Main idea

An answer produced by a model from a repository can assert that a rule exists without showing anything. It is then one more piece of approximate documentation — faster and better phrased than the last one, no more reliable. All you have done is replace a Product Manager's imperfect memory with a model's probabilistic memory.

The requirement that changes something is tying every assertion to a designatable element: the function that applies the rule, the exception that blocks the case, the test that covers the scenario, the merge request that introduced the behavior, the state of the feature flag, the product decision that arbitrated it.

What matters isn't that this evidence gets read. A support agent or a salesperson won't open it, and it can sit at the bottom of the document. What matters is that it has to exist: an answer that must cite can no longer be produced by plausibility. The constraint operates at the moment of production, not at the moment of review — which is what distinguishes it from a traceability requirement, which presupposes a reader to verify it.

Why it matters

This moves verification from an expensive human step, rarely taken, to a property of the request itself. You don't get a guarantee that an answer is right; you get that it cannot be produced out of thin air.

It also gives the shape to give any request addressed to a model over a corpus: demand, for each conclusion, the element of the corpus it rests on. The same device works for analyzing a body of texts, where you ask for the sentence or the paragraph that supports each conclusion.

Nuances and limits

A cited piece of evidence can be badly chosen: a function that exists isn't necessarily the one that executes, and a passing test doesn't establish that the case is covered the way you think. The constraint reduces gratuitous assertion, it doesn't guarantee that the support is relevant.

And a lavishly sourced answer gains apparent authority, which can discourage the very review it made possible.

Open questions

• What distinguishes evidence that genuinely establishes the behavior from evidence merely compatible with the assertion?