Optimize Accuracy First

Use this skill for a performance investigation or an optimization whose benefit needs measurement. Ordinary bug fixes do not need a performance study.

Locate the owner of the work and state the expected improvement before changing it. Preserve the applicable contracts: deterministic results, exact token budgets, Unicode matching, truthful coverage, bounded work, atomic generations, and snapshot consistency. An intentional contract change should be described and tested as such.

Choose evidence that answers the question. Behavioral regressions and differential checks establish correctness; candidate counts can explain work; representative release measurements support latency, CPU, memory, or storage claims. A smaller local timer does not justify worse evidence or more retries. Do not invent agent-success or provider-cost results from retrieval proxies.