EU AI Act · Article 53(1)(d) evidence
Since 2 August 2025, every general-purpose AI provider selling into the EU must publish a public summary of its training content on the Commission's mandatory template, and keep it updated. Those pages are edited in place; the old versions vanish. BotWitness records the disclosure surfaces daily: hashed, RFC-3161 timestamped, tamper-evident.
The point most people miss
When we began recording these surfaces in July 2026, major providers' transparency hubs had yet to link a template-form summary. Every dated capture showing "not published" (or an earlier revision of a disclosure) is a fact about the past. Once the page changes, no one can go back and prove what it said, or didn't say, before. We can.
How it works
01 · Record
We fetch the Commission's template page and each provider's disclosure surface every day and store the exact bytes returned, content, status code, or block page. Dynamic page noise is filtered so the archive keeps the real revisions and discards the churn.
02 · Seal
Each capture is SHA-256 hashed and bound to an independent RFC-3161 trusted timestamp; every check joins a Merkle digest whose root is timestamped too, proof of proof that the monitoring ran without gaps.
03 · Prove
Ask what a provider had published (or not published) on a specific date, and what changed between two dates. The record is neutral: it cuts the same for complainants and for providers.
Who needs it
Complaints and "qualified alerts" to the AI Office turn on what a provider disclosed and when. A dated, tamper-evident revision history of the disclosure (from a neutral third party) is the exhibit.
The same record proves the other direction: that you published on time, updated on schedule, and said what you say you said. Independent corroboration beats your own CMS logs in front of a regulator.
Track how training-content disclosures evolve across the industry, which sources appear, which quietly disappear between revisions.
The deliverable · $199
One branded, filing-ready PDF (plus machine-readable JSON) covering every watched disclosure surface, or just the providers you name.
Included
Every content change of each disclosure surface, dated, SHA-256 hashed, RFC-3161 timestamped, including the entering baseline and what the surface returned on days it published nothing (403s and 404s are dated facts).
Included
Time since each surface's last revision and the longest observed gap, measured against the six-month update yardstick in the Commission's template guidance, with staleness flagged, not editorialized.
Included
Optionally supply up to ten terms (your domain, an imprint, a dataset name) and the report shows whether each appears in the latest published disclosure text, with context excerpts.
Included
Every capture cited in the report can be escalated to a Merkle inclusion proof or a full evidence package, the same archive discipline as the rest of BotWitness.
Live now
Coverage is free to inspect; capture history follows the same access rules as domain history.
« what GPAI surfaces are watched, since when » curl "https://bot-witness.com/watch?kind=gpai" « the dated change history of one surface » curl "https://bot-witness.com/watch/history?kind=gpai&key=ec-template"
Early access
Enforcement begins in August 2026. The archive of what was disclosed before then is being written now, and it cannot be back-dated. Reports are $199 self-serve; for a watchlist extension (your providers, your models) or recurring delivery, talk to us.
