Data Handling

Paper Instruments does not store customer content from the Ultramarine API by default and does not use API inputs or outputs for model training.

Customer content

Customer content includes the prompts, messages, system prompts, images, tool definitions, tool calls, tool results, stop sequences, and URLs you submit to the API, values of a request's metadata parameter, and the reasoning and outputs the API generates in response.

By default, Paper Instruments does not save customer content to databases, logs, traces, or analytics. However, some features you choose to use can involve storage or third-party processing, as described below.

Operational metadata

We retain operational metadata for billing, rate limiting, capacity planning, security, and debugging. Operational metadata describes a request and the account that made it, excluding customer content. It includes:

• Request identity: request, organization, and API key identifiers, endpoint, model, and timestamps.
• Usage and billing: token counts, charges, and balances.
• Request settings and shape: the model, temperature, top_p, top_k, reasoning effort, output limits, response format type, and counts of messages, tools, and other request elements, excluding their content.
• Performance and outcomes: latency, serving provider, HTTP status, and error codes.
• Account data: organization members, roles, API key metadata, and billing contacts.

Exceptions

• Web search. When you use web search, the third-party search service receives the data needed to perform the search and handles it under its own policies. To block web search for your organization, contact us to request zero data retention (ZDR).
• Legal requirements. Paper Instruments or its inference providers may retain customer content where the law requires it, such as under a court order or subpoena.
• Provider abuse prevention. Our inference providers may retain customer content to prevent abuse of their services.