What happens to your data, where your tokens come from, and which of it we can prove rather than assert.
The bar is the retention period. Prompts and outputs have none.
Billing metadata is request id, model, token counts, latency, status and cost — never prompt or output text. Deletion requests complete within 30 days across datastores and backups.
| Data | Default | Retention | Who can see it |
|---|---|---|---|
| Prompts and model outputs | Not stored | — | No one |
| Debug capture (opt-in, per key) | Off | 7 days, then deleted | Your organization; support only with a ticket and a time-boxed grant, logged |
| Billing and routing metadata request id, model, token counts, latency, status, cost | Stored | 13 months | Your organization; our billing and support |
| Account and invoice records | Stored | As required by Singapore law | Your organization; our finance |
| Security audit events | Stored | 12 months | Our security team |
Your inputs and outputs are not used to train or fine-tune any model, ours or anyone else's. There is no setting that changes this and no discount for allowing it.
Authorization headers, your API keys and upstream credentials are filtered at the entry point and discarded — not stored and then redacted.
Export your usage records at any time from the console or the API. Deletion requests are completed within 30 days across datastores and backups.
Every model in the catalogue carries a source label, and the label is part of the API response, not just the marketing page.
| Source | What it means | Status |
|---|---|---|
| self_hosted | Open-weight models on hardware we operate. Weight provenance, version and quantization published per model. | Available |
| vendor_api | The model vendor's own API under our commercial account. | Coming soon |
| cloud | Hyperscaler-hosted first-party models under our enterprise agreement. | Coming soon |
| licensed_third_party | A partner we have a written agreement with, labelled as third party even when the model is a vendor's. | Coming soon |
What we will not do. No subscription account pools. No reverse-engineered or impersonated client endpoints. No self-hosted open weights sold under a closed model's name. A route with no authorization on file does not get enabled.
Selected routes run inside a hardware TEE. The attestation quote carries a measurement of the exact code in memory; your client checks it, exchanges keys end to end, and fails closed on any mismatch. Usage comes back signed.
In preview with selected customers — ask us for the current scope.
The client checks the measurement against a reproducible build before any prompt leaves it.
Offered per route, on request. It proves this hop runs exactly that open code and keeps no logs — it cannot prove anything about an upstream vendor's own retention.
Data processing agreement with SCCs for EEA and UK transfers, sub-processor list with 30 days' notice of changes, and a security overview. Available on request; enterprise agreements may use yours.
Inference can be contracted to stay in North America, Southeast Asia or East Asia. Account and billing records are processed in Singapore.
Security incidents affecting your data are notified within 72 hours. Report a vulnerability to security@cflabs.ai.
Admin, billing and member roles with audited administrative actions. SSO and SCIM for enterprise accounts.
Dedicated compute customers are screened for entity and sanctions exposure before delivery. We do not provide service into sanctioned jurisdictions.
SOC 2 Type II is on our roadmap; we will publish the report date when the audit window opens rather than claim it early.
This page summarises our practices. The binding terms are the Terms of Service, Privacy Policy and Acceptable Use Policy.