From Stochastic Models to Deterministic Assurance: A Strategic Framework for Safety-Critical Artificial Intelligence

The contemporary landscape of artificial intelligence is currently bisected by a fundamental misunderstanding of technical depth. On one side of this divide lies the rapid proliferation of generative interface layers—often characterized as Large Language Model (LLM) "wrappers"—which prioritize rapid deployment and conversational fluidity. On the opposite side is the rigorous discipline of deep AI engineering, a field defined by the integration of formal verification, sensor-fusion resilience, and deterministic safety architectures. For enterprise leaders navigating this transition, the distinction is no longer merely academic. As autonomous systems move from digital environments to high-stakes physical deployments, the limitations of probabilistic, "wrapper-based" approaches have been exposed by a series of high-profile failures.

The incidents involving Uber Advanced Technologies Group (ATG), GM Cruise, Tesla, and Waymo serve as critical empirical data for this analysis. These events represent more than isolated accidents; they are systemic indicators of architectural fragility. The $8.5 million settlement involving Uber ATG after the 2018 Tempe fatality, the revocation of GM Cruise's California operating permit in 2023, and the ongoing National Highway Traffic Safety Administration (NHTSA) investigations into Tesla's Full Self-Driving (FSD) system are all symptoms of a "Perception-Logic Gap".¹ This report analyzes these failures to establish a new paradigm for safety-critical AI—one that positions Veriprajna not as a facilitator of stochastic interfaces, but as a provider of deep, verifiable autonomy.

The Architectural Fragility of Stochastic Perception: Lessons from Uber ATG

The March 2018 collision in Tempe, Arizona, involving an Uber ATG test vehicle, remains the foundational case study in classification brittleness. While the narrative at the time focused heavily on the distraction of the human safety operator, the National Transportation Safety Board (NTSB) findings revealed a much more profound failure in the software's ability to maintain a stable representation of the physical world.³

Classification Oscillation and the Failure of Object Permanence

The Uber ATG system first registered the pedestrian, Elaine Herzberg, approximately 5.6 seconds before impact.⁵ At a velocity of 43 mph, the vehicle was nearly 378 feet away, providing an ample window for a standard Automatic Emergency Braking (AEB) system to intervene.⁷ However, the system's perception logic was characterized by "classification oscillation." In the seconds leading up to the crash, the software repeatedly reclassified the pedestrian—first as an "unknown object," then as a "vehicle," and finally as a "bicycle".⁶

Each reclassification was not merely a change in label; it was a reset of the object's predicted trajectory. In probabilistic systems lacking temporal consistency, the AI treats each frame or cluster of frames as a near-independent event. Because the system could not settle on a persistent identity for the object, it could not calculate a reliable path prediction until it was too late. The system determined that emergency braking was needed only 1.3 seconds before impact—a point at which the laws of physics made a collision unavoidable.³

Technical Debt and the Removal of Safety Redundancy

A critical element of the Uber ATG failure was the intentional deactivation of the vehicle's native safety systems. To prevent "erratic vehicle behavior" and provide a smoother ride for the autonomous system, Uber had disabled the Volvo XC90's factory-installed collision avoidance and AEB features.⁴ The engineering team opted to rely entirely on a proprietary, developmental system that was not yet verified for high-confidence intervention.

Failure Component	Technical Mechanism	Strategic Implication
Perception Pipeline	Classification oscillation (Unknown -> Vehicle -> Bike)	Loss of object permanence disables path prediction.6
Logic Suppression	Manual deactivation of Volvo factory AEB	Removal of hard-coded safety layers in favor of experimental code.4
HMI Interface	Over-reliance on distracted human monitor	Failure to account for "automation complacency".3
Prediction Engine	Static trajectory assumption for dynamic actors	Inability to model non-standard pedestrian crossings.5

This decision-making process illustrates a dangerous trend in AI development: the sacrifice of deterministic safety layers for the sake of "smooth" performance in a stochastic model. The $8.5 million settlement for the 2018 crash reflects not just a legal liability, but a failure to manage the "functional limitations" of the automated driving system.⁴ Deep AI engineering, as championed by Veriprajna, argues against this hierarchy, advocating for a "Safety-First" architecture where the perception layer is bound by formal constraints that cannot be overridden by experimental policy.

Misdiagnosis and the Failure of Post-Impact Logic: The Cruise 2023 Crisis

In October 2023, a GM Cruise robotaxi in San Francisco hit and dragged a pedestrian for 20 feet, leading to a total suspension of the company's driverless operations in California.⁹ This incident moved beyond the problem of initial collision avoidance into the territory of "post-impact reasoning"—a domain where typical LLM wrappers and simple perception APIs fail entirely.

The Front-Runover vs. Side-Impact Fallacy

The Cruise incident was initiated by a third party: a human-driven Nissan struck a pedestrian, launching her into the path of the Cruise vehicle.⁹ The Cruise car hit the pedestrian and initially stopped. However, because the system's "impact detection" logic was insufficiently granular, it misdiagnosed the collision. Despite the pedestrian being pinned under the vehicle, the system's sensors failed to recognize a frontal run-over and instead categorized the event as a side-impact collision.¹

This misdiagnosis triggered a pre-programmed "Minimal Risk Condition" (MRC) maneuver. The system was designed to pull over to the side of the road after a side impact to avoid blocking traffic.¹¹ Because the perception layer had "forgotten" the pedestrian after the impact, the vehicle began to pull over, dragging the victim 20 feet at approximately 7 mph.⁹ The dragging ceased only when the vehicle detected "excessive wheel slip," which it interpreted as a mechanical fault rather than a human obstruction.¹¹

Transparency as a Technical Requirement

The Cruise failure was as much an organizational failure as a technical one. Investigations revealed that senior leadership was "fixated on correcting the inaccurate media narrative" and failed to be transparent with regulators regarding the dragging.⁹ Cruise employees admitted to "letting the video speak for itself" during meetings with the DMV, knowing that internet connectivity issues often prevented the "dragging" portion of the video from playing.⁹

This highlights a critical lesson for deep AI consultancy: the safety of an autonomous system is inseparable from its transparency. Veriprajna's approach emphasizes the development of "Explainable Safety Audits," where every decision made by the AI, especially post-impact, is logged in a tamper-proof, deterministic format that can be audited by regulators in real-time. The subsequent $500,000 criminal fine for submitting false reports to the NHTSA underscores the high cost of treating AI safety as a marketing problem rather than an engineering one.¹

The "Vision-Only" Dilemma and the Limits of Probabilistic Sensing: Tesla FSD

Tesla's Full Self-Driving (FSD) system has become the center of a massive regulatory investigation, with the NHTSA opening over 40 inquiries into crashes between 2024 and 2025.² These investigations, particularly those identified as PE24-031 and PE25-012, focus on the system's failure in "Capability Theater"—optimal performance in clear conditions that collapses in the face of environmental "edge cases".¹³

Environmental Sensitivity and Signal Non-Compliance

The NHTSA's investigations have identified specific patterns where Tesla's vision-only system fails to comply with basic traffic safety laws:

1. Traffic Signal Failure: In 18 separate complaints, FSD-enabled vehicles failed to remain stopped for red lights or failed to detect the signal state entirely.¹²
2. Wrong-Way Maneuvers: The system has been observed entering opposing lanes of traffic or executing turns from through-only lanes, disregarding clear road markings and signage.²
3. Low-Visibility Saturation: A significant fatal collision in 2023 occurred during a state of "sun glare on wet asphalt," where the system failed to detect a pedestrian.¹⁴

Tesla's reliance on a "Vision-Only" architecture—eschewing LiDAR and radar—creates a fundamental vulnerability to "sensor saturation." In conditions of fog, dust, or airborne debris, the optical signal-to-noise ratio drops below the threshold required for safe navigation.¹³ While Tesla utilizes "Occupancy Networks" to predict the 3D geometry of the world from 2D images, the NHTSA reports suggest that these predictions are still too probabilistic to be used as a primary safety layer.¹⁵

Formalizing the Driving Environment

To move beyond the Tesla failure model, deep AI engineering utilizes "Hazard-Driven Envelopes." Instead of vague "features," the system must define explicit Operational Design Domains (ODD). If the "glare saturation percentage" or "fog backscatter index" exceeds a verified threshold, the system must initiate a fail-safe transition.¹⁴

Investigated Failure Mode	Frequency/Impact	Technical Cause
Red Light Non-Compliance	18+ Complaints	Signal state detection failure in vision stack.12
Lane Marking Violation	4+ SGO Reports	Inability to distinguish turn-only vs. through lanes.12
Low Visibility Crash	Fatalities (2023-2024)	Optical sensor saturation (glare/fog/dust).13
Opposing Lane Entry	2 SGO Reports	Failure in 3D reconstruction of lane geometry.12

Veriprajna's philosophy is that autonomy cannot be built on "best-effort" software. The 2.9 million vehicles impacted by the 2025 NHTSA probe represent a fleet-wide risk that can only be mitigated through the implementation of "Assurance Gates"—software locks that prevent the AI from making high-risk decisions when the confidence level of its perception system drops below a deterministic point.²

Multi-Agent Gridlock and Socio-Technical Resilience: The Waymo Experience

Waymo is often regarded as the benchmark for safety, having logged over 56 million miles with significantly lower injury rates than human drivers.¹⁷ However, as the system scales, it has encountered a new class of failure: "socio-technical friction." This involves not just how the AI drives, but how it interacts with the complex, often hostile, human social environment.

Intersection Blockages and Communication Failures

During a 2025 power outage in Los Angeles, dozens of Waymo robotaxis became stuck at a series of darkened intersections. The "Waymo Driver," programmed to treat dark signals as four-way stops, became overwhelmed by the concentrated spike in requests for "Remote Assistance".¹⁸ Because the vehicles were unable to communicate with each other effectively, they entered a state of "multi-agent gridlock," where robotaxis were blocking other robotaxis, creating a backlog that the central command center could not resolve.¹⁸

This failure highlights the "Independence Trap"—the assumption that an autonomous vehicle can operate safely as a solitary agent without a broader, coordinated system.¹⁸ Deep AI engineering must account for the loss of wireless communication and the necessity of "V2V" (Vehicle-to-Vehicle) and "V2I" (Vehicle-to-Infrastructure) protocols that allow the fleet to resolve deadlocks autonomously.²⁰

The Necessity of "Danger Escape Mode"

Perhaps the most significant emerging threat to autonomous operations is public aggression. In early 2025, several Waymo vehicles were attacked by crowds during civil unrest in Los Angeles, with protesters slashing tires and setting vehicles on fire.²¹ The vehicles, programmed for "passive safety," simply stopped when surrounded by people.

This has led to the proposed development of a "Danger Escape Mode." Such a system would use the 360-degree sensor suite to detect "malicious human aggression" and shift the vehicle's directive from "passive compliance" to "active escape".²¹ While the vehicle must never be programmed to cause harm, deep AI providers argue that it should be capable of committing minor traffic infractions (such as driving onto a sidewalk or through a red light) to protect its passengers and escape a volatile situation.²¹ This requires a radical rethink of the AI's "Ethics Engine"—a task that goes far beyond the capabilities of an LLM wrapper.

The Technical Solution: Bird's-Eye-View (BEV) and Occupancy Networks

To address the tracking failures seen in the Uber and Cruise cases, the industry is shifting toward Bird's-Eye-View (BEV) perception. Standard per-camera systems process individual images, which leads to loss of data during "stitching." In contrast, BEV perception transforms multi-view camera and LiDAR data into a unified, top-down 3D grid.²²

Occupancy Networks vs. Standard Sensor Fusion

Traditional sensor fusion attempts to "match" 2D pixels to 3D points, a process that is computationally expensive and prone to projection errors. Veriprajna advocates for "Occupancy Networks"—an architecture that predicts the "occupancy probability" of every voxel in a 3D volume.¹⁶

1. Object Permanence: Because occupancy networks track volume rather than just "labels," the system knows a space is occupied even if it cannot decide if the object is a pedestrian or a bicycle. This would have prevented the Uber ATG classification flip.¹⁶
2. Geometric Fidelity: Occupancy networks capture vertical structures and road debris that are often ignored by 2D BEV maps. This would have allowed the Cruise vehicle to "see" the pedestrian underneath its chassis during the post-impact maneuver.¹⁶
3. Spatiotemporal Consistency: Using "BEVFormer" architectures, the system can use "temporal self-attention" to remember where an object was even during temporary occlusions (e.g., a pedestrian walking behind a parked truck).²⁴

$$X_{BEV} = f_{transformer}(I_1, I_2, ..., I_n, L_{cloud})$$

In this model, the Transformer architecture serves not as a conversational tool, but as a spatial reasoning engine that fuses heterogeneous data into a singular "Shared Canvas".²³ This is deep AI engineering: the use of frontier architectures to solve fundamental physics problems in navigation.

Formal Verification: The Veriprajna Standard for High-Assurance AI

The most significant differentiator between a "wrapper" and a "solution" is the application of formal methods. Traditional software testing relies on "black-box" scenarios; if the system passes N tests, it is assumed to be safe. In safety-critical systems, however, we require a mathematical proof of correctness.

SMT Solvers and Network-Level Reasoning

Tools like Marabou and α,β-CROWN allow engineers to verify the properties of deep neural networks. By representing the network as a set of piecewise-linear constraints, we can determine if there exists any input that could lead to an unsafe output.²⁵

A "Safety Property" might be defined as follows:

For all inputs x within the range of "Low Visibility," the output y (Braking Command) must never be less than k.

$$\forall x \in X_{fog} \Rightarrow f(x) \geq Braking_{min}$$

If an SMT solver like Marabou returns a "counter-example," it has identified a specific, often imperceptible, perturbation that would cause the AI to fail. This allows Veriprajna to "harden" the model during the training phase, a process known as "verification-aware training".²⁸

Pruning for Verifiability

A major challenge in formal verification is the "Curse of Dimensionality." Large networks are too complex for current solvers to analyze exhaustively. Veriprajna addresses this through "Neuron Pruning." By removing redundant neurons and non-linearities that do not contribute to the model's accuracy, we produce a "Pruned Model" that is mathematically easier to verify without sacrificing performance.²⁹

Verification Technique	Methodology	Benefit
Bound Tightening	Symbolic analysis of neuron activation ranges	Reduces search space for SMT solvers.25
Reachability Analysis	Computing the set of all reachable outputs for an input set	Guarantees the AI will stay within a "Safe Polytope".28
Piecewise-Linear Approximation	Replacing complex activations with ReLU-based segments	Enables sound and complete verification proofs.27
Formal Safety Filter	Runtime monitoring of AI commands against a verified baseline	Provides a "Safe Recovery" if the main AI behaves irrationally.31

The Regulatory Horizon: SOTIF and ISO/PAS 8800

For enterprises, compliance with emerging international standards is no longer optional. The landscape has shifted from "voluntary self-assessment" to mandatory adherence to a tiered safety framework.⁵

ISO 26262 vs. ISO 21448 (SOTIF)

While ISO 26262 handles "Functional Safety" (e.g., a sensor failing or a chip short-circuiting), it cannot account for the inherent limitations of AI.³⁴ This gap is filled by ISO 21448, the standard for "Safety of the Intended Functionality" (SOTIF). SOTIF is specifically designed to address hazards that occur when the system is working exactly as programmed but encounters an "Unknown/Unsafe" environment.³⁴

The goal of a Veriprajna engagement is to maximize the "Known/Safe" quadrant of a client's AI system. This involves:

1. Hazard and Risk Analysis (HARA): Identifying non-failure risks such as sensor misinterpretation in heavy rain.³⁶
2. Triggering Condition Identification: Systematically mapping the environmental states that lead to perception errors.³⁶
3. V&V (Verification and Validation): Using high-fidelity simulations to "inject" edge cases that would be too dangerous to test on public roads.³⁶

ISO/PAS 8800: The Future of AI Integration

As of late 2024, ISO/PAS 8800 has become the primary standard for "Functional Safety for AI in Road Vehicles".³⁷ It provides the first global guidelines for managing the AI lifecycle, from "Data Acquisition" to "Post-Deployment Monitoring".³³ Veriprajna ensures that our clients' architectures are not just compliant, but "future-proofed" against the increasing rigor of global AI governance standards like the EU AI Act and the NIST AI Risk Management Framework.³³

The Strategic Path Forward: Veriprajna's Deep AI Mandate

The transition from a "Wrapper" culture to a "Deep AI" culture is a journey from probabilistic hope to deterministic assurance. The $8.5 million Uber settlement, the Cruise suspension, and the 40+ Tesla investigations are not reasons to abandon AI; they are reasons to engineer it correctly.

Veriprajna's consulting model is built on three pillars that address the specific failure modes identified in this report:

1. Perception Resilience: Moving clients from per-camera 2D perception to Transformer-based BEV Occupancy Networks to ensure object permanence and tracking stability.¹⁶
2. Verified Decisioning: Implementing SMT-based formal verification to prove that AI-driven control architectures will never violate core safety properties.²⁵
3. Socio-Technical Hardening: Developing sophisticated "Escape Modes" and V2X communication frameworks to manage the reality of civil unrest and multi-agent gridlock.¹⁸

As the global cost of a single data breach reaches $4.44 million and the cost of an autonomous fatality enters the tens of millions in legal and operational damages, the "cheap" wrapper becomes the most expensive mistake an enterprise can make.¹⁴ Veriprajna provides the deep engineering expertise required to build AI that doesn't just work in the lab—it endures in the world.

The choice for the modern enterprise is clear: continue to wrap probabilistic black boxes and manage the inevitable fallout, or partner with Veriprajna to architect a future of verifiable, high-assurance autonomy. The era of stochastic AI is ending; the era of Deep AI Engineering has begun.

Works cited

1. Cruise Admits To Submitting A False Report To Influence A Federal Investigation And Agrees To Pay $500000 - Department of Justice, accessed February 9, 2026, https://www.justice.gov/usao-ndca/pr/cruise-admits-submitting-false-report-influence-federal-investigation-and-agrees-pay
2. NHTSA Opens Probe into 2.9M Teslas Over FSD Violations - Autobody News, accessed February 9, 2026, https://www.autobodynews.com/news/nhtsa-opens-probe-into-2-9m-teslas-over-fsd-violations
3. HWY18MH010.aspx - NTSB, accessed February 9, 2026, https://www.ntsb.gov/investigations/Pages/HWY18MH010.aspx
4. NTSB Shares Investigation Findings and Recommendations Regarding March 2018 Uber ATG Fatality | Eckert Seamans, accessed February 9, 2026, https://www.eckertseamans.com/legal-updates/ntsb-shares-investigation-findings-and-recommendations-regarding-march-2018-uber-atg-fatality
5. H-19-047 - Accident Data - NTSB, accessed February 9, 2026, https://data.ntsb.gov/carol-main-public/sr-details/H-19-047
6. NTSB releases preliminary report on fatal Uber self-driving car crash - Metro Magazine, accessed February 9, 2026, https://www.metro-magazine.com/news/ntsb-releases-preliminary-report-on-fatal-uber-self-driving-car-crash
7. Death of Elaine Herzberg - Wikipedia, accessed February 9, 2026, https://en.wikipedia.org/wiki/Death_of_Elaine_Herzberg
8. New Details Emerge Regarding Uber Self-Driving Vehicle Accident in Tempe, Arizona, accessed February 9, 2026, https://schwedlawfirm.com/blog/new-details-emerge-regarding-uber-self-driving-vehicle-accident/
9. A Root Cause Analysis of a Self-Driving Car Dragging a Pedestrian, accessed February 9, 2026, https://www.computer.org/csdl/magazine/co/2024/11/10720344/215PD0vqgTe
10. Notes on Cruise's pedestrian accident - Dan Luu, accessed February 9, 2026, https://danluu.com/cruise-report/
11. Lessons from the Cruise Robotaxi Pedestrian Dragging Mishap, accessed February 9, 2026, http://users.ece.cmu.edu/~koopman/pubs/Koopman2024_CruiseMishap_IEEEReliabilityMagazine.pdf
12. Office of Defects Investigation (ODI) Resume - nhtsa, accessed February 9, 2026, https://static.nhtsa.gov/odi/inv/2025/INOA-PE25012-19171.pdf
13. US regulators launch investigation into self-driving Teslas after series of crashes, accessed February 9, 2026, https://www.theguardian.com/technology/2025/oct/09/tesla-cars-self-driving-us-regulators-investigation
14. Tesla FSD Safety Issues: NHTSA Probes & AI Driving Future (Part 6) - PRIZ Guru, accessed February 9, 2026, https://www.priz.guru/tesla-fsd-safety-issues-nhtsa-probes-ai-driving-future-part-6/
15. AI & Robotics | Tesla, accessed February 9, 2026, https://www.tesla.com/AI
16. A Survey on Occupancy Perception for Autonomous Driving: The Information Fusion Perspective - arXiv, accessed February 9, 2026, https://arxiv.org/html/2405.05173v2
17. New Study: Waymo is reducing serious crashes and making streets safer for those most at risk, accessed February 9, 2026, https://waymo.com/blog/2025/05/waymo-making-streets-safer-for-vru
18. On Waymo's Traffic Jams - Stanford Center for Internet and Society, accessed February 9, 2026, https://cyberlaw.stanford.edu/blog/2025/12/on-waymos-traffic-jams/
19. Self-driving cars may create more traffic congestion than they solve, expert says - KJZZ, accessed February 9, 2026, https://www.kjzz.org/the-show/2026-01-07/self-driving-cars-may-create-more-traffic-congestion-than-they-solve-expert-says
20. A Systematic Literature Review on Vehicular Collaborative Perception – A Computer Vision Perspective - arXiv, accessed February 9, 2026, https://arxiv.org/html/2504.04631v2
21. When Robotaxis Get Attacked: Do Waymo Cars Need a 'Danger Escape Mode'?, accessed February 9, 2026, https://aragonresearch.com/robotaxis-attack-waymo-cars-danger-escape-mode/
22. MIC-BEV: Multi-Infrastructure Camera Bird's-Eye-View Transformer with Relation-Aware Fusion for 3D Object Detection - arXiv, accessed February 9, 2026, https://arxiv.org/html/2510.24688v1
23. [AV Vol.3] BEVFusion: Unifying Vision in Autonomous Driving Systems - Medium, accessed February 9, 2026, https://medium.com/demistify/av-vol-3-bevfusion-unifying-vision-in-autonomous-driving-systems-b2190f877c9b
24. A Transformer-based Temporal Feature Fusion Approach for Autonomous Driving BEV Perception | Request PDF - ResearchGate, accessed February 9, 2026, https://www.researchgate.net/publication/395804054_A_Transformer-based_Temporal_Feature_Fusion_Approach_for_Autonomous_Driving_BEV_Perception
25. The Marabou Framework for Verification and Analysis of Deep Neural Networks - Stanford Center for AI Safety, accessed February 9, 2026, https://aisafety.stanford.edu/marabou/MarabouCAV2019.pdf
26. Marabou 2.0: A Versatile Formal Analyzer of Neural Networks - arXiv, accessed February 9, 2026, https://arxiv.org/html/2401.14461v1
27. Marabou 2.0: A Versatile Formal Analyzer of Neural Networks - Stanford CS Theory, accessed February 9, 2026, https://theory.stanford.edu/~barrett/pubs/WIZ+24.pdf
28. Creating a Formally Verified Neural Network for Autonomous Navigation: An Experience Report - CSE CGI Server, accessed February 9, 2026, https://cgi.cse.unsw.edu.au/~eptcs/paper.cgi?FMAS2024.12.pdf
29. Verification of Neural Networks for Safety and Security-critical Domains - CEUR-WS.org, accessed February 9, 2026, https://ceur-ws.org/Vol-3345/paper10_RiCeRCa3.pdf
30. Formal Verification of Neural Networks for Safety-Critical Tasks in Deep Reinforcement Learning, accessed February 9, 2026, https://proceedings.mlr.press/v161/corsi21a/corsi21a.pdf
31. Formal Methods for Trustworthy AI-based Autonomous Systems - NII Shonan Meeting, accessed February 9, 2026, https://shonan.nii.ac.jp/docs/No.178.pdf
32. Formal Verification of Neural Networks-Based Control Architecture for Safety-Critical Autonomous Systems - Frontiers, accessed February 9, 2026, https://www.frontiersin.org/research-topics/74336/formal-verification-of-neural-networks-based-control-architecture-for-safety-critical-autonomous-systems
33. Implementing Responsible AI for Automotive Vehicle Safety - LHP Engineering Solutions, accessed February 9, 2026, https://www.lhpes.com/blog/implementing-responsible-ai-for-automotive-vehicle-safety
34. Functional Safety vs. SOTIF: What Is the Difference and Where Do They Overlap? - MES, accessed February 9, 2026, https://model-engineers.com/en/blog/functional-safety-vs-sotif-differences-overlaps/
35. The Necessity of a Holistic Safety Evaluation Framework for AI-Based Automation Features, accessed February 9, 2026, https://arxiv.org/html/2602.05157v1
36. What is SOTIF? (ISO 21448) - Visure Solutions, accessed February 9, 2026, https://visuresolutions.com/automotive/iso-21448/
37. Safety-Related Systems in Road Vehicles with Artificial Intelligence Are Addressed in ISO/PAS 8800:2024 | UL Solutions, accessed February 9, 2026, https://www.ul.com/sis/blog/safety-related-systems-road-vehicles-artificial-intelligence-are-addressed-isopas-88002024
38. ISO 26262, SOTIF and simulation | Applied Intuition, accessed February 9, 2026, https://www.appliedintuition.com/blog/iso26262-sotif-simulation
39. Introducing ISO/PAS 8800 – Functional Safety for AI in Road Vehicles | SGS Georgia, accessed February 9, 2026, https://www.sgs.com/en-ge/news/2025/04/safeguards-04625-introducing-iso-pas-8800-functional-safety-for-ai-in-road-vehicles
40. NIST vs ISO - Compare AI Frameworks - ModelOp, accessed February 9, 2026, https://www.modelop.com/ai-governance/ai-regulations-standards/nist-vs-iso
41. AI Safety vs AI Security in LLM Applications: What Teams Must Know - Promptfoo, accessed February 9, 2026, https://www.promptfoo.dev/blog/ai-safety-vs-security/