The Verification Gate: A Work Method for Generative AI in Consequential Engineering Practice

Generative AI tools draft, translate, summarise and organise well. They fail at the points where a practising engineer must sign. Five such points recur: a table read from an image, an analogue gauge, a standard clause, a manufacturer fault code, and direct arithmetic. Peer-reviewed benchmarks in building engineering document the first and the last. Manufacturer documentation, a preprint and practitioner reports document the rest. In each case the output is fluent and unmarked. Nothing in the text signals the error. The Verification Gate is a work method for using these tools where an error costs money or safety. It has one axiom and three steps in fixed order. It places one mandatory hold point between tool output and any commitment. It adds four fixed-structure instruments, a five-level task-tagging scale, a tool-free competence test, and a five-step installation procedure. It names seven failure classes (G1–G7) and maps each to the control that catches it. It does not teach tools. It fixes the order in which the engineer and the tool work. This paper fixes the method's vocabulary, states its boundaries, and defines a pre-registered evaluation protocol. No results are reported here, and no data has been collected. The protocol names its fields, its closed lists, and the sample size it needs. Any group that meets the stated criteria and submits verification logs under those fields is a valid application. Version 1.0 will report measured results if and when such logs exist. Files: English working paper (primary) and Hebrew edition. The instruments sit in a separate record under CC BY 4.0. Version 0.91 — working paper. Failure classes F1–F7 renamed G1–G7 (G for Gate). No claim, number, citation or structure changed. Method page: https://www.physical-ai-safety.com/method/verification-gate