The simpler explanations failed
Scalar Φ did not generalize as a precursor or a useful steering target. Most futures dispersed. Bulk mass returned without shape correction. Wound response was strongly state- and direction-dependent.
01 / definition
What we mean by goal-directed behavior
A snapshot can look organized and still be causally empty. The stronger test is what the system does when reality pushes it off course.
Without contraction
Microscopic history remains loud. Tiny differences amplify, nearby branches fan outward, and interventions leave different scars. Novelty may be abundant, but nothing forces those possibilities to agree.
This is what most of our Flow Lenia branches did. They generated new futures cheaply. They did not yet show that any one future had become privileged.
With contraction
A higher-level organization should begin suppressing irrelevant microscopic variation. Different wounds, noise histories, and local routes are pulled toward the same macroscopic endpoint. The endpoint now explains the present because the system acts to restore it.
The transition of interest is from divergent to contractive counterfactual dynamics.
Structural reorganization may precede contraction
Before visible organismality, the effective parts or scale may reorganize. That event may convert an expansive field of possibilities into a basin with one preferred macroscopic future.
Prospective prediction and causal control
The reorganization must prospectively precede multi-wound contraction on unseen families. Inducing it must create the basin; blocking it must erase the basin. Anything less remains a lead.
DisturbPush the state awayUse several genuinely different, severity-matched injuries.
BranchChange the microscopic futureReplay multiple fresh futures from each exact wounded state.
MeasureWatch error and spreadDoes target error shrink? Do the injury fans contract?
InterveneMove the organizationInduce or block the suspected precursor and move the basin itself.
02 / evidence
How the experiments changed the hypothesis
Each stage removed a simpler explanation and made the next experiment more demanding.
-
01No broad signal
Confirm the precursor
In 288 trajectories with 211 complete rows, the equal-family Φ effect was +0.0113, p=.4694. Adding Φ made control log loss worse: 0.686445 → 0.691239.
-
02Timing clue failed prospectively
Test temporal timing
72 trajectories yielded 2,027 valid episodes, 1,660 with Φ. Broad correlations were near zero and all held-out improvements were effectively zero. Although 3/3 measurable core onsets rose beforehand in exploration, the frozen follow-up on 23 matched onsets gave only +0.00320 [−0.10242, +0.12651].
-
03Causal null
Steer Φ broadly
Across 1,920 trajectories in 384 blocks, exact-next-checkpoint Φ separation was only +0.0261 [−0.0186, +0.0679]. All five behavior confidence intervals crossed zero.
-
04Φ separated; selectivity failed
Apply an orthogonal pulse
In 960 trajectories, Φ-R separated by +0.1885 [+0.0726, +0.3133], but the required TDMI nuisance-equivalence gate failed. Restoration reward was −0.2605 [−5.826, +5.319]. Selective Φ manipulation and behavioral benefit were not demonstrated.
-
05Futures dispersed
Demand prospective prediction
In 1,152 trajectories across 144 groups, late novelty was 0.4425, similarity fell 0.85494 → 0.75380, and outcome Y was −0.04475. Adding Φ-shape worsened held-out MAE: 0.0206824 → 0.0216150.
-
06Structural lead
Inspect the decomposition
119 fully estimable groups contained 0 positive labels. The only 2 positives were among 25 boundary groups—both f19, both found post hoc. All 144 historical stochastic partition fans still expanded; mean dispersion grew 0.07345 → 0.28385.
-
07Assay invalid
Injure the putative goal
All 160/160 tomography trajectories closed, but 0/8 states passed the immediate injury gate. Mean D4 was 0.000245, maximum 0.001565. All 128/128 raw repair values and 8/8 contractions were negative, but descriptively only.
-
08Direction achieved
Make connected bites
In 32 calibration runs, fixed 5% wounds passed 9/16 severity gates and 7.5% wounds passed 4/16. All 48/48 directional pairs separated, minimum 0.03006. Equal mass was not equal injury.
-
09Woundability mapped
Match wound severity
464 candidate evaluations produced 16 fresh selected wounds, of which only 9/16 entered the band. Yet all 24/24 direction pairs passed, minimum 0.04106. Direction was reliable; this grid and controller did not produce a common severity shell across the four-state panel.
Zero in the smooth regime. Two at the boundary.
The 119 ordinary groups held no positive discovery labels. The 25 boundary groups held the only two, both f19. But 23 boundary groups were negative, both positives were selected after outcomes were visible, and all 144 fans expanded. This is a candidate crack—not a precursor result.
119 ordinary · 0 positive
25 boundary · 23 negative
2 boundary positives · both f19
The evidence now points from scalar magnitude toward changes in response geometry.
Scalar Φ was neither a broad precursor nor a useful held-out predictor, and the intervention studies did not establish it as a selective control target. The remaining lead is structural: rare canalization coincided with decomposition boundaries, and connected wounds revealed strongly state- and direction-dependent susceptibility.
03 / ladder
A hierarchy of goal-directed behavior
“Recovery” hides a hierarchy. We should reserve teleological language for the upper rungs.
Bulk homeostasis
Refill material after loss. Bulk-mass repair was positive in all eight tomography states, roughly +0.0331 to +0.0362, with candidates and controls nearly identical.
Observed · genericShock absorption
Limit how far damage propagates. The two f19 candidates wandered somewhat less in an invalid post-hoc read, but every historical stochastic partition fan still expanded.
Hint onlyError correction
Detect a genuine structural error and reduce it. Our first wound was unreadable; the matched controller reached the severity band in only 9/16 cases.
Not cleanly testedEquifinality
Erase injury identity: distinct wounds and noisy futures collapse toward one macroscopic endpoint. This is the operational signature of a goal basin.
Not observedGoal memory
Carry a target that cannot be read from anatomy alone—and reveal it through repair, transplantation, or conflicting histories.
Not testedThe system restores bulk matter. We have not shown that it restores a specific organization.
That distinction is the heart of the programme. Passive stability, threshold-driven regrowth, and material redundancy can all mimic weak “recovery.” A protected goal must do more: it must make several wrong presents become the same right future.
Our target is level four, discovered before visible organismality and then pushed causally into level five.
04 / newest result
What the wound experiments revealed
inside frozen D4 band
closest candidate outside band
q = fraction of total matter removed
| State | Major + | Major − | Minor + | Minor − |
|---|---|---|---|---|
| f20-s14-c24 3 / 4 in band |
0.035535pass q 7.25% |
0.036005pass q 2.75% |
0.029935low q 5.50% |
0.035559pass q 4.00% |
| f21-s14-c24 2 / 4 in band |
0.035648pass q 4.25% |
0.028726low q 2.75% |
0.035521pass q 4.25% |
0.042057high q 5.25% |
| f24-s14-c24 4 / 4 in band |
0.033853pass q 3.50% |
0.036254pass q 4.25% |
0.036335pass q 3.25% |
0.037239pass q 4.50% |
| f25-s14-c24 0 / 4 in band |
0.038168high q 3.25% |
0.037696high q 4.00% |
0.038513high q 4.25% |
0.030684low q 4.75% |
Four distinct wound directions
All 24/24 within-state direction pairs exceeded the frozen 0.01 separation gate. The global minimum was 0.041058. These are not cosmetic variants of one lesion.
No common shell in this panel
Seven selections missed the D4 band under the frozen 29-point search. f25 missed in every direction; f24 passed all four. State and orientation jointly control injury magnitude, although the seven misses may still include coarse-grid interpolation failures.
An anisotropic woundability manifold
The system filters damage through its organization before damage becomes morphology. This is structured susceptibility—not yet repair, but precisely the geometry a repair controller would have to master.
What this does not show
The 9/16 result is an outcome-seen engineering confirmation of the wound operator. It does not show causal emergence, recovery, a goal, or that f24 is an organism. Because the frozen all-16 gate failed, the 160-run recovery cohort was correctly not launched.
05 / next tests
Seven decisive experiments
Each experiment tests a different consequence of the hypothesis, with a result that would support it and a result that would end it.
Swing 01Take first
Measure basin formation directly
Hypothesis: agency begins exactly when the response operator becomes contractive—when different errors start losing their identity.
Current basisOur branches mostly dispersed. The newest assay finally generated genuinely distinct directions: 24/24 pairs separated, and f24 supplied four severity-matched wounds.
Decisive testAt dense checkpoints, give several matched wounds, replay common fresh futures, and measure both target-error reduction and cross-injury contraction at D+6, D+7, and D+8.
Decisive resultA state still looks ordinary, yet every valid injury repairs and all injury fans contract toward one future—while matched earlier states keep dispersing.
Swing 02
Track structural reorganization through time
Hypothesis: the basin begins with an abrupt reorganization of parts, boundaries, and effective scale—not with a smooth rise of scalar Φ.
Current basisThe 119 fully estimable groups held zero positives. Both positive groups sat among 25 estimator-boundary cases and belonged to f19. That is a post-hoc anomaly, not proof—but it motivates the test.
Decisive testMeasure every bipartition, minimum-information-partition turnover, entropy boundaries, covariance geometry, and basin contraction at dense checkpoints in unseen families.
Decisive resultThe same structural transition repeatedly appears one checkpoint before contraction begins, even when morphology and scalar Φ remain unremarkable.
Swing 03
Induce and disrupt the candidate mechanism
Hypothesis: if a structural reorganization creates the goal basin, a brief, correctly targeted intervention should install or destroy future-directed control after the intervention ends.
Current basisThe pulse study produced a +0.188 Φ-R separation, but nuisance equivalence failed and restoration benefit was not demonstrated. The next intervention should target the proposed organization directly, not assume that moving its summary statistic is selective.
Decisive testInduce the candidate partition/scale transition in a pre-basin state; block it in a near-transition state; release both; then apply the same blinded wound panel.
Decisive resultInduction creates persistent multi-wound contraction and blockade abolishes it, with matched energy, mass, timing, and sham interventions.
Swing 04
Test for hidden goal memory
Hypothesis: two states can look nearly identical while carrying different target memories. Anatomy may underdetermine the future the system will restore.
Current basisOur branching work showed that small, cryptic differences can become causally important later. The distributed-wound assay barely moved visible D4, yet those hidden differences subsequently amplified.
Decisive testFind morphology-matched states from different histories, blind the selector to history, apply the same calibrated lesion, and ask whether each returns toward its own history-specific endpoint.
Decisive resultSnapshot twins with indistinguishable current anatomy regenerate reproducibly toward different targets—and swapping the relevant history-carrying organization swaps the destination.
Swing 05
Test whether an organizer transfers the target
Hypothesis: a compact region may carry disproportionate information about the whole-system target. Move the region and the destination should move with it.
Current basisConnected bites showed extreme spatial anisotropy: in f20, two nearly equal D4 wounds required 2.75% versus 7.25% mass. Location and orientation can matter more than bulk dose in this panel.
Decisive testLocate regions whose perturbation most changes future contraction; transplant them between donor and recipient states with mass-, geometry-, and damage-matched shams.
Decisive resultThe recipient adopts the donor’s repair destination, while donor removal weakens or redirects its original basin.
Swing 06
Select for multi-wound recovery and replay ancestry
Hypothesis: robust agency may be too rare to await accidentally. Select directly for the behavior, then use the full ancestry to identify when contraction first appears.
Current basisNatural branching gave abundant novelty but vanishingly little reconvergence: just two positive groups in 144, both from one family and both discovered retrospectively.
Decisive testSelect for recovery from four orthogonal matched wounds, common-endpoint convergence across noise futures, environmental robustness, and phenotype maintenance under material turnover.
Decisive resultAn ancestral replay reveals a discrete generation where contraction first appears, preceded by a reproducible structural reorganization that can itself be manipulated.
Swing 07
Test conflicts between learned targets
Hypothesis: if goals are real higher-level causal constraints, incompatible goal memories should interact through dominance, boundary formation, or stable compromise—not merely average locally.
Current basisThe woundability result says organization is directional and distributed. A goal may therefore be separable from current shape yet still compete across the whole field.
Decisive testJoin regions or matched twins carrying different inferred targets, then wound the chimera and map which endpoint, boundary, or hybrid its futures protect.
Decisive resultThe composite resolves toward one target or a stable negotiated form with nonlinear, history-dependent dominance that cannot be explained by local material averaging.
Map the 464 response points
Plot every D4(q) curve. Determine whether the seven misses are coarse-grid interpolation errors or real jumps and forbidden severity regions.
Freeze a state-aware wound controller
Match actual morphological displacement, preserve four directional distinctions, and demonstrate applicability before any recovery outcome is opened.
Locate the first contraction
Measure dense pre/post checkpoints around candidate reorganizations and compare them with matched ordinary states.
Write, erase, transplant
Only manipulation can turn an early warning sign into evidence that the emerging organization produces the protected future.
06 / kill criteria
Falsification criteria
These results would reject the corresponding hypothesis and force a change of direction.
-
No contraction after valid matched wounds
If diverse, visible, severity-matched injuries continue to fan outward across complete common futures, there is no demonstrated goal basin in those states. Stability and regrowth remain the better explanations.
-
Structural reorganization does not predict contraction
If decomposition-boundary events fail to precede contraction in unseen families—or appear just as often without it—abandon this precursor. The f19 anomaly was selection, estimator pathology, or coincidence.
-
Architecture moves but repair does not
If inducing or blocking the suspected reorganization leaves recovery competence unchanged under matched controls, our architecture metrics are epiphenomenal. They describe the transition without causing it.
-
Twins share one injury future
If snapshot-matched states from different histories always recover toward the same endpoint, there is no evidence for cryptic target memory beyond present anatomy.
-
The organizer does not carry the destination
If transplant effects vanish under mass-, geometry-, and damage-matched shams—or alter only local tissue—the target is not portably encoded in that region.
-
Evolution produces only wound closure
If selected systems seal local lesions but fail cross-injury equifinality, environmental generalization, or material-turnover tests, we evolved sophisticated homeostasis—not organism-scale agency.
-
No discontinuity anywhere
If ancestral replay and dense natural-history sampling show only smooth quantitative improvement with no identifiable birth of counterfactual contraction, the “sudden goal” picture is wrong. Emergence may be gradual—or Flow Lenia may not support it.
07 / belief
The decisive observation
We have not observed causal emergence. We have narrowed the required observation and identified the measurements that do not suffice.
Decisive evidence would be the onset of reproducible contraction toward one endpoint across distinct, matched injuries and future perturbations.
Prediction
A structural reorganization should precede the first checkpoint where wounds and noisy futures begin contracting toward one endpoint.
Causation
Inducing that organization should create the protected future; blocking it should erase the protection; transplanting it may move the destination.
Meaning
At that point, the whole earns causal status because it explains what its parts do next. Matter has begun acting on behalf of a state that does not yet exist.
08 / record
Evidence trail and claim boundary
- Pigozzi & Levin · causal architecture before self-replication The motivating paper. Our Flow Lenia measures are adaptations, not software-identical replications of an incompletely specified estimator.
- Pre-injury precursor confirmation · exact report 288 trajectories and 211 complete rows; the broad equal-family Φ precursor did not confirm.
- Temporal precursor exploration · exact report 72 trajectories, 2,027 valid episodes, and the narrow three-onset timing lead.
- Broad causal steering · exact report 1,920 trajectories in 384 blocks; no isolated Φ separation or behavioral effect.
- Orthogonal pulse-and-release · exact report 960 trajectories; a substantial Φ-R shift without demonstrated restoration benefit.
- Prospective branching and canalization · exact report 1,152 trajectories across 144 groups; novelty rose, similarity fell, and Φ-shape did not improve held-out prediction.
- Architecture-quake atlas · exact report 119 ordinary groups with zero positives; 25 boundary groups with the two post-hoc f19 positives.
- Goal-state tomography · exact report 160 closed trajectories; the primary assay was mechanically inconclusive because all eight states failed the injury gate.
- Connected-bite calibration · exact report 32 runs; fixed mass doses produced directionally distinct but severity-unequal wounds.
- Matched-wound confirmation · exact report Exact 16-run, 464-candidate result; report SHA-256 095ea944413557b59cee8df78fb1645d7dd3a99c03560a471576ca5c0bd1e65b.
- The Goal Before the Organism · prior visual synthesis The architecture atlas, tomography narrative, teleological ladder, and earlier next-step logic in plain language.
Claim boundary
We have not found causal emergence, goal-directed repair, autonomous replication, consciousness, or a universal precursor. We found that scalar Φ did not perform as a predictor and was not demonstrated as a selective control target; that rare canalization coincided post hoc with structural estimator boundaries; that a distributed 5% injury was unreadable in the chosen morphology metric; and that connected woundability is strongly state- and direction-dependent. The proposed experiments on this page are falsifiable research commitments built from those results—not results themselves.