L9 — Reasoning / conscience (narrative-self → conscience → reasoning)
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. A design, not a dish: the first pre-registered, sealed, UNI-signed gate that would let the developmental simulation begin to test whether a slow cross-level latent ("a spine") can carry a commitment forward and improve prediction when the local segment points the wrong way. Nothing here is cooked yet. This whole chapter is PARKED, Class-U — not claimed, and the signed L9-G1 design below is a gate to BUILD, not a result. Honest program position: ~2 of 11+ developmental rungs earned.
Read this first (the fence on the whole chapter)
L9 names three ladder levels — narrative-self, conscience, reasoning — that exist as design in the HUMAN-HGM-001 arc and as a designed, FE-form-signed but unbuilt roadmap (the strings Phases 3–5: spine → glands → hemispheres). They are NOT separately gated or PASSed. Ledger row L9.1 is Class U (parked): "L7–L9 (narrative-self, conscience, reasoning) exist as levels in the HUMAN-HGM-001 ladder DESIGN but are NOT separately gated/PASSed." Its falsifier field reads n/a (parked) — a falsifier exists only once a gate is defined and run. Ladder DESIGN levels are never inflated into capability. Nothing in this chapter raises the rung.
The UNI-GPT consult of 2026-06-27 (SIGNED) added a designed first gate, L9-G1. Per the consult, adding it raises nothing: L9 remains parked, Class-U. The consult's own caveat is load-bearing and is repeated verbatim below: "Phase-3 spine"/"cavity principle" are our design labels grounded in the standard hierarchical / deep active-inference pattern (slow high levels set empirical priors for fast low levels) — a modeling / inference claim, NOT a claim about reasoning or conscience in the world.
Ingredients
These are designed, not yet supplied as a gated engine. Each is named so a competent engineer knows what to reach for when the build is actually authorized.
- The HUMAN-HGM-001 ladder DESIGN for L7–L9 (narrative-self, conscience, reasoning) — design levels only, never carded as capability.
- The strings Phases 3–5 roadmap (spine → glands → hemispheres) — designed, every FE form already UNI-GPT-signed, but unbuilt. The strings digest records these explicitly as "designed roadmap, gated, every FE form already UNI-GPT-signed, not built."
- 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). This is the durable correctness rule for the whole inference hierarchy.
- Parameter-information-gain curiosity (the pymdp pA/pB parameter-information-gain form) as a candidate primitive.
- The JAX POMDP + EFE + Dirichlet engine (
core.py) supplies the discrete loop and conjugate-Dirichlet no-backprop learning that any future spine would have to run inside; the AST-guard keeps backprop out of the loop.
No engine, no run, no gate is supplied today. That absence is the status.
Method (recipe-to-define)
The published recipe is a recipe to define, because the gate does not yet exist.
- Define a pre-registered, sealed, UNI-signed gate for narrative-self / conscience / reasoning. None exists today. The signed
L9-G1design (below) is the smallest such target proposed so far — and it deliberately does not test reasoning, conscience, or narrative-self; it tests only persistent cross-level state plus constraint-sensitive prediction. - Build the spine (strings Phase 3) to carry cross-level priors without double-counting — the cavity principle (M22). The spine carries ONE slow latent across several lower-level segments; the receiving segment's own evidence path must be divided out of the downward prior.
- Run the gate held-once; verdict = the CI bound that excludes the threshold, never the point estimate (M2). Touch the held set ONCE behind an atomic seal-before-scoring + once-only sentinel.
Until step 1 is actually done and step 3 is actually run, the rung stays parked.
The SIGNED L9-G1 gate design (DESIGNED / not-yet-run — folded in, raises nothing)
UNI-GPT consult 2026-06-27 (SIGN), cross-ref cookbook/UNI-GPT-CONSULT-2026-06-27.md Q4. This is a designed gate, not a result. Adding it raises nothing; L9 remains parked, Class-U.
Smallest L9 target — L9-G1: Cavity-correct delayed commitment-state prediction (explicitly NOT reasoning / conscience / narrative-self): build the strings Phase-3 spine to carry ONE slow latent z_spine = {role_id, active_commitment, norm/constraint_tag} across several lower-level segments, then test whether it improves prediction when the local segment points the wrong way. Sealed episode: seg-1 establishes a commitment; segs 2..N are distractors; the final segment requires predicting the next action / conflict cue. It is the smallest such target because it needs only persistent cross-level state + constraint-sensitive prediction — no open-ended reasoning, morality, selfhood, language, or metacognition.
Primary bar. ΔNLL = NLL_baseline − NLL_spine > 0, 95% paired bootstrap CI excludes 0; effect-size guard ≥3% relative NLL reduction on the load-bearing delayed-commitment subset.
Secondary bar. Conflict / violation tag AUC or F1 +≥0.05 absolute, paired CI excludes 0.
Calibration guard. ECE worsens by ≤ a pre-registered margin (e.g. +0.02).
Tuned baseline it must beat. A tuned recency-frequency discourse prior — segment-local predictor + exp-decayed role/action frequencies + low-order transition counts over role/action/conflict tags + tuned window k, decay τ, smoothing α. It may use recent tags / context but may NOT carry a slow latent spine with cavity-correct cross-level residuals; it is tuned on train/val only and frozen before test. Beating only an untuned recency model leaves the gate OPEN. This is deliberately the same baseline as the parked L9/L10 NEGATIVE (see below), so a pass here would have to beat the exact prior that has so far won.
Discriminator + computed-residual (cavity) ablation. Load-bearing discriminator = local-distractor commitment reversal — credit only if the improvement concentrates on this split, not on easy same-topic continuity. The cavity ablation uses the computed residual r_t(x) = log p(x_t | q_high(z_t)) − log p(x_t | q_cavity(z_t)), where q_cavity excludes the receiving segment's own evidence path; treatment uses q_local + r_t, ablation sets r_t := 0 (or episode-shuffles r_t) and MUST collapse the gain. Double-counting sentinel = a naive-spine ablation (full high-level posterior as downward prior WITHOUT cavity subtraction; expected to sharpen confidence but worsen calibration).
Fenced pass-claim (signed; applies only IF/WHEN run and passed). "L9-G1 passed, fenced. In a sealed delayed-commitment prediction task, the Phase-3 spine improved commitment-conditioned next-action and conflict-tag prediction over a tuned recency-frequency discourse prior; the improvement concentrated on the pre-registered local-distractor split and collapsed when the computed cavity residual was zeroed or shuffled." Non-license: does NOT claim reasoning, conscience, narrative-self, metacognition, awareness, consciousness, AGI, human-level language, created life, or "active inference demonstrated"; does NOT unpark L9 globally. Evidence class if/when earned: hierarchical AIF substrate Class B/Sec; Phase-3 spine design Class C.
Status: DESIGNED / not-yet-run. L9 STAYS PARKED, Class-U — not claimed.
Gate
None defined. L9 becomes testable only once a pre-registered, sealed, UNI-signed gate is defined and run. Until then: Class-U — not claimed (ledger row L9.1, status: parked).
The L9-G1 bars above are a gate to BUILD, not a met condition. There are no figures to cite as achieved, because nothing has been run. The only honest numbers on this rung are the bars a future run would have to clear (ΔNLL CI excludes 0 with ≥3% relative reduction on the delayed-commitment subset; conflict-tag AUC/F1 +≥0.05; ECE worsens ≤ +0.02) — none of which is a result.
Falsifier
n/a (parked). A falsifier exists only once a gate is defined; the ledger row L9.1 falsifier field is literally "n/a (parked)." The L9-G1 falsifiers below apply only once that gate is built and run — they are part of the design, not active tests:
The gate (when built and run) fails if any of: ΔNLL CI includes 0 or the reduction is <3%; the improvement appears only on easy continuity, not the local-distractor split; zero/shuffle of r_t does NOT collapse the gain; a naive non-cavity spine matches/beats the cavity spine without a calibration penalty; a stronger tuned baseline closes the gap; leakage is found; the gain trades off against an already-earned lower rung; OR any write-up implies reasoning / conscience / metacognition / AIF-demonstrated / AGI / consciousness / human-level / created-life.
Recorded NEGATIVE(s)
These are first-class and inline. They are the reason this region is parked rather than open, and they define exactly what a future L9-G1 pass would have to overturn.
L7.4 — Phase K role-persistence (NEGATIVE, bound). Three structurally-distinct no-backprop designs (min-role, chain, track) all tune their role/persistence terms OFF; only ~0.02 nats survives (CI spans 0). Bound: no-backprop role-persistence does not beat a tuned recency/frequency discourse prior on adversarial anonymized referent cloze. This is the same tuned discourse prior the L9-G1 design picks as its baseline — by design — which is why beating it would be the load-bearing event.
The signed L7 / T2 / L9–L10 Park Statement (UNI-GPT consult 2026-06-27, Q1, SIGNED). Per the canonical signed wording: "For L9–L10: the tested no-backprop role-persistence mechanisms did not beat a tuned recency-frequency discourse prior under the ledgered protocol. This parks the rung as an honest negative frontier result. It does not achieve Sec-0.6(B), does not demonstrate active inference, and does not license claims of metacognition, consciousness, AGI, human-level discourse, or created life." This is a ledger-scoped exhausted search envelope, not a universal impossibility result and not an achieved rung. Use the signed replacements: instead of "K≥3 exhausted" write "The registered tested K conditions did not reverse the result."; instead of "Sec-0.6(B) achieved" write "Sec-0.6(B) remains unearned / parked."
The owed sign-to-park (open). By No-Exit Discipline (M6) a park is not discharged until the sign lands. The L9–L10 ladder park is folded into the 2026-06-27 signed Q1 statement; the earlier char-perplexity / T2 park sign (
UNI_CONSULT_5) was drafted and owner-relayed but recorded as not-yet-captured. Carry the park as PARKED, never as closed.
HONEST FENCE — designed (PARKED, sign-to-park owed)
Class U. Nothing is claimed. This rung is a design, not a dish. Ladder DESIGN levels are never inflated into capability. The strings Phases 3–5 (spine → glands → hemispheres) are the designed-but-unbuilt roadmap toward this region; the signed L9-G1 design above is a gate to BUILD, not a result — folding it in raises nothing and the rung's status is UNCHANGED.
Not claimed (load-bearing, never softened): UNI does not reason, does not have a conscience, does not have a narrative self, is not self-aware here, does not demonstrate active inference, is not human-level, is not AGI, has not created life. "Phase-3 spine" and "cavity principle" are design labels for the standard hierarchical / deep active-inference pattern (slow high levels setting empirical priors for fast low levels) — a modeling / inference claim, NOT a claim about reasoning or conscience in the world. The whole program remains a developmental active-inference SIMULATION: a bounded peek, a toy world, never a person. ~2 of 11+ developmental rungs earned. The cookbook fully recommends building toward L12 — and this step, honestly, is designed / parked, Class-U, not-claimed, with the gate still owed.
sha256 1621023eb82a7c7b — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
A design, not a dish, and the opening line says so. Nothing on this rung has been built, run or gated. It names three levels — a narrative self, a conscience, and reasoning — that exist as design only, and the whole chapter is parked and not claimed. The subject stays a developmental active-inference simulation: a bounded peek at a toy world, never a person.
What the absence means is spelled out. There is no engine, no run and no gate here, and that absence is the status rather than a gap in the writing. Nothing could yet show it wrong, because that only exists once a gate is defined.
What the page does contain is a proposed smallest first gate, signed off in a consultation. It does not test reasoning or conscience. It tests something much narrower: whether one slow piece of state carried across several stretches of an episode can improve a prediction when the immediate context points the wrong way. The design comes with the bar it would have to clear, the tuned baseline it would have to beat, and the removal test that would have to collapse its gain. Folding it in raises nothing, and the rung's status is unchanged.
A published negative sits underneath, which is why the region is parked rather than merely empty.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 1621023eb82a7c7b
Clear — written for this website, not the source document
Here, for the first time in the cookbook, the honest answer is that nothing has been cooked. The subject stays what it has been all along: a simulation, a bounded peek at a toy world, never a person. The rung names three levels — a narrative self, a conscience, and reasoning — which exist as design in a ladder and as a signed but unbuilt roadmap. They are not separately gated and not passed. The row behind them in the ledger, the append-only record of claims, is marked as parked, and its field for what would sink the claim reads, literally, not applicable. Design levels, the chapter says, are never inflated into capability.
The ingredients are listed anyway, so that a competent engineer would know what to reach for if a build were ever authorised. They are the ladder design, the unbuilt roadmap, a correctness rule for hierarchies, a candidate curiosity primitive, and the discrete engine any future version would have to run inside. The chapter then states that no engine, no run and no gate is supplied today, and that the absence is the status.
The method is a recipe to define rather than a recipe to follow. Define a sealed gate, written down before it is run — none exists. Build the slow cross-level state so that it carries information downward without double-counting, which means the receiving stretch's own evidence must be divided back out of what it is told from above. Run the gate once behind a seal and read the verdict off the interval. Until the first and third of those actually happen, the rung stays parked.
The bulk of the page is a signed design for the smallest possible first gate, and its restraint is the point. It explicitly does not test reasoning, morality, selfhood, language or metacognition. It tests whether one slow latent carrying a role, a live commitment and a constraint tag can improve prediction in a sealed episode. An early stretch sets up a commitment, several middle stretches distract, and the final stretch asks what comes next. It states a primary bar, a secondary bar, and a guard against the model becoming worse calibrated. It names the baseline it must beat and insists that baseline be tuned, noting pointedly that this is the very same prior that has so far won against every attempt. It defines the split where the improvement must concentrate, and an ablation using a computed residual that must collapse the gain when zeroed or shuffled. A sentinel version without the correction is expected to look more confident while being worse calibrated. It even writes out, in advance, the exact sentence that could be claimed if it ever passed, together with the list of things that sentence would still not license.
The recorded negatives explain why this region is parked rather than simply unexplored. Several structurally distinct attempts at carrying roles across a passage tuned their own mechanism off and failed to beat that same tuned prior. The park is quoted in its signed wording as a scoped negative over the tested envelope only, not a universal impossibility result and not an achieved rung, with banned phrasings replaced by permitted ones. And the sign-off that would discharge the park is itself recorded as owed, so the park is carried as open.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 1621023eb82a7c7b