SKY-GOV: A Knowledge-Constrained Decision-Intelligence System for Risk-Adaptive Human Oversight of Autonomous Drone Fleets

Abstract

Autonomous drone fleets require a defensible rule for deciding when automated monitoring is sufficient and when scarce human attention must be escalated. This study proposes SKY-GOV, a knowledge-constrained selective decision-intelligence layer that converts multivariate mission-risk evidence into four ordinal oversight states. The system combines a calibrated histogram-gradient-boosting classifier, class-conditional split-conformal prediction sets, auditable symbolic safety floors, and reject-to-human escalation. Unlike the earlier concept formulation, oversight targets are generated by an independent noisy latent process rather than copied from the knowledge rules. Evaluation uses complete-mission group splits, separate training, probability-calibration, conformal-calibration, policy-validation, and untouched test partitions, 10 paired random seeds, a constructed risk shift, ablations, matched-coverage comparisons, paired tests, and a 10,800-row cost/capacity sensitivity analysis. On the in-distribution test, mean decision cost was 2.209 versus 2.313 for calibrated boosting (4.5% relative change), with 22.9% review and 26.1% critical under-supervision among automated decisions. Under shift, the corresponding costs were 2.046 and 2.814 (27.3% relative change); both paired comparisons had Holm-adjusted p=0.0176. Marginal conformal coverage changed from 70.4% to 74.9%, but worst-class coverage fell from 67.4% to 36.4%. A separate grouped-flight experiment on the public ALFA benchmark evaluates the telemetry fault-classification and selective-review subpipeline without inferring unobserved human behavior. Results support the integration of calibrated prediction, conformal ambiguity, symbolic constraints, and human escalation as an auditable oversight-allocation method. They do not establish universal coverage under shift or operational safety. This is a non-peer-reviewed preprint. The accompanying ZIP contains reproducibility code, derived public ALFA telemetry features, results, figures, tests, and documentation. Software is MIT licensed; ALFA-derived data retain CC BY 4.0 attribution.