Privacy policy — draft
This policy is a draft. The operating legal entity (Prefill Systems, LLC) has not been formed yet, public contact channels are still being set up, and the operator has not finalized this document.
It accurately describes how the service works today and the operator's commitments, but it is not an onboarding-ready legal document of a formed entity. Treat it as a disclosure, not as a contract. Our terms explain the legal status in more detail.
1. Who this applies to
Prefill Systems operates an OpenAI-compatible inference gateway that serves the Qwen3.8-27B model (qwen/qwen3.8-27b) to invitation-only evaluation customers ("you"). Distribution through OpenRouter is planned but not yet live; when it is, this policy will be updated to distinguish requests that arrive through OpenRouter (where OpenRouter's own privacy terms also apply to its users).
2. The short version
- Zero data retention: our policy prohibits persistent storage of prompts and completions. Transient in-memory prompt caching supports inference; content is not used for training. Policy scope is described in §3.
- We do retain metadata needed to operate the service: account, session/security, API-key, and usage/accounting records (§4–§7).
- No advertising, no trackers, no third-party analytics on this site, and no sale of your information.
3. Prompts and completions: transient only
Requests travel over TLS to the public gateway and are processed by an inference worker. Inference requires transient prompt, output, and KV state in memory. The gateway does not write request or response content to its database or logs, and streams responses without buffering them to storage.
- Zero data retention: request and response content is not persisted. Necessary operational metadata is retained separately.
- Prompt caching: transient in-memory tokens and KV state may be reused to reduce repeated processing, without persistent content storage.
- Review pending: hosting-provider handling, operating-system crash dumps, and other persistence layers still require review before an end-to-end zero-data-retention claim can be verified. No such verified claim is made today. Rented workers must undergo this review before use.
- No training. Customer content is never used to train, tune, or evaluate models.
4. Account information
Customers join by invitation. We hold: your name and email, a hashed password (never the password itself), your role and account status, and creation/ modification timestamps. Invitation and recovery links are single-use and expire. Tokens are carried in the URL fragment, which browsers do not send in HTTP requests; the acceptance page removes that fragment from the current history entry.
5. Sessions and security metadata
Sign-in uses an opaque random session token stored in an HTTP-only cookie (Secure and SameSite-restricted in production) — it is never exposed to browser scripts. CSRF tokens protect authenticated actions. We keep administrative access-management audit events, credential timestamps and revocation state, and short-lived sign-in throttle counters. Disabling an account invalidates its sessions and stops its API keys from authenticating.
6. API keys
You create keys from your dashboard. The raw key is shown exactly once at creation; only a cryptographic hash and non-secret metadata (name, prefix, timestamps, revocation state) are stored. Each key is owned by one account and usage is attributed to that account and key. Revocation is immediate and permanent.
7. Usage and accounting metadata
Handled inference requests produce accounting records containing: request ID, start and end times, model, backend used and whether it was a fallback, HTTP status and error type, prompt/cached/completion token counts, timing measurements, finish reason, retry count, and the estimated charge at configured rates (plus known backend cost and gross profit where the backend cost is known — internal economics are never shown to customers). This is how your dashboard reports usage and how we operate the service. No content is part of these records.
8. Cookies and local storage
We set one session cookie when you sign in (see §5). We do not use localStorage or sessionStorage for secrets, and the site sets no advertising or analytics cookies. Raw API keys and invitation tokens are kept in page memory only, never in browser storage.
9. How information is used
Account and session data operates the service; usage and accounting metadata runs the service, detects abuse, and (in the future, on the OpenRouter channel) supports billing. We do not sell personal information. Internal logs retain metadata, not content.
10. Sharing and processors
- Cloudflare — DNS and TLS-terminating ingress (Cloudflare Tunnel) for the public hostname.
- Operator-controlled servers — the gateway, PostgreSQL database, and Redis coordination store run on operator-managed hosts.
- Inference processing — operator-controlled workers and Alibaba Cloud Model Studio (QwenCloud). QwenCloud's API documentation describes no persistent input/output storage, transient prompt caching, and operational request metadata retained for 14 days. See data handling and metadata retention. Rented GPU providers require data-handling review before use.
- OpenRouter — not yet connected; planned as a distribution channel. OpenRouter is a separate company with its own terms and privacy policy.
We do not disclose personal information to law enforcement except where legally required, and we do not sell it.
11. Retention
Prompt and completion persistence is prohibited by our operating policy (§3). Sessions and invitation/recovery credentials expire and are invalidated on use or revocation as applicable. Account, security, and accounting metadata remains in the operational database; an automated metadata deletion schedule is not currently implemented. Requests for access or deletion are reviewed manually, including any records that must be retained for legal or accounting reasons.
12. Security
Passwords use slow, salted Argon2 hashes. High-entropy API keys and session tokens are stored as cryptographic hashes. Authenticated mutations require CSRF tokens and same-origin checks; public ingress uses TLS via Cloudflare. Database and coordination services are on the private network. This is an invitation-only pilot; we make no SOC 2, ISO, or similar certification claims.
13. Your choices
- See your usage and keys any time in the dashboard.
- Revoke any key yourself, immediately.
- Change your password (this signs out all sessions).
- Ask for account deletion or a copy of your account data via the contact route.
14. Changes to this policy
This draft will be replaced after the operating identity, contact channels, and data-handling review are complete. Material changes will be published here and communicated to affected account holders through their established contact channel.