Policy-Induced Blind Spots in Symbolic Action-Model Repair

An embodied agent that plans with a symbolic action model can in principle use its own execution outcomes to detect and correct defects in that model. We show this is not always possible, and characterize exactly when it is not. Across three STRIPS domains and six single-literal fault types, we corrupt one action schema and ask whether execution reveals the defect. For a spurious (over-restrictive) precondition, a passive policy – one that only ever attempts actions it currently believes are applicable – exposes the defect in 0 of 200 confirmation-split trials at a fixed interaction budget, because the defect itself is what prevents the policy from attempting the action that would refute it; this is a deterministic consequence of the policy definition, proved as Proposition 1 and confirmed empirically with a 95% Wilson confidence interval of [0, 1.9%]. A diagnostic policy that instead deliberately attempts actions the agent’s own model currently doubts – using only the agent’s own belief model, current state, and hypothesis set, and verified from code to never access the true environment before acting – exposes the same defects in 197 of 200 trials (95% CI [95.7%, 99.5%]) at an identical budget. This replicates, in a repair setting with exact injected ground truth, the established principle that active querying resolves model ambiguity passive observation misses [6]; our contribution is the sharper, budget-independent statement that for this fault type passive exposure is not merely rare but provably impossible, and a matching negative control showing the same diagnostic policy does not improve exposure of a different fault type (delete-effect defects: 9.0% passive vs. 12.5% active, overlapping confidence intervals), confirming the mechanism is specific rather than a general exploration bonus. All experiments use symbolic simulators with 4–7 objects and a single injected fault per schema; we do not validate on continuous or physically embodied execution.