The Ghost Toll: How ML Caught the Errors Silent Monitoring Missed
Abstract
Satellite-based free-flow tolling systems silently accumulate GPS positioning errors that rule-based monitoring never flags, until a pattern becomes a financial liability. This paper walks through a real production incident in which systematic GNSS error went undetected across thousands of transactions, the machine learning error-classification model built from production logs to catch it in real time, and the deployment tradeoffs made to run that model at the roadside edge. The failure is characterised precisely: the errors were statistically masked rather than absent, so every individual transaction looked acceptable and only the distribution was wrong, which is exactly what threshold-based validation cannot see. Readers take away a framework for identifying silent failure modes in location-dependent ML pipelines, an account of why rule-based validation fails at the edge of GPS accuracy and which features actually matter for error classification in positioning systems, and a realistic picture of what edge deployment demands.
- Published in
- Proceedings of the Raptors Conference, Vol. 1 (2026)
- Conference
- Raptors Conference 2026, 21 October 2026, online
- Accepted
- 2026-10-11
- DOI
- https://doi.org/10.68495/rc.2026.C094
- Presentation
- Slides are not yet available for this paper.
- Licence
- © 2026Venkateswarlu Mullangi. Published under Creative Commons Attribution 4.0 International (CC BY 4.0).
- Publisher
- Hackathon Raptors, Community Interest Company 15557917, London, United Kingdom
- Permanent address
- https://proceedings.raptors.dev/mullangi-ghost-toll-ml-caught
Open access. Every entry in the Proceedings is free to read, with no charge to authors and no charge to readers. Slides are served from this site, so they stay available at the paper's permanent address.
Cite this paper
Venkateswarlu Mullangi (2026). The Ghost Toll: How ML Caught the Errors Silent Monitoring Missed. Proceedings of the Raptors Conference, Vol. 1 (2026). Hackathon Raptors, London. https://doi.org/10.68495/rc.2026.C094
@inproceedings{raptors2026-c094,
title = {{The Ghost Toll: How ML Caught the Errors Silent Monitoring Missed}},
author = {Mullangi, Venkateswarlu},
booktitle = {Proceedings of the Raptors Conference},
volume = {1},
year = {2026},
publisher = {Hackathon Raptors},
address = {London, United Kingdom},
doi = {10.68495/rc.2026.C094},
url = {https://doi.org/10.68495/rc.2026.C094},
abstract = {Satellite-based free-flow tolling systems silently accumulate GPS positioning errors that rule-based monitoring never flags, until a pattern becomes a financial liability. This paper walks through a real production incident in which systematic GNSS error went undetected across thousands of transactions, the machine learning error-classification model built from production logs to catch it in real time, and the deployment tradeoffs made to run that model at the roadside edge. The failure is characterised precisely: the errors were statistically masked rather than absent, so every individual transaction looked acceptable and only the distribution was wrong, which is exactly what threshold-based validation cannot see. Readers take away a framework for identifying silent failure modes in location-dependent ML pipelines, an account of why rule-based validation fails at the edge of GPS accuracy and which features actually matter for error classification in positioning systems, and a realistic picture of what edge deployment demands.}
}
TY - CONF TI - The Ghost Toll: How ML Caught the Errors Silent Monitoring Missed AU - Mullangi, Venkateswarlu T2 - Proceedings of the Raptors Conference VL - 1 PY - 2026 DA - 2026/10/11 PB - Hackathon Raptors CY - London, United Kingdom DO - 10.68495/rc.2026.C094 UR - https://doi.org/10.68495/rc.2026.C094 AB - Satellite-based free-flow tolling systems silently accumulate GPS positioning errors that rule-based monitoring never flags, until a pattern becomes a financial liability. This paper walks through a real production incident in which systematic GNSS error went undetected across thousands of transactions, the machine learning error-classification model built from production logs to catch it in real time, and the deployment tradeoffs made to run that model at the roadside edge. The failure is characterised precisely: the errors were statistically masked rather than absent, so every individual transaction looked acceptable and only the distribution was wrong, which is exactly what threshold-based validation cannot see. Readers take away a framework for identifying silent failure modes in location-dependent ML pipelines, an account of why rule-based validation fails at the edge of GPS accuracy and which features actually matter for error classification in positioning systems, and a realistic picture of what edge deployment demands. LA - en ER -