kolesnikov-arch/patchward

Evidence home for the verdict-layer approach: pre-registered, reproducible trust evaluation of AI-generated code changes - silent wrong ships 17/50 → 0/50.

2 best points · 1 comments · 2 Hacker News posts · 1 GitHub stars · View on GitHub

Hacker News discussions

• Show HN: I pre-registered the rules,then measured AI agents shipping wrong fixes (2 points, 1 comments, by kolesnikov-arch, 2026-07-06)
• Of 1190 AI agent closing statements, one reports failure (1 points, 0 comments, by kolesnikov-arch, 2026-07-27)