What problem does it solve? When you replace an operator, fuse kernels, change parallel partitioning, or alter compilation paths, you need to prove the change did not alter results. This Skill provides a rigorous acceptance standard that distinguishes bit-exact equivalence from approximate similarity, preventing silent numerical regressions from being shipped as "lossless" optimizations. ## Core Features & Use Cases - Layered Equivalence Definition: Classifies changes as L1 bit-exact (torch.equal per shape), L2 numerical gate (explicit thresholds with justification), or L3 lossy gate (quality metrics plus visual review), forbidding mixing of levels. - Three-Stage Acceptance Sequence: Enforces same-config rerun bit-exactness, cross-config numerical gates with md5 artifact comparison, and quality gates for lossy tiers, with mandatory evidence records. - Silent Failure Localization: Provides a systematic debugging workflow (CPU ground truth, per-layer divergence bisection, operator isolation, threshold sweeps) for wrong results with no errors or NaNs. - Use Case: You rewrote a deconvolution that takes 97% of decode time and the output md5 changed with max difference 9.5e-7. The Skill tells you this is not lossless, requires explicit L2 downgrade declaration or rejection, and directs you to check index/phase/order bugs rather than precision. ## Quick Start Ask the AI to verify whether your operator replacement is truly lossless and what acceptance evidence you need to collect.
