Visual media
Art, UI screenshots, stylized portraits, social compression and unseen generators can shift model scores.
False-positive handling
A signal can be wrong.
Design for correction.
False positives are expected model failures, not edge cases to hide. Review source context before any consequential decision.
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.
Official trailers
Through the deployed lane, 3.5% of 200 official-trailer frames still drew false warnings. The set covered 70 trailers from 6 distributors unseen in training.
Sampled video
A clip is judged from at most five browser-sampled frames, so one stylized, compressed or graphics-heavy frame can move the aggregate — and moments the sampler never lands on cannot correct it.
Do not infer more than the lane measured.
- Do not treat a raw score as a calibrated probability.
- Provide correction and appeal paths wherever people can be affected.
- Prefer unknown over forcing unsupported content into human or AI.