CARTOGRAPH framework helps AI scientists know when to stop experiments

A new paper from researchers Neel Tushar Shah and Manglam Kartik introduces CARTOGRAPH, a verification layer designed to make AI-driven scientific discovery more reliable. The system couples three key mechanisms: unresolved-subspace experiment steering (select), explicit ambiguity closure (resolve), and residual-based library inadequacy detection (refuse). Under a local linear-Gaussian bridge, CARTOGRAPH-A emerges as the exact unresolved A-optimal rule, with closed-form EIG and Box-Hill serving as local comparators rather than global equivalents.

Across five testbeds, CARTOGRAPH-A beat raw projection with a record of 129 wins, 0 ties, and 15 losses at dimensionality d=8 (p < 10^-21). The framework demonstrated a unique ability: it tentatively identified three out-of-library pharmacokinetic mechanisms but later revoked those identifications as residuals exposed structural misfit, while one perturbed in-library control stayed identified. In low-dimensional pharmacokinetic and filtered EPA settings, near-ties against disagreement matched theoretical predictions. A retrospective audit of 40 positive claims from A-Lab flagged all 4 claims later marked inconclusive while passing 32 of 36 confirmed claims. Code is available on GitHub.