Karpathy's Loop Meets Goodhart's Law: Introducing the Governance Gauntlet for Auto-Research

Andrej Karpathy's auto-research loop works brilliantly. As it spreads from ML training into high-stakes domains, a minimal dual-rubric governance companion becomes necessary. Introducing the Governance Gauntlet — a second-meta-agent audit layer that catches silent metric-gaming that a single-scalar loop cannot. Pre-registered empirical study, open-source reference implementation, and EU AI Act Articles 14/15 mapping inside.