Jev’s score is 0..n-1, not 0–1 — the gotcha that breaks thresholds

If you treat TypeSafe’s Jev score like a confidence float between 0 and 1, your gates will be wrong.

What Score actually returns

Score is the probability-weighted mean of rubric indices. If your criteria are ["Low","Normal","High","Urgent"], indices are 0..3. A live response we copied looked like:

• score: 2.46 (not 0.62)
• legend: {"0":"Low","1":"Normal","2":"High","3":"Urgent"}
• probabilities keyed by index strings ("0", "1", …), never by your labels

So 2×0.5 + 3×0.48 ≈ 2.46 means “between High and Urgent.” Round for the band; the fraction tells you which way the mass leans.

Two siblings that also bite

• Noul has no confidence field — not null, absent. Distance from 0.5 is what you have for a gate.
• Ask several questions in one call: they share one state and run in parallel. Paying for the state three times is the expensive habit.

Try it without the official waitlist

I run an independent relay (not a TypeSafe first-party key): Google sign-in → jv_live_… → POST https://jev-agent.com/api/v1/classify with a ruleset name, or /api/v1/systemone with the official shapes so you can swap host+token later.

Free tier: 5 credits / month (1 credit ≈ 1000 input tokens). No uptime SLA.

Start: https://jev-agent.com/api-access · docs: https://jev-agent.com/docs

Not affiliated with TypeSafe AI.