Seven questions every AI answer should survive

It’s easy to be impressed by an AI system. It’s much harder to know whether you should trust one. We wanted a test simple enough to remember and strict enough to mean something — so we wrote one, and we hold our own work to it before anyone else’s.

Before a recommendation is worth a real decision, it has to be able to answer seven questions.

1. Why was this recommended? The reasoning has to be there to see — not just the conclusion.
2. What evidence supports it? Sources, cited, and open to inspection.
3. What did it assume? Assumptions stated plainly, not buried.
4. What is still uncertain? The unknowns named, along with how much they matter.
5. Who is accountable? A specific person, identifiable before the fact — not located after the harm.
6. How can it be challenged? A real path to question or correct it.
7. How can it get better? What would change the answer, and how it improves over time.

If a system can’t answer these, it isn’t yet worthy of trust — however capable or convincing it sounds.

Notice what the test is not. It doesn’t ask whether the answer is impressive, or fast, or even, by itself, correct. A correct answer you can’t check is still a gamble; a careful answer that shows its work can be trusted even when it turns out to be wrong, because you can see how, and fix it.

We deliberately published the first version early and imperfect — v0.1 — so that we would be the first to be measured against it, and we revise it in the open. We don’t sell certificates for meeting it. We meet it ourselves, in the open, and invite anyone to use it on us.

That last part is the point. A standard you impose on others is a weapon. A standard you hold yourself to is a promise.