citation-check

Citation check for AI output

Paste a document. You get back mechanical convictions with the line each one sits on. No style advice, no rewriting.

Check a document here by hand — or run the same checks in a pull request (GitHub Action) or from an agent (MCP). Same engine, same refusals, same free ceiling.

Seven checks run on every document, and each answers one question that no other answers:

1. Enumeration and manifest — the document against the counts and listings it makes about itself. “Four phases” over two; a listing that disagrees with its contents; a total graded better than its worst part; one marker used for two meanings.
2. Internal reference — a pointer against the thing it points at. “See Section 4” where there is no Section 4.
3. Arithmetic — a stated total against the figures beside it. A sum, average or interval that does not match the numbers in its own sentence.
4. Count against list — a count introducing a list against what is in that list. “Two changes:” over one bullet.
5. Declared bound and form — an instruction the document gave itself against what it produced. “In two sentences:” over three.
6. Ordinal sequence — a named sequence against its own order. First, Second, Fourth.
7. The document against its own principles — a rule the document stated against what it later does with it. A term defined one way and later defined another.

There are three verdicts, not two, and the third is the one worth reading closely: a conviction (a located defect), a _clean_result (the check ran, applied, and found nothing), and a refusal — the check did not judge, and says why. A refusal is not a finding of guilt and not a finding of innocence. Every answer states what did not apply and what was never examined, because a checker that quietly skips something is worse than one that finds nothing.

What it does not do: it does not read meaning. It cannot tell you a claim is false, a principle unwise, or an order wrong. Those are readings, and this pass refuses them by design rather than by omission.

It also reports whether the sources your document cites can actually be retrieved by an automated reader, and why not where not. That is not the same question as whether a claim is true, and the report says so.

Free: 3 checks per day, and 3 documents per day for the source retrieval, counted separately. Up to 200 KB per document. Your document is checked and discarded — never stored. We keep one anonymous daily count of how many documents convicted, and nothing else. These checks are pure parsing: no AI model reads your text on this page.

Check it Or try an example

The example is our own document, built to fail two of the checks — a count whose members do not match, and a reference that resolves to nothing. It is a demonstration, not a typical result.

Weakest Mark · a verdict service for machine-made claims. Need more than three a day? A prepaid pack of the same mechanical checks is on the pricing page. Packs also run thegrounding pass — we fetch the source your document cites and compare the claim against what it actually says. It is under development and we say so; there is a real recorded run on that page so you can judge it before paying.
We ran these checks on a handbook of our own and published everything they raised — including the one finding we got wrong and had to retract.
Impressum · Datenschutz / Privacy