Certifying the Black Box: Continuous Statistical Assurance for Autonomous Vehicles | Zenodo
Skip to main
You are using an outdated browser. Please upgrade your browser to improve your experience.
Published July 19, 2026
| Version 1.0
Preprint
Open
Certifying the Black Box: Continuous Statistical Assurance for Autonomous Vehicles
Authors/Creators
Shah, Neel<br>(Researcher)
Description
June 2026 made autonomous-vehicle certification newly urgent: the UN adopted a Global Technical Regulation on Automated Driving Systems whose In-Service Monitoring and Reporting (ISMR) pillar mandates lifecycle safety evidence but does not yet specify its machinery, and NHTSA withdrew its AV STEP oversight proposal after industry judged it too burdensome and safety advocates judged it too weak. This paper proposes that missing machinery. The foundational tension: even today's best-performing stack - Waymo's largely modular fleet, with 94% fewer serious-injury-or-worse crashes over 220.6 million rider-only miles - cannot be certified by specification, because its safety is demonstrated statistically rather than specified in advance; the shift to opaque Large Driving Models (LDMs) deepens that gap from difficult to intractable. Grounded in a synthesis of AV architectural evolution, the standards landscape (ISO 26262, ISO 21448/SOTIF, UL 4600, UN R157/DSSAD), the certification-methodology portfolio, and the empirical incident record, this paper proposes Continuous Statistical Assurance (CSA): the certificate becomes a machine-readable safety envelope - ODD boundary, SPI control limits, calibrated confidence thresholds, pre-agreed expansion rules - that provider and regulator read as a live, continuously computed compliance verdict. Three layers make learned drivers certifiable: a provable residual-risk bound; calibration-aware confidence gating that provably tightens under measured miscalibration; and post-incident rollout inspection that treats generated explanations as hypotheses to check, not evidence. An incentive architecture (aggregate data-return channel, aviation-style legal privilege) makes participation rational for providers and regulators; a tiered CSA-Baseline path and five reproducible simulations make CSA adoptable below robotaxi scale.
Files
Continuous Statistical Assurance for Autonomous Vehicles .pdf
Files<br>(4.3 MB)
Name<br>Size
Download all
Continuous Statistical Assurance for Autonomous Vehicles .pdf
md5:50e72589bf83715239979a757cb23671
4.3 MB
Preview
Download
Additional details
Related works
Is cited by
Other:
https://medium.com/@neelshah0147/how-do-you-certify-a-self-driving-car-that-keeps-learning-3a802266991b
(URL)
Dates
Created
2026-07-17
64
Views
Downloads
Show more details
All versions<br>This version
Views
Total views
64
64
Downloads
Total downloads
Data volume
Total data volume
34.0 MB<br>34.0 MB
More info on how stats are collected....
Versions
External resources
Indexed in
OpenAIRE
Communities
Keywords and subjects
Keywords
autonomous vehicles
AV safety
functional safety
SOTIF
UL 4600
conformal prediction
safety cases
large driving models
end-to-end learning
safety certification
runtime monitoring
mechanistic interpretability
world models
AV certification
operational design domain
confidence calibration
statistical safety assurance
UN GTR ADS
Details
DOI
DOI Badge
DOI
10.5281/zenodo.21438788
Markdown
[](https://doi.org/10.5281/zenodo.21438788)
reStructuredText
.. image:: https://zenodo.org/badge/DOI/10.5281/zenodo.21438788.svg<br>:target: https://doi.org/10.5281/zenodo.21438788
HTML
Image URL
https://zenodo.org/badge/DOI/10.5281/zenodo.21438788.svg
Target URL
https://doi.org/10.5281/zenodo.21438788
Resource type<br>Preprint
Publisher<br>Zenodo
Languages
English
Rights
License
Creative Commons Attribution 4.0 International
The Creative Commons Attribution license allows re-distribution and re-use of a licensed work on the condition that the creator is appropriately credited.
Read more
Citation
Export
Technical metadata
Created
July 19, 2026
Modified
July 29, 2026
Jump up
This site uses cookies. Find out more on how we use cookies
Accept all cookies<br>Accept only essential cookies