Jev confidence vs probability
Probability is the mass on each option or level. Confidence (Choice and Score only) collapses the shape of that distribution into one number in [0, 1]. A peaked distribution → high confidence. A flat one → low confidence. TypeSafe computes this for you. Their confidence page’s three-option demo uses the approximation (3 × largest probability − 1) / 2. They also say you are not locked into their definition: the full probabilities object is always there.
This page is for builders who already heard of Jev / System One and want a production-minded mental model. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Confirm facts on docs.typesafe.ai. We do not sell keys.
Do not mix the two words. Probability answers “how much mass on billing?” Confidence answers “how peaked is the whole map?” A 0.51 / 0.49 split can share a winner with a 0.98 / 0.02 split and still be a different operational event.
How to use
TypeSafe’s starting pattern: high confidence → act; medium → confirm; low → human or safe default. Official examples use floors such as 0.5 or 0.6 and higher bars (0.85 or 0.9) for irreversible actions. Those numbers are illustrative. Start conservative and tune on your labels. Showing a balance at 0.6 is not the same stake as approving a transfer.
Read probabilities when the runner-up matters. Confidence is too coarse for dual-queue designs. You can compute your own statistic from the map; you cannot replace TypeSafe’s field in the request.
Low confidence is usually still HTTP 200. Hold in code. Persist the number on the decision trace with the pin and the action.
Noul is different
Noul has no confidence. Do not invent one. Threshold the noul, or define bands (auto / review / abstain) on the distance from 0.5. A Noul of 0.5 means yes and no are equally likely — uncertain, not “medium confidence” and not a medium Score.
If you copy a Choice floor onto a Noul, you are mixing peakedness with P(yes). That is a common production bug. See Noul and Choice.
When not to use confidence
- Do not treat high confidence as a warranty on the next ticket. Higher confidence is trained to mean higher accuracy across groups of predictions. Vendor training claim: RLCD. Measure on your data.
- Do not use one τ for read-only display and irreversible side effects. Scale the floor with risk.
- Do not look for confidence on Noul.
- Do not skip the map and keep only confidence if you later need expected Score or a runner-up route.
- Do not treat confidence as a dollar, a SLA, or a legal attestation.
Calibration reminder
Schema-safe ≠ correct. A peaked Choice can still pick the wrong winner. Official pattern notes live at docs.typesafe.ai/confidence and confidence-gated routing. First-party evaluation is still POST /v1/systemone.
Hub: Primitives. Glossary: Confidence, confidence threshold.
See also
FAQ
Showing a balance at 0.6 vs approving a transfer?
Different stakes. Scale the floor.
Can I send my own confidence algorithm?
You can compute one from probabilities. You cannot replace TypeSafe’s field in the request.
Low confidence HTTP status?
Usually 200. Hold in code.
Does Noul include confidence?
No. Threshold noul. Distance from 0.5 is your analogue.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.