Avoid Single-Run Performance Conclusions

PERF-AVOID-SINGLE-RUN-CONCLUSIONS

Summary

Do not land performance conclusions from one short benchmark run. Repeat and contextualize timing evidence because warmup, scheduling, cache state, and background load can make a single result non-reproducible.

Rule

Do not decide performance from one short benchmark run.

Why

One short benchmark run can be dominated by warmup, scheduling, cache state, background load, or measurement noise. Landing a performance change from that evidence risks adding complexity for a result that will not reproduce.

Helps

  • Keeps performance claims reviewable and avoids optimizing noise.

Limits

A single run can be useful as a smoke check or exploration signal. Treat it as a prompt for better measurement, not as landing evidence.

Agent Instruction

Do not decide performance from one short benchmark run because one short benchmark run can be dominated
by warmup, scheduling, cache state, background load, or measurement noise.

Mechanisms

Supported by repeated benchmark runs, Criterion reports, hyperfine, saved benchmark commands, and PR notes that include baseline and changed results.

References