Snotra: Continually Adapting Irrigation World Models under Physical Change

Physical agents must revise imperfect world models as the systems they control change. However, persistent prediction error alone does not reveal which change is identifiable, which revision deserves belief, or whether that revision matters for action. We study this problem in precision irrigation, where partially observed soil–crop–weather dynamics interact with constrained, long-horizon interventions. We present Snotra, an evidence-governed framework that separates world-model revision into three decisions: KNOW, whether competing explanations are distinguishable from available evidence; BELIEVE, whether a candidate survives validation, falsification, physical-consistency, safety, and provenance checks; and ACT, whether an admitted model changes a legal decision and improves an independently evaluated physical outcome. We further introduce SnotraBench, a stage-resolved evaluation protocol for world-model revision under physical change. On 35 matched governance units, Snotra accepts all 10/10 supported revisions while rejecting all 25/25 unsupported ones; Heldout-Only accepts the same 10/10 supported revisions but also 10/25 unsupported revisions. The paired reduction in unsupported admissions is 0.286 (95% cluster-bootstrap CI [0.143, 0.429]). In a one-check-removed replay, identifiability and physical-consistency are the only checks whose removal increases false admissions on this scenario set. Finally, a controlled drift sweep exposes a prediction-to-decision threshold: prediction discrepancy increases as physical drift grows from 0% to 50%, but the irrigation action remains unchanged through 30%; at 50%, revision changes the action by 10 mm and reduces independently evaluated cost by 110 protocol units. These results show that predictive repair, model admission, and decision value are distinct quantities and motivate evaluating adaptive world models by both admission selectivity and prediction-to-decision transmission.