ADR 0053: Prevent restart recipe replay and repeated exact-tree evaluation

Date: 2026-08-06

Context: A long QD campaign repeatedly generated and evaluated equivalent candidates. The sampler's seeded RNG lived only in one scheduler process, so each restart replayed its initial choices. Historical (base, inspirations) recipes were not excluded, and agents could reproduce an inspiration tree exactly. Archive-level executable identity protected MAP-Elites from some duplicate admissions, but it acted after generation and evaluation.

Decision: Derive sampler randomness from the configured seed, island ID, and persistent per-island job ordinal. Persist an order-insensitive recipe hash and cool down a bounded number of recent recipes, with bounded resampling and an explicit unavoidable-reuse flag. Tell planning and coding agents that inspirations are evidence rather than target snapshots. Before evaluation, reuse a prior passed result only for an exact Git tree under the same evaluator name/version and campaign program. Keep evaluator-provided candidate identity as the archive admission invariant.

Consequences: Scheduler restarts no longer replay the start of the random stream, common lineage recipes are less likely to recur, and exact source-tree repeats avoid another benchmark. The cooldown is not a permanent ban: small archives can continue making progress and expose reuse in provenance. Source-distinct candidates that compile to the same binary remain separate before evaluation because source-tree identity is not a substitute for evaluator-defined phenotype identity.