Start with a measured bottleneck and a fixed workload. Change one variable, repeat enough runs to see variance, and keep correctness, validation, and logging intact.

Objective

reduce a measured bottleneck.

Context

profile, baseline, workload, latency budget.

Decompose

reproduce, instrument, form hypothesis, change one variable, benchmark, compare variance.

Criteria

improvement under the same workload without correctness regression.

Security

do not disable validation or logging without explicit review.

Rollback

retain benchmark and revertable change.

Output

baseline, method, result, confidence interval or repeated-run spread.

Handoff checklist

  • Baseline is reproducible
  • One variable changed
  • Variance is recorded
  • Correctness is unchanged
Editorial recipe adapted from the project’s September 2026 research brief. Commands, authority and acceptance tests must be matched to your repository. No production execution is implied.

Use the bounded agent contract before starting ↗