# Gemma scheduling policy pilot
Freeze study.py, this protocol, model identity and seeds before collection. Four paired search seeds, five arms, 4096 full input + generated tokens per arm/seed, charged including cached inputs, invalid operations and repeats. Stop if a complete prompt plus 24 output tokens cannot fit. Same model, temperature, starting policy and task distribution throughout. Independent calls with identical requests may be reused at full logical cost.

Each rule has four integer coefficients in [-2,2]. All arms can express the same 625 genomes. Search scores use the same 16 training jobs sets, nine jobs per instance with machine-dependent durations. Alternate independently sampled machine times and strongly specialized jobs. Behavior signature: assignments of all 54 jobs on six separate probes. Archive parent choice alternates highest fitness with the most behaviorally distant candidate from the leader, breaking ties by fitness. Best-only repeats the leader. Generic rewrites can change any coefficients. The targeted operator cycle requests four one-coordinate edits, a crossover and a migrant. Reject out-of-range, malformed or operator-inconsistent outputs; identical-parent crossover is allowed and counted. None of the arms receives test feedback.

Freeze each training champion before generating 96 independently seeded final instances. Primary contrast: mean held-out optimum/makespan of mutation/archive versus fresh, paired across seeds and instances. Also report all four factorial cells, effect of retention, behavioral coverage, cost, and a hand-built regret-first reference. Inspect transfer by workload family and parent/child differences descriptively. Evaluation calls use the same deterministic scorer and input counts per candidate; total evaluation work varies with number of candidates. Four search seeds limit inference. Test cases come from the same two synthetic families. Findings concern policy search in this restricted language.