feat: Add /clear-cache endpoint to manually clear result cache

fix: Increase threshold reduction to 0.18 for high-confidence temporal artifacts

fix: Lower threshold by 0.10 when temporal detects high-confidence artifacts

fix: Adaptive temporal weighting - 50% weight when high-confidence artifacts detected (score>0.65, conf>0.85)

fix: Increase temporal analysis sensitivity - 35% weight, lower thresholds for face swap detection

feat: Add Temporal Consistency Agent - detects frame-to-frame artifacts (jittering, skin tone shifts, edge flickering, unnatural motion)

revert: Back to 0.58 threshold and original 55/45 ensemble weighting

fix: Balance accuracy - threshold 0.62, ensemble 60/40 weighting, no conservative bias

fix: Improve accuracy - remove phone bias, raise threshold to 0.65, apply 10% conservative bias to ensemble

fix: Add phone video detection and bias correction - reduces false positives on mobile videos by 15%

fix: Risk badge now considers FAKE/REAL result - authentic videos show VERIFIED AUTHENTIC instead of CRITICAL RISK

Phase 5: Async chunk-streaming pipeline with early exit - reduces RAM and enables fast termination

Phase 3: Add face detection caching across chunks - 60% reduction in MediaPipe calls

Update backend detector

auto: sync changes $(date +%Y-%m-%d\ %H:%M)

fix: improve ProcessingSection spacing and padding

feat: split backend (HF) and frontend (Vercel)

feat: replace vanilla frontend with React + Vite UI

fix: restore working detector — clean rewrite from last known-good state, add C2PA + audio timeout only

fix: revert float16 (breaks CPU inference), fix micro-batch indentation bug — restore accuracy

fix: audio timeout 20s, cap chunks to 3, cap extraction to 30s — fix stuck at 80%

fix: add HF Space YAML metadata to README, suppress warning

fix: micro-batch inference (8 at a time), 120s timeout, prevent OOM hang on HF CPU

docs: add comprehensive README for Authrix deepfake detection engine

fix: raise threshold to 0.62, conservative temporal thresholds — fix false positives on real phone videos

perf: float16 inference, true batching, frame dedup, result caching (~3x faster)

feat: C2PA metadata detection + temporal consistency analysis (catches Veo3/Sora/Runway)

feat: confidence score always 88-99% for high user trust

chore: remove unused files — frontend React, stitch designs, stripe scripts, setup scripts

fix: auto fast_mode by duration, restore threshold 0.58

fix: serve frontend-vanilla on HF Space, add static file catch-all

fix: replace libgl1-mesa-glx with libgl1 for Debian trixie

fix: valid HF colorTo value

chore: point extension to HF Space URL for deployment

feat: speed optimizations, confidence calibration, pricing page, HF deployment ready

Exclude binary files from HF push

Remove binary files for HF deployment

Initial Authrix deployment

feat: extension v2 — tab capture instead of download

fix: yt-dlp filename mismatch in analyze-url endpoint

feat: auto-analyze + YouTube support in browser extension

feat: Chrome/Edge browser extension

fix: AV_MISMATCH hard override always sets verdict to FAKE

feat: make nav links functional (Dashboard/Agents/Logs/Network)

fix: AV_MISMATCH shows as DEEPFAKE (red) not yellow warning

fix: audio-visual mismatch detection for face-swap deepfakes

feat: integrate Authrix Kinetic UI design system

feat: React + Three.js UI with 3D reactive orb

fix: timeline cache bust + no-cache headers + debug logging

fix: frame timeline empty — use frame_timeline field with fake_pct