Frozen discovery pilot · 960 trajectories

GARD-inspired causal-emergence test

We moved Φ-R. It did not move restoration.

For 16 passages, we deliberately pushed a Lenia version of integrated predictive information up or down. Then we stopped and watched. The score moved—but later compositional restoration did not.

That is useful state-steering evidence. It is not evidence that causal emergence caused a replicator-like regime.

16 × 12organic founder families × independent supply schedules

5matched policy arms in every block

960completed Flow Lenia trajectories

119blocks still eligible for a new restoration after q25

01 / Setup

The test, without the jargon.

Every matched block began identically. Only the temporary founder-supply policy changed.

q25

Same starting ecology

All five arms shared the exact same history through passage 25.

q26–41

Temporary push

Choose supplied founder material to raise or lower Φ-R or ordinary predictability.

Φ+Φ−TDMI+TDMI−random

q42–100

Hands off

Return every arm to ordinary random supply. Ask whether restoration appears earlier or more often.

Whole TDMI

How much the ecology’s current four-founder composition predicts its next composition.

Excess Φ-R

The integrated part of that prediction, above a shuffled-time null. This is our explicit Lenia adaptation—not Levin’s undisclosed estimator.

Restoration reward

A fixed four-cycle compositional-restoration test after release. Earlier first restoration scores higher; no restoration scores zero.

02 / Result

One link worked. Three did not.

The strong result needed the whole chain, not one interesting number.

Yes

Φ+ raised excess Φ-R

+0.188

At q41, Φ+ exceeded Φ−. The frozen 95% lower bound was +0.073.

The change was not isolated

−0.035 TDMI

The point was near zero, but its interval was far too wide to show TDMI stayed fixed.

Restoration did not improve

−0.261

Φ+ had slightly lower—not higher—mean restoration reward than Φ−. The interval crossed zero widely.

No Φ-specific advantage

−0.437

The Φ contrast was no better than the matched TDMI contrast. Again, the interval crossed zero widely.

In plain English: the intervention could steer the information score for a while. We did not show that it steered only that score, and the organisms did not restore their composition better afterward.

Important support caveat: the q41 Φ contrast used 112 of 192 matched blocks. Score availability was asymmetric at q41 (166 Φ+ arms versus 119 Φ− arms), as were typed uniform fallbacks during the pulse (367 versus 757 events). Missing scores were not filled with zeros, so the +0.188 result applies only to the matched measurable support.

03 / Signal

The Φ-R rise was real—and temporary.

The line is Φ+ minus Φ−. Positive means the Φ+ policy produced more excess Φ-R.

Excess Φ-R across the fixed checkpoints

The policy built a clear separation by the end of the pulse. One passage after release, that separation was gone. This is a controlled, predeclared checkpoint curve; it is not a hand-picked peak.

Could we isolate Φ-R?

The first row moved as intended. The second row had to stay tightly inside the yellow “same enough” zone. It did not.

Excess Φ-R · Φ+ − Φ−+0.188 · [+0.073, +0.313]

−0.350+0.35

Whole TDMI · Φ+ − Φ−−0.035 · [−0.286, +0.240]

−0.350+0.35

frozen bootstrap intervalrequired equivalence zone

Did restoration improve?

Both frozen behavior contrasts pointed slightly the wrong way and were compatible with substantial effects in either direction.

Restoration reward · Φ+ − Φ−−0.261 · [−5.826, +5.319]

−80+8

Φ specificity · difference in differences−0.437 · [−7.859, +6.820]

−80+8

04 / Behavior

The five arms restored at nearly the same rate.

Of 192 matched blocks, 119 were still at risk for a new restoration at q25. The other 73 were excluded by the frozen rule because restoration was already prevalent before the intervention.

Φ+

44 / 11915.50 reward

Φ−

43 / 11915.76 reward

TDMI+

44 / 11915.75 reward

TDMI−

45 / 11915.57 reward

random

45 / 11915.50 reward

Φ+ versus Φ− incidence differed by one event: +0.84 percentage points. Its frozen interval crossed zero. Persistence summaries were survivor-conditioned and descriptive only; they were never allowed to rescue the primary result.

05 / Meaning

What we learned—and what we did not.

This was a prospective discovery test with frozen endpoints. Its negative chain is a result, not a broken run.

We can say

The action policy can temporarily steer excess Φ-R in this Flow Lenia ecology.
The q41 rise was followed by immediate decay after release.
This particular intervention did not improve the frozen post-release restoration endpoint.

We cannot say

That Φ-R was changed independently of ordinary whole-system predictability.
That causal emergence induced restoration, reproduction, or a replicator.
That this rules out every information-theoretic route to reproduction in Lenia.

The clean next move

Do not simply scale this exact policy. First make the intervention genuinely orthogonal: show on outcome-blind calibration data that Φ+ versus Φ− changes excess Φ-R while whole TDMI stays inside the frozen “same enough” band—and that the mirror TDMI policy works too. Only then use a fresh, disjoint cohort to ask again whether the isolated information change leaves a restoration effect after release.

Frozen analysis receipt

status: complete_discovery_only_first_result_analysis
result label: coherent_discovery_chain_not_met
claim class: discovery_first_result
confirmatory paper proof: false
design: 16 families × 12 schedules × 5 arms = 960 trajectories
inference: paired two-way family-by-schedule bootstrap · 2,000 draws
analysis plan SHA-256: 720a9150bc3334cf43d6de17f832b4d91c31762f1d15b313096f4117dc6ff888
analyzed runs SHA-256: 0bdca6407c2119b04b5952d0f516cc069593a9f9ed871df7bd090617c9ebc203
source report SHA-256: 3595116668cc1626c75792568eff4412d7cdd45322870c60e466a19378e0b047