S-L0 - Molecular: genome to zygote (float32 tier)
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 Encyclopedia is the UNI method written out as a reference work: 39 pages, arranged in wings, setting out what the programme is attempting and why it is built the way it is. This is where the ideas are explained in order and in prose, rather than as code, as runbooks, or as dated receipts.
Every chapter is authored against two ledgers and never ahead of them. One records what UNI has built, and the evidence class of each claim. The other records nature's own regularities, kept separate on purpose. That way a fact about biology is never quietly reused as a fact about the software. Where a chapter and a ledger disagree, the chapter is the thing that is wrong. Every chapter closes with an invitation to falsify it, and a recorded negative is published beside the result it qualifies rather than after it.
Read "How to read this work" first. It is the evidence constitution: the classes, the four ledger states, and the rule that a finished chapter is not the same as a working system. Then the calibration ledger, which carries the figures every other chapter is required to use.
What it is not: a description of a person or of a mind. The programme calls itself a developmental active-inference simulation, a bounded peek into a toy world, and its own index prints how much of the developmental ladder has actually been earned — roughly two rungs out of eleven or more. It is also not a report of what is running today. For what ran, and when, go to the evidence record.
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.
This is the first rung of the UNI developmental ladder (HUMAN-HGM-001: conception toward a speaking three-year-old, eleven levels plus a global affect modulator). It is a no-backprop, nested-Markov-blanket developmental SIMULATION, never a person and never a mind. The honest position is printed on the spine and repeated here without softening: roughly 2 of 11+ developmental rungs are earned, and this chapter accounts for the molecular end of one of them.
The honest position of this specific entry is narrow and exact. UNI models a zygote's first division as an exact discrete Bayesian (conjugate) update, anchored to its closed-form value at the float32 tier (~6e-8, ~1e-6 for a single-step filter), not at the float64 tier. That is the entire claim. It is not a claim that anything was born, awakened, or made alive. It is a toy world, a bounded peek at one well-understood piece of math wired into a developmental scaffold.
What this rung actually is
The molecular rung implements the conception prior: the inherited "blueprint" that seeds the simulated organism before any development runs. Three primitives carry it. recombine combines two parental genomes; seed_zygote instantiates the first cell's generative model from that genome; and a no-backprop guard enforces that the only learning rule anywhere in the loop is exact conjugate-Dirichlet count addition (counts plus a learning rate times a sufficient statistic), with autodiff, optax, torch, grad, and backward forbidden in the loop and caught by an AST scan. The first cell division is then expressed as a conjugate Bayesian posterior update: given the prior and the categorical evidence, the engine computes the posterior in closed form and checks it against the analytic value.
This is the ledger's clean, reusable "is the inherited blueprint doing anything?" experiment. The control that travels with it is a marginal-preserving within-family scramble, deliberately not a pooled permutation, so the question "does the genome carry signal beyond chance?" is asked honestly rather than against a strawman.
The exact claim, with its anchor and its tier (L0.1)
Ledger row L0.1, evidence Class A (machine-exact anchor), status proven, records: the zygote first-division is modeled as exact discrete Bayes (a conjugate update); recombine and seed_zygote are in place; the no-backprop guard holds; ontogeny 6/6 passes (test_embodiment_ontogeny.py); and Embodiment Rungs 1-2 are GREEN (uncommitted).
The load-bearing calibration is the exactness tier, and it is the heart of this chapter. The JAX core runs float32. A float32 host cannot hold a value to the <1e-10 EXACT tier. So these anchors hold only to ~6e-8 (a ~1e-6 single-step filter), and they are carded at exactly that tier, never higher. The genuine <1e-10 tier exists only in the NumPy / Rust-f64 path (the cross-language bridge that proves Rust-f64 equals Python); it is not where this rung lives.
This is not a hypothetical caution. A genome docstring once claimed <1e-10 for this very path; it was caught as an overclaim and corrected to the float32 tier. That correction is the rule made concrete: never card a float32 anchor at the f64 tier. The cell-division identity embedding is held to a separate, looser bound, exceeding 1e-9 would trip its falsifier, and that bound is also stated at its real strength, not inflated toward the f64 oracle.
There is no held-out capability gate here and none is implied. Class A is a machine-exact anchor and a live correctness check, not a dev-gate eval (Class C) and not evidence of any cognitive ability. An anchor proves the arithmetic matches closed form; it does not prove a capability, and the interpretation is fenced accordingly.
What is NOT claimed in S-L0 - Molecular: genome to zygote (float32 tier)
- Ceiling. That UNI "created life" or produced a "conscious baby" or anything alive, aware, or mind-bearing is NOT shown, and neither is exactness at the float64
<1e-10tier. The most we claim is exactly L0.1: a float32-tier (~6e-8, ~1e-6 single-step) SIMULATION of a conjugate Bayesian first cell division, withrecombine/seed_zygote, a holding no-backprop guard,ontogeny 6/6, and Embodiment Rungs 1-2 GREEN (uncommitted), at Class A, no higher. - Fences engaged. Red line 2 (never consciousness / sentience / aware, no "conscious baby"). Red line 4 (never "created life" / "digital life" as a CLAIM, north-star framing only, posed as an open falsifiable question, never an answer). Red line 7 (never raise a claim above its source class, Class A stays a machine-exact anchor, not a capability or a held-out gate). Red line 10 (no PII, no patent-level math, textbook-level framing only). The standing exactness fence (FM-2, M10): never label a float32 anchor
<1e-10 EXACT. - Negatives that travel with this claim (cite alongside, never strip). The corrected-overclaim record: the
<1e-10genome docstring was an overclaim, caught and down-calibrated to the float32 tier; that correction is content, not an embarrassment to hide. The tier ceiling itself is the standing negative: the<1e-10EXACT tier is unreachable on this float32 path and lives only in the NumPy / Rust-f64 bridge. The marginal-preserving within-family scramble is the control any "the genome is doing something" reading must clear; it is not a pooled permutation. - Parked / owed. No sign-to-park is owed at L0 (this rung is proven, not parked). Embodiment Rungs 1-2 are GREEN but recorded as uncommitted, the GREEN status is a held check, not a committed milestone, and the chapter carries it that way. The developmental rungs above this (narrative-self, conscience, reasoning, and onward) remain parked or not-yet-built; nothing here advances them.
- One-line honest summary a skeptic could not dispute. UNI reproduces the closed-form posterior of a single conjugate Bayesian cell division to the float32 tier (~6e-8) and passes a 6/6 ontogeny test, and that is all this rung claims.
Falsify this
The lead falsifier, stated operably: run test_embodiment_ontogeny.py; if it drops below 6/6, this rung fails. It also falls if the first-division posterior diverges from the closed-form discrete-Bayes value beyond the float32 tier (~6e-8), if the cell-division identity embedding exceeds 1e-9, or if the no-backprop guard trips (any autodiff, optax, torch, grad, or backward appears in the loop). Any one of these overturns L0.1. Note the asymmetry the tier enforces: a divergence at the 1e-9 scale would falsify the f64 EXACT tier but is well within the float32 anchor's stated ~6e-8 budget, which is precisely why the claim is carded at float32 and not above it.
Sources
- Ledger:
CLAIM-LEDGER.md, row L0.1 and the exactness-tier calibration-down note (Section 1, L0); FM-2 (A-U rubric, exactness-tier honesty) and FM-4 (the not-claimed template) inMASTER-PLAN.md; method rows M10 (exactness honesty) and M13 (one engine, no backprop). - Narrative grounding (PII-redacted):
curated/uni-gpt-digest.md(Embodiment Rungs 1-2 GREEN; float32 exactness correction) andcurated/uni-mind-digest.md(the genome / conception-prior primitive:seed_zygote,recombine, the marginal-preserving within-family scramble control; float32 tier vs the NumPy/Rust-f64<1e-10path). - Archive pointers (no PII):
...-UNI-GPTand...-uni-mind.
sha256 35393423f63de040 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This is the first rung of the program's developmental ladder, and its claim is deliberately narrow. The program models a zygote's first division as an exact discrete Bayesian update and checks the result against the closed-form value, at the precision tier its hardware actually supports and no higher. That is the whole claim. It is not a claim that anything was born, awakened or made. It is a toy world, a bounded peek at one well-understood piece of mathematics wired into a developmental scaffold. The chapter's load-bearing point is the precision tier: the core runs at single precision, so its anchors hold only to that tier, and a docstring that once claimed the tighter tier was caught as an overclaim and corrected. Never record a single-precision anchor at the tighter tier.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 35393423f63de040
Clear — written for this website, not the source document
The molecular rung of the developmental ladder implements a conception prior: the inherited blueprint that seeds the simulated organism before any development runs. Three primitives carry it. One combines two parental genomes. One instantiates the first cell's generative model from that genome. The third is a guard enforcing that the only learning rule anywhere in the loop is exact conjugate count addition, with automatic differentiation forbidden and caught by a static scan. The first cell division is then expressed as a conjugate posterior update and checked against the analytic value.
The control that travels with it is a marginal-preserving within-family scramble, deliberately not a pooled permutation, so the question of whether the genome carries signal beyond chance is asked honestly rather than against a strawman.
The recorded row sits at a stated evidence class, the machine-exact anchor, and the calibration on it is the heart of the chapter. The core runs at single precision, and a single-precision host cannot hold a value to the tighter tier, so the anchors are recorded at exactly the tier they reach and never higher. The genuine tighter tier exists only in a different numeric path. This is not a hypothetical caution: a docstring once claimed the tighter bound for this very path, was caught as an overclaim, and was corrected. That correction is carried as content.
The chapter is explicit that no capability was tested here against data set aside for it, and none is implied. An anchor shows the arithmetic matches a closed form; it does not show a capability, and the interpretation is limited accordingly.
Its not-claimed section rules out life, awareness or a mind, and rules out the tighter precision tier. What would show it wrong is operable. Run the ontogeny test and watch for a drop. Check whether the first-division posterior diverges beyond the stated tier, or check the identity embedding against its own looser bound. Or watch for the guard against gradient learning to trip. The chapter also notes the asymmetry the tier enforces, since a divergence that would break the tighter claim sits comfortably inside the budget this rung actually claims.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 35393423f63de040