Wirewalk Instituteof AI
mob-245 Intermediate Available

Functional Safety for Automotive AI: ISO 26262, ISO 21448 (SOTIF) and ISO/PAS 8800

For safety, ML and validation engineers who must produce the HARA, SOTIF analysis, data and validation evidence and safety argument for an ML pedestrian detector, with the team's LLM assistants classified as tools.

Level
Intermediate
Study time
14 hours
Lessons
16 in 5 modules
Exam
20 questions

About this course

ISO 26262, ISO 21448 (SOTIF) and ISO/PAS 8800:2024 are voluntary consensus standards, not US law. US manufacturers self-certify to the FMVSS, and vehicle and original-equipment manufacturers must report safety-related defects within five working days of determining them (49 CFR 573.3 and 573.6). The standards are how automakers and suppliers show due care and meet supply contracts. This course teaches the work products those standards expect when the safety-critical part of a feature is a neural network, and how the engineers who own them produce and check them.

One case runs through every module. Tallgrass Vision Systems, a fictional Michigan Tier 1 supplier, has signed a development interface agreement with Harlan Motor Company, a fictional automaker, to deliver the FC5 front camera and its PDN-5 pedestrian detector for pedestrian AEB. A late scope change has moved the braking decision into the camera, so the platform's safety assumptions no longer hold. Tallgrass has eight weeks to an approved HARA, an Automotive SPICE assessment in February, a safety release gate on March 30, 2027, a $4.2 million validation budget and a July start of production. Along the way the team finds an LLM-drafted HARA that rates sudden braking as controllable, a test set that leaks into training, an INT8 model that passes on average and fails children at dusk, a validation target that needs about 300,000 failure-free hours (and a ledger that only meets it by counting hours on the wrong build), phantom braking under overpass shadows in the field, and a retrained model that is better on average and worse at night. The honest best outcome is a conditional recommendation with explicit pre-deployment evidence gates, including unresolved hardware boundary and degraded-mode claims.

The course keeps two kinds of AI apart: the ML component inside the vehicle that must be assured, and the LLM assistants the safety team uses, which are tools to classify for tool confidence, never the author of record. AI never assigns an ASIL, accepts residual risk, makes the defect determination or acts as the independent assessor. Standard text is not reproduced; concepts are taught in our own words and parts are cited, not clauses. LLM examples are labeled illustrative case drafts and comparison excerpts; no authenticated model-run or benchmark claim is made. Vendor products are described from their own pages as of October 2026 and are not endorsed. NHTSA rules, FMVSS testing and UL 4600 safety cases are covered in mob-255. This course is educational, not legal advice.

What you will be able to do

  • Evaluate an LLM-drafted HARA for an ML feature: route the work to the right standard and owner, remove invalid events, add missed ones, and rate S, E and C yourself with a sourced rationale
  • Classify each LLM, ML framework and compiler use for tool confidence with its compensating verification, and judge an LLM co-pilot by its measured recall and precision
  • Build a triggering-condition catalog in which every item has evidence, and set SOTIF acceptance criteria for missed pedestrians and phantom braking from risk and baseline data before testing
  • Choose a safety measure and threshold from trade-off data, and write an independence argument that names every shared cause before claiming an ASIL decomposition
  • Write per-ODD-bin AI safety requirements with lower bounds and minimum sizes, and find the coverage gaps, leakage and circular label evidence an Automotive SPICE assessor would raise
  • Calculate the failure-free exposure a target needs, count only independent in-ODD hours on the frozen release software, and decide target equivalence from per-bin bounds
  • Escalate a fleet signal against a concurrent control within the contract window without a defect determination, and stop a retrain whose per-bin results regress
  • Create the safety argument and release recommendation for an ML component, flagging unsupported and stale claims and stating gaps on the summary page

Syllabus

5 modules · 16 lessons · about 14 hours

Module 1: The DIA Is Signed: Scoping the Job, the HARA and the Rules for the Team's AI

Map the FC5 pedestrian detector to the three standards and their owners, run a HARA with an LLM drafting hazardous events but never ratings, and classify the team's LLM uses before they touch a work product.
3 lessons
  1. Reading the DIA: What ISO 26262, ISO 21448 and ISO/PAS 8800 Each Ask of the FC5 Pedestrian Detector
  2. A HARA for FC5 Pedestrian AEB With an LLM Drafting the Hazardous Events
  3. Before the Assistant Touches a Work Product: Tool Confidence for the Safety Team's LLMs
  • Decision Lab · 12 minThe Eight-Week HARAChoose how to produce an approved in-context HARA for FC5 pedestrian AEB in eight weeks, what the team's LLM assistant may draft, and how to answer the assessor's findings.

Module 2: Seeing the Child at Dusk: The SOTIF Analysis and the Architecture Around the Detector

Build the triggering-condition catalog from FC4 miss logs, set acceptance criteria for missed pedestrians and phantom braking alike, and design a safety measure that does not share the detector's blind spots.
3 lessons
  1. Triggering Conditions From Evidence: Building the FC5 SOTIF Catalog From FC4 Miss Logs
  2. Monitors, Radar Confirmation and Degraded Modes When the Detector Cannot Carry the Claim
  3. Phantom Braking Is a Hazard Too: Acceptance Criteria for Both Directions
  • Decision Lab · 14 minThe Dusk ClusterBuild FC5's triggering-condition catalog, set acceptance criteria for both missed pedestrians and phantom braking, choose the safety measure around the detector, and respond when a new condition arrives from the field.

Module 3: The Recall Number Nobody Should Trust: AI Safety Requirements, Data and Labels

Turn safety goals into per-bin AI safety requirements and a dataset specification, remove leakage and circular label evidence, and prepare ML evidence for an Automotive SPICE assessment.
3 lessons
  1. From Safety Goal to Dataset Specification: AI Safety Requirements, ODD Coverage and Leakage
  2. Label Specs, Vendor Audits and Auto-Labels Without Circular Evidence
  3. Preparing the ML Team for an Automotive SPICE Assessment: MLE.1-MLE.4 and SUP.11 Evidence
  • Decision Lab · 12 minThe Recall That Was Too GoodDecide what PDN-5's AI safety requirement says, how to fix a leaky test split, who labels the test set, and how to answer an Automotive SPICE assessor who pulls one thread.

Module 4: Evidence for Gate S3: Validation Targets, the Target SoC and the Pipeline FMEA

Size and source the failure-free exposure the acceptance criteria need, prove the INT8 model on the target SoC per bin, and build the ML pipeline FMEA with an LLM as a recall check, not an author.
3 lessons
  1. Three Hundred Thousand Hours: Zero-Failure Targets and Where the Exposure Comes From
  2. Does the INT8 Model on the KS-800 Still Meet the Requirements? Per-Bin Equivalence and Hardware Faults
  3. The ML Pipeline FMEA With an LLM Co-Pilot, and Tool Confidence for the Compiler
  • Decision Lab · 14 minThree Hundred Thousand HoursPlan the exposure for FC5's unintended-braking criterion on a $4.2 million budget, decide whether the INT8 build is equivalent, choose how to build the FMEA, and take a position at Gate S3 when a lineage check removes 30,500 hours.

Module 5: Keeping the Argument True: Field Events, Retrained Models and the Safety Case

Escalate a fleet cluster within the contract's notice window without prejudging the defect decision, catch a retrained model's hidden regression before the OTA, and assemble the safety argument and release recommendation.
4 lessons
  1. Forty-One Complaints: Triaging Fleet Events Into an Escalation Without Making the Defect Decision
  2. A Retrained Model Before the OTA: Change Impact Analysis That Catches Hidden Regressions
  3. The Safety Argument for the ML Component: GSN, Stale Evidence and the Assessor
  4. Capstone: The FC5 Safety Release Recommendation
  • Decision Lab · 12 minForty-One ComplaintsDecide when and how to notify Harlan of an FC4 phantom-braking cluster, what the memo says, how to contain it, and what data to send, without making the defect determination.
  • Capstone simulation · 35 minGate S3 to SOPRun FC5 from approving the 2027 V&V plan through Gate S3, SOP and the first field issue: choose the exposure plan, word the claims, choose who confirms the case, decide what ships and how to answer the field.

Exam and certificate

The exam unlocks once you have completed every lesson and played each Decision Lab to an ending. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 40 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.

Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.

The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Synopsys (Ansys medini analyze), MathWorks, NVIDIA, Foretellix, Applied Intuition, Kognic, Anthropic, Jama Software, or IBM. Product and company names are trademarks of their respective owners.

More in AI in Transport, Logistics and Mobility

All 14 courses in this track
mob-001 Free intro Available

AI in Transport, Logistics and Mobility: The One-Hour Briefing

A free one-hour briefing built on ten real transport jobs, not an overview: ranking late loads by predicted ETA, classifying an invoice to HTSUS, coaching from AI dash-cam events, turning aircraft-health alerts into work orders, coding safety reports, acting on machine-vision track findings, watchkeeping with an AI lookout, planning a CII-aware voyage, retesting a perception model and triaging satellite conjunctions. Each lesson shows where AI helps, where the person who signs must decide, and the rule that governs that task.

1 h · 9 lessons Free intro
mob-210 Intermediate Waitlist

Building and Deploying AI in Logistics and Supply Chain

Building and Deploying AI in Logistics and Supply Chain: an intermediate course in the AI in Transport, Logistics and Mobility track. Syllabus in preparation.

h · 0 lessons Join the waitlist
mob-120 Basic Waitlist

AI for Aviation Professionals

AI for Aviation Professionals: a basic course in the AI in Transport, Logistics and Mobility track. Syllabus in preparation.

h · 0 lessons Join the waitlist