Open · Last updated 2026-09-21

Jev for data teams

Data teams should treat Jev as an external labeling/feature function with a model id column, not as a source of numeric truth.

Unofficial data-team notes — not a warehouse product and not TypeSafe’s AutoResearch cookbook (that remains outbound). Confirm facts on docs.typesafe.ai. Open-cluster replicas are not Jev and not endorsed. We do not issue keys.

Warehouse hygiene

Store model, question git SHA, full probability maps, and the action taken. Mark tables as non-deterministic hops. Do not join as if Nouls were identities.

What not to ask Jev

Aggregates, unique counts, date diffs, and “expected revenue from a Score.” Those are SQL.

Feature store pattern

Write Jev outputs as a narrow table: entity_id, as_of, model, question_sha, noul_or_choice, probabilities (JSON), action. Downstream models can consume them (AutoResearch-shaped). Do not overwrite source-of-truth facts with a Noul.

Calibration is your job. TypeSafe trains for calibrated decisions — that is a vendor claim. Plot reliability on your labels before you use a Noul as P(churn).

Columns the narrow table still needs

Make that feature-store row append-only and add the fields audits actually ask for:

Churn flags, revenue, unique counts, and date diffs stay in SQL. Jaggedness: jev-1.13 is the wrong engine for math and date order. Joins that treat Nouls as identities will invent cohorts. Pin jev-1.13.0 if any consumer treats a number as a floor; jev-latest can move under you.

Calibration is yours

TypeSafe trains for calibrated decisions — label that a vendor claim. Plot reliability on your labels before you use a Noul as P(churn). Methodology: evaluation guide and offline evaluation. Official: questions are independent; do not average Nouls across ids into one “model quality” feature.

Noul 0.5 = yes/no equally likely. If a feature store publishes that as “50% churn,” product will believe it. Pair with a human sample. Deterministic pipelines still own exact compute — vs deterministic pipelines.

Operate failure modes

• Silent overwrite. A nightly job replaces orders.revenue with a Score expectation. Undo is a warehouse incident, not a model ticket.
• Mixed processors. SemIf/OpenJev rows in the same table without a processor column. Those interfaces are not Jev and not endorsed.
• Token blow-ups. Unused CRM fields still bill (vendor list price $0.042 / M input, confirm models page) and hurt jev-1.13 accuracy.
• Streaming without caps. Vendor RPM/TPS figures on the models page are vendor and can change. See vs streaming.

When not to use this page’s advice

Do not ask Jev to compute aggregates “as a feature.” Do not treat this table as ECE numbers we invented. Do not skip humans on high-stakes labels. Schema-safe answers are still not factual correctness.

What this page does not claim

• No ECE figures invented.

FAQ

Can Jev replace our labeling vendor? Only for snap judgments you can define. Still sample humans.

Streaming features? Call from workers with RPM caps. See vs streaming.

May I upsert the Noul onto the entity fact table? No. Append a hop table. Facts stay facts.

Is AutoResearch required? No. It is an official cookbook shape. Your classical model is still 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.

Open-cluster pages are independent field-guide notes. Replicas and third-party interfaces mentioned anywhere on jev.pro are not Jev and not endorsed. Hub: Open. Siblings: evaluation guide, offline evaluation, vs deterministic pipelines. Canonical: https://docs.typesafe.ai.