CB-02 — The shared pantry (engines and primitives)

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 → learn with exact conjugate-Dirichlet no-backprop learning, surprisal/VFE carried only at the textbook level F[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.py exposes the discrete POMDP loop (active_inference_step, exact_posterior_discrete, EFE/policy selection); learning is counts + lr * sufficient_stat for 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

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 .bin brains 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_state bottleneck 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.