Contract Cards for Auditable Private Conformal Prediction

We propose contract cards for auditable private conformal prediction. Private conformal methods provide coverage and privacy guarantees, but deployment audit also requires explicit records of which requirements were imposed, which configuration was selected, and what evidence supports the decision. A contract card pairs user-specified coverage and privacy requirements with either an infeasibility report or a certified private conformal configuration, together with diagnostics that describe operational cost. We instantiate the idea with a transparent differentially private adaptive prediction set backend. This instantiation supplies the formal quantities recorded in the card: a finite-sample coverage lower bound, a calibration privacy guarantee, and a certificate-width diagnostic that tracks privacy-induced threshold inflation. Experiments on image-classification data show that the recorded fields satisfy the formal checks and vary predictably with calibration privacy and sample size, making contract cards useful evidence for lightweight audit.