A lightweight, reproducible PoC series for studying recursive self-improvement under endogenous yardstick drift, using Gemma 3 local models to separate proxy-score drift, proxy gaming, delayed audit, and strict-confirmed improvement.
ai proof-of-concept replayability ai-evaluation ai-governance recursive-self-improvement auditability yardstick-drift benchmark-drift evaluator-drift proxy-metrics proxy-gaming delayed-audit
-
Updated
Mar 17, 2026 - TeX