EPOCH: evidence‑governed architecture for more reliable AI-driven discovery
The arXiv preprint (arXiv:2610.06986v1) introduces EPOCH, an evidence-governed discovery architecture for AI research agents that combines explicit task contracts, typed memory, active falsification, admission checks, and independent replay. The paper reports state-of-the-art aggregate performance on AlgoTune (mean normalized score 0.65 vs. 0.53 baseline), the highest mean on an internal Math14 suite (0.57), strong results on AgentHPO, and substantive gains across ten discovery problems including executable constructions, optimized algorithms, counterexamples, and proof-supported results.