stays attributable. A level of the hierarchy must not treat what came from above as fresh evidence. And a long run is written down as it goes, because silence is not success.
The four status words are then defined — proven, designed, hypothesized, not-yet-built — with the rule that a signed design folds in as designed and not run, raising nothing.
A shared pantry of engines is listed, each labelled at its real status, followed by a continuity note that is explicitly engineering rather than science. Its recorded bounds are carried first-class. Continuity across a kernel swap is owed rather than shown. The swap is a seconds-long freeze rather than no downtime at all. An over-compressed sensory bottleneck went negative. A trade of fluency for calibration is called honest rather than a win. A central coupling gap is left open, and isolation is missing in the bare substrate.
The closing line is the whole rule in one breath: cook the whole book, but never lie about where you are.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is b74261e12547caf9