Before the Replicator · GARD-inspired Flow Lenia study
The behavior appeared.
The signal test did not.
We found genuine episodes where a four-founder mixture returned after repeated passage and regrowth. But this census did not produce the fair positive-versus-negative comparison needed to ask whether an early whole-system information signal foretold them.
That distinction matters: Φ did not fail. It was never calculated on this cohort.
01 / The path
How this follows the Levin paper.
The paper’s striking idea is not merely that replicators exist. It is that a whole-system information quantity changes early enough to predict which chemical systems will later settle into self-replicating compositions.
GARD observation
Integrated predictive information tracks the approach to recurring chemical compositions.
Lenia state
Represent the ecology by the relative amounts of four tracked founder lineages.
Lenia event
Ask whether that mixture returns across repeated growth-and-passage cycles.
Outcome gate
First obtain repeated future positives and negatives inside the same families.
Early signal
Only then compare prefix-only Φ with mass, growth, motion, fragmentation, and other controls.
This keeps information that may exist only in the coordinated whole.
minus
The proposed Φ-like score is the extra next-state information left in the whole.
Important: this would be an integrated predictive-information measure inspired by the paper. By itself it would be an observational precursor—not proof that “causal emergence” causes recurrence. That stronger claim eventually needs an intervention that deliberately moves Φ.
02 / The event
What “comes back” means here.
This is not ordinary Lenia fission. We impose a GARD-like cycle on the ecology and follow composition, not silhouette.
Parent mixture
The relative amounts of the four founder lineages immediately before passage.
→
Passaged daughter
A finite-molecule partition removes roughly half the matter and can disrupt the mixture.
→
Threshold return
A small external founder parcel is added; Flow Lenia regrows until the passage threshold is crossed again.
One window = four cycles
A single lucky recovery is not enough. We slide a four-cycle window through every run and ask whether the mixture repeatedly returns.
A window passes only if…
- the passage meaningfully disturbed the mixture;
- the return stayed close to the parent every time;
- the median return was at least 0.95 similar;
- at least three of four cycles improved over the daughter.
PREFIX · steps 0–128
FUTURE · steps 129–512
step 0frozen cutoff: 128step 512
Strict positive: an informative failure in the prefix, followed by a passing window wholly in the future. Strict negative: an informative prefix failure, no later pass, and at least two informative future failures. An early pass, a cutoff-crossing pass, or too little evidence stays unlabelled.
03 / Mechanics
One real passage cycle.
This is the first cycle of the first fixed family/schedule pair—chosen before outcomes, not selected for appearance.
Parentbefore passage
F1 26%F2 32%F3 22%F4 20%
→
Daughterimmediately after finite-molecule passage
F1 26%F2 36%F3 18%F4 20%
→
Returnafter reservoir-fed regrowth to threshold
F1 32%F2 28%F3 13%F4 27%
What you are seeing: total-matter fields from c01 / p31001, independently centered and peak-normalized. Brightness therefore shows shape, not total mass; the printed mass values show the loss and regrowth. The colored bars show the four founder fractions. The external reservoir makes this reservoir-fed compositional recurrence, not autonomous self-reproduction.
04 / Accounting
The behavior was real—but labels were scarce.
We retained every trajectory and every frozen rule. Passing windows occurred in 22 runs. The much stricter prefix-versus-future classifier produced only eight binary outcomes.
completed cycles
2,781
four-cycle windows
2,565
compositionally valid
2,027
meaningfully challenged
262
passing recurrence windows
56
LR · later recurrenceNR · no later recurrencePR · prefix recurrenceCA · cutoff ambiguousIP / IF · insufficient evidence
Do not read the grey area as failure. Most runs were unlabelled, not negative. A run could be excluded because recurrence appeared too early, crossed the cutoff, or lacked enough informative evidence before or after it. Nine runs had at least one future passing window; only four satisfied every strict positive rule.
05 / Every run
The complete 12 × 6 outcome map.
Rows are biological families. Columns are common random passage/reservoir schedules. Nothing is hidden.
| family | p31001 | p31002 | p31003 | p31004 | p31005 | p31006 |
|---|---|---|---|---|---|---|
| c01 | IP | IP | NR | IF | IF | IF |
| c02 | IF | IF | NR | IF | IF | IF |
| c03 | IP | IP | LR | IF | IP | IF |
| c04 | IF | IP | CA | IF | PR | IF |
| c05 | IP | IP | PR | NR | IF | IP |
| c06 | IP | IP | LR | PR | IF | IP |
| c07 | IP | IP | PR | PR | IF | NR |
| c08 | IP | IP | IF | PR | IF | IP |
| c09 | IF | PR | LR | PR | CA | IP |
| c10 | IF | CA | LR | IF | IF | PR |
| c11 | IP | IF | PR | IF | PR | IP |
| c12 | IP | IP | CA | PR | IF | IP |
LR · strict positiveNR · strict negativePR · happened too earlyCA · crossed cutoffIP / IF · not enough evidence
The pattern to notice: all four strict positives occurred under common schedule p31003. Two strict negatives also occurred there, but each family contributed at most one binary label. A positive-versus-negative Φ comparison could therefore mistake family or schedule structure for an early warning signal.
Four strict positives
- c03 · p31003first pass 377–425
- c06 · p31003first pass 408–456
- c09 · p31003first pass 382–429
- c10 · p31003first pass 380–426
Four strict negatives
- c01 · p310032 future informative failures
- c02 · p310034 future informative failures
- c05 · p310043 future informative failures
- c07 · p310062 future informative failures
06 / The gate
Why Φ stayed closed.
The frozen rule deliberately demanded variation inside families. Without that, “prediction” can be nothing more than recognizing which family or schedule a run came from.
Each with at least 2 later-recurrence and 2 no-later-recurrence trajectories.
vs.
No family contained even one example of both strict classes. Selected families: none.
+ 0− 1
need +2 and −2+ 0− 1
need +2 and −2+ 1− 0
need +2 and −2+ 0− 0
need +2 and −2+ 0− 1
need +2 and −2+ 1− 0
need +2 and −2+ 0− 1
need +2 and −2+ 0− 0
need +2 and −2+ 1− 0
need +2 and −2+ 1− 0
need +2 and −2+ 0− 0
need +2 and −2+ 0− 0
need +2 and −2Predictive-information phase · unopened
No Φ value exists for this census.
Had the gate passed, the next stage would have used only the prefix to estimate how much the whole four-founder state predicts its next state beyond its best split into parts.
ΦR = information in the whole − information in the best-separated parts
We would then ask whether ΦR predicts later recurrence beyond mass, recent growth, fragmentation, motion, component count, and other plain controls. That test was not run, because its outcome comparison was not identified.
07 / Paper ↔ Lenia
Same research track. Different ontology.
The analogy is useful only if we keep the differences visible.
| Levin / GARD | Our Flow Lenia translation | Important limit |
|---|---|---|
| Molecular composition | Relative mass of four tracked founder lineages, q(t) | We ignore most spatial morphology in this first state description. |
| Assembly grows and fissions | Finite-molecule passage, external founder reservoir, then Flow Lenia regrowth | The division and feeding are imposed operators, not autonomous behavior. |
| Recurring compositional attractor | Four-cycle return of founder composition after meaningful disruption | We call it recurrence, not a self-replicating creature. |
| Early integrated-information trajectory | Planned prefix-only whole-versus-parts predictive information | Not computed until a fair within-family outcome contrast exists. |
| Φ-directed interventions | Later: matched interventions that raise or lower the proposed signal | Needed before making a genuinely causal claim. |
Paper followed: Pigozzi & Levin, “Causal Architecture Dynamics Prior to Arrival of Self-replicators…”. Their implementation leaves several estimator details unspecified; our eventual Lenia estimator must therefore be fully declared and frozen rather than presented as an exact software replication.
08 / Meaning
What we learned—and what we did not.
This is a useful intermediate result: the phenotype exists, while the current sampling design does not identify its precursor.
We can say
- The reservoir passage operator worked mechanically in all 72 trajectories.
- Reservoir-fed founder-composition recurrence genuinely occurred: 56 passing windows across 22 runs.
- Four runs showed a strict transition from prefix failure to future recurrence.
- This cohort cannot support a clean whole-versus-parts precursor comparison.
We cannot say
- Integrated predictive information rose before recurrence.
- Integrated predictive information failed to predict recurrence.
- Flow Lenia has—or lacks—causal emergence.
- These systems autonomously reproduce, or that p31003 caused recurrence.
The clean next move: fix the contrast, not the answer.
- Use this census only to design a new, prospectively frozen outcome-yield study.
- Run more fresh schedules per family and a longer pre/future observation horizon, while keeping the recurrence thresholds unchanged for the first extension.
- Require repeated positives and negatives inside at least three families. Do not select only p31003, relabel grey runs, or loosen thresholds.
- Then freeze a separate fresh confirmation cohort and open the prefix-only Φ-versus-controls analysis exactly once.
- If a precursor survives confirmation, perturb it with no-op, random, mass-matched, and spatially matched controls before using causal language.
09 / Receipts
Why trust this stopping decision?
The mechanics, outcome pair, and arithmetic amendment were independently reconstructed and cryptographically bound.
72 / 72mechanically clean, complete trajectories
2,803division events with exact identity-mass checks
0identity-mass mismatches
closedpredictive information, causal emergence, and Φ
The arithmetic amendment, in plain English
The first analysis attempt stopped before producing any outcome labels. It had compared two totals accumulated in different—but both legitimate—Float32 orders: one global sum and four founder-bucket sums. Seven late events in one trajectory differed only because floating-point addition is order-sensitive.
The stage-local amendment did not widen a tolerance. It required the incoming founder identity vector to bit-match the recorded event, reconstructed the grouped mass through the exact same arithmetic path, and authorized only the already identified events. All 2,803 division events passed before outcomes were opened. No simulation state, recurrence threshold, history window, episode rule, label, or frozen input changed.
Independent outcome reconstruction
The saved mechanical audit was replayed from the raw artifacts. Every cycle, four-cycle episode, status, family count, and final decision reconstructed exactly: four later-recurrence labels, four no-later-recurrence labels, zero qualifying families, and stop_without_phi.
Final claim boundary: Outcome-only census of founder-composition recurrence across externally imposed growth-passage cycles; not autonomous reproduction, causal emergence, or Phi.
Bound artifact hashes
| Stream records | b4a291bde013b86b3e2bb0e919ef35836422d3f263172c45aba641a0a6aa47d3 |
|---|---|
| Outcome report | 46e6def179715ad72cbb4ed5ef39d411278b4b0327d157a0261da93f3c9f6d4c |
| Analysis amendment | b6471a1f73d95b1206eea4199ad237a2ab73d47f6009df81fe63234a94466da1 |
| Mechanical audit | ef2e2fdd5bd4bd6d253dbe69401e738e97c139c1f08cf108180e70dc85ba0eb8 |
| Freeze | 6fdbeef181fc939c1597f1beba8b625c0d4945a99a53ac3af815d41afe70d9a0 |
| Manifest | a2dd6760eeff20207898056ece8dfd09d410ece00308a1157edd240467d03685 |
This standalone page is explanatory output generated after the sealed result. It is not itself part of the prospective freeze.