Decide in code

Choice / Score confidence is a shape statistic: peaked is high, flat is low. The full probabilities ship in the response, so you can compute your own. Flavio copies the docs’ three bands: low (e.g. <0.5) to a human; mid confirm or gather more; high auto-act. Destructive work gets a higher bar (e.g. >0.9). Examples, not defaults.

Do not ask “is this good.” Split the ruler; mix weights in code. Jev also will not invent text to type: browser-use calls a small LLM only on TYPE_TEXT. Dates, counts, money, permissions stay in ordinary code — OmniaKey’s checklist.

Jev’s output contract is concrete… but the public evidence does not yet support the strongest intelligence story. Treat Jev as a fast, cheap decision layer that still needs labels, thresholds, human escalation and a separate generator when the workflow must write words.

• After you tune thresholds, pin jev-1.13.0. Aliases move. Log the resolved model.
• Langfuse’s pattern: a Noul is P(true); P(false) = 1 − noul; compare to your THRESHOLD.
• Apidog: assert on choice / noul / score so a behavior change fails a test instead of silently misrouting.

Exercise

A vague instruction. Read intent.choice and confidence. On paper: <0.5 to a human; approve_transfer below 0.9 asks first. Do not let the model “also execute.”