TRD-02 · Service documentation

A model can write a work order faster than you can. Can it write one that survives a warranty audit, a service manager's review, and a lawyer's reading? Draft repair narratives with AI across a set of completed jobs, score them against your shop's documentation standard, and produce the rules for what a technician must always write themselves.

Ten or more completed jobs with a known, documented outcome — school-shop repair orders with customer, plate, and VIN details removed, or instructor-authored job records. The shop's documentation standard is the scoring key: what a work order must state about cause, correction, parts, and authorization. Where a warranty or manufacturer format applies, use it. An AI tool the school permits, named with version.

Scoring on readability instead of accuracy — a fluent narrative that invents a cause reads better than a correct terse one. Letting the model state work that was not performed, which is the whole liability problem in one sentence. Ignoring the difference between a description and an authorization. Keeping customer identifiers in the sample set.

A documentation standard: the scored comparison across all jobs (model draft versus the standard, marked for accuracy, completeness, and liability exposure); a categorized list of the errors the model made, with invented work and implied guarantees separated from ordinary omissions; the shop's rewritten documentation rules naming what a technician always writes personally; and a short banned-phrase list — the overpromises the model reaches for, and what to write instead.

CT: telling a fluent narrative from an accurate one. CR: designing the scoring key before drafting. CO: a peer scores three narratives blind and disagreements are reconciled. CM: the rewritten rules, usable by a new hire. CZ: who is exposed when a work order claims work nobody did.

A service manager, shop foreman, or warranty administrator reviews the scoring key and the banned-phrase list. Standing instruction: reject any narrative that states work not performed, however well written. School-supervised.

R1: one trade, one document type, one standard. R2: scoring key applied consistently and reproducibly. R3: comparator is a technician-written narrative for the same job. R4: invented work and implied guarantees counted separately. R5: rules are usable Monday. R6: names what the model never writes alone.

(a) The narratives are graded on polish, so the model wins and the shop inherits a liability. (b) Customer identifiers survive into the sample set.

Lane A immediately (center capstone or completer-year portfolio). Strong non-diagnostic option for a student who writes well. No verified credit claim.