The Generator Is the Shift: A Cost Audit of Conformal Molecular Search

Optimizing a molecular surrogate changes the distribution on which its uncertainty must be useful. This need not worsen prediction errors, and a valid but uninformative bound does not establish that the predictor deteriorated. We audit this distinction with a controlled stochastic proposer over four molecular lookup pools, holding the predictor and outcomes fixed while varying selection pressure. Across 120 random-forest and 60 extra-trees realizations, we compare source calibration, exact-ratio weighted conformal prediction, fresh proposal calibration, and a finite-budget tolerance audit. Random-forest optimism risk falls on all four pools, but weighted conformal bounds become infinite on 44–50% of proposal mass at maximum pressure. Extra trees reverse the risk trend on two pools. Fresh-audit and draw-matched uniform-calibration deployment rates cross across tasks and learners. The audit either certifies the existing lower bound, recalibrates within a declared slack limit, acquires more audit labels, or stops. Its classical order-statistic certificate controls optimism risk simultaneously over the six tested proposal laws and four audit looks. The contribution is an empirical accounting of prediction risk, calibration support, usable bounds, and paid evidence—not a new conformal guarantee. All methods, proof, data identities, numerical results, and diagnostics are contained in this paper. Experiments are fixed-library simulations; no de novo synthesis or therapeutic efficacy is claimed.