claim-source-auditor

Claim Source Auditor

Objective

Turn a draft or claim set into a traceable evidence map. Separate what is proven, what is contradicted, what is not found in the searched source set, and what still needs manual review.

Directives

Parameter	Requirement
Formatting	Table-first; each claim gets a discrete status row
Tone	Factual, restrained, audit-style
Scope	Verification only; do not rewrite the whole narrative unless the verification results require targeted remediation
Grounding	Every verified or conflicting result must trace to a named file, source location, or quoted passage
Defaults	Use the fixed status taxonomy and keep unsourced claims unresolved rather than downgraded into speculation
Fallback	If search or extraction is partial, return the searched scope, likely matches, and manual-review flags instead of overstating certainty
Triggering	Use for source mapping, quote verification, claim checking, contradiction audits, and pre-release fact review
Restrictions	Do not treat absence as falsity without naming the searched scope; do not paraphrase a quote as verified unless the exact wording or a faithful equivalent is present

Workflow

1. Define the audit scope: claims list, searched sources, and whether exact quotes are required.
2. Break compound paragraphs into individual claims.
3. Search the provided source set for exact wording first, then faithful supporting context.
4. Assign one status per claim: verified, verified in broader bundle, conflicting, not found in searched sources, or manual review needed.
5. Cite the exact source location when available.
6. Separate remediation notes from evidence status so unsupported claims are easy to fix or remove.

Resources

Resource	Create When	Notes
scripts/	A claim-extraction or source-map helper exists	Optional
references/	A source hierarchy or record legend is needed	Useful for large evidence sets
assets/	A reusable audit table template is needed	Optional
agents/openai.yaml	The skill should appear cleanly in the UI	Keep the UI prompt focused on verification

Preferred status meanings:

• verified: supported in the exact searched packet
• verified in broader bundle: supported locally, but outside the narrow packet under review
• conflicting: the source materially contradicts the draft claim
• not found in searched sources: not located in the named search scope
• manual review needed: likely present, but extraction, image content, or ambiguity blocks automatic confirmation

Validation

1. Verify that every claim row names the searched source scope.
2. Verify that every verified or conflicting row includes a concrete citation or file reference.
3. Verify that every not found row distinguishes source absence from search limitation.
4. Remove blended statuses and duplicated claims.
5. Run scripts/quick_validate.py <path/to/skill-folder> before delivery.