CB-02 — The shared pantry (engines and primitives)
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: the stocked shelf every later recipe reaches into. Not a new dish: the single set of engines and primitives, named, fenced, and labeled by real status, so that when L0 calls for "the embodiment ontogeny" or L6 calls for "the count/cache World-C reader," you know exactly which jar it means, what it actually proves, and what it must never be read to prove.
One engine, many scales. That is the load-bearing fact of this whole book. The recipes do not each ship their own model; they reuse one discrete-categorical active-inference engine and swap only the observation/generative model and the world it couples to. The pantry below names the engines the recipes call by name. Each recipe's Ingredients list points back here.
The kitchen rules carried in (the constitution travels with the food)
Two governance patterns from the master plan's kitchen rules (status: method, never a capability
claim) define what every pantry engine is and is not:
- WORLD ⊥ BODY ⊥ MIND (M12). Two typed Markov blankets. Interoception is hardware signals. The
mind runs a discrete POMDP loop
perceive → EFE-plan → act → learnwith exact conjugate-Dirichlet no-backprop learning, surprisal/VFE carried only at the 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 (active_inference_step,exact_posterior_discrete, EFE/policy selection); learning iscounts + lr * sufficient_statfor the A/B/D/E Dirichlet tensors; an AST-guard enforces no autodiff / optax / torch / grad / backward in the loop, and isolation is a whitelist (world.<attr>), not a blacklist.
These are status: method (Class method/E). They are not evidence that active inference is
demonstrated. "Active inference" and "free energy principle" appear here only as the framing lens,
at textbook level (Parr, Pezzulo, Friston, Active Inference, MIT Press 2022; plus the unrefereed
preprint, Polzin et al. 2026, Zenodo DOI 10.5281/zenodo.19785799, MIT, foundation only). No AIF loop
exists in the Rust crate; the live UNI.OS loop is a separate reimplementation, not gate-matched.
Reading the FENCE label. Every recipe closes with one of four words, drawn from the ledger and never inflated:
- proven — a held, sealed, UNI-signed PASS exists (Class A/C), 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.
The shared pantry
| Pantry item | What it supplies | Real status (this pantry) |
|---|---|---|
The JAX POMDP + EFE + Dirichlet engine (core.py) |
The discrete POMDP loop: active_inference_step, exact_posterior_discrete, EFE/policy selection; conjugate-Dirichlet learning counts + lr*sufficient_stat for A/B/D/E; AST-guard no-backprop. |
proven as a method (M13, Class E); float32 host — anchors hold to ~6e-8 (~1e-6 single-step filter), NOT the <1e-10 f64 tier. |
| The Rust-f64 / NumPy deep-reader path | The genuine <1e-10 EXACT tier; the cross-language Rust-f64 == Python oracle (M10/M11). |
proven (the only path carded at the f64 tier). |
| The count/cache World-C reader | A no-backprop COUNT reader + multi-level cache; the one citable empirical PASS family (L6.1: +0.081 nats/char, CI [0.0736, 0.0890]). | proven (Class C). Explicitly NOT active inference. |
| The Z affect modulator | Global [energy, arousal, valence, fatigue, pain, threat, safety, inflammation] → sets precision / preferences / habits / learning-rate / horizon (L4.1). |
proven functional (Class C). Affect modeled, never felt. |
| The embodiment ontogeny | recombine / seed_zygote, conjugate first-division, test_embodiment_ontogeny (6/6), Embodiment Rungs (L0.1). |
proven (Class A, float32 tier). |
| The metabolism organ | Standing-metabolic-drive interoception organ; viability edge; :pb_seed strong-Dirichlet seam (L2.1). Additive + genome-gated (default byte-identical). |
proven as a foraging/crafting driver (Class C); recorded NEGATIVE as a building driver, G6 OPEN. |
| The motor hierarchy | Proprioceptive diagonal-A prior (posterior 0.0→0.75), continuous servo + reafference; live RCON craft-chain bridge; motor-ablation collapses harvest ~700× (L5.3). | proven (Class A posterior shift / C live RCON gate / E 277 tests). |
| The precision labs (Precision / Echo / Loop / Cell / Heart) | Browser-native one-engine-many-scales demos; three precision knobs (gamma_a, gamma_b, softmax_temperature); 2D bifurcation map; inline-engine + canonical-TS + parity-test triad (L1.1, L3.1, L6.3). |
proven (Class C / E with parity tests). |
| UNI.OS (the embodiment substrate) | "Lab-in-a-box on metal": body→mind 7-modality categorical sensorium ([M=7, O_max=4]), deterministic replay, fail-closed transport, ASK-mode Dirichlet policy-prior, gated control-MCP (C1, C3, C4, C5). |
proven as engineering/substrate evidence, NOT general-AIF evidence. |
The exactness fence (M10, load-bearing here). The JAX core that core.py lives in is float32.
Its machine-exact anchors hold to ~6e-8, not the <1e-10 tier. A genome docstring claiming <1e-10
was caught as an overclaim and corrected. Only the NumPy / Rust-f64 deep-reader path earns the f64
tier. Any recipe that calls the JAX engine inherits the float32 ceiling; never card a JAX anchor at f64.
The count-baseline fence (L6.1, L6.6). World C is the program's one citable empirical PASS, and it is a COUNT baseline win — NOT active inference, NOT comprehension, NOT "talking," NOT beats-LLMs. It sits ~10–15% behind backprop LLMs on char-perplexity by a chosen design trade. The standing ceiling on this ingredient cannot be raised; "World C" in any Ingredients list means a count/cache reader and nothing more. Engine-attached published bounds (cross-ref): Phase G (L6.4) — five structurally-distinct within-segment-structure designs all NEGATIVE-with-discriminator, only the L5 cache (World C) wins; Phase F (L6.5) — the World-C gain is diffuse (nothing to gate). These are first-class ledger negatives on the reader-as-engine; full treatment lives in the L6 recipe.
The substrate fence (continuity sub-ladder, §2). UNI.OS supplies a real body→mind sensorium, a bit-for-bit process-restart continuity (Stage-1 / C1, Class C engineering), a measured ASK-mode Dirichlet learning shift, and a single-approval control gate — all live on metal. This is engineering / substrate evidence, orthogonal to L0–L12, NOT general-AIF evidence. It does not imply any science gate is met. Stage-2 (the mind surviving a real kernel swap) is owed, not shown (C2 NEGATIVE/owed). And the "embody only what's proven" coupling is the program's central open gap (C10, PARKED, Class U): UNI.OS has no no-backprop Dirichlet learning, no exact info-gain EFE, heuristic isolation, and a different dev model than the one that earned the bars. When a recipe lists UNI.OS, it is borrowing a substrate, never a science result.
Honest floor (C7, first-class NEGATIVE). A single-box swap is a SECONDS-LONG FREEZE, not zero-downtime — true zero-freeze needs the second node carrying the platform — and external media legs (RTP/SIP/kernel-mode rtpengine) are NOT preserved across kexec. Two more UNI.OS-attached negatives are carried here by cross-reference and treated in full in CB-continuity: the EDAIT trade (C9) — 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); and the multi-tenancy gap (C14) — the bare substrate is missing tenant/namespace isolation and temporal count-decay (the product build is the wrapper, not the core).
The affect fence (L4.1). The Z modulator is proven functional — it sets precision and flips the
pragmatic↔epistemic balance, and a Z-ablation collapses the effect. Affect is modeled, never felt;
phenomenal feeling and sentience are explicitly disclaimed wherever Z appears.
Gate (what "the pantry is stocked" means)
The pantry has no single empirical gate — it is the shelf, not a dish. Its pass condition is fidelity: every ingredient above is carded at exactly its ledger class and no higher, and every recipe that lists an ingredient inherits that ingredient's fence. Concretely:
- the JAX engine ships float32 anchors at ~6e-8, never
<1e-10(M10/M13); - World C is cited only as a +0.081 nats/char count-baseline PASS, flagship CI [0.0736, 0.0890], seeds 0–4 (L6.1), never as active inference or a beats-LLMs result;
- UNI.OS is cited only as substrate/engineering evidence (§2), never as a science gate;
- the AST no-backprop guard (M13) holds across every engine that claims it.
Falsifier. The pantry fails if any ingredient is carded above its ledger class anywhere in the book: a float32 JAX anchor printed at the f64 tier; World C described as comprehension, talking, active inference, or beating LLMs; UNI.OS substrate work read as a met science gate; or the no-backprop guard shown tripping where an engine claims it holds. Where this pantry and the ledger disagree, the ledger wins and this pantry page is wrong.
Recorded NEGATIVES (first-class, inline). Scope contract: engine-attached negatives are carried here (metabolism −14% / G6 OPEN, the 2-modality bottleneck C8, A3 #2, mean-field-rejected, plus the fenced C7 / C9 / C14 and L6.4 / L6.5 cross-refs above); per-rung negatives (L1 honest losses, the L7 central wall + J.attribution_caveat) live in their own recipe chapters by design, not here.
- The metabolism organ is a foraging/crafting driver, not a building driver. In the same first 12 h
live RED that gave +135% / 2.35× tool-crafting and +19% mining (L2.1), building went WORSE (−14%)
and G4 allostasis never separated (L2.2). The plateau-break gate G6 stays OPEN. A read-only
counterfactual-EFE audit on real hoarder
.binbrains diagnosed epistemic_starvation — NOT γ-runaway (γ≈7.8, unsaturated), NOT a curriculum ceiling (L2.3). Stock the organ for foraging; never spin the +135% as "breaks the plateau." - The over-compressed sensory bottleneck went NEGATIVE. A 2-modality
deep_statebottleneck lost on held data versus the full 7-modality contract (C8) — keep all 7 modalities. The same negative recurs on the science side as A3 Design #2 (slow Z-bottleneck), held Δ −0.091, CI [−0.134, −0.055], a single sealed motor negative (K-negative = 1, no Section 0.6(B) bound owed; L5.2). - The exact joint posterior is mandatory. A factored mean-field posterior variant was implemented and rejected as lossy (the cavity principle, M22). The exact joint is used everywhere; mean-field is not a shortcut a recipe may take.
HONEST FENCE — proven (as a method-and-engine shelf), with named designed/parked items.
The pantry as a set of method patterns (M12, M13, M10) and held engines (L0.1 Class A; L1.1, L2.1,
L3.1, L4.1, L5.3, L6.1, L6.3 Class C/E; C1, C3, C4, C5 substrate) is proven at its stated classes,
falsifiers live. Folded in as DESIGNED / not-run (raising nothing, leaving every rung's status
unchanged): the SIGNED consult designs the later recipes reach for — L2's build_epistemic_frontier
organ (G6 stays OPEN), L5's Design #3 Proprioceptive Servo Bridge, L9-G1, L11-R1, the L12 bound, and the
C10 Dirichlet-first port. Each is a typed spec or signed-in-principle design, not a result.
Not claimed: this pantry does not demonstrate active inference, does not create life or a synthetic organism, does not make UNI conscious / aware / self-aware, and does not show UNI is human-level or AGI. The JAX core is float32 (anchors ~6e-8, not f64); World C is a count baseline (~10–15% behind backprop LLMs, not active inference, not beats-LLMs); UNI.OS is engineering/substrate evidence, not general-AIF evidence. The whole program remains a developmental active-inference SIMULATION — a bounded peek, a toy world, never a person. Honest position: ~2 of 11+ developmental rungs earned.
sha256 1c81292170dfda92 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
A shelf, not a dish. It lists once, in one place, the engines and building blocks that every later recipe in the cookbook reaches for. When a recipe names an ingredient, you can come here to see exactly what that ingredient is, what status it holds, and what it must never be read to mean. All of it belongs to a simulation, not to a person.
The recipes do not each ship their own model. They reuse one engine and swap only the world it is coupled to. The label attached to an ingredient here travels with it wherever it is used, and a recipe inherits its ingredients' limits.
Most of the page says what its ingredients must not be read to mean. A number computed at one level of precision may not be printed at a finer one. A reading result is a counting baseline and nothing more — not comprehension, not active inference, not a win over other systems. Hardware and substrate work is engineering evidence, never evidence that a science gate was met. Negative results are set out inline rather than tucked away. And it closes with a list of what is not claimed: no demonstration of active inference, no created life, no awareness, no human-level ability.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 1c81292170dfda92
Clear — written for this website, not the source document
This chapter is the stocked shelf the rest of the cookbook reaches into. It exists so that a later recipe can name an ingredient in one line and a reader can look up here what that ingredient actually is, what its status really is, and what it must not be read to prove. The load-bearing idea is stated at the top: one engine, many scales. Recipes do not each bring their own model; they reuse one discrete engine and swap only the observation model and the world it couples to.
Two governance rules travel with the food. World, body and mind are kept as separate typed layers, with signals from the hardware treated as the body's own senses. And learning is done by adding counts, with a guard that forbids gradient methods inside the loop. The chapter is explicit that these are method patterns rather than evidence, and that the language of active inference appears only as a framing lens at textbook level.
The centre of the page is a table of ingredients, each with what it supplies and its real status. There is a discrete planning engine, a higher-precision reading path, a counting reader, and a modulator that sets how strongly signals are weighted. Then a starting-point routine for a first cell division, an organ for metabolism, a motor hierarchy, a set of small browser demonstrations, and a substrate that runs the whole thing on real hardware.
Then come the fences — the sentences saying what each ingredient must not be read to mean — which take up more room than the table. The precision fence says that anchors computed at the coarser numeric setting must never be printed at the finer tier, and it records that a claim of that sort was caught and corrected. The counting fence says the reading result is a baseline win by counting and nothing more. It sits behind systems trained by gradient methods on one measure, by a chosen design trade, and it points to two published negatives attached to the same reader. The substrate fence says that engineering work on the body does not imply any science gate is met. Continuity across a real kernel swap is owed rather than shown. And coupling the body only to what has been earned is the programme's central open gap, which is parked. The affect fence says the modulator is functional and that feeling is modelled, never felt.
There is no single experiment that passes this page. Its condition is fidelity: every ingredient labelled at exactly its class and no higher, and every recipe inheriting its ingredients' fences. What would sink it is stated as the mirror image: the pantry fails if any ingredient is ever labelled above its class anywhere in the book.
Recorded negatives close the chapter, carried inline by design. The metabolism organ helps foraging and crafting but made building worse in the same live run, and a gate on that stays open. An over-compressed sensory bottleneck lost on data it had not seen, so the fuller set of channels is kept. A simplified factored posterior was built and rejected as lossy, so the exact joint version is used everywhere. The final fence repeats what is not claimed: no demonstration of active inference, no created life, no awareness, no human-level ability — a developmental simulation, a bounded peek, about two of eleven-plus rungs earned.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 1c81292170dfda92