Something changes early
A whole-system information measure or causal decomposition reorganizes before the later phenotype.
We looked for the moment a system stops merely changing and starts correcting deviations toward a future of its own.
Replication may not be the beginning of organismality. A system may first acquire a higher-level causal organization—a primitive “self”—and only later become able to copy it.
The radical claim in Pigozzi and Levin’s paper is not simply that chemical replicators appear. It is that the causal architecture of the whole system may reorganize before recognisable self-replication arrives. If a macroscopic organization starts carrying information about, constraining, or steering its microscopic parts, then the future organism is already casting a causal shadow.
This matters because it turns “when does life begin?” into a measurable dynamical question. Do we see a new level of control before we see a copier? Can that reorganization predict which histories will later become canalized, self-maintaining, or replicative?
The ambitious possibility is a goal before a body: a target state that already governs change before there is an obvious organism to own it.
That is the connection to biological regeneration. Planarian tissues and Hydra are compelling not because they are static shapes, but because they can act across many local routes to rebuild a species-typical whole. In operational terms, teleology means error correction toward a future state: disturb the present, then ask whether different damaged histories converge on the same destination.
A whole-system information measure or causal decomposition reorganizes before the later phenotype.
Changing that organization changes the future, beyond matched physical disturbance and shared random futures.
Distinct injuries are driven back toward a common macroscopic state: active equifinality, not mere persistence.
Question: can founder composition recover after injury? Early confirmation did not establish a recovery signal. That was a failure of this phenotype and assay—not a verdict that Flow Lenia lacks causal emergence.
Result: recurrence was too sparse and uneven to support the intended causal-emergence comparison. We stopped before manufacturing a Φ story from weak base rates.
Result: across 1,660 estimable valid episodes, there was no broad scalar-integration precursor. Yet every one of the three measurable core onsets rose beforehand. That gave us a narrow timing hypothesis, not a general law.
Result: 960 trajectories showed that the Φ-like quantity could move under perturbation, but the motion was not isolated from total dynamical information and did not buy recovery. At the key checkpoint, Φ-R moved +0.188 [ +0.073, +0.313 ], while the restoration reward for Φ-positive versus Φ-negative families was −0.261.
Result: 1,152 trajectories across 144 branching groups. Adding Φ-shape made held-out prediction slightly worse: baseline MAE 0.0206824; augmented MAE 0.0216150; gain −0.00093256. Only one of four sealed families improved. But 25 groups crossed the Φ estimator’s validity boundary. That failure pointed away from a smooth rise and toward a structural event.
Stop watching the gauge. Look for the architecture to crack and re-form.
We reopened the sealed branching runs and asked a different question: did the decomposition itself reorganize, even when scalar Φ did not rise smoothly?
Instead of compressing each state into one number, the atlas recovered all seven bipartition cuts, the minimum-information partition, cut turnover, entropy-normalizer boundaries, covariance geometry, and morphology. It reused all 1,152 trajectories—no new outcomes were generated.
The broad answer was still no. Across ordinary finite-estimator states, there was no positive canalization outcome. Most structural correlations with the later outcome were weak. There was no universal “quake signature.”
There were 119 groups in which all eight windows were fully estimable and 25 groups that crossed an estimator boundary. The 119 ordinary groups contained zero positive outcomes. The only two positive outcomes in all 144 groups—f19-s13-c24 and f19-s15-c16—were both f19 boundary cases.
This does not make an estimator failure causal. Boundary cases were not unique to success, and the two positives were found after outcomes were visible. But it sharpened the prospective hypothesis: the relevant precursor may be a reorganization of decomposition geometry, not a rise in scalar Φ.
Step one did not find causal emergence. It found the first exact place worth cutting open: two f19 states where successful canalization coincided with a decomposition boundary.
If those boundary states had acquired a goal, damage should create an error—and the future should reduce it.
The eight-state discovery panel contained the two f19 positives, two checkpoint-matched f19 negatives, two boundary negatives from other families, and two ordinary finite-estimator negatives. Every arm within a state began from the exact same pre-challenge state, and each of its four future seeds was shared across all five arms.
The injury happened immediately after the original division partition. Fresh future randomness began only at the next division. We measured normalized, recentered D4/Hellinger morphology at once and again at divisions D+6, D+7, and D+8.
The runtime did exactly what it was asked to do. The scientific assay did not.
The square-root, normalized, recentered 32×32 spatial matter-density metric was almost blind to the wound. Removing matter from thousands of microscopic locations barely changed the coarse normalized density pattern. That is compatible with distributed spatial mixing, redundancy, and passive robustness in this representation.
It is not evidence of repair. Repair requires a detectable initial error followed by its reduction. Here the readout said, in effect, “the target shape was barely displaced” at the first moment we looked.
The hidden differences did not disappear. They amplified. Every one of the 128 injury/future rows had a larger late distance than immediate distance. Their raw “repair” scores were therefore all negative, ranging from −0.258632 to −0.006151.
And the different injuries did not converge on a common future. Cross-injury contraction was negative in 8 of 8 states, from −0.1590 to −0.0630. The fans widened. That is the opposite of morphological equifinality.
The cleanest reading is cryptic internal damage followed by dynamical amplification—not a system recognizing and correcting a morphological error.
A second, post-hoc read of the same sealed runs separates three very different things that the word “recovery” can hide.
The total-matter difference shrank. Raw bulk-mass error repair was positive in all eight states, about +0.0331 to +0.0362. But the f19 candidates were not special: pooled candidate +0.03446 versus controls +0.03465. This can follow from generic threshold and regrowth dynamics; it does not by itself demonstrate an error-sensing controller.
Mass error is |log(Minjured/Muntouched)|, compared at D and the paired D+6, D+7, and D+8 observations.
observed, genericThe coarse normalized spatial pattern barely moved under a distributed wound. The f19 states also showed a faint raw hint of less late divergence. Yet representation-level insensitivity and resistance can both be passive, and every injury fan still expanded.
hint onlyFounder-composition repair was negative in all eight states, and founder-composition contraction was negative in all eight. Candidate matched repair favored f19 in only 4 of 6 comparisons. No distinct injuries were pulled back to one higher-level target.
Founder-composition error is squared Hellinger distance over normalized four-founder mass shares, compared at D and the paired D+6, D+7, and D+8 observations.
not observedThe system can refill its tank. We have not shown that it knows what structure the fuel is supposed to build.
A little less divergence is not the same thing as repair.
Pooling the invalid raw rows, the two candidate f19 states had a mean raw repair score of −0.05985, versus −0.08189 in the six controls—a difference of +0.02203. All six candidate-versus-matched-control raw comparisons favored the candidate.
The candidate contraction was also somewhat less negative: −0.07850 versus −0.09571 in controls. If we were looking only for dynamical buffering, that would be worth remembering.
With essentially zero valid starting displacement, raw “repair” is dominated by how far the futures eventually wandered. The absolute sign was still wrong in every candidate: both f19 injury fans expanded. Their contraction contrasts against controls split three positive and three negative.
Nor did founder-targeting show consistent specificity. Across the 16 state-by-founder comparisons, exactly eight favored targeted injury over its mass-matched sham and eight did not. So there is no coherent evidence that the putative f19 organization selectively restored morphology after founder-targeted damage.
Bold interpretation: the f19 boundary states may be slightly better dynamical shock absorbers. Honest interpretation: this assay cannot tell whether that hint is real, and it shows no goal-directed correction.
The atlas converted an unhelpful scalar result into a precise candidate event: a decomposition boundary in two f19 states. That is progress. It says where our information decomposition becomes singular—and therefore where a higher-level organization might be forming, or where the estimator may simply become ill-conditioned. Either way, this is now an exact site to interrogate.
A normalized 32×32 spatial matter-density map can look nearly unchanged after a distributed 5% loss because its material is intermixed or redundant. That is passive robustness—or readout insensitivity. A goal is more demanding: it must notice a macroscopic error and spend dynamics to remove it.
We injured founder-labelled matter but judged success by a normalized, recentered 32×32 matter-density map. The transform discarded absolute mass and founder identity while retaining coarse normalized internal density. The internal composition changed substantially while this chosen goal variable barely moved. Later, the hidden damage expressed itself as divergent trajectories. The experiment is therefore a powerful demonstration of delayed amplification—but not a clean regeneration test.
We have not seen multiple wounded paths being drawn back toward one future. We have seen the reverse: tiny cryptic differences flowering into different futures. In this assay, history remained causally loud.
The sealed-bundle exploratory check is now done: bulk matter returned, founder composition did not. Preserve it as a hypothesis generator; do not reopen the morphology decision or promote this unfrozen secondary analysis into confirmation.
On outcome-blind, unselected states, run a tiny immediate-only calibration of one coherent spatial wound. Freeze an operator that reliably produces a D4 displacement between 0.01 and 0.05 without extinction. This is engineering the measurement, not testing the hypothesis.
Keep the untouched, matched-sham, and common-random-future logic. Ask whether distinct visible injuries contract toward the same future, and whether the f19 boundary prediction beats sealed negative controls. Do not launch the previously imagined 320-run confirmation until the challenge is demonstrably capable of moving the claimed goal-state.
This is an outcome-seen, eight-state discovery panel. It cannot confirm causal emergence, autonomous replication, consciousness, biological universality, or even absence of goal-directedness in Flow Lenia. The mechanically authorized conclusion is narrower: the morphology-repair assay was inconclusive because all eight states failed the prespecified immediate-injury gate.
The exact challenge, state identity, future pairing, and dose accounting all passed closure. The negative raw repair and contraction patterns are scientifically informative descriptions, not valid primary hypothesis tests.