Before the Replicator · Flow Lenia causal steering v1b Complete · primary chain not met
Frozen prospective result · 1,920 trajectories
Flow Lenia · GARD-inspired information experiment

An encouraging pattern. Not a causal-emergence result.

The maxPhi worlds produced more restoration onsets than minPhi, and two secondary trajectory measures separated cleanly. But the two predeclared primary links did not: the first post-action Phi shift and the fixed-horizon restoration reward both remained uncertain.

So the honest result is: something in the policy changed the compositional dynamics, but we did not show that integrated predictive information predicts or causes restoration.

00 / In 20 seconds

What survived the full test?

The result separates into three clean statements.

Not supported

Natural precursor

No

Past-only Phi did not reliably rise before untouched restoration, and adding Phi did not improve held-out prediction beyond ordinary state controls.

Not supported

Causal steering

No

The primary maxPhi − minPhi Phi and restoration intervals both crossed zero. The required chain did not close.

Supportive only

Trajectory divergence

Yes

Average later Phi and same-cycle restoration separated. Those are real findings, but their conditioning makes them unable to authorize the main claim.

01 / Primary evidence

The two links we needed.

Matched copies shared the same first 25 passages. From passage 26 onward, one policy chose the reservoir parcel with the highest local Phi score and another chose the lowest. Positive estimates favor maxPhi.

Controlled intervention

Same past.
Opposite policy.

maxPhi versus minPhi in the same family × schedule block.

Policy begins at division 25; outcomes begin later.

Link 1 · no

First offline Phi checkpoint

+0.0261
95% interval −0.0186 to +0.0679 · n=294 blocks
−0.040+0.08
Link 2 · no

Fixed-horizon onset reward

+1.91
99% interval −2.28 to +6.64 · n=342 at-risk blocks
−40+8
Bottom line: both point estimates lean in the hoped-for direction, but both uncertainty intervals include no effect. In addition, only 294 of 316 intervention-eligible blocks had both checkpoint-26 Phi measurements. The predeclared causal-steering claim therefore fails three separate requirements.
02 / The tempting pattern

Why the result still looks interesting.

The raw onset counts are directional. They are also not the final inference: schedules and families are clustered, and maxTDMI looks almost as good.

maxPhi − minPhi onset incidence

+5.85 pp

99% interval −3.08 to +16.07 percentage points

The raw difference is substantial, but the family × schedule resampling does not resolve it.

maxPhi − maxTDMI onset reward

−0.50

99% interval −4.28 to +3.25

There is no evidence that targeting Phi beats targeting ordinary whole-system predictability.

Rates above are direct counts from the receipt-bound run stream. The paired intervals—not the bar lengths—govern the frozen claim.

03 / What did resolve

Two positive secondary results.

These tell us the policy was not inert. They do not tell us that Phi caused unconditional restoration onset.

Separated

Mean later excess Phi-R

+0.1325

95% interval +0.1126 to +0.1528 · 384 blocks

Averaged across all available later action-linked checkpoints, maxPhi was clearly higher than minPhi.

Separated

Same-cycle compositional restoration

+0.001380

99% interval +0.001010 to +0.001836 · 372 blocks

On cycles valid in both matched arms, maxPhi recovered slightly more parent composition than minPhi.

Why later Phi cannot rescue the claim

It averages different post-treatment follow-up sets. Before outcomes were opened, amendment v1a demoted this metric to descriptive and replaced it with the exact first post-action checkpoint.

Why same-cycle ΔH cannot rescue the claim

It only uses cycles observed and valid in both arms. That survivor-conditioned comparison cannot stand in for the unconditional onset reward, where every no-onset terminal remains zero.

The useful interpretation: parcel choice has real leverage over later information structure and local compositional recovery. We still do not know whether Phi is the operative ingredient, or whether these policies are exploiting some broader dynamical feature.
04 / Natural prediction

No Levin-style precursor appeared.

In untouched random-policy trajectories, past-only Phi neither tracked imminent restoration nor added held-out prediction beyond ordinary controls.

Prospective questionEstimate95% intervalResult
Does excess Phi predict the next ΔH?−0.000465−0.001342 to +0.000378No
Does Phi rise just before first onset?+0.003201−0.102420 to +0.126505No
Does adding Phi improve held-out onset prediction across families?−0.003619−0.014620 to +0.006125No
Does adding Phi improve held-out onset prediction across schedules?−0.004533−0.018236 to +0.008326No

For the two prediction rows, positive loss improvement would mean that adding Phi helped. Both point estimates are slightly negative. All 12 families and all 32 schedules were represented in the onset-prediction target set.

05 / What was measured

The experiment in normal words.

This is about restoration of a four-founder mixture after external passage—not a creature splitting itself.

qP

Parent

The four-founder composition just before passage.

qD

Daughter

The composition just after external passage disrupts it.

qR

Return

The composition when externally fed growth reaches the next threshold.

ΔH = similarity(qP, qR) − similarity(qP, qD)

What Phi-R asks

How much next-state prediction is lost when the four-founder composition is split into parts? Higher excess Phi-R means more predictive structure belongs to the whole mixture beyond shuffled-future fits.

What onset asks

After not being restored at passage 25, does the lineage later achieve a sustained, predeclared run of composition-restoration episodes before the campaign ends?

The colored bars are explanatory cartoons, not campaign frames or data.

06 / Cohort geometry

Same campaign, different panels.

The restoration and manipulation questions use declared but non-identical subsets. Their conjunction is not a mediation test.

384 blocks
12 vetted families × 32 fresh schedules; five arms per block = 1,920 trajectories.
342 at risk
Used for incident onset and fixed-horizon reward. The other 42 were already restored at passage 25.
316 intervention-eligible
Both policies genuinely applied finite-score extrema at the first active action.
294 jointly measured
Both checkpoint-26 Phi values available; all 12 families and 28 of 32 schedules represented.

1,708

trajectories reached passage 100

212

ended by capacity exhaustion

0

terminal extinctions or step-cap terminals

07 / Claim ladder

Exactly how far we can climb.

A yellow rung is useful evidence. It is not a substitute for a failed red rung.

Policy leverage

The five reservoir-choice rules produced different later Phi and same-cycle restoration dynamics.

Supportive

Natural association

Past-only Phi did not reliably track later restoration or onset in untouched random arms.

Not supported

Generalizing precursor

First-quarter Phi did not add held-out onset prediction beyond ordinary controls across unseen families and schedules.

Not supported

First-step manipulation

The fixed checkpoint-26 Phi contrast was positive but unresolved, with incomplete joint coverage.

Not supported

Causal steering

The required conjunction of first-step Phi separation and fixed-horizon restoration advantage failed.

Not supported

Phi-specific steering

maxPhi did not beat maxTDMI on restoration reward.

Not supported
08 / Boundary

What this result actually says.

We can say

  • The 1,920-run campaign completed and the receipt chain verifies.
  • Reservoir parcel choice has measurable leverage over these externally maintained compositional lineages.
  • The maxPhi direction produced an interesting onset-count pattern worth a new, separately frozen hypothesis.
  • The strongest resolved effects were later average Phi and survivor-conditioned same-cycle ΔH.

We cannot say

  • That Phi naturally foreshadows restoration in Flow Lenia.
  • That raising Phi caused unconditional restoration onset.
  • That the effect is specific to Phi rather than general predictability or another policy-correlated feature.
  • That this is autonomous self-replication: feeding and passage are supplied from outside.
  • That Phi-R is a universal or primitive cause of causal emergence.
The clean next question: can a fresh cohort, designed around the resolved long-run policy divergence but with a common unconditional observation schedule, distinguish Phi-specific steering from general predictive control? This page does not answer that question retroactively.
09 / Receipt

Frozen outputs, directly linked.

The HTML is an explanation. The JSON artifacts are the scientific source of truth.

Full analysis reportSHA-256 eab15f3c894edf4c8865f8ebc22f7ee84a9052b630d1376c593ae29967ca1815
Publication receiptSHA-256 06d19ac1e2f28bd00999531aad5f5c8c8946c99ec0f01a1d94e3afa1e8b32bd0
Raw input manifestSHA-256 8b6e98d77ae847077b7ce587865aec2294273dc9cd9d8aaeae0399b6b960cfa5
Receipt-bound record streamSHA-256 97040f3859a6b258dc5df804eb278e4b17f074987b260e2b35d7991e1e44beab
Outcome-blind v1a estimand amendmentSHA-256 458a3d4ff1d5747424070807b2408be2a67af4c5413d4ed7ee7d3f4c53fd3362
Mechanical analysis-opening amendmentSHA-256 386197a801936cb0d5a2c40e428fb1a9f4f8e94dfee3462c60bb007cb2fff3ff
Exact provenance bindings
result receipt schema: flowlenia_reservoir_causal_steering_analysis_receipt_v1b
result status: immutable_append_only_v1b_analysis_publication
raw scientific input tree: fbbb3ea968eaafac15649e93ae0ecf7aa0f41e8371c5f9b99eda2a0348bd67a6
scientific input leaves: 13,440 across 1,920 run bundles
closed cohort audit: c7be9bb85c668cf0d964e8a2b4cc5eaf05b5155c6447aa0780d992fa8db5570e
mechanical erratum verifier: 7cbcb2433c9169f8699ed0d76e3f21345c883d1d0723382b916db80b6bbda73a
v1b analysis runner: 01b58e92b9b96147fe2762bcfdd90be1e2c999744f75bb336b4e4af75f00df95
frozen protocol: fd1c08b354c3ebc720ba77917f171b6b74a1908bedb6ab0a6061619b023703df
frozen declaration: 4af657488551bc0678e993a1d7b8c9a59cf6442241b023a2fca43d07f3895fa0
frozen manifest: 2ac8135ea233056676d03cb259c23b28280ccac3d8b4b5cdb9ed7507ab1b8dde

All displayed numbers were copied or directly counted from the immutable v1b report and receipt-bound record stream. Output hashes, amendment hashes, the closed audit hash, and the 13,440-leaf scientific-input tree binding were independently recomputed before this page was populated.