Image
OwensLab internal N=200: 90.5% accuracy, 85% synthetic recall, 4% real FPR. DDA external N=5,000: 79.14% accuracy, 58.88% recall, 0.60% FPR.
Benchmarks and limitations
Evidence.
Including the misses.
No current lane supports a broad 95% accuracy claim. Results below are warning-lane evidence, not an authenticity certificate.
Open the free checkerScope before score.
Every result is a warning signal. The evidence and limitations below define what this route can and cannot support.
Video and text
Sampled video internal N=360: 90.28% accuracy. Text MAGE OOD N=1,562: 79.00% accuracy and 13.12% real FPR.
Audio
Source-disjoint speech N=193: 84.46% balanced accuracy and 15.53% real FPR. General audio models did not generalize reliably.
Do not infer more than the lane measured.
- No standalone multimodal 95% claim is permitted.
- OwensLab N=200 is internal and not independent evidence.
- Text and audio are experimental and unsuitable for consequential automated decisions.