A proposed AI monitor identified the same procedural deviations documented by investigators in reconstructions of two historical air accidents. In the reported tests, its detections came 31 and 29 seconds before the Überlingen collision, and 19 seconds before the impact in a wrong-runway-takeoff case.
The work is a methods preprint submitted to IEEE for possible publication. Its evaluation compares monitor outputs with constructed expectations, expert annotations and official investigation reports; it does not measure accident reduction or controller workload.
From speech to a time-stamped trace
The framework is designed to check spoken air-traffic-control procedures by combining controller-pilot exchanges with surveillance data and observations from aircraft. It parses radio communications into events linked to the relevant entities, then merges those events with the other sources in a time-stamped trace.
Those traces are assessed with metric temporal logic, a formal rule language that records when observations occur and sets limits on how much time may pass between them. A violation is not returned as a bare warning: the system reports the breached obligations and the observations supporting the verdict.
A clean test of the rule engine
Because safety-critical configurations are rare in recorded traffic, the researchers built synthetic situations around real corpus events that satisfied a monitored formula's trigger, then applied controlled perturbations. From 255 trigger-satisfying events, they created 1,495 situations: 664 compliant and 831 violating, exercising 50 forms across five of the eight formula families.
In this part of the evaluation, synthetic situations assessed implementation correctness rather than agreement with expert judgment in real traffic. The monitor logic returned the expected verdict in every one of the 1,495 situations derived from the public corpora.
The harder question: real traffic
The real-traffic comparison used ATCO2 and TartanAviation, two public corpora of recorded traffic. To establish the reference judgment, the team fully annotated the busiest hour on each of three days of KAGC traffic from 2022, blind to the monitor's output; the set contained 12 violations.
The qwen3.8-27b parser in its no-thinking configuration was reported with a precision of 0.79, recall of 0.92 and F1 score of 0.85. In ordinary terms, precision reflects how often flagged events matched the reference, recall reflects how many annotated violations were found, and F1 combines the two measures.
Warnings with limits
In the two historical accident reconstructions, the monitor identified the same procedural deviations documented in official investigation reports. In the Überlingen case, detections at 21:35:01 and 21:35:03 preceded the 21:35:32 collision by 31 and 29 seconds. In the wrong-runway-takeoff reconstruction, detection fired at 06:06:16 as ground speed entered the takeoff-roll regime, 19 seconds before the 06:06:35 impact.
In those reconstructions, the monitor produced the documented deviations and detected them before impact. That is a replay-based validation, not evidence that the system would prevent an accident in operation.
The study's scope is narrower than the headline numbers might suggest. The synthetic exercise covered five of eight formula families, while the public corpora exercised 37.6% of routine atom schemas, the system's predefined building blocks for routine events. The real-traffic result also combines speech recognition, parsing, trace construction and formula evaluation, so its F1 score cannot isolate any one stage.
The authors say operational deployment would still require human-in-the-loop testing of workflow integration, alert presentation and calibration under controller workload. The evidence therefore concerns implementation and selected reference cases, not reduced operational risk or a replacement for human oversight.
Paper data and sources
Original title: Formal, Executable and Explainable Runtime Monitoring of Spoken Air Traffic Control Operational Procedures
Authors: Roberto Luvini, Giacomo Longo, Alessandro Armando, Enrico Russo
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text