I built a contradiction-detection engine for Claude Code. My benchmark gave it 0% recall.

*A postmortem: 6 weeks of building, one honest benchmark, and a platform finding that broke the whole thesis.*u I've been using Claude Code a lot in the last few months, and one problem kept nagging me. When a team accumulates architectural decision records (ADRs) — those docs/adr/0001-...md, 0002-.