Yield accuracy is not enough: diagnosing greenhouse world models across four evidence axes

Yield accuracy can make greenhouse AI appear decision-ready even when management actions are ignored. In public Autonomous Greenhouse Challenge records, a cherry-tomato yield model reached an R² of 0.851 after future actions were removed, and 46 of 50 shuffled-action tests matched or improved the full-action baseline. We use this schedule-memory trap to motivate a four-axis evidence contract: endpoint prediction, state rollout, action response and generalization. Across three hourly greenhouse records spanning cucumber, cherry tomato and dwarf tomato, this contract provides a reproducible template for auditing greenhouse AI before it enters production decisions.