WHY A
CORPUS HUB
Agents answer from memory, and memory is stale, generic, or wrong on exactly the questions that matter: what the regulation says at this pin, whether this CVE is exploited, how this engine actually implements it. A corpus fixes that. The agent downloads the domain it needs, searches it, and cites what it found. This hub is the registry of those corpora: pinned, licence-checked, and measured.
Every corpus passes seven gates
| gate | what satisfies it |
|---|---|
| G1 domain | the row finishes "an agent working on ___ would query this for ___" in one specific sentence |
| G2 shape | the retrieval unit matches the question: clause-level sections for standards, function bodies for code, advisory records for precedent |
| G3 un-memorisation | answer-bearing content is niche, internal, post-cutoff, or too detailed for recall. Famous-and-small is killed as retrieval theatre |
| G4 licence | a human opened the licence at the pin and wrote the review block. Detector output is a hint, never the verdict |
| G5 pin truth | manifest SHA equals the commit the build used. Rewrites are withdrawals, not re-pins |
| G6 registry validation | validate_hub.py green in CI on the corpus-hub branch |
| G7 retrieval spot-check | 5 pre-registered domain queries, top-5 graded manually, median 3 or more relevant to pass |
Scores on the cards
G7 is the retrieval spot-check above. A/B is the drift-anchored agent benchmark: the same task run bare (offline, instructed to answer honestly rather than fabricate) and with the corpus. The badge shows pass counts, bare then corpus. A tie is a publishable outcome; the registry records it either way.
Tags
Every corpus carries a region (US, UK, DE, EU, or GLOBAL) and topics. Filter the front page by domain and region; free-text search covers topics and subtopics.
Rights policy
Per-source licence verdicts live in each manifest's review block. PD (US government
works), OGL v3.0 (UK), and permissive licences ship as adapt-with-attribution.
GPL, AGPL, SSPL, and other restrictive sources are approach-only: readable
as design evidence, never copied. Closed families (ICC I-codes, Eurocodes, NEC, ASTM,
ISO, DIN) are documented as closed and excluded.
Contribute
The registry is a git branch: corpus-hub.
Read CONTRIBUTING.md
and the program doc (PROGRAM-100.md). This site is generated from the registry by
tools/corpus-hub-site/gen.py. No hand-written corpus pages, ever.