What problem does it solve? When you dispatch a subagent to fact-check a document against sources but hand it an abridged or partial source dump, the agent confidently reports claims as "unsupported" or "fabricated" — even though the supporting text was in the section you trimmed. This Skill prevents those false positives and shows how to reconcile flagged claims against complete sources. ## Core Features & Use Cases - Prevention protocol: Pass complete primary sources (full files, cloned repos, retrievable originals) to verification subagents instead of hand-written summaries, or explicitly declare which sections were cut so the agent marks affected verdicts UNVERIFIABLE rather than INACCURATE. - Response protocol: Before accepting an "unsupported / fabricated" verdict, check the agent's report for hedges about partial sources and reconcile the flagged claim against the complete source yourself. - Parallel fan-out variant: Covers the case where N parallel verifiers each receive a different source subset, causing cross-cutting claims anchored to out-of-slice sources to be falsely flagged — with mitigations like shared provenance indexes and inline source-anchor tags. - Use Case: A planning-review document is fact-checked against two cloned repos and a hand-transcribed Google Doc dump; the agent flags one claim as fabricated, but the supporting text was in the abridged section — the claim stands and the lesson is to attach the real document. ## Quick Start Before dispatching a subagent to fact-check a document, attach the complete original sources and ask it to mark any claim touching missing sections as unverifiable rather than inaccurate.
