How MissingProof Works

Paste an AI answer and the documents it should be grounded in. Here is exactly what happens next.

1. Your answer becomes claims

The answer is decomposed into atomic claims, and each claim is further split into its facts (the individual assertions) and its reasoning links (the relations asserted between them — causal links, attributions, implications, certainty levels). This decomposition is the foundation: fabrications overwhelmingly live in the links, not the facts, and systems that check whole sentences miss them.

2. Facts are checked against your evidence — with quotes

Every fact is scored against your documents, and support comes with the verbatim evidence span that establishes it. No support means the fact is flagged; opposing evidence means it's marked contradicted. You never have to take a score on faith — the quote is right there.

3. Reasoning links are checked two ways

Each link is scored separately for direct support (does any passage in your documents explicitly state this relation?) and inferred support (does it genuinely follow from the supported facts alone?). Two true facts sitting near each other never validate a causal claim between them — that distinction, scored explicitly, is what catches the failures aggregate tools launder.

4. Every failure gets a locus

Flagged claims are diagnosed, not just graded: PREMISE (a fact lacks support), BRIDGE (facts fine, reasoning fabricated), or CONTRADICTION (your documents say otherwise). The locus determines the fix — retrieval and documentation problems route differently from over-synthesis problems.

5. The missing proof

For every fixable gap, MissingProof generates the exact statement your sources would need to contain to make the claim supported, plus the document type where such a statement would plausibly live and who in an organization typically owns it. Contradicted claims are marked unfixable-by-evidence — the claim itself must change. Your report ends as a work plan, not a grade.

What it costs to run

Anonymous visitors get 3 free analyses a day. A free trial key by email unlocks 25 analyses over 14 days — no card. Paid plans start at $49/month for 500 analyses; details on the pricing section. Your pasted content is processed for analysis and not shared; analysis requires sending text to our model provider under their API terms.