Jev data handling and customization
Two sentences from the models page matter more than most vendor slogans: Jev is not trained on customer requests or responses, and Jev is not fine-tuned or LoRA-adapted with customer data. The same weights serve every account. Jev is TypeSafe’s first public System One model — unstructured state in, typed decisions out — not a per-tenant trained classifier.
This page is for builders who already heard of Jev / System One and need a data-handling mental model before security review. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Confirm docs.typesafe.ai and the legal index. We do not sell keys. This page is a map, not a DPA.
Definition: shared weights, request-time specialization
You specialize behavior through the request: put proprietary material in state, encode domain rules in instructions and criteria, and compose answers in code. The AutoResearch cookbook is the documented path if you want a downstream classical model on top of Jev’s probabilities.
There is no “upload a LoRA for my tenant” in the public docs. If you need a private encoder, that is a different product class. Jev remains the same jev-1.13.0 (or current pin) for every account.
Schema-safe ≠ correct, and “not trained on requests” is not the same claim as zero data retention.
Mental model: three different promises
| Claim | What TypeSafe points at | What not to collapse into it |
|---|---|---|
| Not trained on customer requests/responses | Models page | ZDR, deletion SLAs |
| Not LoRA-adapted per account | Models page | “We fine-tune Jev for you” |
| Zero data retention (ZDR) | Legal index / enterprise contract | A request flag this site invented |
TypeSafe points to its Legal index for the Data Processing Agreement, Privacy Policy, and zero data retention (ZDR) for enterprise customers. We will not invent which plan includes ZDR or how it is enabled on the first-party API.
Vercel’s AI Gateway changelog describes Zero Data Retention and No Training as per-request Gateway options when you call typesafe-ai/jev through the Gateway. That is a Vercel product surface, not a substitute for TypeSafe’s own legal terms.
Read docs.typesafe.ai/legal and your contract.
What you should still filter
Send only the state slice the questions need. Text only. English primary. Unused fields still tokenize (vendor price $0.042/M input, output free — confirm there) and 1.13 accuracy falls as unrelated detail grows. Hard policy and allowlists stay in code (policy-aware decisions).
jev.pro does not see your API traffic. We do not store your traces. We do not issue keys.
Customization that is in-bounds
- Named fields in state plus backtick paths in instructions.
- Criteria that close the answer set (Choice / Score).
- Compose weights in code; change a coefficient without a new model.
- Pin
jev-1.13.0so thresholds stay attached to known weights.
Out of bounds: asking Jev to generate a redacted document, to crawl a file URL, or to remember last week’s tickets without you putting that slice in this request. No hidden session memory.
See also
FAQ
Can I upload a LoRA for my tenant?
Public docs say no per-account LoRA. Customize via state and questions.
Does “not trained on requests” mean ZDR?
Those are different claims. Read the legal pages.
May jev.pro store my traces?
We do not see your API traffic.
Is OpenJev the same data-handling story?
No. Open replicas are not Jev and are not endorsed. Their training data is their authors’ problem.
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.