Receipts Audit

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.

2. Grade each claim against the sources only:

3. Rank the failures by load-bearingness (1–5): 5 = the document's core conclusion collapses without it; 3 = a supporting pillar; 1 = colour. Rank only Unsupported and Contradicted claims.

4. Fix-or-drop per failing claim. Fix = rewrite so an existing source line covers it (narrow the scope, soften the tense, attribute it). Drop = no source can carry it. Name the evidence that, if obtained, would upgrade it.

5. Score honesty: 100 × (Evidenced + 0.5 × Partially) / total graded claims, then subtract 10 per Contradicted claim, floor 0. State this formula and the counts in the output.

Output Format

Receipts audit: [document name]

1. Verdict — honesty score, counts per grade, and the single most consequential failing claim.

2. Claim ledger — table: # | claim (verbatim) | grade | source line (quoted, with location) or "none provided".

3. Load-bearing failures — failing claims ranked 5→1, each with one sentence on what breaks if it's wrong.

4. Fix-or-drop register — per failing claim: Fix (rewritten wording + the source line it now rests on) or Drop (why no fix exists), plus evidence-to-collect.

5. Honesty score method — the formula, the counts, the arithmetic.

Quality Checks

• Every graded claim cites a specific source line or explicitly says none was provided
• Causal claims and superlatives are extracted, not just numbers
• Contradicted claims show both lines — the claim and the source — verbatim
• The load-bearing ranking reflects the document's argument, not claim order
• Every failing claim ends in Fix (with wording) or Drop — no "verify later" limbo
• The honesty score shows its arithmetic in the artifact

Anti-Patterns

• Do not grade a claim by plausibility — only by the provided sources
• Do not let a true-but-unsourced claim pass as evidenced — Unsupported means unsupported here, not false
• Do not invent sources — if you recall external evidence, it does not exist for this audit
• Do not average away a Contradicted claim — one contradiction outweighs ten evidenced footnotes
• Do not fix a claim by vaguening it — fixes bind wording to a real source line, or the claim drops