The corpus as a dataset

The complete capture index is published as machine-readable JSON: one record per stored version with provider, model, capture time, source URL, SHA-256, size, and provenance status.

ledger.json — 150 records since 11 Aug 2026, regenerated on every build. Metadata is CC0; the full corpus (raw bytes, manifests, OpenTimestamps proofs, event log) lives in the public repository.

Fields

Top-level keys: generated_utc, license, source (site origin), repository, records.

• source_id, provider, model — what the capture belongs to
• kind — one of provider-live, provider-page, aial-archive, watch-page, regulatory, cop-doc (labels as on the site: provider site / provider page / AIAL archived copy / watched page / official document / Code of Practice doc)
• captured_utc — record-write timestamp (UTC, ISO 8601); it can trail the HTTP fetch by seconds (for fanned-out captures, up to ~90s) — the version page's Fetched field is authoritative
• url — the fetched location; expiring signature tokens are masked with …, so masked URLs are records, not resolvable links
• sha256, size_bytes, text_sha256 — content identity; text_sha256 is null when no text could be extracted
• ots, wayback — provenance status
• permalink — the version page serving the stored bytes and proof (site-relative; prepend source)

Scope of the changes feed

The changes feed covers content changes to already-tracked summaries (same target, new content). Newly published summaries appear on the status page when discovered.

Cite this archive

Cite version permalinks directly — they are stable. Suggested form: GPAI Ledger, “<Model> training-data summary, version of <date>”, <permalink URL>, accessed <date>.

For legal or licensing matters requiring verifiable copies of specific captures, write to contact@gpailedger.com.