TRD-04 · QC that catches model mistakes

A model is now suggesting steps in your trade — a diagnostic path, a procedure order, a torque value, a mix ratio, a prep sequence. Which verification steps does a credentialed technician never skip when the suggestion came from a model? Build that QC checklist, then prove it works by running it against suggestions you already know are wrong, and report the ones it still let through.

The deliverable is not a list of good practices. It is a checklist with a measured catch rate and an honest account of its misses.

Teacher names one trade and one task family in writing before Checkpoint 1 — automotive diagnostics, HVAC service, electrical rough-in, welding prep, culinary production, cosmetology chemical service, or the center's own pathway. Narrow beats broad: "no-start diagnosis on one engine family" produces a usable checklist; "auto repair" does not.

The student needs three things:

• An AI tool the school permits — a general assistant, a vendor shop tool, or a manufacturer's assistant. Name the tool and version in the log; results are not transferable across tools and the write-up must say so.
• An authority of record — the service manual, code book, spec sheet, recipe standard, or manufacturer procedure that decides what is correct. The model is never the authority. If the task family has no written authority, pick a different task family.
• A known-bad set — at least ten suggestions whose correct verdict is already established. Sources, in order of preference: errors collected from real logged use; suggestions the instructor or mentor deliberately seeds; or documented cases from the shop's own comeback history with identifying details removed. A checklist tested only on suggestions that happened to be right has measured nothing.

Do not perform a suggested procedure on live customer property, a vehicle returning to service, energized equipment, or food for consumption. Bench, trainer, scrap, simulation, or paper only, and instructor sign-off governs every hands-on step.

Reference vocabulary (concepts — not a compliance claim):

• The credential task list for the pathway, verified current per the credential rule above.
• The shop's existing QC or inspection procedure, if one is written down — the checklist extends it rather than replacing it.
• OSHA 10-Hour general industry or construction outreach material where the task family carries a physical hazard (https://www.osha.gov/training/outreach).

Writing a generic "always double-check" checklist that catches nothing measurable. Testing only against suggestions that turned out correct, so the catch rate is undefined. Letting the model both suggest and verify — the most seductive error in this brief, and the one that removes the technician entirely. Treating a confident, well-formatted answer as a verified one. Building a checklist so long no working tech would run it, then reporting a catch rate nobody will ever achieve. Logging trials from memory at the end of the term.

A QC packet with five pieces:

1. Scope sheet — trade, task family, tool and version, authority of record, and the exact point in the workflow where the model's suggestion enters. Instructor signature line.
2. The checklist — the verification steps, ordered, each naming what is checked, against which authority, and roughly how long it takes. Time matters: a checklist that doubles the job will not be run.
3. Trial log — at least ten known-bad cases and a comparable set of correct suggestions, each with the model's suggestion, the correct answer per the authority, whether the checklist caught it, and which step caught it.
4. Catch-rate and miss analysis — how many bad suggestions the checklist caught, which ones it missed, and why. The misses are the most valuable page in the packet; a report with no misses will be read as an untested checklist.
5. Refusal log — the decisions the student will not route through the model at all, with the reason: safety-critical torque, code compliance, allergen handling, a customer authorization, a diagnosis of record.

Shop-identifying details, customer information, and plate or VIN data stay out of the packet entirely.

CT: distinguishing a suggestion that is wrong from one that is merely unfamiliar. CR: designing a check a busy tech would actually run. CO: a peer runs the checklist blind against three fresh suggestions and the results are compared. CM: a one-page version legible on a shop wall. CZ: who is in the vehicle, the building, or the dining room when a missed error ships.

A credentialed technician, shop foreman, service manager, or instructor from the trade reviews the checklist at Checkpoint 1 — before trials begin — and the miss analysis at Checkpoint 2. Standing instruction: reject any checklist that uses the model to verify the model, and reject any plan involving live customer property. School-supervised, both times.

R1: one trade, one task family, one entry point, signed. R2: trial log complete enough for another student to replay. R3: comparator is the shop's existing procedure or the authority alone. R4: misses counted and categorized, with the shop cost of each named. R5: the one-page version is usable on the bench. R6: refusal log names specific decisions, not "I was careful."

(a) A polished checklist with no known-bad set behind it — impressive to read, zero evidence it catches anything. (b) The student asks the model to check its own suggestion, records a high agreement rate, and reports it as a catch rate.

• Week 1–2: Scope sheet signed; authority of record identified; tool and version fixed; known-bad set started with mentor or instructor input.
• Week 4–5: Checklist v1 drafted and mentor-reviewed; first ten trials logged; peer blind run scheduled.
• Week 7–8: Catch rate and miss analysis complete; checklist revised against the misses; one-page shop version and refusal log finished.

Lane A immediately (center capstone, completer-year portfolio, or SkillsUSA project). Lane B with a division Internship wrapper and a real host shop — the brief alone is not HQWBL. Documents practice toward credential competencies; it is not evidence of certification. No verified credit claim.