Repair Before Veto: Repair-Augmented Constraint Learning for Contextual Decisions

Researchers introduce Repair-Augmented Constraint Learning (RACL), a machine learning framework that decides whether to repair constraint violations before rejecting candidates, rather than applying hard vetoes immediately. The method achieves significantly lower false-veto rates (0.25%) compared to baseline approaches (26.4%) on real-world airline data, with applications to automated decision systems.

RACL addresses a fundamental gap in how machine learning systems handle constraint violations in real-world decision scenarios. Traditional constraint-learning approaches treat violations as terminal rejections, but deployed systems often have repair options available—such as schedule modifications, configuration changes, or service upgrades. This research formalizes the decision to repair before rejecting, embedding repair operators directly into classifier semantics rather than handling them outside the decision framework.

The innovation matters because many automated decision systems operate in environments where violations are recoverable. Airline scheduling, resource allocation, and configuration systems frequently offer modification options that could transform infeasible candidates into acceptable ones. By learning when repair is worthwhile, RACL prevents false vetoes that unnecessarily reject feasible solutions. The theoretical contributions separate the identifiability of binary labels from learnability of decision rules, providing capacity and calibration bounds for real-world implementations.

The experimental results demonstrate substantial practical impact. On synthetic benchmarks, RACL recovers intended credit and repair structures. On raw airline data, validation-selected RACL reduces false vetoes to 0.25% (10 of 4,039 cases) compared to 26.4% for the strongest baseline repair-search method. This 100x improvement has direct economic implications for domains where false rejections carry significant costs—missed sales, customer friction, or operational inefficiency.

Looking forward, adoption depends on integration into existing constraint-learning pipelines and domain-specific validation of repair operator sets. The framework's explicit false-veto/error-detection-rate trade-off enables practitioners to calibrate systems for their specific cost structures, making it particularly valuable for high-stakes decision automation.

• →RACL embeds repair operators into classifier semantics, learning whether to repair before rejecting constraint violations.
• →Achieves 100x reduction in false-veto rate (0.25%) compared to baseline methods on real airline scheduling data.
• →Separates binary-label identifiability from decision-rule learnability, providing new theoretical foundations for constraint learning.
• →Makes the false-veto/error-detection-rate trade-off explicit, enabling practitioners to calibrate systems for specific cost structures.
• →Applicable to automated decision systems with known repair options: airlines, resource allocation, configuration management.