jev current

Use this version with "engine": {"name": "jev", "version": "current"}. current runs whichever Jev version the provider serves. Every answer records the epoch of model behaviour that produced it; a new epoch begins only when we detect a change, and then we re-check quality against the conformance suite.

Limits


Context	32,000 tokens for the state plus the longest question
Whole request	64,000 tokens
Choice options	255, including the `none_of_the_above` escape option
Score levels	2 to 10
Questions per request	32
Tokenizer	approximate; counts can differ slightly from the provider’s

Judging on this version is billed per judgment, by size class, the same as on every engine.

Data handling


Hosted by	TypeSafe AI, in the US
Region	us
Subprocessor	TypeSafe AI operates the model. The full list of subprocessors is in the data processing agreement.
Customer data	The compiled context of each document it judges is sent to the engine provider’s API. After a change of model, the stored contexts of the answers your recent outcomes labelled are sent once more, to the same provider on the same terms, to re-fit your calibration for the new model; a document you have deleted is not. Every subprocessor it passes through is listed in the data processing agreement, and how long each keeps it is set by that provider’s terms there: see where your data goes.
Version retirement	Version `current` is not pinned: it is whatever model the provider serves now, and that can change without notice. Every answer records the epoch of model behaviour that produced it: `engine_version` is `current+` the date the epoch began and its number, such as `current+2026-09-24.1`, or `current+pending` only briefly, before the very first epoch has been recorded. A deployment or restart keeps the current epoch. A new epoch begins only when we detect a change. We monitor its behaviour continuously and start a new epoch, with the next number, when it changes, usually within minutes; every answer from then on records it. Answers already computed keep the epoch they were computed under. Calibration is re-fitted for the new epoch, usually within the hour, from a replay of the answers your recent outcomes labelled: see when the engine changes. `current` itself is never retired: a change of model starts a new epoch instead. If the provider stopped serving Jev altogether, judgments on it would fail with `engine_version_unavailable`, and your organization’s owners and admins would get an email; we never re-pin a judgment silently.

Conformance

Our conformance suite ran on 2026-09-27 against golden-v1. Because current can change, we run the suite on it again every day and after every change we detect, and GET /engines returns the latest run with its date.

Type	Accuracy	Expected calibration error	Log loss	Option-order flips
bool	82.6%	0.017	0.395
choice	80.6%	0.108	0.902	3.0%
score	33.6%	0.302	2.189

Score accuracy counts only the exact level. The answer was within one level 75.4% of the time, 0.94 levels off on average. What this means for each type: how far to trust each type. On choice items where no option fits, it answered none_of_the_above 70.7% of the time, and on the others 0.7%. The golden set is deliberately hard: 40% easy, 40% medium and 20% hard items, where the hard ones are those people disagree on or that were built to mislead. Accuracy by difficulty:

Type	Easy	Medium	Hard
bool	95.0%	85.0%	53.0%
choice	92.5%	81.5%	55.0%
score	44.0%	30.5%	19.0%

Latency	p50	p90
Batch 1	210 ms	288 ms
Batch 32	301 ms	412 ms

Get started

Before you build

Concepts

Guides

Relations

Engines

Reference

Limits

Data handling

Conformance

​Limits

​Data handling

​Conformance

Limits

Data handling

Conformance