Product Positioning
JevTypeSafe lets users try the Jev and Laya System One models in the browser. Its core purpose is to make structured judgments about text, rather than generate free text or explain reasoning. Users can ask questions and view the model's returned judgment results and their probability distributions, and can also reuse configurations in their own development workflows.
Main Capabilities
The platform supports three judgment question formats:
- Choice: Define several candidate answers, and the model returns one of the named answers, suitable for classification and routing scenarios.
- Yes/No: Answer binary questions, suitable for determining whether a certain condition is met.
- Score: Score the input on a defined scale, suitable for measuring degree rather than category.
One question is run at a time. The results show the selected answer and the corresponding probability. The probability reflects the model's judgment weight for the given input and question, and does not constitute a guarantee of correctness. Users can retain the raw response to inspect the specific returned content.
The platform provides preset scenarios, including real inputs and judgment criteria, for example:
- Routing support tickets to the Billing, Technical, or Account team;
- Determining whether a message describes an issue that needs to be handled today;
- Assessing customer frustration level (1 means frustrated but polite, 2 means very frustrated).
These examples show how the output changes when the same text is paired with different questions. It also provides several illustrative workflows, such as categorizing feedback as a bug report, feature request, or general comment, ranking articles by topical relevance, or routing requests to different handlers.
Model Description
- Jev: TypeSafe's flagship System One model, supporting rich documents, with question length up to 8,000 tokens.
- Laya English: 512 tokens, ultra-fast and lightweight, designed for high-throughput short text, comment moderation, and real-time LLM guardrails, with latency under 0.35 seconds.
- Laya Multilingual: 1,024 tokens, positioned similarly to Laya English.
Both Jev and Laya models make structured judgments and do not generate free text or reasoning explanations.
How to Use
- Open jevtypesafe.net in a browser, start from a preset scenario, or directly fill in the text, question, and answer definitions.
- Choose the Jev or Laya model to run, and on the pageView the results.
- Keep the judgment configuration, export the configuration, or copy the corresponding API example.
- Log in to your account and create a JevTypeSafe key.
- Follow the API, CLI, MCP, or skill instructions to integrate the key into your own code or Agent environment.
The CLI, MCP, and skill clients are open-source projects that share the same judgment configuration and account quota. Switching invocation methods does not create a separate quota. Reusing the configuration in code does not bypass the service's output guarantees.
Billing and Data
Registration comes with 100,000 input tokens, no credit card required. The Playground and API share this balance, which applies to Jev and Laya; output tokens are free. Once the balance is exhausted, you can purchase token packages or subscription plans on the Pricing page, and the plans likewise grant input tokens. The platform does not persist input or result history unless explicitly performed, and only records minimal usage and billing metadata; drafts are saved in the current tab session. Published examples pair fictional inputs with recorded responses, and users' own inputs and responses are not made public by default.
Use Cases
- Text classification and intent routing, such as ticket assignment and request triage.
- LLM guardrail testing and real-time moderation.
- Scoring or ranking inputs according to a defined scale.
- Applications that need named answers directly processable by code rather than parsing long text.
Limitations and Caveats
- The model returns judgments, not guarantees; probabilities do not prove correctness.
- Ambiguous cases require human review, and whether to trigger an action is determined by application code and permission checks.
- The model does not write customer replies, nor does it explain its reasoning process.
- The specific supported platforms, integration details, and prices are not yet provided.
