Skip to main content
On this page

Zero retention

zdr.sh serves requests on two kinds of supply, and it states which is which rather than averaging
them into one claim:

  • Partner routes are admitted only under a written zero-retention agreement. A provider that will
    not commit in writing does not serve a request on these routes.
  • Rented GPU routes are machines the platform rents to serve the same models. The
    platform keeps no copy of the request or the answer there either, and the container and its disk
    are transitory. The machine's operator is a third party, though: it holds root on that machine and
    can transiently see memory while a request is being processed, so this tier is labelled
    transient: it is not zero retention, and the platform does not claim that it is.

What each tier carries#

Partner routes: zero retention, contract-backed. The platform keeps no copy of the request or the
answer, and the provider's written commitment not to keep them is the other half. That commitment is
an artefact, not a request option: the gateway cannot prove a signature, and a zdr: true field in a
request body is not a contract. It is what governs the provider's side, including any narrow
inspection carve-out the provider's public terms reserve.

Rented GPU routes: transient, disclosed. The platform's container logs no prompt and no answer,
keeps no store and is destroyed when the rental ends. The host operator is an anonymous third party
with root access to the machine; it can transiently read memory, and the platform cannot give a
zero-retention guarantee for it. Traffic is routed here only when the request does not demand zero
retention.

The price does not depend on the tier. The catalog publishes one price per model and a request is charged that price whichever tier serves it. Requiring zero retention changes which routes may answer, never what the request costs.

The answer says which tier served it. A settled answer carries x-token-retention, the class of the route it ran on: zdr for a partner route, transient for rented capacity. It is a header on an assembled answer and a trailer on a stream, because the class is known when the request settles.

Restricting a key to zero retention. A key can be restricted so that its requests are served only
by partner routes; such a key never reaches a transient rented route. The switch is zero-retention
routes only
, on the key's settings page: on for the key that must never leave the tier, off for one
where rented capacity is fine. A single request can ask the same with the :zdr suffix on the
model id (glm-5.3:zdr). The platform fails closed: when no partner route serves the model, the
request answers 503 zdr_unavailable rather than falling back to a transient route.

What is not stored, on either tier#

  • The request and the answer. No prompt, file, image or generated text is written to a store of
    ours. There is no trace system, no traffic sample kept for quality and no dataset built from
    requests, so there is no opt-out to look for.
  • The logs. A request leaves a line with its id, the account, the key, the model, the token counts,
    the latency, the cost, the status and the time. The body of the request and the body of the answer
    are not in it.
  • A partner route's serving side. The agreement we hold with the provider obliges it not to retain
    the request or the answer once it has been served.

Two things a request can carry are kept on its usage row because they are yours to analyse and not
content: the end-user id and the metadata entries the client sent. Neither is forwarded to a model.
Chat completions has the fields.

The three places text can rest#

Three features hold text on purpose, and each is a choice rather than a default:

WhereWhat it holdsFor how long
Memory vaultThe notes you write to it yourselfUntil you delete them
Response cacheAn answer to an exact repeat, when a key turns the cache onTen minutes
Idempotency replayThe answer to a request sent with an Idempotency-Key, so a retry is answered once24 hours from settlement

None of the three is a record of traffic: nothing enters them unless a key asks for it, and none is read to train, score or review a model. Optimizations describes the cache, and Chat completions the idempotency key.

The memory vault rests in the service database. Its note text is not encrypted at rest today, and neither is the host's disk: the protection there is access control on the host and the database, not encryption. Ask support if you need the current technical details for a compliance review.

The web tools are not model routes#

A search sends the query to the search service, and a browse is read by the platform's own reader. Neither keeps the text on this platform. The web search and web browse pages say what each one sends and keeps.

If that ever changes#

Which tier serves a request is chosen per key or per request, never globally. A change to what either tier promises would be a change to the product, and a model that could not meet the rule would leave the list rather than serve under a weaker promise. The privacy policy states what is processed, why and for how long. The terms of service section 2 states the contractual position.

This site uses only the technical cookies the sign-in and the security need. No analytics, no advertising, no tracking. Cookie policy