acceptodds
Desk rejected at ICLR 2027

AmendGuard: Certify-Then-Decide Amendment Triage for Drafted Clinical Trials

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

An operationally attractive trial amendment can be scientifically inadmissible, while an admissible package can still be too weak to justify changing a drafted protocol. AmendGuard turns this distinction into a three-action structured decision: it searches coupled eligibility, visit-burden, and endpoint-specification edits, certifies searched packages against reconstruction, power, safety, and local-support gates, and surfaces the highest-utility package only when its lower-confidence benefit clears a separate intervention threshold; the other outcomes are protocol preservation and search-limited refusal. In method-blinded paired adjudication of 72 package, preserve, and refuse decisions, reviewers accept 59 AmendGuard outputs versus 47 retrieval outputs and prefer AmendGuard in 42 cases versus 12. On 117 historically amended trials and 39 matched no-amendment controls, AmendGuard attains 74.2% trial-macro directional concordance, surfaces packages for 3/39 controls, and has no surfaced package fail the encoded-gate recheck. Against decision-matched precedent retrieval, paired gains are 3.6 direction points (95% CI 0.8–6.7), 6.7 exact-match points (1.9–11.4), and 0.06 Jaccard overlap (0.02–0.10). Removing only the power gate gains 1.9 direction points but surfaces 19/156 packages that fail the locked power criterion; restricting the policy to any single edit family loses more than 21 points. The ordering persists in two disjoint later cohorts, where AmendGuard again records no encoded-gate violations.

Primary area:
causal reasoning
Keywords:
biomedical and clinical ai, AmendGuard

Related papers

Discussion (0)

Sign in to comment.