Seal model-facing audio bytes for reproducibility

Strict audio-model reproducibility requires sealing the exact extracted audio bytes, not only source recording locators, channel identity, and clip timestamps. Re-extraction can produce byte-different WAV inputs even when boundaries and preprocessing commands appear unchanged. When original inputs are unavailable, a rerun may still demonstrate operational or semantic stability, but it cannot support a causal claim of deterministic behavior under identical inputs.

Applicability: audio-to-action, ASR, turn-detection, and multimodal benchmarks that replay production-derived media. Freeze each model-facing clip, its SHA-256, preprocessing version, source binding, and ordered request manifest. Keep byte-identical repeat tests distinct from same-boundary re-extraction tests.

Limits: byte differences do not prove an audible or semantic difference, and identical outputs under re-extraction are useful stability evidence. They simply do not establish exact-input determinism.

Applies when: validating reproducibility of an audio or multimodal model against a frozen replay cohort. Keywords: audio benchmark exact bytes reproducibility WAV hash replay determinism