Restore stable fusion verdict (a9fce62).

Fix false ELEVATED on broadcast and exam log crash.

Add full-frame pixel forensics for AI detection.

Add per-channel examination bands.

Add face-guided regional scoring pipeline.

Rebalance fusion bands for AI vs broadcast.

Harden findings assembly and isolate exam-log failures.

Cap visual-only broadcast spikes to INCONCLUSIVE.

Sync teacher-primary examination pipeline from monorepo.

Detect Kling and visual-only AI exports as SUSPICIOUS or ELEVATED.

Fix AI bare-metadata band classification and return all analyzed frames.

Add per-second timelines and friendly provenance labels for device and software stamps.

Speed up HF CPU analysis for short clips.

Stop weak heuristics from flagging edited news as SUSPICIOUS.

Downgrade re-encoded broadcast news from SUSPICIOUS to INCONCLUSIVE.

Reduce false ELEVATED on re-encoded broadcast news.

Add audio teacher and fused visual+audio EXM assessment.

Integrate Deep-Fake-Detector-v2 visual teacher and exam logging.

Dense frame analysis and court assessment band.

Forensic tiers: composite DNA, provenance, report package.

Frame-specific findings and peak-frame DNA.

Fix heatmap false positives and add neural patch heatmaps.

Per-frame heatmaps, visual fix, GRU audio loader priority.

Fix audio CNN+GRU in backend/app (Docker COPY path).

Fix audio CNN+GRU model and return all analyzed frames.

Remove Python cache files from HF Space repo.

Add real inference pipeline for visual and audio models.

Deploy Detect-X API with Docker and models

initial commit