Position: Toward a Computational Theory of Scientific Discovery

As AI improves, which scientific problems become solvable next, and when? We argue that a computational theory of scientific discovery should connect capability growth to target success, the availability of a suitable procedure, and validated completion. A worked assay scenario shows the consequence: the first procedure to become available can have a lower chance of delivering a result by a common deadline. Two retrospective studies test what information improves target predictions. Across 73 models and two reasoning families, object count improves pooled held-out prediction but worsens larger-task deduction; it does not isolate computational complexity. In 550 chemical-table replays, early outcome summaries help only in some endpoint–budget settings. A retrospective diagnostic finds that separate budget-specific fits do not remove the longer-budget reversal, while accurate aggregate success rates hide opposing errors across sources. Exact conditional calculations explain which measurements constrain availability dates and why validation and evaluation costs matter. Together, these comparisons motivate a research programme for choosing informative measurements and comparing research procedures. They establish conditional consequences and retrospective prediction boundaries; prospective completion dates remain unvalidated.