SaneCite
The checker

A citation is a claim about a source. SaneCite reads the source.

SaneCite takes a paper and checks its citations one by one. Does the source exist? Does it say what the paper says it says? Every verdict quotes the passage it rests on, and when it can’t tell, it says so.

How a check works

Four steps, every citation.

  1. 01

    Find

    Pull every citation out of the paper and give it a stable ID. The whole reference list, every time.

  2. 02

    Locate

    Find the passage in the cited source that the claim points at, by meaning, not keyword luck.

  3. 03

    Judge

    Compare the claim with the passage and give a verdict, quoting the passage either way.

  4. 04

    Review

    Disputed verdicts go to people. The finished record keeps its audit trail, mistakes included.

Verdicts

Three answers, and one of them is “I don’t know.”

✓SupportedThe source says it. Here is the passage.
✕Not supportedThe source says something else. Here is what it says.
?Cannot verifyThe source isn’t open, or the passage isn’t there. No guess.
What runs it

Models picked by a public test.

The reading runs on models served by Cloudflare. No model is trusted on its word: each candidate is scored against an answer key, and the winner is picked for accuracy first, then speed, then cost.

For the Phase 2 pilot, 12 models were tested on 118 citations labelled by a separate AI model (Claude), not by people. The best single model scored 93%. A panel of three (Cloudflare’s Clef, TypeSafe’s Jev and Kimi K2.6) scored higher, so the panel reads every citation and the majority wins. On 73 more citations that Claude labelled blind, the panel agreed 96% of the time. Human labellers check a fresh sample next. The full scorecard is public.

That keeps the cost of reading close to nothing. The expensive part, and the part your support pays for, is people checking the machine.

Open source

A checker you can’t check is just another authority.

So there is no black box. Every dataset is free. All the code behind Campaign One’s numbers is public under the MIT license, with the data, on GitHub: the Phase 1 count, the Phase 2 reading panel and the code that calls the models. The general checker’s own code follows under the same license when Phase 2 is finished. Don’t trust my verdicts. Re-run them.

SaneCite was first proven in security compliance, answering questionnaires only from approved evidence and declining whenever the evidence wasn’t there. That engine now reads the published record, with the same rule: cite the evidence or say so.

Status: being rebuilt with search by meaning instead of keywords. Its first public job is Phase 2 of Campaign One.