Simulation-Based Verification and Bounded Repair of Adaptive Inventory Controllers

Introduction: Adaptive inventory controllers must respond to stochastic demand while satisfying service, capacity, budget, and stability requirements. Outcome-level evaluation can reveal poor performance, but it does not necessarily identify which controller decisions generated a failure or whether a bounded rule correction will remain effective on unseen stochastic runs.

Methods: We evaluate a simulation-based verification protocol in a stylized inventory benchmark with eight inventory units, a 120-period horizon, stochastic demand, and explicit replenishment-capacity and financial constraints. The controller is represented by finite symbolic rules, diagnostic predicates, explicit priorities and conflicts, and machine-readable explanation logs. Five experiments vary the replenishment envelope or remove a stockout-priority or order-smoothing rule. Failures are diagnosed on five seeds; selected edits from a library permitting rule addition, deletion, and priority revision are then evaluated on twenty held-out seeds using empirical 95th-percentile statistics. Metrics include stockout rate, service loss, tracking error, working-capital excess, order volatility, and budget, capacity, and conflict violations.

Results: Restrictive and relaxed replenishment envelopes produced stockout, service-loss, and tracking failures that were not corrected by the tested stockout-increase edit. After removal of the stockout-priority rule, the same edit improved service loss and tracking error but slightly worsened stockout rate and raised the conflict rate above its guardrail. Removal of the smoothing rule increased held-out order volatility to 20.62. When the coaching selector prioritized volatility, it added a smoothing rule and reduced this statistic to 11.46, below the declared threshold of 12.00. Stockout rate and tracking error nevertheless remained above threshold, so the repair was locally successful but the controller was globally rejected.

Conclusion: Explanation-guided finite revision can distinguish resource-induced failures, unsuitable edits, and locally repairable rule defects. The protocol supports bounded verification under declared simulation conditions; it does not establish controller optimality or deployment-domain validity.