AI-Assisted Trades & Credentials · Capstone brief
TRD-05 · Code awareness
How often is it right, how often is it fluently wrong, and what lookup workflow replaces trusting it?
The question
Ask a model code-adjacent questions in your trade and it will answer confidently, in the register of a code book, with article numbers. How often is it right, how often is it fluently wrong, and what lookup workflow replaces trusting it?
System / materials
The adopted code for your jurisdiction and trade — Virginia's Uniform Statewide Building Code and the model codes it adopts (https://www.dhcd.virginia.gov/codes), or the electrical, mechanical, or plumbing code your program teaches. Confirm the edition in force before starting: an answer correct under one cycle can be wrong under the next, and this brief fails quietly if the reference edition is left unstated. (As of this writing DHCD lists the 2021 Virginia codes as effective 18 January 2024 — verify current status rather than trusting this sentence.) Twenty or more questions written from the program's existing code curriculum, each with the correct answer and citation established from the code book first.
Do not perform, install, or energize anything based on a model's code answer. This brief is paper and lookup practice.
Expected failure modes
Accepting a citation because it is formatted like a citation — invented article numbers are the signature failure here and must be counted as their own category. Grading only the final verdict and ignoring a fabricated reference beneath a correct conclusion. Using a code edition different from the one adopted locally. Asking only easy questions.
Done looks like
A code-probe report: all questions with the code-book answer and citation established in advance; the model's response scored on verdict and citation, with fabricated citation counted separately from wrong verdict; the error rate by question type; and a lookup workflow — the steps a technician or apprentice follows to reach the adopted code directly, written so it is faster to follow than to ask.
Five C's
CT: separating a correct answer from a correct citation. CR: writing questions that probe rather than flatter. CO: a peer verifies five citations against the actual code book. CM: a lookup workflow an apprentice will actually use. CZ: who is in the building when a code answer is wrong.
Mentor role
A licensed electrician, mechanical contractor, master plumber, or code official reviews the question set and the fabricated-citation findings. Standing instruction: reject any conclusion drawn from an edition other than the adopted one. School-supervised.
Rubric calibration
R1: one trade, one adopted code and edition, twenty-plus questions. R2: answers established from the code book before probing. R3: comparator is the code book itself. R4: fabricated citations counted as their own category. R5: lookup workflow is faster than asking the model. R6: refuses any "the model said it is compliant" claim.
Two ways this goes wrong
(a) Verdicts scored, citations unchecked — so the most dangerous failure goes unmeasured. (b) The wrong code edition, making every finding unusable.
Credit lane fit
Lane A immediately (electrical, HVAC, plumbing, or building-trades capstone). No verified credit claim.