Write a PREreview

Purpose. Persistent health AI may produce a proposed write from an authorized longitudinal-health disclosure and attempt to persist that proposal after clinical state, consent, role, or policy has changed. Rechecking only the current record and current authorization does not establish which exact governed disclosure supplied the evidence and governance context used during inference. We evaluate whether that disclosure can remain a verifiable dependency of later write admission. Methods. GLHS implements a disclosure-to-commit contract. A Task-Bounded Health State Snapshot (THSS) persists the actor-, purpose-, task-, temporal-, consent-, and evidence-scoped disclosure supplied to inference. A Governed State Transition (GST) admits a persistent proposal only after revalidating the bound disclosure and current state/governance. Primary evidence comprises a deterministic conformance matrix, a matched exact-binding ablation over 320 frozen logical schedules, 12 governance/state time-of-check-to-time-of-use (TOCTOU) schedules repeated with jitter, and a profile-global versus entity-partitioned versioning contention check. Model-context utility is secondary and is evaluated in prospectively frozen synthetic cohorts. Results. In the matched ablation, an otherwise governed arm without exact disclosure binding admitted all 256 binding-specific invalid schedules, whereas the exact-binding arm rejected all 256 and admitted all 64 clean controls. This result is interpreted as deterministic mechanism/conformance evidence on the frozen attack families, not as an estimate of real-world attack prevalence. Across 600 repetitions of 12 governance-writer schedules, no forbidden commit was observed; 10/12 logical schedules matched their frozen classifications in every repetition and two retained classification mismatches. Entity-partitioned versions eliminated the tested false-stale mechanism for disjoint partitions through 128 synchronized writers. In a later 384-subject synthetic comparison, Strict THSS versus full authorized history differed by $-0.781$ percentage points (95% bootstrap CI $[-6.77,+5.21]$), which does not establish the prespecified $\pm2$ percentage-point equivalence margin. Mean prompt tokens decreased from 3,280 to 412; measured provider latency also decreased, but that latency effect is benchmark-specific rather than a controlled systems-speedup estimate. Conclusions. The strongest supported claim is a software-governance property: on the evaluated snapshot-bound path, accepted persistent proposals remain tied to a specific persisted governed disclosure while current state and governance are rechecked. The experiments do not establish clinical benefit, regulatory compliance, optimality of the THSS representation, correctness for arbitrary unbounded schedules, or universal downgrade resistance across every downstream workflow.