Skip to main content
School of Engineering JudgmentThe verification gateA mapping call, at no cost
WhatsApp058-4887876
A mapping call, at no cost

All articles

The Verification Gate: a hold point for AI output

Published

Check every item against a source before signing. One blocked item holds the output at draft.

The Verification Gate is a work method for using generative AI where an error costs money or safety. It puts one mandatory hold point between the tool's output and any commitment. Every item in the output is either traced to an authoritative source or blocked. The method does not make the output correct. It makes the check countable.

The problem it answers

A split air-conditioning unit shows fault code E3. A technician asks a tool what it means. Maker A lists E3 as actuation of the high-pressure switch. Maker B lists it as low pressure. Maker C lists it as a remote-controller transmission error. A technician who acts on the wrong one adds refrigerant to an overcharged system, or recovers charge from a system that is already short.

The tool is not weak. One language model passed two national engineer exams for building mechanical equipment on all five attempts (Choi, Lee, and Kim 2025). Its mean scores were 80.6 and 81.25 out of 100. Capability is not the gap.

The gap is that no step in the workflow stops the output before signature. Engineering already has such steps. A checker's signature precedes drawing release. A pressure test precedes commissioning. AI output has no equivalent, so it can enter a signed report without any check.

The axiom

A number that was not verified does not exist.

In practice, a value that originates in tool output has no standing as data until someone checks it against an authoritative source. Until then it is a draft. It may not enter a signed or submitted document, or a decision.

An authoritative source is a document with an owner and standing. That means a manufacturer manual, a standard in its official text, an approved procedure, a signed drawing, a direct measurement or a manufacturer data table. A tool is never an authoritative source. The output of a second tool is not one either.

Three steps and one gate

The order never changes, and nothing gets skipped.

  1. Do it yourself. A first pass with no tool: an estimate, a manual reading, a written hypothesis. This takes five to twenty minutes. The purpose is a recorded independent judgment, and it does not have to be an accurate one.
  2. Ask the tool. A structured prompt with six parts: role, context, task, constraints, source, and the missing instruction: do not invent, write MISSING.
  3. The gate. Compare every item in the output with an authoritative source, find the gap, and log the result.

Why does order matter? An estimate made after seeing the answer anchors to it. The motivation comes from a field experiment with nearly one thousand secondary students (Bastani et al. 2025). Students with open access to a chat tool later scored, without it, below students who never had it. The study is in secondary mathematics, not engineering practice, and we take it as motivation only.

Four results, no fifth

Every check ends in one of four.

Partial verification is not one of the results. One blocked item holds the whole output at draft. Only output that passed in full, with a complete log, counts as a deliverable. Nothing else may carry a signature, go out as a submission or feed a decision.

What makes it countable

Four documents with a fixed structure.

The level follows the consequence of an error, and difficulty plays no part. The default for an untagged task is level 0, no tool. A simple calculation that enters a signed survey is level 0. On a routine work order, three fields can be enough: source, value checked, and who checked.

What it catches

The method names seven recurring failure classes. They share one property: the output carries no external marker of error. So the check cannot depend on suspicion. It runs on every item and leaves a record.

The evidence is not equally strong, and you should know that. Table from image and direct arithmetic rest on peer-reviewed research. Analogue gauge rests on a preprint, and standard clause on a professional report. For fault codes, the manufacturer tables show the conflict, and the tool's error is an inference. Right answer - wrong reasoning is a field observation, and nobody has measured it.

The gate is the primary control for six of the seven. The seventh passes a source check, and Step A and the error-hunt quiz are what catch it. That whole mapping is a design claim, not a measurement.

What it is not

The method does not detect hallucinations, tune prompts or rank tools. It does not teach tools either. It fixes the order in which the engineer and the tool work.

It does not promise that the deliverable is correct. It promises that someone checked every item against a source and recorded the result. If the source is wrong, the deliverable is wrong, and the log shows exactly where.

It does not promise speed. Step C adds time per deliverable. Whether that is less than the cost of a signed error is an empirical question.

It does not replace professional competence, and responsibility stays with the signer. ASHRAE and ASTM take the same position (ASHRAE 2025; ASTM 2025).

It does not apply everywhere either. With no document that has an owner, there is nothing to verify against. A creative task that leads to no commitment is level 1, and the gate does not apply. Certified engineering calculation software is not a tool in this sense.

Where the evidence stands

The working paper is version 0.92, and it is not peer-reviewed. We have collected no data and recruited no group. The external evidence measures the failures the method targets, and none of it measures the method.

The paper defines a pre-registered evaluation protocol: roughly 300 logs, about 30 participants submitting about 10 each. Version 1.0 will report measured results if and when such logs exist. The claim we defend is narrow: under the method, every item traces to a source or stays blocked, and you can count the trace. Data could falsify it.

Try it on one deliverable tomorrow

You do not need an organisational rollout to see how this works. Take one deliverable a tool touched this week.

  1. Write down your own estimate before you look at the output again. If you have already seen it, note that.
  2. Mark every item in the output: number, part number, standard clause, fault code, reported action.
  3. Check each one against an authoritative source, and record the source and the location in it.
  4. Count: how many match, how many gaps, how many blocked, and how many minutes it took.

You now hold a figure of your own, on a deliverable of your own. If one blocked item remains, the deliverable is still a draft.

In an organisation the method goes in through five steps, and training is step 4: map, tag, embed in existing templates, train and measure. The decision rule, the scale and the log sit together on the method page.

References