Verification methodology

Statistical detectors estimate how likely a passage is machine-written from patterns like word predictability. They produce a probability, not evidence, and they misclassify non-native writing and heavily edited text often enough that no serious decision should rest on them alone. We run one such detector locally and, on request, cross-check with Pangram — and we say so plainly: agreement between detectors is a stronger signal, never proof of authorship.

Provenance means a marker was deliberately attached at the moment content was produced. It can be verified rather than estimated. Its weakness is the opposite of detection's: absence proves nothing, because a marker can simply never have been applied, or have been lost in re-encoding.

When the Wmarks detector and an optional Pangram cross-check flag the same passage, that's a stronger signal than either alone — but still not proof of authorship. When they disagree, that's meaningful too: different models with different training data get different passages wrong. Either case is a reason to look closer, not a verdict.

Wmarks does not claim a detector finding proves authorship. It does not claim that the absence of provenance evidence means a human wrote something — most AI-generated content carries no verifiable marker at all. And it does not claim to have 'understood' why a model flagged a passage when no provider supplies a reason: what's visible is only which detectors flagged a spot, never why.

A watermark is a pattern embedded in the content itself — in token choice, pixel values or an audio signal. It survives some editing. Verifying one requires the key or detector held by the system that applied it, which is why several of our vendor checks report 'Not verifiable'.

Metadata is data alongside the content: EXIF, XMP, container tags. It records which device or tool wrote a file. It is easy to read, easy to edit, and routinely stripped by platforms on upload. Useful as a lead, weak as proof.

C2PA defines a signed manifest attached to a file describing how it was created and edited. Because it is cryptographically signed, a valid manifest is strong evidence. We report whether a manifest is embedded; validating its signature chain against the C2PA trust list requires a validator service, and we label results accordingly rather than implying we did more than we did.

• Verified — a technical fact or validated claim was directly established. A verified formatting or container fact does not automatically indicate synthetic origin.
• Declared — the content or producing tool states an origin. It is useful evidence, but editable metadata is not cryptographic proof.
• Forensic — a model may report a pattern or probability. This remains an indication and is never converted into verified provenance.
• Not verifiable — the required public verifier is unavailable, the credential was not validated, or the check could not complete. Unknown is not negative, and absence is not human authorship.

We do not offer watermark removal, provenance stripping, metadata scrubbing for evasion, or rewriting intended to defeat detection. Where a signal can be located, we show where it is — so it can be understood, not erased.