Scientific discovery systems usually optimise experiments within a fixed hypothesis space, which creates a failure mode: when all candidate models miss the same mechanism, candidate disagreement can vanish even though the model class is systematically wrong. We therefore formulate experimental model‑class revision: a discovery policy jointly proposes a structural edit and a diagnostic experiment to test whether that edit is needed. The method couples a class‑level distinguishability objective with anytime‑valid sequential evidence that triggers structural revision only after the current class is rejected. In other words, a single shared parameterisation must explain all selected experiments.
On 400 held‑out controlled dynamical environments the joint policy achieves 89.5% exact recovery with a budget of only 32 real experiments, improving the strongest matched baseline by 10.0 percentage points while using fewer executed experiments and candidate fits. The learned revision‑experiment pairing transfers to unseen mechanism combinations, held‑out but expressible primitives, parameter extrapolation and shifted experiment costs; when the true mechanism lies outside the edit grammar it detects library insufficiency in 88% of cases with a 5.5% false‑support rate. Revision gains also transfer to ODEBench and ODEBase model‑library tasks as well as DiscoverPhysics worlds. These results support a view of scientific discovery in which deciding which mechanisms a theory should make expressible and where to collect evidence are treated as a single sequential decision problem.
Review