Minimum Necessary Intervention: Constraint-Based Adaptive Assistance with Intervention Debt and Recovery-Aware Withdrawal

Abstract

Artificial intelligence can improve task completion, yet stronger assistance can also reduce opportunities for independent recovery. We present Minimum Necessary Intervention (MNI), a control framework that treats assistance as a graded resource rather than a binary feature. MNI estimates user struggle, predicts recovery across an ordered intervention ladder, selects the least costly action satisfying a recovery-probability constraint, verifies recovery, escalates only when necessary, and withdraws support following recovery. We operationalize Intervention Debt as intervention burden weighted by the absence of modeled independent recovery. A reproducible synthetic benchmark spanning six digital-use domains compares MNI with no assistance, static help, and maximum assistance. Across 20 random seeds with 1,200 users per seed, MNI achieved mean completion of 0.703 (95% CI [0.698, 0.709]) with mean intervention burden of 0.632, compared with 0.769 completion and 1.000 burden under maximum assistance. MNI reduced modeled Intervention Debt from 0.785 to 0.443 and increased modeled independent recovery from 0.215 to 0.392 relative to maximum assistance. Pareto, longitudinal, ablation, heterogeneity, domain-shift, sensing-noise, and adversarial stress tests reveal a consistent trade-off rather than universal dominance. These results support MNI as a recovery-aware framework for providing sufficient assistance without defaulting to maximal automation.