Breeding spatial economies: the final v3 harvest

Post 12 · two islands from the v2 archive · generations 1–12 complete, including final selected reseeds. Baseline: g1–7; changed-configuration continuation: g8–12. Source: audited final harvest.

1 · What C9 measures

v3 adds a “spatial economy” proxy to the v2 complexity score. Its factors reward used empty space, localized coupling, episodic bonds, and persistent phenotype diversity. They do not directly test regional chemical composition.

Any measured zero factor makes C9 zero; the alive gate also rejects dead worlds. I_v2 is the preserved 0–100 v2 score. Metric definitions and constants give the exact implementation.

Partial and full C9 are different assays. The continuation measured partial C9 only: surface locality was not measured. The baseline includes 13 full-C9 creative screens in generation 1, so its C9 statistics are split below. Six full-C9 atlas imports are outside the creative-screen comparison. Missing C9 is not zero. Partial C9 can earn the structured descriptor, but cannot satisfy the full-s9 economy class rule.

2 · Comparable results

Unless marked otherwise, qualification means status=ok, C9 ≥ 0.4, and common interest ≥ 60 at W9 = 0.40. The audit applies the same blend to preserved measurements; it does not rerun metrics or compare raw scores made with different weights. “Qualifying” is not a test for biological cells or the strict economy class.

Creative screens only

One screen per emitted candidate-name unit within each island. Failed and missing-C9 screens stay in the yield denominator. C9 means and maxima use only the measured rows shown; imports and selected reseeds are excluded.

Selected reseeds: a separate distribution

The best selected assay is p1g12_049_s3, an add_chan child on island 1. Its genotype scored 0.6737 / 0.8854 / 0.8895 in the screen / seed-2 / seed-3 observations. The aggregate selected-reseed mean fell, and island 2's best selected C9 remains the older 0.8048 result. A higher maximum is not uniform improvement or a guarantee on a new seed.

Descriptor breadth, not physical cells

Count each exact MAP-Elites descriptor key once if at least one creative screen qualifies at common W9 = 0.40. The baseline observed 51 bins (48 restricted to partial C9); the continuation observed 45 (all partial), including 14 new relative to the baseline. Their union is 65 (62 using partial C9 only). These are observed behavior categories across runs, not physical cells, verified archive incumbents, or compartment counts. The earlier 76 was a pooled screen-plus-selected-reseed bin count at W9 = 0.25, not archive occupancy.

3 · What actually changed during the search

Two islands searched from the v2 archive with a spatial-class descriptor axis. The following are requested creative slots per island per generation, not emitted counts or a nominal population size.

The classical-merge loop requested 24 slots in g1–7, then 20. Retry failures can drop targets. g8 emitted 81 + 82 screens; g9–12 emitted 560 in total. Continuation SIC emitted 45 screens in g8 and 25 in g9–12, not 24 per island throughout. Immigration fell from 20 to 5 requested slots; duplication and shared-channel merge targets went to zero. C9 backfill and archive reblending also changed parent selection.

Spatial-IC composition and operator yield

merge_spatial_ic (SIC) keeps one parent's chemistry and composes two parents' developed fields into soft-masked spatial regions as the initial condition. It does not create region-specific chemistry. Confirmation jobs omit that composed IC and start from soup.

For the historical W9 = 0.25 comparison, SIC's direct screen yield was 52/112 = 0.4643. The old 0.82 was (52 screens + 40 selected-confirmation observations) / 112 screen lanes: 92 observations in the numerator, not a hit probability. Counting each of the 112 bases once, 23 had at least one qualifying same-genome confirmation, whether or not its screen qualified; 12 passed the threshold in all three assays (11 partial-only). Those assays did not replicate the composed spatial phenotype.

These are descriptive operator yields. Weight, operator mix, retry limits, and parent archive changed together. There is no causal W9 or SIC comparison, and an unchanged-configuration continuation was not tested.

4 · Six GPU re-simulation films

These are not the original measured trajectories or score reconfirmations. All six captured requests match their actual result-row and job genomes and are outside the 11 anomalous links (identity checks). The C9 labels below belong to the original assays. Replays regenerate soup initial conditions; runtime and numerical-path differences can change trajectories.

Frames are 250 tu apart. This coarse time-lapse does not establish an interaction mechanism. The films' inherited “species” panel labels mean activator fields u_i, not inferred phenotype species or compartments. Posters show replay times T/2, 3T/4, and T. Colors overlay activator fields. Positive excess above each activator's baseline is scaled by its maximum across the three shown frames, separately per film and activator. The scale is fixed within a strip, not comparable across films.

5 · Before another campaign

Do not start another campaign until identity and the intended measure are validated.

• Make candidate and evaluation identities collision-safe. Carry exact genome, initial-condition, seed, and protocol provenance through confirmations and archive migration.
• Validate a regional-composition measure against distinct-composition and identical-composition controls, including moving regions and grid-origin changes.

The h9 review found that v0 is a fixed-grid label-association diagnostic, not a compartment-phenotype ranker. Its toy counterexamples can reward identical compositions or lose separation under translation. Do not add v0 to C9, choose a biological cutoff, or scan the full harvest to hunt for one. The next step is identity safeguards and measure validation, not a larger search justified by a proxy maximum.

Details and reusable tables: audit JSON · screen CSV · operator CSV · selected-assay maxima. The historical harvest and raw outputs remain preserved.