Which Conditions Matter? Task-Conditioned Criticality for Task Plan Veriffcation and Repair

Embodied plan verification must determine not only whether a condition is violated, but also whether that violation matters for the current task. We introduce TCCV, a framework for task-conditioned plan verification and repair. TCCV learns condition criticality from matched simulator interventions and combines it with current environmental evidence to decide whether to execute, observe, repair, or reject a plan. In controlled AI2-THOR experiments, we use structured simulator state to isolate criticality estimation from perception. On 282 score-independent ALFRED valid unseen contrast pairs, TCCV achieves 99.3% decision accuracy with zero false acceptances, compared with 90.2% for adapted VeriGraph and 73.4% for adapted VerifyLLM. A matched-plan diagnostic shows that performance depends on joint task–condition semantics: removing either input reduces ranking to chance. TCCV-guided repair completes 98% of tasks and fixes all 50 injected failures while preserving 96% of initially successful plans. These results support task-conditioned criticality as an auditable interface for embodied plan verification and repair under structured evidence.