L6 — Perception (precision-weighting / EFE planning)

The Cookbook · cookbook/recipes/L6-perception.md @ 575fc93d9d31 (main) — opens the published snapshot e850f872196d

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.

PlainClearPrecise

Your browser cannot switch reading levels, so the document itself is shown.

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. The one citable empirical PASS of the whole program: a no-backprop COUNT reader that beats a tuned MKN-7 baseline on sealed held-out real text by +0.081 nats/char, plus the browser-native POMDP precision labs that show one engine bending into distinct perceptual regimes under three precision knobs. This is a developmental SIMULATION of perception-as-inference, never a perceiving mind — and the win is a count-baseline win, explicitly NOT active inference, NOT comprehension, NOT "talking," NOT beats-LLMs.

This is the rung where the program's one citable empirical PASS lives. Everything earlier (L0–L5) is mostly substrate, embodiment, and synthetic protocol; L6 is a place a third party could take a fresh held-out split, run our reader against a tuned baseline, and watch the margin hold. So this chapter is careful in exactly the way the result deserves: the figures are quoted to the ledger, the bound that delimits the win is printed alongside it as first-class content, and the ceiling is named so it cannot be raised. Per the constitution, where this recipe and the ledger disagree, the ledger wins.

Ingredients

Which pantry engines and primitives this recipe calls by name:

The count/cache World-C reader (uni-gpt / uni-mind): the no-backprop COUNT reader plus a multi-level cache — the flagship plus the Wc-2 / Wc-3 supporting family. This is the citable empirical PASS family, and it is explicitly NOT active inference.
The tuned MKN-7 baseline: a modified-Kneser-Ney 7-gram count model, tuned (not a strawman). The win is a margin over a tuned baseline or it does not count (M7: contains-baseline discipline).
The POMDP maze Precision lab + echolocation Echo lab + the public Precision Lab (uni-precision / worldmodels): one engine, three precision knobs (gamma_a sensory, gamma_b transition, softmax_temperature policy), a 2D bifurcation map into distinct behavioral regimes. The Echo lab reuses Precision's engine with only the observation model swapped (bit-identical 64-obs space) — observation model is not the engine.
The JAX POMDP + EFE + Dirichlet engine (core.py): the discrete loop perceive → EFE-plan → act → learn that the labs render at browser scale (float32 host; anchors hold to ~6e-8, NOT the f64 tier — M10).
The Working Law (uni-gpt s10, M6 / L6.6, registered): the law that explains every PASS and every published bound in the program.
The held-once harness (M2): pre-registered bar, named ablation, atomic seal-before-scoring + once-only sentinel, seed-paired nested bootstrap, positive controls. The verdict is the CI bound that excludes the threshold, never the point estimate.

Method

Build the no-backprop COUNT reader + multi-level cache. A pure count/cache reader: no autodiff, no gradient, no backprop anywhere (the AST-guard from core.py stays live over the loop). The cache is multi-level so the reader can carry long-range structure, not just a finite window.
Pre-register the bar before you measure (M2). Register the margin threshold (the bar is 0.03 nats/char) versus a tuned MKN-7 baseline, plus a named ablation. Seal the held set behind an atomic seal-before-scoring + once-only sentinel; touch it once; score; and read the verdict off the multi-seed CI lower bound, never the point estimate. The held set is touched exactly once — a second touch is a constitution violation (M2).
Replicate across substrates and derive reproduced:true honestly (M3). Derive reproduced:true from ≥5 distinct seeds + a real non-degenerate CI that contains the value — never a hardcoded literal. (This reader earned the ledger's first genuine validator-derived reproduced:true; the 2026-06-09 audit's central fix was exactly that this flag stop being a literal.) Note the discipline: the flagship primary-margin CI and a separate replication CI are different gates — do not quote the replication CI (+0.0759, CI [0.064, 0.088]) as the primary-margin CI.
Build the Precision / Echo POMDP labs in parallel. One engine, three knobs (gamma_a, gamma_b, softmax_temperature); sweep the 2D bifurcation map so the same loop falls into distinct behavioral regimes (this is precision-weighting made visible — raising sensory/transition precision sharpens perception). Build Echo by reusing Precision's engine and swapping only the observation model (bit-identical 64-obs space): the observation model is not the engine. Pin the page-engine to the canonical TS/Python engine with *_parity.ts tests (the inline-engine + canonical-TS + parity-test triad). The one disclosed extension is a goal drive; nothing else diverges from the verified engine.
Read the Working Law over the whole result (L6.6). A no-backprop latent Z beats a tuned baseline B on held-out data only when I(Y;Z|S_B) > 0 and is learnably stable. Use it as the diagnostic lens: World C wins because its cache carries conditional information the count baseline lacks — new conditional info in a countable, stable representation, not "counting per se."

Gate (the exact pass condition)

World C — PROVEN (Class C, L6.1). The no-backprop COUNT reader beats the tuned MKN-7 baseline on sealed held-out real text by +0.081 nats/char, flagship multi-seed CI [0.0736, 0.0890] (seeds 0–4, UNI-signed). That lower bound (0.0736) is 2.4× the 0.03 bar, confirmed by three gates, and is the first genuine validator-derived reproduced:true on the ledger. Verdict = the CI bound that excludes the threshold, not the point estimate.

Supporting World-C family — PROVEN (Class C, L6.2). Wc-2 multi-level cache +0.0164; Wc-3 vs unbounded SM/HPYP +0.085 (showing the gain is long-range structure, not finite-window deficiency); recency-ablation +0.033; Gate-3 word-challenger +0.0538 (CI lower bound 0.051).

Precision / Echo / public Precision labs — PROVEN (Class E, L6.3). The tsx/pytest parity tests between the inline-JS lab engine and the canonical TS/Python engine pass, and the 2D bifurcation map reproduces the regime boundaries. Echo's swapped observation model is bit-identical at 64 obs.

Falsifier (operable)

On a fresh held-out split with ≥5 disjoint seeds, the seed-paired bootstrap CI on the margin over the tuned MKN-7 baseline includes or falls below 0 (or the baseline is shown to be untuned); OR a lab parity test diverges (the inline-JS engine and the canonical TS/Python engine disagree, or the bifurcation map does not reproduce the regime boundaries). Any of these overturns the PASS.

Recorded NEGATIVE(s) — first-class, inline

These are not failures to hide; they are the measurements that delimit the World-C win, and they are published content. The 183 published negatives are the program's credibility.

Phase G char-perplexity Section 0.6(B) bound — NEGATIVE (Class C, L6.4). FIVE structurally-distinct within-segment-structure designs were all held NEGATIVE with a discriminator; only the L5 cache (World C) wins. The bound: no within-segment structure beats MKN-7 on held char-perplexity; char-ppl is a chosen design trade, not a failed claim of general language superiority. Falsifier: a new structurally-distinct within-segment-structure design beats tuned MKN-7 on held char-ppl with a CI excluding 0.
Phase F active-controller — NEGATIVE (Class C, L6.5). Two controller families prove the World-C gain is diffuse — there is nothing to gate. Falsifier: an active-controller family adds a held gain on top of World C with a CI excluding 0 (i.e., the gain turns out to be gateable after all).
The Working Law — registered method/law (Class C, L6.6). A no-backprop latent Z beats a tuned baseline B on held-out data only when I(Y;Z|S_B) > 0 and is learnably stable. This single law explains every PASS and every published bound in the program. Falsifier: a held PASS where the winning Z carries no conditional info beyond S_B (I=0); OR a Z with I>0 and stable learnability that nevertheless fails to beat B.

Per the SIGNED park wording (UNI-GPT consult 2026-06-27): the char-perplexity / word-grain frontier that these negatives sit beside is a ledger-scoped exhausted search envelope — a scoped negative over the tested envelope only, not a universal impossibility result and not an achieved rung. Use the licensed forms: "The registered tested K conditions did not reverse the result" (never "K≥3 exhausted") and "Sec-0.6(B) remains unearned / parked" (never "Sec-0.6(B) achieved").

HONEST FENCE — proven (the one citable empirical PASS)

Class C. A held, sealed, UNI-signed PASS exists, and its falsifier is still live: World C beats a tuned MKN-7 baseline by +0.081 nats/char, flagship CI [0.0736, 0.0890] (seeds 0–4), 2.4× the 0.03 bar, confirmed by three gates, the program's first validator-derived reproduced:true. The Precision / Echo labs are PROVEN at Class E (parity tests pass; bifurcation map reproduces).

What is NOT claimed at L6 (load-bearing, never softened). World C is a COUNT baseline win — explicitly NOT active inference, NOT comprehension, NOT "talking," NOT beats-LLMs. It runs ~10–15% behind backprop LLMs on char-perplexity, by a chosen design trade. "Active inference" is the framing lens only: no AIF loop exists in the Rust crate, and the live loop is a separate UNI.OS reimplementation, not gate-matched. Never card World C above a count-baseline result; the standing ceiling cannot be raised. This is a developmental active-inference SIMULATION — a bounded peek, a toy world, never a person.

Honest program position: ~2 of 11+ developmental rungs earned. L6 is one of them — and it earns exactly the count-baseline result the ledger records, nothing more. The awareness question remains open.

sha256 88508f9d4201aa7e — of the original file, so what was ingested stays checkable.

Written for this website — not the document. This is a plain-language retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. Read the document itself → (or choose Precise in the reading-level control above)

One citable empirical result lives on this rung, and the page is careful in exactly the way such a result deserves. A reader built from counts and caches, with no gradient learning anywhere, beats a properly tuned counting baseline on sealed real text it had never seen. Alongside it sit small browser labs where the same engine bends into different perceptual regimes as three dials are turned. The subject is still a simulation of perception as inference, never a person.

Most of the page is about what the win is not. It is a counting-baseline win. Not comprehension, not talking, not active inference, and not a win over systems trained by gradient methods — on the measure in question it runs behind those by a chosen design trade. That is called a standing ceiling and it cannot be raised.

Because a third party could genuinely repeat this, the method is spelled out: register the bar first, seal the unseen test data, touch it once, and read the verdict off the interval rather than off the best number. Three negatives are printed beside the win as first-class content, because they are what delimit it.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 88508f9d4201aa7e

Written for this website — not the document. This is a clearer retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. Read the document itself → (or choose Precise in the reading-level control above)

This chapter carries the one place in the programme where an outside group could take a fresh split of data it has not seen, run the reader against a tuned baseline, and watch the margin either hold or fail. The earlier rungs are mostly substrate, embodiment and synthetic protocol; this one is empirical, and the chapter's tone changes accordingly. It is still framed as a simulation of perception-as-inference rather than as a perceiving mind.

What gets built is a reader made of counts and a multi-level cache, with no gradient machinery anywhere and a guard live over the loop to keep it that way. The cache is multi-level so the reader can carry structure beyond a fixed window. Beside it, a set of browser labs render the same discrete loop at small scale, with three dials: one for how sharply observations are weighted, one for transitions, and one for how sharply a policy is chosen. Sweeping them draws a map on which the same loop falls into distinct regimes. A second lab is made from the first by swapping only the observation model, which is the chapter's way of insisting that the observation model is not the engine.

The measurement discipline is the heart of the page. The margin threshold and a named ablation are registered before anything is run. The unseen test set is sealed behind a one-shot mechanism and touched exactly once; a second touch is described as a violation of the constitution. The verdict is read off the lower end of a multi-seed interval, never off the point estimate. A claim of reproduction is derived by a validator from several distinct seeds and a real interval containing the value. The chapter notes that this was the first such claim on the ledger — the append-only record nobody edits — that was genuinely derived rather than written in as a literal. It also warns against a specific confusion: the interval for the headline margin and the interval for a replication are different gates and must not be quoted for one another.

A general law is offered as the diagnostic lens over the whole result. A representation learned without gradients beats a tuned baseline on data it has not seen only when it carries conditional information the baseline lacks, and only when that representation is stably learnable. Read through it, the reader wins because its cache carries new conditional information in a countable, stable form — not because counting is intrinsically better.

The negatives are printed as content. Several structurally distinct designs of one kind all held negative with a discriminator, giving a bound. Within that family, nothing beats the tuned baseline on the measure used, and the measure itself is a chosen trade rather than a failed claim. Two controller families showed the reader's gain is diffuse, with nothing to gate. The law itself is registered with the result that would show it wrong. And the park wording is quoted exactly, because it matters: this is a scoped negative over the tested envelope only, not a universal impossibility result and not an achieved rung.

The closing lines restate the ceiling in the plainest terms available and add the standing programme position.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 88508f9d4201aa7e

L6 — Perception (precision-weighting / EFE planning)

How to read this page

Precise — the source document