Jev Cuts Without Text: Games Ready, Cautious Company

TypeSafe offers Jev, a model that does not generate text but makes type-based decisions accompanied by probabilities. Access is only via API, with no explanations provided and documented position effects necessitating a structured implementation. Video games are already leveraging it, and projects are linking it to computer vision.

Scalable API, no images, and no explanation: necessary safeguards

Jev returns probabilities but no justification, complicating its use in contexts that require explainability. The service is accessible only through API, with no public weights or self-hosting options, and each request falls outside the scope. Documentation indicates an alias jev-latest that evolves with versions and dynamically adjusted rate limits. Jev is text-only and does not process images. A position effect has been measured, with the same option varying between 0.84 and 0.96 depending on its location, and less effective calibration on more difficult and newly generated problems. Similar to image classifiers known for their overconfidence, the requirement for calibration on held-out data applies here as well.

Data processing: contractual conditions and signal use

TypeSafe claims not to train Jev on client queries and responses, which does not preclude the possibility of storage. No dedicated retention option is offered through a separate agreement, but a DPA is available. Outside of the DPA, one must assume possible storage for operational reasons and read this framework before any actual submission. In practice, sending derived and minimized fields is recommended, as is obtaining prior written commitments regarding the retention of personal data. Jev's probabilities should be treated as a signal and calibrated on internal data, with a generative model retained to document escalated cases. To reduce the observed position effect, mixing the order of options and making two average calls is advised when the stakes are high.

Video games: quick decisions, rules in code, and initial results

Games provide a natural framework for Jev: choices bounded by the engine, rules held by the code, and tight latency where around 200 ms is acceptable while 10 seconds is not. Since the game state is synthetic and not personal, API outsourcing is less sensitive. One example is identifying which NPC a player is addressing: Jev can produce scores such as innkeeper 0.81, guard 0.12, bard 0.03, with a threshold of 0.6 triggering an appropriate response, weighing contradictory clues (sword vs room request). A community benchmark on 79 difficult statements reports an F1 score of 0.96 and a precision of 1.0, compared to 0.82 for a heuristic, and 0.93 with misspelled names. Used as a chess engine, Jev achieves about 37% for the best move, an Elo rating close to 950, and a median of 166 ms per move, provided that positional facts are pre-calculated in the code. Some projects encapsulate the model by exposing only valid actions to it, with the model making a judgment while the code maintains the workflow. In Unreal, world revision labeling allows for the exclusion of late responses.

Technical integration: HTTP API, SDKs, and Unreal plugin in alpha

Jev is called via an HTTP API and official Python and JavaScript SDKs. On Unreal Engine 5.8, a public alpha C++ plugin adds a decision task triggered upon entering a state and stamps each request with a revision number to ignore responses that arrive after a context change. For Unity, Godot, or Three.js, the same scheme applies via HTTP. An operational rule is imposed: do not embed the API key in the build and interpose a small gateway server.

Features and performance announced by TypeSafe and third-party measurements

Jev does not generate text: it reads a state and answers type-based questions in a single parallel pass. Reported call times range from 70 to 500 ms. Limitations include text input only, up to 64k tokens per request. The announced rate is approximately $0.04 per million input tokens, with no output cost. Supported questions cover multiple-choice, a score, and yes/no, with a probability distribution for each output. TypeSafe attributes calibration to a post-training RLCD whose recipe is not public. Across four workflows, TypeSafe reports 67.8% accuracy compared to 67.9% for GPT-5.6 Terra, at about 1/75 the cost per case and approximately 25 times faster, while noting that larger reasoning models perform better. On the third-party measurement side, a calibration error of about 0.03 on 1,200 MMLU items is reported.

Vision + language: OCR pipeline and zero-shot labels before decision

Computer vision has been returning a probabilistic distribution on fixed labels for years; Jev transposes this response format to language. In the absence of image support, the practical architecture involves having a vision model produce a typed state, then submitting questions to Jev. This chaining associates two successive probabilities. Community projects are already applying this scheme: jev-eyes executes OCR, retains layout on a character grid, adds zero-shot labels from a SigLIP model as needed, and then forwards to Jev for the final decision.

Open self-host alternative: Laya from Convai under Apache 2.0

When sending data outside the scope is prohibited, Laya, published by Convai under Apache 2.0, offers an alternative based on the same idea and can be hosted internally. The model has 421M parameters, with a multilingual variant at 322M, runs on CPU in a few hundred milliseconds, but comes with explicit caveats: in its raw state, it approaches randomness on type-based decisions, its probabilities are overly confident before temperature tuning, with a reported calibration error dropping from 0.47 to 0.08 after adjustment, and accuracy collapsing beyond about 20 options. Laya thus serves as a foundation to refine on internal labels rather than an immediate replacement.