Comparisons · Last updated 2026-09-21

Jev vs probabilistic pipelines

A probabilistic pipeline combines many uncertain nodes (Bayes nets, factor graphs, stacked calibrators). Jev emits probabilities for questions you asked. It is one node, not the net.

Independent. TypeSafe trains for calibrated decisions — vendor training claim. Measure calibration on your labels. docs.typesafe.ai.

Comparison scope

If you already have a net with known CPTs, keep it. Add Jev only where a node’s evidence is language.

Criteria that decide the architecture

Decision quality and control

Do not multiply Jev Nouls as if they were independent truths. Official: a Choice is relative; Nouls are absolute and can all be low.

Integration trade-offs

Treat Jev outputs as observations with a measured error model. Enforce identities in code if you need them.

TypeSafe’s public models page lists jev-1.13 at $0.042 per million input tokens with output tokens free — a vendor claim, not a jev.pro measurement. Confirm on the models page before you bid.

When each approach fits

Prefer Jev when

• A node is a snap language judgment

Prefer Probabilistic pipelines when

• You need a coherent joint distribution you specify

Your graph; Jev as a leaf; code enforces constraints.

What this page does not claim

• No ECE numbers invented.

• Schema-safe output is not the same as factual correctness.

• No independent bake-off numbers live here.

When not to use Jev

Do not replace the Bayes net with a pile of Nouls. Jev is not a factor graph. Official jaggedness: a question and its negation need not sum to 1; Noul and yes/no Choice need not match. If you need those identities, enforce them in code.

Do not treat TypeSafe’s “calibrated” training claim as your ECE. Measure on your labels (calibration glossary).

Do not backprop through Jev. Public docs we used are not a training API. Use outputs as features or observations.

Skip Jev when every node already has a CPT from structured sensors — language is the only reason to add a leaf.

Pairing-specific failure modes

• Naive independence. Multiplying three Nouls as P(A∩B∩C) without an error model. They can all be low together.
• Date/count nodes. “P(overdue)” from a due date is arithmetic, not a Noul.
• Confidence vs option probability. TypeSafe documents confidence as distinct. Mixing them in a likelihood is a silent bug. Read confidence.
• Alias as a new prior. Pin jev-1.13.0 after you fit the leaf’s error model; jev-latest invalidates the CPT you attached to the observation.

Production sketch

1. Keep the net (or stacked calibrator) you own. Identify leaves whose evidence is language.
2. POST those questions only. Pin jev-1.13.0. Log response.model.
3. Map Jev outputs to observations with a measured confusion / ECE on your set — not a vendor slogan.
4. Enforce any required identities (negation, mutually exclusive parents) in code after the call.
5. Downstream action uses the joint you specified. See uncertainty handling.

Hub: Comparisons. Siblings: vs deterministic pipelines, glossary calibration.

FAQ

Is confidence a probability? TypeSafe documents it as distinct from the option probabilities. Read confidence.

Can I backprop through Jev? Not a training API in public docs. Use outputs as features.

Is a calibrated Noul a Bayes net? No. It is one typed leaf. The joint stays yours.

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.