# AI Answer Verifier

> Split an AI answer into claims and see which ones your sources actually support. Paste an AI-generated answer and the sources it should be based on (text or public URLs). Each sentence becomes a claim; check-worthy claims (numbers, dates, names, superlatives, causes) are flagged, and every claim is matched to the best-supporting source sentence with BM25 scoring and number checks.

Part of usefulweb.site (https://usefulweb.site/). Results are deterministic calculations from the inputs and the listed assumptions; the API returns exactly what the web page shows.

## Use it
- Human interface: https://aianswerverifier.usefulweb.site/ (interactive tool at https://aianswerverifier.usefulweb.site/tool)
- JSON API: POST https://aianswerverifier.usefulweb.site/api/v1/calculate with a JSON object of inputs (schema: https://aianswerverifier.usefulweb.site/api/v1/tool)
- OpenAPI: https://aianswerverifier.usefulweb.site/openapi.json
- MCP (Streamable HTTP): https://aianswerverifier.usefulweb.site/mcp - tool `ai_answer_verifier`

## Inputs
- answer (textarea, required): AI answer
- sources (textarea): Source texts - Paste one or more sources. Separate sources with a line containing only ---
- sourceUrls (list): Source URLs (one per line) - Public pages fetched once (max 5).

## Method
- Claims = sentences of the answer with at least 15 characters.
- Check-worthy if the claim contains numbers, years, multi-word proper names, superlatives/absolutes or causal wording.
- Each source is split into sentences; the best match for a claim is the source sentence with the highest BM25 score (k1 = 1.2, b = 0.75) over content words.
- Coverage = share of the claim's distinct content words found in that sentence. supported: coverage ≥ 0.7 and every number in the claim appears in it; partially-supported: coverage ≥ 0.4; contradicted-number: coverage ≥ 0.5, a number from the claim is missing from the evidence and the evidence contains a different number; otherwise not-found.

## Limits
- This is lexical evidence matching, not truth verification: "supported" means similar wording exists in your sources, not that the sources are right. Paraphrases can be missed.