modelgenius-sdk is designed so that you can verify its behavior by reading the source:

• Local-only. The library makes no network calls. Capture writes JSONL files to a local spool directory; validation reads local files. The only network traffic in your process is your own, unmodified provider call. Uploading snapshots to ModelGenius is a separate, deliberate step performed outside this library.
• No telemetry. The library collects and transmits nothing about you, your usage, or your data.
• No credential access. Provider clients are duck-typed and passed through; the SDK never reads or stores API keys. Transport/auth kwargs (extra_headers, extra_query, timeout) are stripped from records before they are written.
• Fail-open by design. Capture errors are swallowed (logged once per error class) and never alter your application's calls, return values, or exceptions.

Snapshot files contain your prompts and (when captured) model outputs — which may include end-user content or PII. Recommendations:

• Point MODELGENIUS_CAPTURE_DIR at a directory with appropriate filesystem permissions, and apply the same retention and access policy you use for application logs containing user content.
• Use the redact hook on CaptureConfig to scrub or drop sensitive records before they are written, and request_filter / sample_rate to limit what is captured at all.
• MODELGENIUS_CAPTURE=0 disables capture entirely.

Please report suspected vulnerabilities privately via GitHub Security Advisories ("Report a vulnerability" on the repository's Security tab). Do not open a public issue for security reports.

Include what you can: affected version, reproduction steps, and impact assessment. You will receive an acknowledgment within a few business days, and we will coordinate a fix and disclosure timeline with you.