The organism hardens without narrowing.
We froze four predeclared experiments to pin down what commitment means in a developing Flow Lenia creature: a fresh-cohort confirmation of the fine-shape result, a steerability profile across developmental age, a basin-width measurement under initial-condition variation, and a history-disambiguation curve. The push stopped working on schedule, the sibling fan kept opening, the population of look-alikes emptied out, and the headline we came to confirm did not confirm.
Four plans, frozen first, opened once.
Every experiment below wrote its plan.json before a single outcome run existed, in the campaign's write-once style with SHA-256 bindings between code, tests, and upstream artifacts, and each primary is a single frozen rule that secondaries cannot rescue. Nothing fresh remained on either allocation axis of the compact sweep, so the plans say so out loud: the steerability, basin, and disambiguation experiments reuse the f33-f48 solo block with the reuse disclosed in every plan, and the confirmation experiment generated genuinely fresh material by extending the original sweep with 512 new seeds under byte-identical search parameters.
| Experiment | Frozen primary | Result | Verdict |
|---|---|---|---|
| fresh-cohort-fine-shape-confirmation-v1 | Φ-forward minus Φ-backward fine-shape distance over passages 6–12, family bootstrap CI entirely below zero | −0.0019 [−0.0076, +0.0038] | not confirmed, interval straddles zero |
| steerability-age-profile-v1 | pre-turn minus post-turn push divergence positive, turn computed from the past alone | +0.0038 [+0.0018, +0.0054] | met |
| basin-width-init-variation-v1 | early basin width minus late basin width positive | −0.0378 [−0.0436, −0.0329] | wrong direction, decisively |
| history-disambiguation-curve-v1 | matched-pair future-divergence ratio rises with matching age | +0.0066 [−0.0257, +0.0380] | null; look-alike count secondary met at ρ = −0.988 |
The fine-shape compression did not confirm.
The synthesis report's own frozen protocol demanded a confirmation run: a fresh unseen cohort, an endogenous predeclared peak rule, five matched arms with the same local dose and the same future seeds, and one sole primary, the translation-normalized pairwise fine-morphology distance over future passages 6 to 12. We executed it literally. The fresh cohort came from 512 new sweep seeds under the original search parameters, filtered by the same three mechanical gates the original roster used, ordered by a fresh authority-bound hash, and formed into 12 quartet families whose composite arenas, growth-division machinery, past-only structural peaks, probe scoring, and morphology extraction all reuse the campaign's frozen code by import.
Cyan whiskers mark the predeclared passage 6 to 12 window; grey whiskers the earlier passages. The dashed line is the original 19-family point estimate. Point estimates go negative exactly where the original result said they would, and every interval straddles zero.
Sole primary, Φ-forward minus Φ-backward, mean fine-shape pair distance over passages 6 to 12, 95% CI [−0.0076, +0.0038], 12 families split six negative and six positive. The frozen criterion required the whole interval below zero. predictionPassed = false.
Forward minus blind −0.0009 [−0.0069, +0.0050], forward minus orthogonal −0.0023 [−0.0082, +0.0041], forward minus untouched −0.0005 [−0.0048, +0.0043], and the effective-shape estimator repeats the same pattern. No arm distinguishes itself from any other.
What this run rules out is any large effect: compressions stronger than about −0.008 per pair distance are outside the interval. What it cannot do is distinguish the original −0.0040 from zero, because twelve fresh families buy an interval half-width of about 0.006, so a true effect of the original size would fail this criterion more often than not. The late-passage shape of the profile matches the original finding while its magnitude, at half the original point estimate, does not clear it. The fine-shape compression now stands as an unconfirmed hypothesis with a direction-consistent replication attempt on record, not as a result.
Disclosures
The fresh cohort's appearance review was not repeated by a human: the plan replaces the appearance rubric with the three mechanical gates and records appearanceRubricReplacedByMechanicalGates: true, scientificOutcomesInspected: false.
One post-freeze mechanical fix: analysis.py read the passage key under the wrong name and died on a KeyError before computing any value; the key name was corrected after freeze, with frozen and corrected SHAs recorded in generated/post-freeze-fixes.json. No metric, window, contrast, or bootstrap rule changed.
The original −0.0040 was a predeclared secondary with a post-hoc-timed peak arm on 19 families; this run promoted it to a sole primary on fresh material, which is the harder test the protocol demanded.
The push stops working on schedule.
Steerability is the interventional face of commitment: take 64 organisms, push each one with the same local swap at ten ages spanning development, four action seeds per age, and measure how far the pushed future drifts from the untouched twin 120 steps later using the same translation-normalized shape distance throughout. The frozen prediction was that divergence is lower after the organism's structural turn than before it, with the turn computed from the untouched baseline alone as the earliest strict interior local peak of the windowed whole-minus-sum information score, declared before any outcome opened.
The amber curve is the mean divergence produced by an identical push at each age. The violet bars count where each organism's past-only structural turn landed: 45 of 60 turned at age 140, right where the curve begins to fall.
Family mean of pre-turn minus post-turn divergence, 95% CI [+0.0018, +0.0054], 16 families, 60 of 64 organisms with a turn, 4 excluded by the frozen no-peak rule. predictionPassed = true.
Family-mean Spearman correlation of divergence with age, 95% CI [−0.713, −0.562]. The same push loses about two thirds of its effect across development, falling from 0.0088 at age 60 to 0.0027 at age 780.
This is the cleanest positive of the program, and it is the interventional definition of commitment the campaign has been circling: the organism's future becomes progressively harder to move with a fixed dose, the loss of movability is aligned to a turn computable from the organism's own past, and no future information enters the turn rule.
The fan keeps opening while the grip tightens.
If commitment were convergence, siblings should agree more and more. We measured that directly: eight fresh initial-condition seeds per genotype, 512 developments with no interventions at all, and basin width defined as the mean pairwise shape distance among the eight siblings at each age. The frozen prediction said the basin contracts with age. It does the opposite, monotonically, with the entire bootstrap mass on the wrong side. The disambiguation experiment then asked whether a present state pins down a future less as history accumulates: its ratio curve stayed flat, but its predeclared secondary resolved sharply, because the number of look-alike pairs under a fixed matching radius collapses from 6,144 at age 60 to about 1,650 by age 700.
Width grows from 0.032 to 0.079 across development. The frozen primary predicted the opposite sign and failed with CI [−0.0436, −0.0329], all 20,000 draws negative territory.
Cross-organism pairs whose coarse states match within the frozen radius: 6,144 at age 60, about 1,650 by age 700, Spearman ρ = −0.988, predeclared secondary met.
The predicted coupling between late basin width and turn age, 95% CI [−0.289, +0.242], 60 organisms. Null: how early an organism turns does not predict how wide its sibling fan ends up.
The matched-pair future-divergence ratio across all ten ages. Flat: pairs matched on present coarse state diverge at the same relative rate whether matched young or old, so the primary stayed null while the pool of matchable pairs emptied out.
Read together, the two misses carry the program's real finding. Development in this substrate does not funnel siblings toward a shared template, it walks each one into a region of shape space increasingly its own, which is why look-alikes become scarce with age while the width of the sibling fan grows. Commitment, measured interventionally in the steerability profile, coexists with divergence measured observationally here: each organism becomes rigid about its own course precisely while the courses spread apart.
Commitment is individuation plus rigidity, not convergence.
The word commitment suggested a narrowing, a possibility space closing around a target form, and the basin experiment was designed to watch the narrowing happen. What the four verdicts support instead is a two-part statement: after a turn that the organism's own past announces, an equal push moves the future less and less, and at the same time nothing about the population converges, since siblings spread and look-alikes vanish. The organism does not settle into a mold. It becomes a specific individual that holds its own course against the same perturbation that once redirected it.
What is causally supported
Push-induced future divergence falls after a past-only structural turn, as a frozen primary with the interval clear of zero, on 60 organisms across 16 families.
An identical dose loses about two thirds of its effect across development, monotonically by rank.
The look-alike population under a fixed matching radius collapses with age, as a predeclared secondary at ρ = −0.988.
Basin width grows with age. The prediction failed, and the failure is itself a decisive measurement with every bootstrap draw on one side.
What is not yet settled
The fine-shape compression by Φ-forward pushes. The fresh-cohort sole primary straddles zero, direction-consistent at half the original size, and no arm separates from blind or orthogonal controls.
Any link between turn age and basin width. The predicted coupling came back null.
Whether a matched present state pins the future less as history accumulates. The ratio curve is flat where the count curve is steep, and the two disagree about what identity means here.
Whether Φ-selected pushes differ from arbitrary equal-dose pushes at all. Five matched arms produced zero separations.
Stop here and think
A creature that cannot be pushed off course, whose siblings do not converge, and whose look-alikes disappear is not narrowing toward a template. It is becoming harder to argue with about a future only it is heading into. That is what commitment turned out to mean when we measured it four ways at once.
Large enough to see the profile. Frozen enough to trust the misses.
The confirmation experiment reused the campaign's frozen machinery by import: the composite arena and growth-division operators, the 65-state past-only peak selector with eligible checkpoints 18 to 28, the four-step probe scoring under the SHA-pinned Gaussian information model, and the zero-padded D4 translation distance with its bandwidth-free effective-shape estimator. The steerability, basin, and disambiguation experiments ported the same shape distance to the 900-step solo engine at stride 4 and declared every age grid, window, threshold quantile, and drop rule in their plans. Verdicts live in each experiment's results.json beside the plan that predicted them.