Effective 2026-08-29
Data policy
What we do with prompts and completions, in plain terms: we do not train on them, and we do not keep them.
Zero retention of inference content
Lore Hex does not retain prompts or completions. The body of an inference request, and the body of the response we return, exist in process memory for the duration of the call. They are not written to disk, not written to a database, not placed on a queue, and not shipped to a logging system. When the request completes, they are gone.
This applies to every model in our catalog and to every customer, on every plan. It is not a paid tier and not an opt-in setting.
No training on customer traffic
We do not use prompts or completions to train, fine-tune, distill, evaluate, align, or otherwise improve any model — ours or anyone else’s. We do not build datasets from customer traffic. We do not sell, rent, or share inference content with third parties for their own use.
What we do keep
We keep the minimum operational metadata needed to bill accurately, keep the service up, and stop abuse. For each request:
- Timestamp and duration
- Model id
- Input and output token counts
- HTTP status and error class
- API key identifier and account id
- Source IP, for rate limiting and abuse prevention
No prompt or completion content appears in this record — not in truncated form, not hashed, not sampled. Token counts are integers, not text.
Operational metadata is retained for 90 days and then deleted. Aggregated, non-identifying usage totals (tokens per model per day) are kept longer for capacity planning and financial reporting.
Errors and debugging
Error traces are scrubbed of request and response bodies before they leave the serving process. If we cannot diagnose a problem without content, we ask the customer to send us a reproduction case explicitly — we do not reach into traffic to get one.
Abuse handling
Because we do not retain content, an abuse investigation works from metadata: request rates, token volumes, error patterns, and account signals. If we suspend an account we can tell you the pattern that triggered it; we cannot show you the prompts, because we do not have them.
Legal process
We respond to valid legal process, and what we can produce is bounded by what exists. We cannot produce prompts or completions, because they are not stored. We will produce operational metadata where legally required, and we will notify the affected customer unless we are legally prohibited from doing so.
Questions
Data protection questions go to privacy@lorehex.co. See also the privacy policy and the terms of service.