Receipts Audit Skill

A document earns trust one sourced claim at a time. This skill takes a document plus the evidence behind it, extracts every factual claim, and grades each strictly against the provided sources — never against plausibility. The output is a claim ledger, a ranked list of the unsupported claims that matter most, and an honesty score whose method is shown, not asserted.

What This Skill Produces

• A numbered claim ledger: every factual claim, its grade, and the specific source line that supports or fails it
• Unsupported and contradicted claims ranked by how load-bearing they are to the document's argument
• A fix-or-drop recommendation per failing claim, with rewritten wording where a fix exists
• An honesty score (0–100) with the calculation method stated in the artifact

Required Inputs

Ask for these if not provided:

• The document — the report, deck text, post, memo, or page under audit
• The sources — the data, citations, quotes, or evidence the document claims to rest on. If none are supplied, run the extraction anyway, grade everything Unsupported (no sources provided), and say so plainly at the top
• Audience stakes (optional) — who reads this and what they'd decide from it; used for load-bearingness ranking. If absent, infer from the document and label the inference

Grading Framework

1. Extract claims. Every checkable factual assertion: numbers, comparisons, dates, quotes, causal statements ("X drove Y"), and superlatives ("fastest", "first"). Opinions and hedged forecasts are out of scope — note them separately, don't grade them.