No training on requests
The Tsubasa application does not use customer prompts or outputs to train models, build advertising profiles, or sell behavioral data.
PRIVACY MODEL
Tsubasa processes request content long enough to produce a response. The application does not intentionally write prompt, completion, source-code, repository, patch, or tool-call bodies to usage history.
0
request-content fields in the Tsubasa usage-event schema
CONTENT HANDLING
OPERATIONAL METADATA
These records authenticate requests, reserve and settle credit, enforce limits, diagnose failures, and reconcile payments. They are not a place to copy free-form request content.
The Tsubasa application does not use customer prompts or outputs to train models, build advertising profiles, or sell behavioral data.
Request content exists in process memory and must reach the configured inference infrastructure. Provider retention is a separate deployment boundary, not covered by an absolute claim.
Structured application logs redact content-bearing and credential fields. Raw request and response logging must remain disabled across every deployed proxy and observability hop.
RETENTION STATUS
The repository verifies a metadata-only usage schema, hashed API keys, bounded billing metadata, and application-log redaction.
A complete production retention schedule, backup-deletion propagation, and self-service account deletion are not yet implemented. Public copy therefore does not promise universal immediate deletion of every operational record.
If an optional digital-asset invoice is used, transaction data may be visible on the relevant network by design. Tsubasa keeps the invoice and settlement evidence needed for accounting.
Report a privacy or security concern without including prompts, API keys, access tokens, or customer content.