WHAT DOES A DIFFERENT COUNTERFACTUAL COST? LOSS-CALIBRATED AUDITS OF LEARNED COUPLINGS

Fitting outcome distributions does not determine individual counterfactual predictions. We audit that gap in an already trained flow. Gaussian-source rotations carried through the frozen flow change its pairings while preserving the complete conditional density path, and an exact quadratic identity prices finite alternatives under the original flow-matching objective, allowing query-specific search and independent evaluation without retraining. On a Kang stimulation task the released scCBGM-FM editor closes most of the distance between unedited and stimulated cells; an alternative gives up under one percent of that gain, costs a fraction of a percent of the fitted objective on both the global mixture and the edited condition, and still moves a majority of a 233-cell held-out panel by at least a tenth of a control standard deviation. Every cell keeps a positive predicted increment, so an aggregate reporting only the direction of effect records nothing, while the mean effect and the responder fraction both move. A costlier alternative, qualifying only at the secondary budget, reverses a responder call that four independently trained fits agree on. Learned Gaussian experiments check the accounting against a closed-form boundary price, which a strict reversal approaches without attaining, and show that the query determines which alternative is found and that optimizing an estimated price understates it. Good task performance and agreement across refits can thus coexist with sensitivity in individual predictions: audit the prediction the decision rests on, not only the aggregate reported beside it.