Jev Decisions (TypeSafe)

Jev is NV oOS's decision provider — a System One model that returns typed, probabilistic answers (choice / score / noul) over supplied state instead of generated prose. Use it whenever the task is a judgment, not writing.

When to use Jev

• Routing / triage / classification against a fixed set of options
• Scoring content against a rubric (2–10 ordered levels)
• Yes/no judgments with a calibrated probability (noul)
• Safety screening with typesafe_guardrail

When NOT to use Jev

• Generating prose, summaries, or code — use the assistant's chat model
• Arithmetic, counting, date ordering — do those in code
• Any decision that must carry a written rationale

Question design rules

1. Batch everything into one call — parallel questions cost tokens, not latency.
2. Add an explicit other option to every choice question so the model can abstain.
3. Keep 2–10 ordered, distinct levels on score rubrics.
4. Use noul criteria (true/false descriptions) when the yes/no boundary is subtle.
5. Send only the state fields the questions need — irrelevant state degrades accuracy (context rot).
6. State is adversarial: user-controlled content can steer answers. Never route consequential actions on Jev output alone — keep them behind code gates.

Confidence: the three-path pattern

• confidence ≥ your high bar → act automatically
• middle band → confirm, review, or gather more information
• below the floor → route to a human or fall back

Use min_confidence on typesafe_decide to flag answers below a floor (they come back flagged below_threshold, never dropped), and weights to combine score answers into a composite locally — no extra API call.

Model pinning

• jev-latest follows new releases — pin jev-1.13.0 once thresholds are tuned.
• The response always reports the concrete model id that answered; log it when auditing decisions.

Guardrails

• typesafe_guardrail returns advisory verdicts (pass / review / block) per hazard category. It never blocks anything itself — enforce the verdict in code.

Billing & latency

• Input-only billing ($0.042 per million input tokens; output free).
• Identical repeated requests can be served from the decision cache when the site enables it (marked cached: true).
• Trim state to the fields the questions need — that is both cheaper and more accurate.