What problem does it solve? On shared multi-card Ascend NPU machines, performance numbers drift across measurement windows, so naive comparisons produce false conclusions about whether an optimization actually helps. This Skill defines the acceptance criteria that decide whether a measured speedup is real and which numbers are allowed into the summary report. ## Core Features & Use Cases - Two-state discipline: Separates exploratory readings (single-run, few-step, must be tagged [探索], never enter the summary table) from acceptance-state results (same-window control arm, repeated hot runs, full evidence) that may be written into reports. - Same-window A/B authority: Enforces that only same-window differences count as gains, with A/B/A drift correction when differences fall inside the noise floor, plus explicit noise-floor thresholds for when to declare "indistinguishable". - Triple-evidence rule: Every reported number requires HTTP 200 status, artifact byte count, and a server-side log line; md5 plus key-tensor dump dual evidence for lossless claims, with fidelity comparison against an independent ground-truth implementation to close the md5 blind spot. - Use Case: A teammate reports the new decoder made inference slower (15.005s vs 14.585s yesterday). Use this Skill to reject the cross-window subtraction, decompose the difference into drift versus structural cost, and require a same-window A/B or A/B/A run before any conclusion enters the report. ## Quick Start Ask the AI to verify whether a specific optimization's measured speedup on the shared NPU machine is real and eligible for the summary report.
