Article 50 provenance does not survive the real world — we measured it

The experiment

ProvBench is a pre-registered harness. We embed a provenance manifest in an asset, then apply a sequence of ordinary transforms — a re-encode, a resize, an edit, a re-upload. We inspect the asset after each transform for whether any recognisable provenance survived.

The signed result

We embedded 20 manifests. Across every measured transform chain, 0 of 20 survived. That is not a claim about our skill at embedding; it is a statement about how fragile the current article-50-homologous marking surface is once content is really free to move.

This is the kind of finding that has to be a number before it can be a policy. The full set, with fingerprints, is published as part of the ProvBench instrumentation and the signed result record.

What it means

• Provenance that only lives in the file header is not protection against the copies that matter.
• If the rule is "AI content must be machine-readable as AI content," then a marking that does not survive a resize is not meeting the rule in the real world.
• Measurement is the only honest way to know which marking scheme is worth relying on.

Boundary

This is a measurement of our test assets under our transform set, not a legal opinion and not a finding about any deployed system. It is, though, the sort of evidence the Article 50 obligation will eventually be argued against.