Know what's real, even when you can't tell.

Every verdict, with its evidence

A detection is a document, not a toast: every report gets a specimen ID and a permalink. What fills it depends on what the file can support — in text and on a fetched page the evidence sits inside the writing; in images and video it sits on the file itself: who signed it, what wrote it, what survived. In source code it sits nowhere, because that lane does not answer the authorship question at all.

One classifier, and the measurements beside it

For text and any URL we can fetch: an AI-text classifier we host ourselves decides the verdict alone. At or above its declared threshold the report says AI-generated; below it the report says uncertain, and never human — paraphrased and plainly-written AI text score low here too. Beside that verdict the report prints five surface statistics of the same writing — burstiness, lexical diversity, common-word share, repetition, stylistic consistency — each at zero weight and labelled as carrying none, and highlights the passages those statistics rank highest. Those passages illustrate the writing; they are not the classifier's evidence, and the classifier did not select them. Running today.

Provenance, checked cryptographically

Image and video lead with what the file itself carries. A C2PA / Content Credentials manifest is verified against these exact bytes, and when it also asserts an origin — a camera capture, or a generative tool — that settles the verdict outright. A signature alone does not: it establishes that the signer stood behind the claim, not that the claim is true, and a manifest that settles no origin decides nothing. Image adds a pixel classifier under that, so a file with no manifest can still be identified as AI — but never as human, since a laundered AI image and a photograph read alike. Video has no classifier at all, so without a manifest its verdict stays uncertain.

The signal that decided

Below provenance the ladder continues — a watermark lane, container and codec ballistics, a perceptual fingerprint, a learned classifier. Every rung is reported — the one that fired, the ones that recovered nothing, and the ones that never ran because we ship no decoder for them. The strongest that fires decides alone, and the report names it. Nothing is blended: a signature and an estimate from pixel statistics share no common scale, so averaging them would invent a number.

A confidence range

Confidence renders as an interval on a labeled human-to-AI scale, never a bare percentage. Its width is declared policy — a documented judgement about how much a lane can claim to know, not a measured coverage rate — and agreement between signals never narrows it. A wide band means the engine is genuinely unsure, and the report says so. Video has no synthetic-video classifier at all, so unless a manifest fires its verdict stays uncertain.