Governed negotiation for human-centric disruption-aware railway dispatching in flatland. Methodology and pilot validation

Abstract

Major rail disruptions create dense conflicts that are still largely resolved manually under time pressure. We present a reproducible, non-learning-based dispatch methodology for human-supervised decision support that combines (i) short-horizon reservation negotiation, (ii) deterministic conflict arbitration with explicit anti-starvation guarantees, and (iii) a bounded governance layer with auditable parameter updates. The design supports bounded mixed-initiative control: dispatchers can inspect local rationales and steer policy parameters within explicit limits while preserving deterministic behavior and traceability. In a five-scenario Flatland pilot suite (paired seeds, bootstrap confidence intervals (CIs)), the stress-test BASELINE fails in all scenarios (strict deadlock 1.0 everywhere; throughput 0.0 in four scenarios and 0.015 in one), while NEGOTIATION and FULL resolve all scenarios (strict deadlock 0.0, throughput 1.0). On the four hard scenarios, NEGOTIATION_FIFO (First-In-First-Out arbitration) improves makespan with unchanged robustness, indicating that reservation plus anti-starvation is the dominant robustness driver and local ranking is a secondary efficiency lever. FULL-minus-NEGOTIATION paired deltas quantify the net effect of enabling governance under the configured policy, while governance intervention tables provide descriptive evidence of when and how updates are applied. Component-level governance causality (adaptive scheduling vs. patch policy vs. source) is not isolated. Ablations and a global Nstarve sweep provide mechanism evidence, including a monotonic efficiency-fairness trade-off. Evidence remains simulator-level and does not include real signaling, timetable constraints, user studies, or quantified human-factors impact.