Giorgio Gilestro
6f8cef1ac5
Removed from main (all preserved on the dev branch): the arXiv build and
its sources, design documents (blueprint, results summary, review responses,
essay drafts), tasks/ and CLAUDE.md, the cover letter and reference tooling,
two unused manuscript figures, and every experiment that feeds no figure or
number in the paper: the collapse null, the sexual-vs-asexual lineage, the
NK speciation variant, the 0.5B single-seed LLM prototypes, the compose and
society experiments with their calibration and pilot runs, and their
configs, runners, tests, figure scripts and PBS jobs. Their result bundles
are moved to results/_archive/ (ignored) so the parquets stay on disk.
Also: plot_llm_speciation reads the s{seed}/ layout; the mating-breadth
plot writes under its bundle name; Makefile targets reduced to the kept
experiments; REPRODUCING.md and README point to dev for the rest.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y64o8FKP7rCuXzC48pxpMm
|
||
|---|---|---|
| .. | ||
| fig4_society_ablation.pdf | ||
| fig4_society_ablation.png | ||
| manifest.json | ||
| README.md | ||
| resolved_config.yaml | ||
E11 — the dynamic Lamarckian society: the vertical claim (C3)
Claim tested. The culmination. Everything before was a single operator; the society's load-bearing, un-preempted claim is that composing them makes general capability climb over generations while specialties are re-earned, and that every operator is necessary — remove one and it breaks. This is blueprint C3, the "vertical."
Setup. A finite population of N=60 agents (genotypes of L=12 biallelic loci) evolves on a
rugged Kauffman NK landscape (K=8) that is reality. Each generation composes: grounding,
directed recombination (sex), quality-diversity selection, and mutation. Grounding is made
load-bearing by the consensus-conformity (self-consumption) mechanism — selection acts on
g·true_fitness + (1−g)·conformity, where conformity = agreement with the population's own consensus,
so at g=0 the society optimises fitting-the-crowd rather than reality. A 4-arm ablation, 12
replicate landscapes, 80 generations.
Symbols
- agent = a model (a genotype); reality = the NK fitness landscape; grounding
g= share of the selection signal that is real fitness vs conformity. - directed sex = recombine many parents + keep the fittest offspring (E10). quality-diversity = select for capability and novelty, so specialties persist.
- conformity − true fitness = how far the population's mutual agreement exceeds its real capability (the self-consumption "delusion" signature).
The three panels (over generations)
- The vertical climb. Best real capability: the full society (green) climbs to the global
optimum (≈0.78 of 0.79), while no_grounding (red) collapses to a fit-looking but poor
consensus (0.48).
no_sexandno_diversityplateau below the full society. - Specialties maintained. Population diversity: the full society explores longest (diversity
decays slowly as it searches), while every ablation collapses diversity fast —
no_groundingalmost immediately. - Self-consumption signature. Conformity minus true fitness: largest for
no_grounding— the population strongly agrees while being wrong, the signature of training on its own consensus.
Takeaway — every operator is load-bearing
- full → climbs to the optimum with diversity maintained. The vertical claim.
- no_grounding → self-consumption collapse to an unfit consensus (train on the crowd → regress to a confident, wrong mean).
- no_sex → can't recombine complementary specialists to escape local optima → plateaus below full.
- no_diversity (greedy) → collapses diversity fastest, gets stuck at a worse local optimum.
This integrates the whole study — E1–E6 (collapse/grounding/selection), the learning kernel, and E7–E10 (sexual transmission) — into one system, and shows the Lamarckian society needs all of grounding + directed sex + diversity to climb without collapsing. On a rugged landscape you need diversity to explore basins, sex to recombine them, and grounding to select on reality; remove any and you fail differently. Falsifier (not triggered): if any ablation had matched the full society, or the full society had failed to exceed every ablation, the integration claim would fail.