# Replacing crossover with independent migrants

Freeze before inference. Gemma 4 26B-A4B QAT, Ollama 0.40.13, temperature 0.8, thinking disabled, original policy language and fallback. Use the previous 16 training and six behavioral probe instances. Run eight new paired search seeds (113,127,149,167,181,199,223,241), two retention policies (best-only and best plus a behaviorally distant parent), and two operator schedules.

Control cycle: four one-coefficient mutations, one crossover, one independent migrant. Treatment replaces alternate scheduled crossover slots with the same parent-free migrant instruction. Alternate which slot is replaced first across the eight seeds. Keep the parent selector, mutation/migrant/crossover validity checks, per-step sampling seeds and all other generation parameters identical. Histories may diverge following the first substitution.

Every arm/seed receives 4096 full prompt-plus-output tokens. Invalid attempts, cached prompt tokens and shared calls are charged in full. Same 24-token reserve, 64-token maximum output, actual cost and remainder reported. Cheaper prompts can buy another attempt. A request with uncertain consumption halts collection. Freeze training-selected champions before generating the 192 new test schedules, half from each original workload family, disjoint from training, probes and the previous 96 test cases.

Primary effect: treatment minus control in mean held-out quality (100 times optimum/makespan), paired by seed and parent policy. Report each retention stratum and the equally weighted mean across strata. Secondary outcomes: distinct assignment signatures per search, actual operator counts, candidate count, spent tokens, and seed-level differences. The same test schedules are reused across arms. Eight independent search seeds limit generalization. An operator intervention can affect diversity, score and later search history; a score difference alone cannot establish diversity as its cause.