About this instrument

The AI Safety Corpus is a living, weekly-snapshotted collection of AI safety research writing — Alignment Forum, LessWrong, the EA Forum, arXiv, lab blogs, and government publications — behind a hybrid search index and a grounded assistant. It is a research instrument in private beta, not a product.

How answers stay honest

• Retrieval-only synthesis. The assistant answers exclusively from passages retrieved out of the corpus. It has no other knowledge source at answer time.
• Verified citations. Every sentence carries citations whose quotes are checked, server-side, to be exact substrings of the stored corpus text. A citation that fails the check is marked unresolved.
• Flagged sentences. A sentence with no verified citation is displayed with a visible unverifiedunverified treatment — never silently blended into the answer.
• Abstention. When retrieval can’t support an answer, the assistant abstains and says why, rather than guessing. The full reasoning trace (queries, hits, evidence, model pin) is attached to every answer.

Permalinks and snapshots

The corpus is an append-mostly series of dated snapshots. Permalinks — /d/ documents, /u/ passages, /a/ answers — resolve from the published store independently of deploys. A superseded document redirects to its current version; removed documents render an honest tombstone, never a silent 404.

Logging and budget

Queries are logged anonymously: hour-bucketed query text and result counts, with no accounts, cookies, IP addresses, or any identity. The assistant runs under a hard daily spending cap; when it is exhausted the assistant pauses until 00:00 UTC while search and dossiers stay available.

Licensing and takedown

Corpus text remains the copyright of its authors and venues; the instrument quotes short passages with links back to the source. If you author an indexed document and want it removed or corrected, use the takedown contact in the footer — removals propagate at the next snapshot and tombstone the permalinks.