Blind-taste-test workflow: two Gemini models grade Claude Code's video edits

aryanXmahajan · x · 2026-11-22

A practical QA workflow for AI-edited video audio: render 4 loudness-matched anonymized versions, send them to two independent Gemini models that rank them and flag exact failure timestamps, then have Claude Code verify each timestamp claim against word timings and real cut points, discarding hallucinated gaps and keeping the consensus best version. Blind listeners preferred the lightest noise reduction — heavy cleanup sounded watery and hollow despite better loudness metrics. Ship what sounds human to a cold ear, not what the meter prefers.

Original post →

More from coding & agent

coding & agent channel →