Jev 1.13 Release Notes: Behavior Changes & Migration

TL;DR: Jev 1.13 is the September 2026 release of TypeSafe AI’s System One judgment model, demoed through OpenRouter at $0.0462 per 1M input tokens. This page is an honest snapshot: it tracks publicly available OpenRouter information, marks what is unverified, and gives a migration checklist (model slug, defensive confidence parsing) rather than invented diff tables.

Version snapshot

What is confirmed vs what is not

Jev is in a fast iteration phase, and this page deliberately does not invent a diff table. What we can state:

• The three question types and the typed output shapes used across this site’s fixtures (answer, confidence, rationale, and scale for scoring) are consistent with what OpenRouter currently exposes for typesafe/jev-1.13.
• Anything finer-grained — system-prompt sensitivity, exact confidence calibration, per-version accuracy shifts — is not published in a form we can quote. Treat claims about specific behavior changes as unverified until you test them yourself (see “How to self-test” below).

Check the official announcement post and typesafe.ai docs for authoritative behavior notes.

Migration notes

Check the model slug

If you pinned an earlier slug (or an alias), verify typesafe/jev-1.13 is still what the OpenRouter model page lists. Pin the exact string; do not rely on a bare typesafe/jev alias resolving to 1.13.

Parse the confidence field defensively

Version upgrades are exactly when field behavior can shift. In your client:

• Treat a missing, non-numeric, or out-of-range confidence as a fallback condition (human review), never as an approval or publish signal.
• Do not hard-fail on an unexpected rationale length or formatting — log it and continue.
• Keep your fallback threshold in configuration, not in code, so recalibration after a version bump is a config change.

Watch your confidence distribution

After any version change, compare the confidence histogram of your traffic before and after. A shift in the distribution is the earliest practical signal that scoring behavior moved — even when answers look unchanged.

How to self-test behavior changes

Since we will not hand you a fabricated benchmark table, build your own — it is small:

1. Freeze a fixture set: 20-30 inputs covering all three question types, with expected answers and acceptable confidence ranges recorded per input.
2. Run the set against the version you are migrating from, and save the raw outputs.
3. Run it against Jev 1.13 (typesafe/jev-1.13).
4. Diff the answers, confidence values, and rationales. Investigate any input whose answer flipped or whose confidence moved outside your recorded range.
5. Only then promote the new slug in staging, and last in production.

Related guides

• Cost and latency — what 8-call batches and two-call pipelines cost at this price point.
• Troubleshooting — unstable scores and confidence anomalies, start here.
• Confidence and fallback — the threshold patterns this page references.
• First API call — if you are setting up the OpenRouter call for the first time.

FAQ

Is 1.13 the latest version of Jev?

This page was updated 2026-09-21 and tracks 1.13. Jev iterates quickly — check the OpenRouter model page for the slug currently served before assuming 1.13 is the newest.

Did pricing change in 1.13?

The price listed on OpenRouter as of September 2026 is $0.0462 per 1M input tokens. Pricing can change independently of model versions; verify on the OpenRouter model page before committing to a budget.

Where is the official Jev changelog?

The authoritative source is TypeSafe AI — start with the announcement post and typesafe.ai docs. This page only tracks what OpenRouter publishes publicly.