When May a Model Replace the Experiment? Audits, Licenses, and the Price of Trust in Surrogate-Driven Design

The Problem with Accuracy

In fields like materials science and drug discovery, machine learning surrogates are increasingly used to replace expensive experiments (oracles). A common, yet flawed, practice is to trust these surrogates based on aggregate accuracy metrics like R^2. The authors demonstrate that such metrics are structurally blind to decision quality: a model can have near-perfect accuracy while consistently selecting the worst possible candidates. This occurs because optimization pressure specifically targets the surrogate's errors, leading to a "selection tax" where the surrogate systematically over-predicts the value of its chosen candidates.

The Separation Principle

To ensure reliable results, the paper introduces an architectural rule: the "proposal" pipe (which uses surrogates to suggest candidates) must be strictly separated from the "certification" pipe (which determines if a candidate is truly good). The authors prove that admitting surrogate predictions into the certification process—treating them as if they were experimental measurements—leads to a deterministic failure mode where the model essentially confirms its own biases. Conversely, if certification relies solely on true oracle evaluations, the search is mathematically guaranteed to be safe, regardless of how the surrogate behaves.

Audits and Economic Efficiency

Since trust cannot be established by accuracy, it must be purchased through "selection-aware audits." These audits simulate the actual deployment process by letting the surrogate pick a champion from a candidate pool and then verifying only that champion with the oracle. The authors prove that this audit design is optimal in terms of query complexity. While auditing is an added cost, it enables a "dichotomy of economics": for generative search tasks, audited surrogates can reduce the total number of expensive oracle evaluations by a factor of up to 25 compared to un-audited or purely oracle-driven approaches.