Scalar integration
Do four founder-composition channels predict their next state better as one coordinated whole than after we split them into parts? Our excess Φ-R score was the first instrument.
01 / idea
The Levin clue was bigger than one number
The GARD paper asks whether a chemical system becomes organized at a higher level before recognizable self-replication arrives. We translated that into Flow Lenia in three increasingly direct views.
Decomposition architecture
Which split is weakest, when does that split change, when does the estimator cross a boundary, and how does the covariance geometry reorganize? The parts themselves can move even when the scalar barely does.
Future repertoire
From one exact state, launch several matched futures. Their pairwise similarities form a map. If a structural event precedes a change in that map, the system’s organization is beginning to constrain possibility itself.
Φ-R was a lens. The shape of possible futures became the object.
The earlier work becomes legible in that frame. The scalar experiments showed that information organization can be driven; the branching and wound experiments exposed structure hidden from one magnitude; the prospective repertoire experiment joined those lessons into a fresh prediction.
02 / journey
Every attempt changed the question
Each experiment redrew the object: recurrence replaced silhouette fission, intervention replaced correlation, and the geometry of several possible futures replaced one scalar gauge.
Find a phenotype
- Injury and recovery A founder lineage could be depleted and sometimes return, but the early Φ-R effect changed sign across creature families.
- Reservoir recurrence Four-founder compositions really did recur across repeated growth and passage. That gave us a cleaner GARD-like event than silhouette fission.
- Temporal exploration Broad recovery did not track scalar integration, yet all three measurable strict onsets showed a pre-onset rise. Timing became the lead.
Push the information
- Broad steering Policies changed later compositional dynamics, but the intended first-step Φ separation was uncertain.
- Orthogonal pulse A cleaner five-arm intervention moved excess Φ-R by +0.188 at q41. One passage after release, the contrast was −0.013.
- New object We had found a controllable, actively maintained information mode—not a self-sustaining restoration switch.
Open the geometry
- Branching futures 1,152 trajectories showed that most future fans expanded. Scalar Φ-shape did not predict the rare canalizing cases.
- Architecture atlas The only two positive canalization outcomes occurred among 25 estimator-boundary groups; none occurred in 119 ordinary groups.
- Connected wounds Equal removed mass produced unequal injury, while nearly equal morphological wounds sometimes required 2.64× different mass. Susceptibility had direction.
Map possibility
- Repertoire atlas Eight matched futures became an 8×8 similarity kernel: a direct picture of how one state distributes later possibilities.
- Prefix-only peak Boundary changes, minimum-information-partition turnover, and effective-rank shifts selected one structural peak before outcomes opened.
- Prospective hit At those peaks, the later future map reorganized more than at neighboring checkpoints in 14 of 19 fresh families.
03 / discoveries
Four things survived every change of framing
These are now properties of the system worth studying in their own right, not consolation prizes for an earlier endpoint.
Controlled information mode
We can drive Φ-R up—and watch it relax
The orthogonal policy built a clear Φ+ versus Φ− separation during sixteen passages of active control. The effect vanished immediately after release. That gives us a response curve, a relaxation timescale, and a clean way to ask about dose and hysteresis.
Structural boundary
The rare canalizing cases lived where the decomposition broke form
Among 144 branching groups, all 119 fully estimable groups had negative canalization outcomes. The only two positive outcomes appeared among the 25 groups crossing an estimator boundary, both in family f19. That post-outcome anomaly gave us a precise structural place to look.
119 ordinary · 0 positive23 boundary · negative2 boundary · positive
Anisotropic woundability
The system filters injury through state and direction
All 24 within-state direction pairs separated, but one global mass dose could not put every lesion on the same morphological shell. In f20, two nearly equal D4 wounds required 2.75% versus 7.25% matter removal. Damage is not a scalar input here; it is geometry.
04 / turn
Measure how history reorganizes possibility
The prospective experiment compared two histories under the same ten later random events. It asked whether a structural peak visible in the prefix marks a local change in the relationship among eight sampled futures.
Past-only means past-onlyThe structural peak was chosen from the history. Future morphology was still sealed.
Matched future pressureThe c and c+2 maps used the same ten absolute later random events, so the history changed while future draws matched.
One family is one voteSchedules, checkpoints, eight futures, and 28 future pairs did not inflate the sample above 19 families.
05 / result
The first prospective hit
At the frozen 32×32 spatial scale, 14 of 19 fresh families showed more future-map reorganization at the prefix-only structural peak than at its two neighboring checkpoints. The fixed-panel replication rule was 14.
14/19primary fixed-panel replication passed
The mean family contrast was +0.0036619 Q. Q is the root-mean-square change between two 8×8 future-similarity maps, so the raw number is small by construction; its sign asks whether the selected structural peak sits above its local temporal background.
32×32 primary16×168×8
32×32 primary · frozen replication rule passed
16×16 · mean direction preserved
8×8 · mean direction preserved
The clean reading: the prospective fixed panel reproduced a prefix-to-future-map relationship at the primary scale, and its average direction survived two coarse-grainings. The separate 14-of-19 rule was not met at 16×16 or 8×8, so the result is directionally stable rather than fully scale-robust. Twelve families were positive at all three scales.
05B / morphogenesis
Before the creature, the whole moved first
We returned to the most literal Levin-inspired question: can an integrated predictive configuration appear while a Flow Lenia seed is still becoming an organism? In 64 fresh source specimens, the frozen whole-over-parts signal rose before an independently certified sustained break from the artificial square.
+0.364fresh pre-event rise · 95% interval +0.231 to +0.502
The state was the square-root mass composition of four spatial quadrants. The primary score compared their joint next-step predictive information with the sum available after the best independent bipartition. Its event-near mean exceeded its own earlier pre-event mean in 39 of 54 specimens with complete 32-step windows.
Development reproducedFour-fold development gave +0.507 in 44/59 measurable specimens. The next 64 source organisms were frozen before replay.
Fresh breadthFifteen of sixteen fresh four-specimen family means were positive. The direct Φ-R companion independently rose by +0.171 [+0.084, +0.255].
Event stayed independentThe morphology detector used only occupied-body rectangularity: eight sustained frames below 0.7. One non-transition by step 900 remained right-censored.
The constructive reading: a causal-emergence-like whole-over-parts predictive configuration prospectively precedes a reproducible Flow Lenia morphogenesis event. The detector is conservative—it marks committed non-square organism formation, not the first faint internal texture. Ordinary whole-state predictability had no consistent direction, while the formal whole-over-parts-minus-TDMI contrast was positive on average but noisy.
06 / meaning
Organization may be an event, not an amount
The experiments now fit one sharper picture. A Flow Lenia system can become more predictable as a whole than as separable parts before committed morphology arrives; later, its information modes, effective parts, wounds, and possible futures expose further layers of organization. The interesting object is the transition architecture connecting those layers.
Before the creature is fully visible, the whole can already be gaining predictive reality.
BirthWhole-over-parts predictive information rose prospectively before a certified morphogenesis event.
InformationPolicy can keep founder dynamics information-bearing; the scalar contrast itself relaxes fast.
ResponseWound magnitude depends strongly on state and direction, revealing structured susceptibility.
PossibilityA prefix-only structural peak prospectively predicted local future-map reorganization.
07 / parallel read
A hidden information-dynamical mode
A fresh descriptive read of the 960 pulse trajectories found something more interesting than a lasting scalar Φ-R shift. From the same prefix, policy direction changed whether the founder-composition history stayed rich enough to carry the score at all.
Φ+ versus Φ−
+24.48pp
Policy kept the information-bearing coordinates alive.
The paired q41 difference favored Φ+ in 15 of 16 families and 11 of 12 schedules: 54 matched blocks were scorable only under Φ+, seven only under Φ−, 112 under both, and 19 under neither.
Same support+0.1883 → −0.00845
Among the 106 blocks scorable in both arms, the excess-Φ contrast was present at q41 and gone one division later. The within-block change was −0.1967.
W64 historyMore path, less freezing
Φ+ versus Φ− changed composition path length by +0.1209, exact frozen founder coordinates by −0.2448, and minimum coordinate range by +0.01847.
Coarse phenotypeAlmost unchanged
At q41, Φ+ minus Φ− was −0.017 effective founders, −0.44 components, and −0.007 field radius. The mode was informational before it was visually obvious.
After release
The scalar vanished first; the state memory faded later.
The Φ-R contrast collapsed at q42. The availability gap decayed through q49 and was absent by q57.
+24.5q41
+16.1q42
+17.2q43
+9.4q45
+5.7q49
goneq57
| Policy arm | q41 scorable | Mean frozen coordinates | Mean Hellinger composition path |
|---|---|---|---|
| Φ+ | 166 / 192 | 0.135 | 0.696 |
| Φ− | 119 / 192 | 0.380 | 0.575 |
| TDMI+ | 117 / 192 | 0.385 | 0.537 |
| TDMI− | 178 / 192 | 0.073 | 0.724 |
| random | 137 / 192 | 0.286 | 0.645 |
What is genuinely new: from an identical prefix, the controller changed how often the ecology entered a frozen-coordinate condition, robustly across families and schedules. Score availability is defined by those coordinates varying, so the frozen-coordinate audit explains that availability; it is not an independent confirmation. This is a descriptive post-result discovery, ready for a closed-loop-versus-replay causal test.
08 / response surface
Φ postponed turnover; it did not simply add diversity
A fresh 320-trajectory experiment varied Φ-control strength and pulse duration on the same matched worlds. The clearest effect was temporal: Φ-directed supply kept all four founder identities in play longer before the first sustained loss.
32-action pulse
+9.35
passages before first sustained identity loss
The paired delay favored the Φ arm in 23 of 31 event-bearing blocks. The effect was not universal, but it was broad: the full-strength 16-action contrast was positive in six of eight families, while the other two family means were close to zero.
DoseStep, then saturation
At fixed 16-action duration, the paired delays versus random were +3.06, +6.09, +4.06, and +6.00 passages at 25%, 50%, 75%, and 100% control. The response looks threshold-like around half strength, not smoothly proportional.
What changedTurnover timing
At q41, full Φ versus random changed effective founder number by only +0.016. By q100 it was −0.055. Φ did not leave a more diverse endpoint; it altered how long the composition resisted losing one of its identities.
Family breadth6 of 8 positive
The full-strength loss delay by family was +9.75, +11.5, −0.25, −0.5, +9, +8, +8, and +2.5 passages. This is a widespread temporal effect with real ecology dependence, not one creature carrying the mean.
After release
A short memory survived the score.
Full Φ versus Φ− differed little in the scalar at q41 (+0.024) and not at q42 (−0.015). But the founder-composition path stayed longer under Φ+: +0.112 at q41, +0.093 at q42, and +0.037 at q45, before crossing near zero by q49.
| Later event | Matched events | Mean lead from susceptibility peak | Peak came first |
|---|---|---|---|
| First sustained founder loss | 223 | +9.29 passages | 74% |
| Dominance succession | 146 | +5.00 passages | 70% |
| Founder reactivation | 222 | +22.34 passages | 95% |
| Persistent-fission candidate | 4 | +5.75 passages | 25% |
The new object is compositional resilience. Φ-control changed when a founder coordinate vanished, not the final number of effective founders. The earlier dramatic missing-score asymmetry also did not generalize wholesale: at q41 this fresh panel was scorable in 29/32 Φ+ runs and 28/32 Φ− runs. When missingness separates, it still reports coordinate freezing; it is not itself the universal phenomenon. Five persistent-fission candidates occurred across all 320 runs—far too few to connect Φ to fission. The clean prospective target is now founder-identity loss, with susceptibility peak timing frozen in advance.
09 / next
Now move the architecture
Prediction got us to the doorstep twice. The decisive experiments should now move the early whole-over-parts precursor, test whether organism birth moves with it, and ask whether the later structural event changes the future map.
The observation we are now trying to earn
An integrated predictive configuration helps create the organism; later structural transitions reorganize its counterfactual futures. Advancing or blocking those events moves the biology.
10 / ledger
The complete research trail
Every stage is accounted for here. Engineering and sealed aborts remain visible, while the narrative above gives scientific weight to the experiments that changed our model of the system.
Full experiment and analysis ledger
| Stage | Scale | What happened | What it unlocked |
|---|---|---|---|
| Recovery engineering v1–v3 | Three nine-run plans; v1 and v2 stopped on run 1, v3 completed all 9 | Built and audited founder-lineage depletion, timing, and recovery mechanics. | A usable intervention ontology without opening Φ or scientific outcomes. |
| Perturbation recovery outcome v1 | 72 trajectories · 24 matched triplets | Eight recovered and ten did not among 18 eligible external-arm runs. | Enough variation to open a separate past-only predictor protocol. |
| Preinjury predictor v1 | Reused the labelled pilot | The intended primary feature was unavailable under the frozen support rule. | A fresh confirmation cohort and cleaner feature capture. |
| Preinjury confirmation v2 | 288 trajectories · 211 complete rows | Family effects were +0.376, −0.520, and +0.177 SD; adding Φ-R changed held-out log loss from 0.686445 to 0.691239. | Scalar Φ-R was not a family-general injury precursor; representation and phenotype needed widening. |
| Reservoir recurrence census v1 | Mechanical stage | Sealed before outcome inspection after an operator-audit abort. | A corrected reservoir adapter and preserved provenance. |
| Reservoir recurrence census v2 | 72 trajectories · 2,781 cycles | Found 56 passing four-cycle recurrence windows and four strict positive plus four strict negative runs. | A real GARD-like compositional recurrence event, though within-family contrast remained sparse. |
| Temporal precursor exploration | 2,027 valid episodes · 1,660 with Φ-R | Broad association was near zero. All three measurable core onsets rose by +0.101, +0.112, and +0.215. | A concrete prospective timing hypothesis. |
| Reservoir causal steering | 10-run smoke · 1,920 full trajectories | Policies separated later trajectory summaries, while first-step Φ and fixed-horizon restoration remained uncertain; the natural onset follow-up was +0.00320 [−0.10242, +0.12651]. | Policy leverage was real, but the target and intervention were coupled too broadly. |
| Orthogonal full-campaign engineering | Layout, concurrency, storage, compact-evidence qualifications | Technical aborts and verifier defects were preserved and corrected. The large campaign was not launched. | A smaller five-arm discovery pilot with the same causal logic and far less machinery. |
| Orthogonal pulse-release pilot | 960 trajectories · 192 matched blocks | Φ+ − Φ− reached +0.1885 [+0.0726, +0.3133] at q41 and collapsed after release; restoration reward was −0.2605 [−5.826, +5.319]. | A controllable transient information mode and a direct relaxation question. |
| Pulse dynamics exploration | Reused all 960 trajectories · descriptive post-result read | At q41, Φ+ remained scorable in 166/192 blocks versus 119/192 under Φ−: +24.48pp, positive in 15/16 families and 11/12 schedules. The same-support score contrast fell from +0.1883 to −0.00845 by q42. | A hidden controller–ecology mode: policy altered the persistence of information-bearing founder dynamics before coarse phenotype visibly separated. |
| Information-mode response surface | 320 fresh trajectories · 32 matched ten-arm blocks | Φ-control delayed first sustained founder loss by +5.22, +6.31, +6.00, and +9.35 passages for 4-, 8-, 16-, and 32-action pulses versus random. Terminal effective founder number did not increase. | Compositional resilience—not endpoint diversity—became the next causal phenotype; susceptibility timing now has a concrete prospective event target. |
| Branching canalization pilot | 1,152 trajectories · 144 groups | Similarity fell 0.85494 → 0.75380; mean late novelty was 0.44251. Adding Φ-shape worsened held-out MAE 0.020682 → 0.021615. | Future fans themselves became the phenotype; 25 boundary groups demanded inspection. |
| Architecture-quake atlas | Reused all 1,152 trajectories | Zero of 119 ordinary groups had positive canalization. The only two positives occurred among 25 boundary groups, both f19. | A structural-event hypothesis based on boundary, MIP, and covariance geometry. |
| Goal-state tomography | 8 states × 5 arms × 4 futures = 160 | The 5% wounds removed thousands of cells yet moved normalized D4 only 0.000245 on average; bulk mass still returned. | The first wound and chosen morphology readout had missed each other. |
| Connected-bite calibration | 32 runs | All 48 directional pairs separated; fixed 5% and 7.5% doses could not create one shared severity shell. | Directionally distinct wounds and evidence that equal mass is not equal injury. |
| Matched-wound confirmation | 16 selected wounds · 464 candidate evaluations | All 24 directional pairs separated; 9 of 16 wounds entered the common D4 band. | An anisotropic, state-dependent woundability map. |
| Synthesis v1 | Read-only synthesis across nine stages | Joined the scalar, branching, and wound programs into one response-geometry account. | The explicit pivot from “more Φ” toward decomposition change, future repertoires, and goal-basin tests. |
| Future-repertoire atlas | Reused 1,152 trajectories · eight futures per group | Defined pairwise morphology kernels, effective repertoire, and overlap-aligned Q without launching new outcomes. | A direct metric for reorganization of sampled possibility. |
| Prospective future repertoire | 2,736 trajectories · 342 groups · 19 fresh families | Fourteen family contrasts were positive at 32×32; mean +0.0036619. Mean direction stayed positive at 16×16 and 8×8. | The first prospective fixed-panel replication of a structural precursor to future-map reorganization. |
| Prospective organism-birth signal | 64 development · 64 fresh specimens · 32 source families total | The frozen whole-over-parts rise replicated at +0.3638 [95% +0.2311, +0.5021] in 39/54 complete windows; 15/16 fresh family means were positive. The direct Φ-R companion was +0.1706 [+0.0840, +0.2548]. | A direct Levin-style precursor to committed Flow Lenia morphogenesis—and a concrete early intervention target. |
Preserved technical attempts and why they stopped
- Recovery engineering v1: an occupied-support assertion fired before challenge exposure; no biological outcome opened.
- Recovery engineering v2: a Float32 dose-accounting reduction disagreed after state installation; the state itself was intact.
- Recovery engineering v3/v3a: deterministic mechanics completed, then the audit window was repaired prospectively without rerunning the biology.
- Reservoir census v1: a grouped-mass reduction mismatch stopped the stage before outcomes; the corrected v2 replay closed 2,803 division events.
- Steering smoke attempts 1–3: Float32 JSON encoding and a post-fork prefix digest were corrected before the final ten-run qualification.
- Layout qualification v1: 2-channel, 10-kernel founders met a borrowed 3-channel, 20-kernel base; v2 switched to the founders’ native 2c10 base without resampling genotypes.
- Compact evidence v1/v1b: two verifier assumptions—raw JSON key order and cross-accumulator mass equality—were replaced by canonical-record and native-accumulator checks.
- Concurrency qualification v1: all 40 mechanical runs finished, but the diagnostic receipt could not certify the system-log slice; v2 used a hash-closed classifier and passed exact 8-way parity.
- Connected-bite calibration: one preserved preparation abort preceded the successful 32-run calibration.
- Matched-wound confirmation: preparation and execution aborts were preserved; the final 16 selected wounds completed and produced the directional map reported above.
Pigozzi and Levin
The GARD causal-emergence paper supplied the core provocation: organization may become measurable before overt self-replication. Our estimator and Flow Lenia ontology are explicit adaptations, not a claim to reproduce their undisclosed implementation exactly.
Where we stand
We have prospective prediction of future-repertoire reorganization in a fixed panel. We have not yet shown that the structural peak causes that reorganization, that the changed map is contractive, or that an autonomous replicator results. Those are now concrete experiments rather than generic caveats.