v0.3.0

0.3.0 (2026-10-13)

⚠ BREAKING CHANGES

• EvolutionDeploymentDecision.Outcome is EvolutionDeploymentOutcome, not string.

Features

• classify the v1 public api and enforce it (V1-82) (#206) (9cfb696)
• cli: a live, self-refreshing run report (V1-59) (#196) (95d85ba), closes #172
• cli: run openevolve configs and evaluators unmodified (V1-56) (#207) (de8810c)
• enforced resource limits and multi-host distributed evaluation (V1-61) (#205) (d220e82)
• engine: drive deadlines and delays from a replaceable clock (V1-80) (#190) (ac1c68e)
• engine: store large and binary artifacts in full (V1-58) (#188) (5b1fbb8)
• programs: feed experience lessons into proposals (V1-52) (#194) (58cfe7b)
• programs: model provider parity with openevolve (V1-51) (#189) (1396ee0), closes #164
• programs: promote metrics and the score to archive descriptors (V1-55) (#186) (feefca0)
• programs: weighted model ensembles (V1-50) (#195) (389c735)
• ship net462 and netstandard2.0 assets for AiDotNet.Evolution and Surrogates (#224) (29917c4)

Bug Fixes

• resume a budget-truncated batch exactly by carrying its evaluator calls (#210) (b813204)

Performance

• checkpoints: segmented, streamed checkpoints for long runs (V1-73) (#209) (d821729)
• engine: cheaper checkpoint saves, and checkpoints beyond 16 MiB (V1-74) (#198) (08ac70b)
• engine: keep a slow model busy with overlapping proposals (V1-75) (#199) (9feb16d)
• engine: keep per-evaluation cost flat as the archive grows (V1-72) (#201) (0de1a75)
• engine: stop rebuilding search genomes on every validation (V1-71) (#197) (9e27171)
• engine: valid overhead intervals against openevolve, and a state hash that stops pinning memory (V1-70) (#200) (65ef0d2)
• programs: sandbox overhead and warm python workers (V1-76) (#208) (59bd34b)