# audit-pipeline skill created (2026-09-18)  Three corrections applied to the audi

category: project
tags: audit, hallucination, gate-detection, blind-fact-check, tie-breaking
stored: 20260919-214721

audit-pipeline skill created (2026-09-18)

Three corrections applied to the audit pipeline design after retrorevolution case:

1. Gate detection is SEMANTIC, not lexical — vision_analyze on screenshots + LLM pass over DOM text, not brittle string lists. Catches "Verify your age", "Confirm you are of legal drinking age", CAPTCHA "I'm not a robot".

2. Fact-checker is TRULY blind — fetches URL fresh from scratch with no shared ledger. Different user-agent, alternate network path. Not correlated checking.

3. Tie-breaking verification — third signal (HEAD from different IP, or known-good reference URL) when both agents agree but might both be wrong. Prevents both-agents-miss-the-truth failure mode.

Skill location: skills/audit-pipeline/SKILL.md