validate-scores

What problem does it solve?

This Skill removes the manual effort of comparing model scores across multiple backends by automating parity checks, score aggregation, and numerical difference reporting.

Core Features & Use Cases

• Multi-backend validation: Runs the same reference and distorted inputs through every enabled backend and collects per-feature scores.
• Parity analysis: Computes pairwise absolute differences and ULP distances to expose floating-point drift between implementations.
• Regression verification: Helps confirm whether a new SIMD path, GPU backend, or hot-path change preserves expected scoring behavior.
• Use case: Use it when a backend update needs proof that score outputs remain within the accepted tolerance budget before merging.

Quick Start

Ask the assistant to validate a reference and distorted clip pair across the selected backends and report any score differences that exceed the allowed ULP threshold.