SaneCite · The open checker · Being open-sourced
A citation is a claim about a source. We read the source.
SaneCite takes a paper and checks its citations one by one. Does the cited source exist? Does it actually say what the paper claims it says? Every verdict shows its evidence — the passage, quoted. And when the evidence isn’t there, SaneCite says so: not supported, or cannot verify. It does not bluff. The refusal is the product.
How a check works
Step 1 — Extract
Every citation is pulled out of the paper and given a stable ID. Nothing is sampled. The whole reference list, every time.
Step 2 — Retrieve
Semantic search finds the passage in the cited source that the claim is actually pointing at — by meaning, not by keyword luck.
Step 3 — Judge
Claim against passage. Supported · Not supported · Cannot verify. The passage is quoted as the evidence, either way.
Step 4 — Review
Contested verdicts enter human review states. The finished record exports with its full audit trail — who checked what, and what the machine got wrong.
Built for questionnaires. Rebuilt for the literature.
SaneCite started as a commercial tool with one rule: answer security questionnaires only from approved evidence, cite the evidence for every answer, and refuse when the evidence doesn’t support an answer. The SaaS business is retired — nobody buys citation-honesty software. But the engine was always the interesting part. So it is being open-sourced, and pointed the other way: not at a company’s own documents, but at the published record. Same machinery — chunked corpus, scored retrieval, gated verdicts. Opposite direction.
The upgrade
Keywords out. Meaning in.
The original engine scored retrieval with hand-rolled keyword matching, because that was what the platform offered when it was built. The rebuild replaces exactly that one layer: passage-level embeddings (Cloudflare Workers AI embeddings + Vectorize) retrieve by meaning first, then the honesty gate judges support. Retrieval quality is the ceiling on a checker’s accuracy. This raises the ceiling. The parser, the chunking, the gate, the review states, and the audit trail carry over untouched.
A checker you can’t check is just another authority.
So: no black box. SaneCite is free forever, and open source — the checker, the campaign kit (manifest format + certification-pass runner), and every dataset it produces. Don’t trust our verdicts. Re-run them. Every claim a campaign publishes carries a stable address, so a result can be cited, reproduced, or refuted line by line. That is not a feature. That is the entire trust model: the house’s only moat is that anyone can verify it did the work.
Its first public job is already assigned: Phase 2 of Campaign One, classifying the context of every open-access post-retraction citation in biomedicine — in the open, where its mistakes will be visible too.
Status: In refactor — open-source release ships with Campaign One, Phase 2