Data and security
Data controls and zero data retention
What is stored, how to turn storage off, and how to keep requests on zero-retention and no-training providers.
このページはまだ翻訳されていません。英語版を表示しています。
Three separate facts decide whether a request is private: what we store (the prompt capture switch), which providers it may reach (the no-training and zero data retention policies) and what each provider does (its published posture). The serving path itself keeps no content: the request ledger records metadata only.
What is stored, and for how long
| Store | Holds | Kept | Capture switch applies |
|---|---|---|---|
| Request ledger | Metadata: ids, model, provider, tokens, cost, timing, status. Never content | Indefinitely (the billing record) | No, always recorded |
| Captured prompts and responses | The request and response as relayed | While capture is on; an unpaired prompt is deleted after 2 hours | Yes |
| Responses API continuations | Prior turns behind a response id | 30 days from the last turn | Yes |
| Batch files | Input, output and error JSONL | 48 hours if unused, else 24 hours after the job expires | No |
| Idempotency replay | The original response bytes | Briefly, in memory only | No |
Requests routed under zero data retention are never captured, even with capture on.
| At a glance | |
|---|---|
| Ledger | Metadata only, always |
| Capture default | On for new organizations |
| Turn capture off | Pro |
| No-training routing | Every plan |
| Zero data retention routing | Pro |
The prompt capture switch
One organization switch, capture_prompt_content, controls captured prompts, captured responses and continuations together. Admins change it in Settings > Data controls. Turning it off needs Pro (or an organization that was verified before capture became the default); turning it on has no gate.
With it off, nothing new is written, and previous_response_id cannot be resolved: send the full conversation every turn. Stored content on Free plans is part of our Data Licensing Program (de-identified); on paid plans only by opt-in. Content that is not stored is never used. See the privacy policy.
/api/orgs/{org_id}/telemetry-settingscurl -X PUT https://api.experientiallabs.ai/api/orgs/$ORG_ID/telemetry-settings \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"capture_prompt_content": false}'A Free organization gets 403 with needs_subscription.
Provider policy
The provider policy decides which upstream providers your requests may reach. It is applied per rung at routing time, including rungs on your own keys. A rung that fails the policy is removed before dispatch, so it is never tried, not even as a fallback.
| Field | Effect | Plan |
|---|---|---|
require_no_training | Only providers that do not train on API data | Every plan |
require_zdr | Only providers that keep no content | Pro |
allowed_processing_regions | Only endpoints with reviewed processing in these regions | Every plan |
allowed_providers | A provider allowlist | Enterprise |
zdr_continuation_storage | Allow previous_response_id under ZDR (we keep the state, never the provider) | Pro |
prompt_caching | false strips cache markers and cache placement | Every plan |
When no rung survives, the request is refused with 403 model_not_granted and a message naming the setting and a fix.
/api/orgs/{org_id}/provider-policycurl -X PUT https://api.experientiallabs.ai/api/orgs/$ORG_ID/provider-policy \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"require_zdr": true, "require_no_training": true}'PUT is partial: omitted fields keep their stored value.
The three ZDR states
Every rung resolves to one of three zero-retention states, from each provider's published terms for its API.
| State | Meaning | Under require_zdr |
|---|---|---|
| ZDR | The provider keeps nothing by default | Kept, tried first |
| ZDR on request | The route keeps nothing when the request carries a zero-retention constraint | Kept, and every dispatch carries the constraint |
| Not ZDR | The provider may retain content | Removed |
A model whose every rung is ZDR is a "ZDR model" (retention: zdr_all_rungs). A model with some ZDR rungs (zdr_enforceable) is served only through those under ZDR. Some model families have no zero-retention route anywhere because their maker requires retention.
{
"id": "glm-5.3",
"data_policy": {
"zdr": true,
"zdr_on_request": false,
"no_training": true,
"providers": [
{ "provider": "experiential_cloud", "zdr": true, "zdr_on_request": false, "no_training": true },
{ "provider": "fireworks", "zdr": true, "zdr_on_request": false, "no_training": true }
]
}
}The ZDR endpoint list is public: GET/api/v1/endpoints/zdr
Per request: provider.zdr
One request can demand zero retention without an organization setting: add a top-level provider: {"zdr": true} (the OpenRouter field shape) on Chat Completions, Responses or Messages. It only tightens, never loosens, an organization policy, and it needs Pro like the toggle.
{
"model": "glm-5.3",
"provider": { "zdr": true },
"messages": [{ "role": "user", "content": "Summarize this contract." }]
}x-gateway-zdr: trueOne switch for all three
The account-wide switch turns storage off and both routing requirements on in one action. In the app it is Settings > Data controls; for agents it is one call. GET reports enabled, partial (a fact drifted since it was turned on) and the models that keep a zero-retention route under your policy.
/api/orgs/{org_id}/zero-data-retentioncurl -X PUT https://api.experientiallabs.ai/api/orgs/$ORG_ID/zero-data-retention \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"enabled": true}'Verify after the fact
Every completion names the rung that served it, in headers and in the body's provider field. x-gateway-zdr is the gateway's reading of that rung's posture, not a signed attestation from the provider.
| What we do not offer today | |
|---|---|
| Signed per-response ZDR attestation | Headers are our reading of published terms |
| Per-attempt disclosure of failed rungs | Visible to operators only |
| Whole-platform data residency | Regions govern inference, not storage or support |
Step by step: Data controls guide.