Ramp selection and sharper target-sensitivity bounds

Develop an optimal ramp-selection rule and derive sharper sensitivity bounds for smooth shared-target reliability certificates.

Background

The smooth shared-target certificate replaces the expectation-only Markov contamination allowance with an additive concentration radius based on a registered ramp score. The ramp widths control a trade-off: wider ramps reduce target sensitivity but can lower favorable-score rates, whereas narrower ramps approach hard thresholding but may produce a larger concentration allowance.

The paper does not resolve how to select ramp widths optimally or whether sharper sensitivity bounds can substantially improve the certificate. These issues are explicitly identified as open in the limitations discussion.

References

Ramp selection and sharper sensitivity bounds remain open.

— ProxyGuard: Direct Reliability Inference for Randomized Data Release Mechanisms with Shared Targets  (2608.18643 - Mahato et al., 19 Aug 2026) in Section 5, “Limits of the claim”

The current experiments do not establish when it improves on the expectation-only certificate.

— ProxyGuard: Direct Reliability Inference for Randomized Data Release Mechanisms with Shared Targets  (2608.18643 - Mahato et al., 19 Aug 2026) in Appendix, Section “A two-axis block certificate”