None was the right answer

This was a desk test of Jev, TypeSafe’s typed-decision layer. Not a chat. You hand it a paragraph and a closed set of options; it returns a label. The primitives we used were Noul (yes/no with a probability) and Choice (one option from a fixed list).

The question we wanted to stress: what happens when the options are the wrong size for the piece.

We took four English desk writeups of the Sep 17 unsealing in the New York Times case against OpenAI and Microsoft. First Choice: does the paragraph sit in a US, China, or Russia frame, or none?

All four came back none.

We didn’t ask for an essay. The first call was a Choice with four options and short criteria — US / China / Russia / none — each described as “the paragraph sits in this power frame,” with none for everything else. Four desk pieces on the Sep 17 filing all returned none.

Then two Nouls on the same state: theft frame, news-substitute frame. Each had a one-line what and a not_for (opinion without the claim; fair-use defense alone). Both landed yes on all four. On TechCrunch the theft Noul was high enough that code would have gated without a human — that’s the System One pattern: software owns the path; the model only answers the atom.

substr('theft') would light up on TechCrunch either way — Microsoft’s line and a dry fair-use paragraph that only mentions the word. The Noul didn’t look for the string. It scored the frame against a not_for (opinion / fair-use defense alone). Same closed set of outcomes; different job than membership.

That label was right. A copyright filing about training data is not a great-power story. The global set was the miss.

So the experiment’s point is small and usable: typed labels are sharp when the options can sit in the paragraph. They go quiet — correctly — when the set is bigger than the claim.