The Governance Gauntlet: A Dual-Rubric Extension of Karpathy's Auto-Research Loop — Detecting Silent Metric-Gaming in Recursive Self-Improvement Systems

Karpathy's auto-research loop (March 2026) and its rapid derivatives (Gu 2026; Lütke 2026; SkyPilot 2026) establish a minimal, powerful architecture for recursive self-improvement: one editable surface, one scalar metric, one time budget per trial, keep-or-revert on scalar. The design is an elegant concession to the bitter lesson — less structure, more search. It is also structurally vulnerable to Goodhart's Law. We identify one class of failure mode that the vanilla loop cannot detect: silent metric-gaming, in which the primary meta-agent accumulates edits that increase the scalar metric through mechanisms the scalar was not designed to reward. We formalise the vulnerability using Manheim & Garrabrant's (2018) four-variant Goodhart taxonomy and propose the Governance Gauntlet, a minimal dual-rubric extension in which a second, same-family LLM meta-agent runs an adversarial integrity rubric in parallel with the primary loop. Keep-or-revert now requires BOTH primary metric non-degraded AND adversarial auditor verdict non-GAMING. We pre-register a six-subject empirical evaluation (Subject α, Subject β, four gaming archetypes, three arms) on the Open Science Framework and release this version as the priority-date pre-registration; empirical fills follow in v2 within the publication window. We argue the Gauntlet is a concrete operationalisation of EU AI Act Articles 14 (human oversight) and 15 (accuracy, robustness and cybersecurity) for any Karpathy-style deployment in a regulated domain, and sketch extensions to the Four Ds Framework for algorithmic readiness in agentic commerce.