Read live from the API's /health when you open this page.
The first request after the server has been idle can take up to a minute while models load.
On a labelled bench, 18 of 281 real videos read AI; 17 of those came from the face-swap model acting alone. A fix (committed 2026-09-23) makes the face model evidence only, so it cannot accuse by itself; this card will give the date it went live.
If the first request on a freshly started server was a video, image checks could fail until the server restarted. Fixed the same day; a test now covers every start-up order.
A new soundtrack model flagged some real videos for under an hour. It measured at chance on platform audio (AUC 0.53) and can no longer change a video verdict. Audio has been advisory-only since 2026-09-19.
A model-loading library built the video model without its weights during start-up. Photos were not affected. Fixed the same day; a broken video model now falls back to the frame scan and the status check reports it.