From upload to verdict: what Tayanch actually checks, and why you can trust an honest "uncertain" more than a fake certainty.
01
Send data
Upload an image, video, or audio file — or paste a public link (a post, an article, a direct image URL) — to the /analyze endpoint family.
02
Model scores it
Vision-transformer embeddings and physical-consistency checks (camera physics, compression traces) combine into a single calibrated score. Signed Content Credentials (C2PA) are verified first — cryptographic proof beats pixels.
03
Get a verdict
A clear verdict with the confidence behind it. Ambiguous or heavily-processed inputs return an honest "uncertain" instead of a confident wrong call.
Built to be honest.
Detection is probabilistic — pretending otherwise is how people get falsely accused.
Calibrated thresholds
Verdicts are tuned per use case, so you can trade recall for a lower false-positive rate. A real photo being wrongly flagged is the failure we optimize hardest against.
Degradation-aware
Screenshots and messenger re-compression destroy forensic signal. Tayanch detects laundering and widens its uncertainty band instead of guessing.
Scope, stated
Video analysis covers face manipulation and carries a frame-level advisory for generated footage — and says so, instead of overclaiming.