When Should an AI Scientist Stop? Verifiable Experiment Steering and Refusal for Autonomous Discovery

In three linesCARTOGRAPH is a verification layer for autonomous AI scientists, combining experiment steering, explicit ambiguity closure, and library inadequacy detection. Across five testbeds, CARTOGRAPH-A outperforms raw projection (129W/0T/15L, p<10^-21). The refuse guard correctly flags all 4/4 claims later marked inconclusive in manual A-Lab reanalysis.

Summary generated by Claude — human-verified