Reading the evidence

Once you have the conversations, the failure mode changes. It is no longer that you asked the wrong questions; it is that you read the answers the way you wanted to. So separate the reading into two physical columns, kept apart: what they said, verbatim, and what you think it means. Anything that starts "I think they meant" belongs in the second column and must never migrate into the first.

Signals that count: they have already spent money on the problem, or built a workaround, or assigned a person to it, or lost a customer over it. Anything that has already cost them something is real. Signals that do not count: interest, enthusiasm, "you should definitely build this", introductions offered but not made, and any sentence with "would" in it.

The disconfirming quotes are the ones to keep most carefully. If nobody in twelve conversations said anything that made you uncomfortable, you either asked leading questions or you edited the record. A file with no inconvenient quotes in it is a file that has been curated, and a reviewer will read it as one.