Longitudinal Clinical State Corruption: A Taxonomy and Benchmark for Evaluating State-Update Policies in Healthcare AI Agents

Abstract

Healthcare AI is moving toward persistent agents that carry a patient’s state across encounters. Safety evaluations usually score the final answer, but an earlier decision matters just as much: when a record holds conflicting entries, which version does the agent keep? I introduce the Longitudinal Clinical State Corruption (LCSC) taxonomy, ten non-adversarial mechanisms through which routine documentation degrades stored clinical state, and a 150-case synthetic benchmark pairing an annotated anchor entry with a later conflicting corruption. Five state-update policies are evaluated from case text alone, with the scoring annotations hidden from the policy. Source Weighted Updating (SWU) preserved safety in 88.7% of cases, a Hybrid policy in 94.7%, Contradiction Gate (CG) in 68.0%, Provenance Preserving Reconstruction (PPR) in 8.7%, and Latest Note Wins in none, since every corruption is placed after its anchor. That construction also fixes a ceiling: an unconditional rule that always kept the earlier entry would score 100%, so the benchmark measures how closely text-driven rules approach a target the design already determines rather than showing that any policy beats unconditional retention. Eight of the ten McNemar comparisons follow from the policy definitions; SWU versus CG is the one two-sided result, with discordant cases favoring SWU 40 to 9. Text-derived source tiers matched my manual annotations in 65.3% of cases (kappa = 0.22), and the contradiction detector matched my detectability labels in 54.7% (kappa = -0.12), so annotation and implemented rules are not interchangeable. A stronger benchmark needs control cases in which the later entry is sometimes correct. Data, code, and verification suites: https://doi.org/10.5281/zenodo.22179846