Build the ADAS calibration documentation package a plaintiff's attorney would subpoena
The situation
Our program teaches students to run a calibration, hand over a scan report and call it done. That is not what a shop needs and not what survives a lawsuit. The most common documentation failures in the field are all things we never assess: no OEM procedure reference, no pre-repair scan, no record of tire pressure or floor level, no named technician, no verification that it passed.
Steps
-
Assign a real vehicle and make students find the OEM requirement themselves
Give a concrete scenario — a 2022 sedan with a replaced windshield, a bumper R&I. Students must locate the OEM's actual calibration requirement: static, dynamic, both, or scan-only, plus environmental preconditions. No generic ADAS charts.
What you only learn by doing it: Assign three different OEMs across the class and compare in debrief. Students who only look up one manufacturer come away believing there is a general ADAS procedure. There is not — and the moment a student sees two OEMs contradict each other, the “just ask the AI” impulse dies on its own.
-
Ask the chatbot for the procedure, then grade the chatbot
Any free tier plus the OEM document from step 1
Students prompt for the calibration procedure and preconditions for their specific vehicle, then build a two-column table: what the model said, what the OEM document says. Every discrepancy gets flagged.
What you only learn by doing it: The failure mode is specificity, not vagueness — which is what makes it dangerous. It will give a target distance to the tenth of a metre and a fuel-level requirement, both invented, both formatted exactly like the real thing. Tell students to expect plausible precision and have them mark the highest-confidence wrong answer they found.
-
Run the calibration with the documentation form open, not filled in afterward
NC3 Certification Center (Snap-on pathway)
Require the pre-repair scan before anything is touched and record environmental conditions live. The verification step — the report validating alignment, target placement and OEM compliance — is the one programs treat as an afterthought.
What you only learn by doing it: Make students photograph the tire-pressure gauge, the floor-level reading and the target setup as they go. Every cohort tries to reconstruct these numbers from memory at the end of lab, and every cohort's numbers are wrong. Reconstructed documentation is what the industry is actually failing at.
-
Assemble the package and have the AI attack it
Any chatbot, in an adversarial role
The package needs: OEM procedure reference, pre- and post-repair scans, recorded conditions, named technician and date, vehicle-specific detail, pass/fail verification. Then prompt the model as an insurer disputing the bill and respond to every objection in writing.
What you only learn by doing it: This adversarial prompt is the one place in the unit where the model is reliably good, because it is a completeness check against a structure rather than a factual lookup. Give students the role prompt verbatim — “review this document” produces flattery, “you are the party who does not want to pay this claim” produces a usable list.
-
Peer-review against the documented failure modes and resubmit
A printed checklist; I-CAR RTS articles
Students swap packages and score each other against the industry failure list: missing OEM reference, incomplete scans, undocumented conditions, missing technician ID, generic template, unverified results. The resubmission is what is graded.
What you only learn by doing it: Ban the generic template explicitly and check for it, because students will find one online and fill it in. A template without fields for this vehicle's specific preconditions produces a document that looks complete and proves nothing — which is precisely one of the documented field failures.
Where this breaks down
Be straight with students and your dean: the AI content here is small and mostly cautionary. It is a document-completeness assistant and a hallucination demonstration, not a diagnostic or calibration tool.
Do not under any circumstances let students calibrate to a spec a chatbot produced. The models generate fabricated target distances, preconditions and part numbers with full confidence and correct formatting, and a miscalibrated forward camera is a safety-of-life failure, not a grading error.
The equipment barrier is severe and unevenly distributed, so design the assignment so the documentation half stands alone — a program without funding can run steps 1, 2, 4 and 5 using a documented calibration from a partner shop.
OEM procedures change. Anything documented this term may be superseded, and the habit of checking the current procedure is more of the lesson than the procedure itself.
Provenance: the documentation standards and certification structure are documented — the failure modes come from published industry guidance, the pre/post-scan requirement from OEM position statements collected on I-CAR's RTS portal, and the CHECK/PLACE/CALIBRATE/PROVE structure from NC3's ADAS badge. The AI steps are extrapolated; we found no published account of a collision program using an LLM this way.