Main idea

The questions used to delimit a reference model's ground do not stop at framing. Once an area is described, you take them one by one and look for the answer in the cards alone — no opening the code, no calling a colleague, no relying on memory. What is not found there does not exist for the person using the reference model.

The test measures neither the volume written nor the beauty of the model: it measures the one thing the reference model promised. And it surfaces a flaw of a particular kind — the rule written in general terms where the question expected the detail. "Certain operations require validation" looks like a rule, occupies the line of a rule, and leaves the question unanswered.

That is the most likely flaw in this kind of work, and the hardest to see from the inside. Counting cards would never have shown it.

Why it matters

It provides an acceptance procedure a third party can apply without knowing the domain: they only have to ask the questions and observe whether the answer comes out. The writer's judgment of their own work does not enter the measurement.

It also fixes the right unit of control: a question satisfied, not a card written. The two do not move together.

Nuances and limits

The test says nothing about what is missing outside the scope of the questions chosen: it verifies a promise, not coverage. A reference model can answer all its framing questions and ignore whole areas of the product.

And a question left unanswered does not always indict the reference model: it may point to an area where the product itself has never decided.

Open questions

• When should the set of questions itself be reopened, given that the product and its uses evolve faster than the corpus used to frame them?