What is wrong, what is predicted to fix it, and what that does and does not buy

Written 2026-08-19. Every figure here was executed, not asserted. Where a number is hand-derived rather than measured it says so. This document exists because a defect that could not be seen survived for months.

0. The goal, stated so it can be checked by someone who does not trust us

"Full biological parity that can ship and be proven — or disproven — by others." That decomposes into four things, and only the first is ours to finish alone:

The honest headline: nothing in this document moves biological parity. All of it is P0/P1 work — computational integrity and equation/implementation. It is the floor that makes a parity claim meaningful, not a parity claim. An engine whose beliefs are provably frozen cannot support any statement about inference, so this work is a precondition, not progress up the ladder.

Biological parity lives in the flagellum programme, and its status is unchanged: P3 as an executed activity, with P4 (transfer) the first unsatisfied rung, X16 FULL_BIOLOGICAL_PARITY = FAIL, and G10/X12 (does a bacterium implement active inference?) NOT_ESTABLISHED with zero discriminating interventions. See docs/EXTERNAL-DOORS-ACQUISITION-CHECKLIST.md in the flagellum tree for what each external door needs.

1. THE ENGINE'S FREE ENERGY WAS NOT A BOUND — REPAIRED b645421

Wrong. The likelihood was raised to the sensory-precision power and never renormalised: F = −ln Σ_s p⁻(s) Π_m A^m[o_m|s]^γ_m. So γ parameterised no probability model, ∂F/∂γ > 0 everywhere, and F → 0 as γ → 0 — the objective the engine minimises paid maximally for going blind. Measured: 50.5% and 60.5% of live factor-cycles violated F ≥ −ln p(o) in two runs with different sense streams.

Repair. Normalise the tempered column, p_γ(o|s) = A[o|s]^γ / Σ_o' A[o'|s]^γ.

Predicted, and verified. Outcome mass Σ_o e^(−F(o)) = exactly 1 at every γ (was 4.19 at γ=0.1 — the engine was manufacturing probability out of nothing). F → ln(n_o) as γ → 0 instead of collapsing to 0. The argmin over γ moves off the clamp, so precision becomes derivable rather than pinned — the precondition for it ever being model-adjustable.

Correction to my own claim. I said this would take violations to zero. It goes 50.5% → 22.3% against the untempered comparator, and no repair could have zeroed that: the normalised tempered model is a different, legitimate model, so that comparison is a crossing point, not a floor. Against its own model the repaired F is a proper bound — 0 of 12 000.

Untested where it matters. Every live factor was degenerate (§2), so this repair has never run on a non-degenerate likelihood. The frozen-factor RED is its first real test and can only VOID it, never confirm it.

2. TWELVE OF TWENTY-ONE FACTORS ARE FROZEN BY SYMMETRY — NEXT REPAIR

Wrong, and it is the largest defect standing. For every exteroceptive factor the likelihood has identical columns, the transitions are the identity, and the prior is uniform. Relabelling the hidden states therefore changes nothing, and every update rule in the engine reads state-indexed values, never a state index — so the whole update map is equivariant under that relabelling. An equivariant map applied to a point invariant under the full permutation group returns a point invariant under the full permutation group.

The belief is uniform at every step, forever, as a matter of group theory. Measured max|qᵢ−qⱼ|: 4.4×10⁻¹⁶ (400 ticks), 3.0×10⁻¹⁵ (2 000 adversarial), 1.8×10⁻¹⁵ (600 live steps) — float dust on an algebraic zero.

What that costs, all measured. The epistemic term is exactly 0 for these factors, forever. Their per-action value spread is 1.1–5.0×10⁻¹⁶, so they contribute nothing to any decision at any planning depth. A constant added to all actions cancels in the softmax and update_policy is shift-invariant, so every strategist preference override and the entire curriculum are behaviourally inert. On the default genome all twelve are frozen, every action ties exactly, and behaviour reduces to softmax(ln E) + a random draw — a Pólya urn. That is the mechanism behind the measured absorbing action-lock (one action 95.2% of ticks, action entropy 0).

Two things this is NOT. It is not the learning rule — under a uniform belief an observation carries information about which outcome, never which state; a rule that broke this symmetry would be manufacturing information provably not in the data. It is not ns == no — measured at ns=5, no=4 the pin still holds at 1.7×10⁻¹⁶. The code comment naming no==ns names where the defect was noticed, not what causes it.

The actual seam. The initialiser sets Â₀ = 1/n_o — the mean of its own symmetric Dirichlet prior, which is precisely the one point in the simplex that every permutation fixes. The defect is choosing the prior's mean as the starting point, not the prior.

Repair, predicted. Sample the prior instead of evaluating its mean: Â₀[:,s] ~ Dir(1·1_{n_o}) per state-column, drawn once, deterministically seeded from the existing SP.Determinism on a committed domain string. E[Â₀] = 1/n_o exactly, so the concentrations are byte-identical in expectation — the prior is not changed by one iota; only its per-agent realisation stops sitting on the symmetric point. No engine kernel changes; the Python oracle is untouched and cross-language parity is immune by construction.

Predicted effects. Beliefs track their own observations; the epistemic term becomes non-zero for the first time; the agent predicts its next observation better on a prequential (predict-then-see) score formed from the forward prior before inference and before learning.

Falsifiers, pre-registered.

• Agitation, not inference — beliefs move but predictive score does not improve beyond the assay's own null half-width. Does not ship.
• Worse — predictive score degrades materially. Then the deliberate choice at genome.ex:91-94 is vindicated for exteroception and must be reported in those words.
• Covenant breach — two different pre-registered draws give materially different results, meaning the specific realisation carries information and a prior was smuggled in. Does not ship.

Knowingly necessary but not sufficient. With identity transitions, q(o) = A·B^u·q is still the same for every action at birth. Action-differentiation waits on the per-action transitions diverging through learning, which is slow. Failing the action criteria is PARTIAL, not a pass — and any write-up claiming "the curriculum now works" on this evidence is overclaiming.

Precedent, and it matters. A signed lab-team review already ordered init_a: :diagonal for two exteroceptive vision factors, calling a stuck-uniform exteroceptive factor "an inert smuggled feature." That order was never fully carried out; its gate is still PENDING. The recommended fix here is more conservative than the one already signed, because a diagonal likelihood asserts that hidden cause k emits codec bin k — a claim about the world — whereas a draw from the unchanged prior asserts nothing.

Operator gate. genome.ex:5-9 and docs/UNIVERSE.md:138-141 state as covenant that the DNA encodes no world knowledge. The fix is built gated off by default, so it changes no existing lineage; enabling it for a scored lineage is the operator's call, not an agent's.

3. THE REST OF THE STANDING DEFECTS

4. THE FLOW — in order, with what each buys

1. §1 free energy — done, committed b645421. Buys: γ names a probability model; precision becomes derivable. Moves P0/P1. Moves no parity rung.
2. §2 frozen factors — designed, RED pre-registered, built gated-off. Buys: the agent's world-facing beliefs respond to the world at all. This is the precondition for every behavioural claim the engine has ever made. Moves P0/P1.
3. §3.9 re-read the record — once §2 lands, every historical behavioural result on those factors must be re-examined. Cheap, and it is the honest debt of §2.
4. §3.4/§3.5 precision — the operator's rule, now implementable on a bounded F.
5. §3.1–3.3 decay — the absorbing lock and the habit ratchet. FE-touching; needs review.
6. Everything above is P0/P1. The parity ladder moves only at docs/EXTERNAL-DOORS-ACQUISITION-CHECKLIST.md: P4 transfer needs an independent commensurate cohort; P5 intervention needs the discriminating wet-lab experiment (raise sensory ambiguity at zero gradient and see whether the motor acts to resolve it — active inference predicts yes, every matched alternative predicts no); P7 needs a lab independent of us and of this software.

5. What "shippable and checkable by others" requires that we can still do here

• Every gate re-runnable from a clean clone, with the command in the receipt.
• Every number traceable to a file:line or to a shown computation.
• Every adverse result retained and stated first — including that a plain lognormal still out-predicts every mechanistic flagellar model on held-out data.
• No claim above its receipt. X16 stays FAIL until X01–X15 all pass; no averaging, no relabelling.
• The instruments audited too. The test named "the upper-bound property holds" never compared F to −ln p(o). A vacuous test is worse than no test, because it certifies the thing it never checked.

Every defect above was found by looking. The ones we have not found yet are in the parts we have not made visible.