Main idea

An answer about how a product behaves is almost always partial. No test covers the case asked about; the repository holds two behaviors depending on the state of a feature flag; that particular customer's configuration isn't accessible; the written product decision states a rule the code seems to contradict; the support documentation states a third one.

An answer that keeps silent about those zones stays usable, readable and wrong in places — with nothing to say which ones. Whoever receives it, a support agent facing a customer or a salesperson in a meeting, uses all of it, including where it didn't hold. The risk hasn't gone down, it has changed hands: it sat with the person who knew they didn't know, it now sits with the person who thinks they know.

Stating where it gets stuck is therefore part of the answer, not an admission of weakness. It is what a serious human does — flagging the doubt and saying when a test is needed — and it is what lets the real objective be met: reducing the cost of research without reducing the traceability of what is asserted.

Why it matters

This sets an acceptance criterion for an automated answering system that bears neither on its speed nor on its accuracy rate: can it produce "I can't determine that"? A system that always answers completely is disqualified by that very completeness.

It also gives value to what looks like failure: flagging that a PDR and the code diverge, or that no test covers a case, is product information in its own right, sometimes more useful than the answer asked for.

Nuances and limits

The exercise has a threshold. An answer that hedges every sentence becomes unusable again, and the reader goes back to the habit of testing for themselves — the cost you meant to remove returns intact.

And the declaration of ignorance is no more reliable than the rest: a model can confidently claim not to know something it could have established, and that error goes unseen too.

Open questions

• How do you check that a system flags its real zones of ignorance, rather than the ones it finds easy to name?