TRW: An Auditable Consistency Contract for World Models as Materialized Views

TRW frames learned world representations as materialized views over a changing physical world whose base-state reads are costly, delayed, heterogeneous, and fallible. It proposes a commitment-level validity abstraction, consequence-conditioned adaptive view maintenance, dependency-scoped transaction-style compensation for commitments invalidated after authorization, and append-only provenance for exact replay. The Flood-SAR evaluation treats sensing as physical data acquisition and reports measures including freshness, verification cost, stale reads, recovery scope, restoration failure, and replayability across six preregistered questions with held-out seeds. The paper presents a systems contract rather than a new predictive model.