Jev data retention policy, read from the documents

Does TypeSafe train Jev on your API inputs?

No, and it says so in three places. The privacy policy promises not to train or fine-tune any model on your prompts or other input, and not to disclose input to anyone but its service providers. The models documentation says Jev is not trained on customer requests or responses. And there is no per-customer copy to train: Jev is not fine-tuned or LoRA-adapted with customer data, and one set of weights serves every account.

That last point cuts both ways. Nothing you send changes the model, and nothing you send can be used to teach it your labels either. Customisation lives in the request, which is why the state and criteria carry so much on the API reference.

How long does TypeSafe keep Jev requests?

No document gives a number. The Data Processing Addendum, last updated 2026-04-24, says customer personal data is kept for as long as necessary given the purpose of the processing and the law; the privacy policy says as long as reasonably necessary to provide the service or for business purposes. Neither separates request text from account data.

The trust center is more concrete about where the text goes. Of 13 listed subprocessors, Amazon Web Services stores and processes customer information for live requests in databases and caches, while Modal, Nebius and CoreWeave process prompts on their compute nodes but do not store them. Nebius and CoreWeave were added on 2026-10-03. If a review needs a figure in days, ask privacy@typesafe.ai and put the answer in your contract.

How do you get zero data retention for Jev?

Two ways, with different paperwork:

TypeSafe offers ZDR to enterprise customers through sales@typesafe.ai. OpenRouter is the route without a contract: its provider record for TypeSafe says prompts are neither retained nor used for training, and on 2026-10-09 Jev 1.13 was one of 940 endpoints on its ZDR list. To make that a guarantee rather than a default, ask for it per request:

{"model": "typesafe/jev-1.13",
 "provider": {"zdr": true},
 "state": "...", "questions": {...}}

We checked that the flag is enforced on OpenRouter's decisions endpoint and not just accepted. With it, Jev answered as usual, while GPT-6 Luna Decisions was refused with a 404 that no endpoint matched the zero-retention policy; without it, GPT-6 Luna answered. The ids and endpoint are on Jev on OpenRouter.

Does TypeSafe sign a DPA for GDPR?

Yes: the DPA states that it forms part of the agreement between TypeSafe and the customer. What it commits to:

• You are the controller and TypeSafe the processor, with US state-law service-provider terms alongside.
• EU transfers under the Standard Contractual Clauses, modules 2 and 3, governed by Irish law; a UK Addendum and Swiss terms; the Irish supervisory authority for the EEA.
• Notice of a security incident without undue delay and within 72 hours.
• One audit a year at your cost, and 15 days to object to a new subprocessor after notice.

The trust center adds a SOC 2 Type II report for 2026 on request. Whether that is enough for your processing is your data protection officer's call.

Can you keep Jev data in the EU?

Not from anything TypeSafe publishes. The privacy policy says the service is hosted in the United States, and every subprocessor with a listed location is in the USA, so EU data leaves the EU under the clauses above. If residency is a hard requirement, the option is an open-weight decision model you run in your own region, such as those on open-source Jev alternatives.

What should you strip from the state before sending it?

Anything the decision does not need, because data you never send needs no retention policy. Contact details are the easy case. We sent a support ticket with and without its sender's email address and phone number: both versions routed to billing at 1.00 and read a churn threat at 0.98, and the redacted one was 23 tokens cheaper.

import re

def redact(text: str) -> str:
    text = re.sub(r"[\w.+-]+@[\w-]+\.[\w.]+", "[email]", text)
    return re.sub(r"\+?\d[\d\s().-]{7,}\d", "[phone]", text)

Names, account numbers and addresses take more care, because some decisions do depend on them. Redact, then check on a few hundred of your own items that the answers do not move.

What does this site keep when you call Jev through it?

Not the text. Calls through this site go to TypeSafe first and to OpenRouter only if TypeSafe fails, and the text is forwarded and dropped, never written to a database or log here. What is stored is a call log without the text: time, endpoint, question names, credits and tokens. The full list is in the privacy policy.

The Jev data retention policy: common questions