Patpat Performance
When invoked directly, read the operating protocol and execution graph in full. Do not load the router.
Read proof over proxy, fix root causes, and smallest safe change. Apply the performance playbook.
Measure the reported surface before reading toward a preferred cause. Keep workload, environment, warm-up, sample policy, and units comparable. Locate the dominant cost, change one causal variable, and preserve correctness and safety invariants.
When the objective requires sustained improvement toward a numeric target rather than one regression fix, apply the metric hillclimb playbook through patpat-run.
Report the baseline, post-change result, measurement noise, artifact locations, and remaining tradeoffs. Do not claim improvement from a single incomparable or proxy measurement.
Proof closure
Close repository mutations through: