v0.4.0

0.4.0 (2026-10-16)

Features

• bench: rank the cells we lose, which is v1-07's third criterion (#302) (e220730)
• bench: the execute_cell campaign_runner has never had (#289) (9a6c64d), closes #116
• bench: the live pilot runner, with the two refusals that stand before a bill (#294) (51d5959)
• bindings: TypeScript reference worker, worker-overhead harness and CI-run upstream evaluator (V1-25) (#239) (f744ba5)
• both pilot arms behind one seam, each accounting for what it actually spent (#292) (e5b7c18)
• cli: make a run's quality legible, and say when it is not trustworthy (#272) (934909c), closes #169 #170 #42
• cli: notice an imported model configuration that can stall a run silently (#267) (5b3cbce), closes #169 #42
• core: draw each proposal's model from weights on the run's own stream (#163) (#231) (444b5a2)
• deployment: have the promotion sweep emit the evidence it asserts (#284) (14e0fbb), closes #43
• dev-partition ablation for llm feedback, and the two defects it exposed (#246) (d8a6eda)
• eng: add the story-checkbox audit script and its measured baseline (#108) (#230) (ccb430f)
• engine: have the re-measurement sweep emit the evidence it asserts (#285) (b193f20), closes #24
• eng: measure the delivery boxes instead of annotating them by hand (#300) (ccd64cf), closes #108
• expose the judge knobs that decide whether feedback can act (#253) (3529e46)
• five v1 gaps the audits found, each measured rather than asserted (#301) (c5c4ecb)
• gate: require a 95% interval above 1.0 and a per-task regression limit, not just the mean (#297) (dda6bce)
• let a caller declare the environment a sandboxed evaluator sees (#245) (270ca44)
• measure whether the proposer can solve the ablation task at all (#256) (9e44328)
• one task both pilot arms can run, with the held-out split kept out of selection (#288) (0570c8c)
• paired ablation for meta-prompt evolution, and the two gaps it exposed (#249) (d41d8e1)
• programs: evolve the system instructions alongside the programs (#166) (#233) (818b642)
• ptx: evolve PTX kernel source with verified correctness and timing (#218) (ca96524)
• release: port the US-12 release claim gate to main (#240) (65ced88)
• seal a held-out split, score it independently, and measure whether the family can carry a claim (#282) (1b3c9fb)
• ship the observer that credits an evolving instruction (#250) (6fb9d24)

Bug Fixes

• bench: run the declared track panel instead of silently running all six (#287) (9189f22), closes #286
• ci: stop each merge to main cancelling the previous merge's run (#299) (a875c8d), closes #108
• engine: refuse a declared constraint violation in the default archive too (#259) (0a1a561), closes #21 #108
• evidence: stop an unrelated merge turning main red, and refresh the census (#277) (1c68d2c), closes #22
• execution: report the address-space backstop's kill as a memory violation, not a generic failure (#235) (b84f675)
• hand the openevolve evaluator a path a python literal can hold (#247) (55010f7)
• make a failing evaluator's own output reachable from the run report (#258) (1bae398)
• programs: stop the proposal options clone from dropping the declared descriptors (#243) (3adaec0)

Performance

• work: append-only group-committed journal for durable delivery (#241) (4cda0c9)

Documentation

• claude_code_quickstart runs unmodified with the real provider (#257) (b791d8b)
• correct the judge-score derivation, which used the wrong blend coefficient (#255) (568f356)
• engine: warn where the capacity is chosen, not only where it is tested (#273) (5e94baa), closes #24
• evidence: audit the parity stories, where 21 of 23 are closed with unticked criteria (#298) (0424022), closes #108
• evidence: re-measure our overhead column at main, meeting the last V1-70 target (#175) (#242) (de05cb5)
• evidence: re-measure the story-checkbox audit, and what the fall from 90 is not (#260) (e8f21c9), closes #108
• evidence: record the audit after the merges, 72 to 68 (#281) (03e6114), closes #108
• evidence: record the distributed run V1-61 asks for, measured (#174) (#238) (c458617)
• evidence: record what process isolation costs per evaluation (#181) (#229) (f0b1aee)
• evidence: the V1-01 checkbox audit at 0873a7d, 68 to 66 (#295) (0f8cc16)
• give the readme a path to a first report, and time it on a clean machine (#275) (b8996a4)
• parity: document each example's real-run command, and measure the claim about it (#296) (e981bcf), closes #170
• parity: record what the artifacts example needs, without unsmoking it (#244) (0c3a15a)
• record what V1-01's 72 open boxes actually are (#266) (0c6faf3)
• two more real-provider runs, and what each one actually optimised (#265) (ed3e786)