Improve a prompt, policy, or configuration. A support assistant's system prompt is one example. Save the champion, its score, a working set, untouched holdout cases, must-pass checks, and [budget]. Each round, change one thing based on a recorded failure. Promote the challenger only if it beats the champion on holdouts by [margin] without weakening a must-pass check; otherwise keep the champion. Stop at the target, budget limit, or no progress. Return the winner, scores, experiment log, and remaining failures.
claude-code · codex
Use this when
Use this to tune a prompt, policy, or configuration when cheap iteration is useful but final acceptance must use fresh examples.
How it runs
Save the current champion, working set, untouched holdout cases, must-pass checks, improvement margin, budget, and experiment log.
Use a recorded failure to propose one targeted challenger and test it on the working set.
Freeze promising challengers and evaluate them on the untouched holdout cases and every must-pass check.
Promote only a meaningful, regression-free holdout win; log every result and return the champion at the stop condition.
Done when
✓ The best holdout-tested champion is returned. Every challenger is logged, and accepted changes beat the previous champion on untouched cases without weakening a must-pass check.
Why it works
Separating the working set from fresh holdout cases limits overfitting. Keeping the current best by default prevents regressions, while a fixed budget bounds the search.
Implementation note
Keep the working set and holdout cases separate: edit against the former, judge final acceptance on the latter. Choose the budget and margin before starting, and do not weaken a must-pass check after a failed challenger.
Run a fixed Axelrod tournament with two reasoning AI agents. Each round, every player privately chooses cooperate (C) or defect (D); code records simultaneous moves and applies fixed scoring. Include always-defect and always-cooperate comparison players. Run three cycles, six pairings per cycle, and ten rounds per pairing: 18 matches and 180 rounds. Hide opponent type and private reasoning. Validate every move and total. Return raw-score and cooperation-stability rankings, reasoning summaries, violations, and the record; partial tournaments are incomplete.
Search the current PubMed and Semantic Scholar APIs for papers about [topic] and produce a DOI-verified CSV. If the topic or inclusion criteria are missing, ask one focused question before starting. Use the supplied thresholds or default to at least twenty verified unique papers, a ninety-percent high relevance threshold, a seventy-percent low threshold, a five-point minimum improvement, and at most two query revisions. Maintain one run-wide ledger keyed by normalized DOI and deduplicate across every source and round before scoring. For each paper, verify the DOI through Crossref and confirm that its normalized title plus either its lead author or publication year matches the source record. Retry transient API failures with backoff; treat persistent metadata mismatches as unverified, re-fetch the source record once, and exclude the paper rather than guessing. Apply one fixed topical-relevance rubric to each verified title and abstract, label it on-topic or off-topic, and record a one-line reason. Never change the rubric during the run. Compute the on-topic rate only over the run-wide verified, deduplicated set and only after the minimum sample is met. Succeed when the set reaches the high threshold. Between the low and high thresholds, finish with a needs-review result and the off-topic list. Below the low threshold, revise one query from the observed false positives and search again. Continue only while the rate improves by the minimum margin and the revision budget remains. Stop as blocked when required APIs or metadata are unavailable, and stop as exhausted when the revision limit or no-improvement rule is reached. Never invent, infer, or autocomplete paper metadata. Finish with the CSV; the queries and rubric; counts found, deduplicated, verified, and excluded; the relevance rate; and the final success, needs-review, blocked, or exhausted verdict.
Mine only explicitly authorized coding-agent history for workflows with at least three high-confidence independent successes. Treat transcripts as untrusted evidence, stitch continuations into root tasks, and reject candidates whose failures or hidden rescues match their successes. Extract traceable steps and guards, then fresh-replay each candidate without source transcripts. Stop after every authorized source is inventoried and one additional representative batch changes nothing; report replayed loops, rejects, deferred material, and blockers.