What you are building: not a dish and not an ability — the skeleton of the one developmental organism, and the signed method that makes its hierarchy honest. This chapter is the newest work in the book. As of this writing the skeleton is built, sealed, and has passed its ten acceptance runs (A1–A10) — but read the fence carefully: that is an architecture-and-method result. No organism has been born, nothing has learned, and no rung has moved. Where this chapter and the claim ledger disagree, the ledger wins.
You can watch the real skeleton — and the gap between what is built and what is only designed — in the live 3D lab at /lab (private preview). Every number it shows is read from the sealed files described here; there is no black box.
One organism, grown — not one engine, borrowed
The recipes so far reuse one demo engine across scales. DEV-SPINE-0 is a different commitment: it is the first buildable piece of the one hierarchical model (NESTED-ONE, owner decision D99) — a single pure active-inference organism that is grown through development, not assembled from parts. Its first skeleton lives in a dark womb: self-made retinal waves that wire the visual pathway before any light. Concretely it is a 72-site retina, 18 relays, 3 early-visual factors, and 1 homeostat that regulates the waves — 10 of 45 designed factors active, the other 35 registered but dormant (designed, no live state). Per step it carries 92 model units, 214 links, and 36 messages. That is the whole of the living skeleton today; the rest of the 16-layer body is the aim, not the build.
This is not the letter reader. The reader that recognises glyphs is a separate, engineered comparator with a hand-written eye — prior R&D, explicitly outside the organism. DEV-SPINE-0 does not read, and nothing here says it does.
The one hierarchy: temporal ownership (R1)
A hierarchy is only honest if each fact has one owner. The signed rule (C61 TEMPORAL-OWNERSHIP / R1) is: only the root of each active tree carries a transition model — memory of the last step. A factor that has a parent carries no independent temporal prior; its current-state prior is its parent's registered conditional. Inference is one pass up (children hand evidence to parents) and one pass down (parents hand predictions to children) — sum-product on a tree, no more.
The echo, found and repaired
The earlier schedule broke that rule, and the break had a name: the echo (E1). When every input was masked, a non-root factor's own temporal prior climbed the hierarchy as if it were fresh evidence — so the hierarchy could become confident while seeing nothing. This is the oldest fence in the kitchen (the cavity principle, M22: never treat an upstream prior as fresh evidence) failing in a new place. The Architect's audit found it; the science consult confirmed it (C61) and signed the repair; and the sealed acceptance run below now holds it to that repair forever.
The repair is a filtered two-slice statistic, learned only at roots:
ξ_k(s′, s) ∝ L_k(s′) · B̄[s′, s] · q_{k−1}(s), normalised over both indices,
where L_k is the root's upward likelihood message (not q_k divided by its prior), B̄ is the column-normalised transition mean, and no transition count is taken on a factor's first inference (rule B2). A standing check, V41 (NO-INPUT MARGINALIZATION), keeps it honest forever: replace every external likelihood with a state-independent one, marginalise every unobserved descendant, and each root must sit at the exact prior predictive of the generative model — no descendant may create information merely by being latent. V41 is one of the checks that passed in the acceptance run.
Precision, stated exactly
Trust must sharpen evidence without smuggling in a belief about the world. Tempering a categorical likelihood by a precision ω is the column-normalised A^ω (C61 LIKELIHOOD-PRECISION):
ln A_ω(o | s) = ω · ln A(o | s) − ln Σ_{o′} A(o′ | s)^ω,
so trust does not act as a prior over states. The nat is dimensionless and γ = 1 at this gate: precision is asleep in DEV-SPINE-0 (a fixed hyperparameter), to be woken only later. This is the definition; it does not change the proven Class-C affect-as-precision result at L4, which is a legacy-engine finding with its own fence.
What kind of inference this is (the honest label)
DEV-SPINE-0's family is DISCRETE-TREE-BP / ONLINE FILTER / DIRICHLET-MEAN PLUG-IN — and, said plainly, it is not variational message passing over uncertain parameters. It uses the Dirichlet mean Ā = a / Σa as a point likelihood (E_q[ln A] is not used at this gate); it is filtered, not smoothed. Its exactness is fenced precisely: exact state filtering conditional on the registered plug-in parameter matrices, not exact joint Bayesian inference over the Dirichlet parameters. Learning is still count addition, no backprop — "fire together, wire together" is literally A − A0, the learned Dirichlet concentration minus its birth value. Unlike the legacy JAX demo engine (float32, anchors ~6e-8), the skeleton runs float64 (CPython 3.12.10 / NumPy 2.3.5, single thread), with sealed-fixture gaps at the ~1e-12 tier.
Status (the whole point of the fence)
- Signed, in design: the 16-layer charter (C58), the DEV-SPINE-0 architecture (C59), and the echo repair (C61) are all signed. C61 is a SIGNED METHOD REPAIR — and it moves no rung.
- Built and SEALED: a generic runtime and an independent verifier exist; the closure manifest is v1.2.2-sealed — the runtime, verifier, mutation manifest, environment and genome fingerprints (T-SPINE-0, T-SPINE-0-ACT, FX-M13D-1, FX-M13D-1B) are all fixed before A1. "Sealed" means the thing being tested can no longer move underneath the test.
- Acceptance PASSED (A1–A10): on the sealed build, A1–A10 all PASS, with V01–V41 all PASS on the acceptance lineages, and every planted defect of the sealed mutation manifest caught by the check aimed at it (317/317). That is C59's pass rule — "A1–A10 with the checks PASS and every mutant FAIL" — extended to V41 by C61 and closure 1.2.2.
- Not born: there is no UNI-MAIN. DEV-SPINE-0 is an architecture test; a PASS shows the skeleton and its method are correct and honest, not that UNI learns anything, that the homeostat regulates anything in the world, or that any developmental rung has been climbed.
By this book's own vocabulary, a passed acceptance on a sealed architecture is a proven architecture-and-method result — the strongest thing a skeleton can earn before there is an organism to test. It is not a proven ability and not a learning result. It raises no rung.
HONEST FENCE — architecture sealed, acceptance passed; no organism, no ability, no learning. A signed architecture and a signed method repair (C59, C61); a sealed skeleton whose A1–A10 acceptance PASSED (317/317 mutants caught); and still no organism born, no ability demonstrated, no learning result, no rung moved. The developmental organism is a bounded SIMULATION, a dark womb with a first visual pathway — never a person.
Falsifier (first-class): any claim here carded above architecture-and-method — in particular any statement that DEV-SPINE-0 has been born, has learned, has demonstrated an ability, or has moved a rung; any A1–A10 or V01–V41 result that fails to reproduce bit-for-bit on the sealed build (manifest 1.2.2-sealed); or any hierarchy that passes V41 while a latent descendant manufactures confidence from nothing.