Sage-Taiyi Metacognitive Agent Tops AI Scientist-v2 on Taiyi-Bench, Pushes Lab Reaction Yields to 84%

Sage-Taiyi splits scientific search into an object level (target tree, idea graph, evidence state) and a meta level (cognitive state, cognitive repertoire, monitoring/control) with ten methodological heuristics. The architecture scores 54.0 overall on Taiyi-Bench against 47.9 for AI Scientist-v2, posts a 27.8% novelty jump, and wins 66.3% of head-to-head matches. In five expert-guided wet-lab collaborations it lifted cobalt-catalyzed carbonylation to 84% yield from trace, pushed photocatalytic allylic amination from 13% to 69%, and drove a magnesium alloy to 342 MPa tensile strength.