Changelog
Model versions and product updates
Every report records the model version that produced it. When the model changes, it ships as a version bump with an entry here, so an older report stays interpretable.
Initial preview release.
Four-class audio-visual forensic analysis: real or synthetic, scored per track.
Per-segment timeline with a sync-coherence score.
Calibrated confidence band on every result, stated before the class.
Every report includes tagged frames, attention heatmaps, and the analysed spectrogram.
REST API with polling, plus PDF and JSON export.
Single-track fallback when only audio or only video is usable; an explicit unclassifiable state when neither is.
