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.

Prefer the smallest ownership change supported by the evidence. Remove repeated work or reuse an existing indexed primitive before adding caching, concurrency, or a new abstraction. Keep a simpler design when a proposed optimization has no demonstrated benefit.

Measurement references

Consult only the material relevant to the investigation:

• [Measurement matrix](references/measurement-matrix.md): workload and metric

choices for search, context, reads, indexing, or caching experiments.

• [Repository measurement tools](../../../docs/measurement.md): existing

profilers and retrieval datasets.

• [Benchmark commands](../../../benchmarks/README.md): frozen comparisons and

optional paired-agent research.

• [Development checks](../../../docs/development.md): test ownership and CI.

Report the observed result and its limits. Run checks appropriate to the change; broaden them when a failure, risk, or unresolved question warrants it. Missing optional research data limits performance claims, not permission to finish a validated correctness fix.