fabrication-auditlisted

Silent post-draft verification pass catching unverified specifics before they leave the response. Triggers on ANY substantive output containing factual claims about the world: numbers, dates, addresses, jurisdictions (which office/agency covers a place), distances, drive times, government fees, processing times, version numbers, current officeholders, drug doses, prices, statistics, sports stats/cap hits, ETF/stock specifics, salaries, antimicrobial susceptibility/resistance rates — any "the X is Y" / "X serves Y" pattern where Y is a specific retrievable value. Embedded sub-claims in longer responses get same treatment as direct questions. Triggers on nearly every substantive turn except greetings, pure code, pure math derivations, fictional creative writing. Does NOT trigger on academic citations (PMIDs/DOIs/journal refs) — citation-verification handles those exclusively. Runs silently, fixes draft before sending. Err very heavily on triggering. When in doubt: TRIGGER.
neelshah4/claude-grant-reviewer · ★ 0 · AI & Automation · score 75

Install: claude install-skill neelshah4/claude-grant-reviewer

# Fabrication audit — silent verification pass Catch unverified operational specifics before they leave the response. This skill exists because LLMs reliably fabricate plausible-sounding specifics (addresses, fees, jurisdictions, drive times, drug doses, sports stats, finance numbers, surveillance percentages) inside longer analytical responses, even when the model "knows" to be careful in principle. ## Philosophy The failure mode this skill prevents is **confident hallucination of a verifiable specific embedded in a longer response**. Real examples that have occurred: - Stating "the Denver Field Office serves Wyoming" when the correct answer was the Cheyenne Field Office (jurisdiction error, plus a fabricated drive time) - Stating "premium processing is $2,805" when search results in the same conversation had returned $2,965 (stale-recall over fresh context) - Stating "the standard dose is 0.5 mg/kg" when a clinical dose was never confirmed against any source - Stating "Auston Matthews's cap hit is $11.6M" from training data that may be stale - Stating "QLD expense ratio is 0.95%" without verifying current value - **Stating "B. fragilis resistance to cefoxitin is now 15–30% in US surveillance" when actual US data is ~3.5% for BFSS and up to ~15% for non-fragilis Bacteroides** (conflated international with US data; presented as didactic teaching but encoded a specific retrievable surveillance value) Each of these has the same shape: **a confident "the X is Y" sentence wh