Jev Question Design
The capability answers typed questions over a supplied state and returns a
typed answer with the model's probability. It generates no text and decides
nothing: the answer is a labelled model signal a person reads, never a
completion predicate, a merge approval, a queue mutation, or any other
decision that is hard to undo. A narrow exception is the todo --auto cycle's
own candidate ranking, the auto-scoped ranking exception, which sets
selection order only; it is an in-process consumer behind the default-off
gate. Good questions keep that contract; bad
questions waste the call or manufacture false confidence. Four rules govern
question design.
Rule 1 — Compute the question in code
The question must be computed from the caller's inputs by the caller's code, not improvised as free-form prose at call time. A computed question is reviewable, stable across calls, and testable: the same inputs always compose the same question, and a change in phrasing shows up as a diff. Compose question text with the project's language's string formatting from named inputs — never hand-type a question into a prompt-shaped literal that drifts from the code that acts on its answer.
Rule 2 — Always admit a no-match answer
A choice question whose option set forces a pick manufactures confidence the call cannot have. Every choice question carries an explicit no-match option ("none of these fit", "no listed option") so the answer can say the honest thing. Same discipline for noul questions: ask only when both YES and NO are answers a reader would treat as informative — a question with only one meaningful answer is rhetoric, not a question.
Rule 3 — Keep the state small
Input is charged per token and output is not, so cost tracks state size, not question count. Several questions over one state in one call keep answers mutually consistent and cost nothing extra. Supply the minimum denatured facts the question needs — the axis being judged, the options, the counts that matter — and nothing else. Oversize payloads are refused unsent; a state that must be truncated to fit is a signal to redesign the question, not to trim the payload by hand.
Rule 4 — Phrase both noul polarities, and record which one you asked
A noul question has two phrasings for every underlying question ("is the premise alive" vs "is the premise dead"), and the same YES means opposite things under them. Compose the polarity deliberately, record it next to the question id in the caller's own output, and never re-label an answer into the opposite polarity after the fact. An answer read under the wrong polarity is worse than no answer: it is a confident wrong signal attributed to a model that never made that claim.
What an absence means
Every non-answer is a typed unavailability (gate off, no credential, endpoint unreachable, oversize, secret refused unsent). Each is a NO SIGNAL, never a negative answer: an unavailable judgment means the consumer proceeds without it, exactly as before the call was written. Silent defaulting — treating an absent answer as either polarity, or as "no difference found" — is the failure mode these rules exist to prevent.
