EXP-RULE-MULTIGEN

v2 (exp-rule-proposal-v2) showed axis-diversification working but acceptance rate weak (1/6 across 2 runs). Single-generation testing cannot reveal compositional growth — by definition the rule lineage is depth ≤1. Multi-generation extension: rules survive across generation boundaries. Gen N+1 starts with rules accepted in Gen N. Lineage depth tracks how many prior rule modifications shaped the context when a rule was proposed. Falsifiable Bedau OEE test: by gen 5, max axis-depth ≥3 AND ≥2 axes show depth chains ≥2 → compositional growth (positive). Pre-registered predictions: - Positive: max_depth_overall ≥3 by gen 5 (compositional rule modifications) - Positive: rule space EXPANDS — different rule combinations explored across gens (vs collapse to single attractor)

• Positive: rules stabilise within gen but evolve across gen boundaries - Negative: rule space collapses to single attractor by gen 3 (no diversity) - Negative: rules cycle without depth (depth stays ≤1, oscillation only) - Negative: zero proposals accepted in any gen (acceptance rate too weak even with carryover)

Discriminating tests: - Mirror: rules return to training distribution by gen 5 (depth saturates at small) - Generator: depth grows monotonically; novel rule combinations not in training persist (fitness via interaction, not retrieval)