User interviews: how many before you trust the pattern?

I talk to five users and hear five completely different things.

I talk to five more and two themes start repeating, but I can't tell if that's signal or just the loudest people finding each other.

I've shipped features for one vocal user and regretted it for a year. I've also ignored a quiet pattern for months and paid for it when a competitor shipped exactly that.

There's a version of research where you keep going until nothing surprises you — but that's 40 interviews and I have a product to run. There's a version where you trust your gut after three — but my gut has been wrong in expensive ways.

How do you separate noise from truth, and at what number do you stop asking and start building?

Do you ever feel confident, or just less wrong than last time?