L0 — Molecular / genome → zygote (the conception prior)

One recipe, one engine, shown at the smallest scale: the very first cell division, modeled as an exact conjugate Bayesian update with no gradients. A developmental active-inference SIMULATION of conception, never a person, never a baby, never life.

What you are building. A float32-tier, no-backprop instantiation of the conception prior: encode a genome as a categorical prior, seed_zygote the initial Dirichlet tensors, and model the zygote's first division as an exact discrete Bayes (conjugate) update that matches the closed-form value to the float32 tier, gated by test_embodiment_ontogeny.py at 6/6 with a live no-backprop guard.

---

Ingredients (pantry engines / primitives, called by name)

• The embodiment ontogeny (recombine / seed_zygote, the no-backprop AST-guard, ontogeny 6/6, conjugate first-division, the cell-division identity embedding) — from the JAX engine (uni-mind / uni-gpt).
• The JAX POMDP + EFE + Dirichlet engine (core.py): the conjugate discrete-Bayes update (exact_posterior_discrete); learning is counts + lr * sufficient_stat for the A/B/D Dirichlet tensors. float32 host — anchors hold to ~6e-8 (~1e-6 single-step filter), NOT the f64 tier.
• The Rust-f64 / NumPy deep-reader path: the genuine <1e-10 EXACT tier and the cross-language Rust-f64 == Python oracle. Named here only to mark the boundary you must NOT cross when carding L0.
• Method constitution called by name: M13 (one engine, no backprop — AST-guard, world.<attr> whitelist isolation), M10 (exactness honesty — never card a float32 anchor at the f64 tier), M12 (WORLD ⊥ BODY ⊥ MIND — module-exact only at the single-step categorical interface).

---

Method (numbered build steps)

1. Encode the genome as a categorical prior. Represent the inherited blueprint as categorical parameters; seed_zygote instantiates the initial A/B/D Dirichlet tensors from it. Keep the model/process split sacred: the agent owns only the MODEL and couples through a typed Markov blanket; it never reads world hidden state.
2. Model the zygote first-division as an exact discrete Bayes (conjugate) update. Use exact_posterior_discrete from core.py. No gradients: the only update is conjugate count addition.
3. Hold the no-backprop AST-guard live over the whole division step (M13). An AST scan must find no autodiff / optax / torch / grad / backward in the loop; isolation is a world.<attr> whitelist, not a blacklist.
4. Run test_embodiment_ontogeny.py for the 6/6 ontogeny gate. Verify the cell-division identity embedding holds. (Embodiment Rungs 1–2 are GREEN but uncommitted — the figures below are the held gate values, carded at their real class.)
5. Optional control (reusable "is the blueprint doing anything?" experiment). Compare recombine against a marginal-preserving within-family scramble (NOT a pooled permutation): the scramble preserves marginal statistics so any surviving effect is attributable to inherited structure, not to the marginals. Treat any positive dev signal from this as DEV-only, never as a held PASS.

---

Gate (the exact pass condition — EXACT ledger figures, ledger row L0.1)

ontogeny 6/6 GREEN; the first-division posterior matches the closed-form discrete-Bayes value to the float32 tier (~6e-8, ~1e-6 single-step filter); the cell-division identity embedding ≤ 1e-9; and the no-backprop guard never trips.

Exactness honesty (M10, load-bearing). Card this at float32, NEVER at the <1e-10 f64 tier. The JAX core.py host runs float32, so these anchors hold only to ~6e-8 (~1e-6 single-step filter). The genuine <1e-10 EXACT tier lives ONLY in the NumPy / Rust-f64 path. Recorded negative (inline): a genome docstring claiming <1e-10 was caught as an overclaim and corrected to <1e-6 (float32 tier) — the precision-tier line is the one carded here.

Ledger class for L0.1: A (machine-exact anchor), at the float32 tier. Status: proven.

---

Falsifier (operable)

Any one of these falsifies L0.1:

• test_embodiment_ontogeny.py drops below 6/6; OR
• the first-division posterior diverges from the closed-form discrete-Bayes value beyond the float32 tier; OR
• the cell-division identity embedding exceeds 1e-9; OR
• the no-backprop AST-guard trips (any autodiff in the loop).

---

Recorded NEGATIVE(s) (first-class, inline)

• The <1e-10 docstring overclaim (caught and corrected). A genome docstring asserted the f64 EXACT tier (<1e-10) for a float32-host result. This was caught as an overclaim and corrected to <1e-6 (float32 tier). It is recorded here not as a hidden failure but as the load-bearing reason this rung is carded at float32: never card a float32 anchor at the f64 tier (M10). The negative is the precision discipline working as designed.

---

HONEST FENCE — proven (Class A, float32 tier)

A held, sealed value exists: ontogeny 6/6 GREEN, first-division posterior matching the closed-form discrete-Bayes value to the float32 tier (~6e-8, ~1e-6 single-step filter), identity embedding ≤ 1e-9, the no-backprop guard live. Its falsifier (above) is still live.

Not claimed at L0 (the exact not-claimed line): not "we created life", not a "conscious baby", not exact at f64 — it is a float32-tier developmental SIMULATION of a conjugate Bayesian first division. No awareness, no sentience, no person; functional self-awareness is not even in scope at this rung (that vocabulary is reserved for L8, and phenomenal sentience is disclaimed program-wide). "Active inference" here is the framing LENS at textbook level (Parr/Pezzulo/Friston, Active Inference, MIT Press 2022), not a demonstrated AIF loop; the genuine exactness tier is the Rust-f64 path, not this one.

Honest program position (printed, never softened): ~2 of 11+ developmental rungs earned. The whole program is a developmental active-inference SIMULATION — a bounded peek, a toy world, never a person. Where this recipe and the ledger disagree, the ledger wins.