q1729 is a local research repository, with no deployed network service. It reads supplied run files and optionally sends them to the NVIDIA NIM API. Those files can contain environment or other user-supplied information.

• NVIDIA_API_KEY is the runtime credential, read from the environment at call time. .env is gitignored but not loaded automatically; never log the key.
• CI may also use CODECOV_TOKEN; it is separate from the runtime credential.
• The narrator sends the full JSON object, including system/environment metadata, not just performance numbers or GPU names. Review the payload before sending.
• Narration is unchecked model output. It is not trusted executable input or validated scientific evidence. Shared semantic validation now rejects malformed run inputs; narration itself remains unchecked.

CI uses read-only repository permissions. The tag-triggered release workflow has write permission only for publishing after preflight and a fresh reusable quality workflow. Preflight checks tag/version/notes and fetched main ancestry. Dependency currency follows AGENTS.md; it does not guarantee dependency safety.

Architecture rendering uses Node.js/Mermaid CLI and its browser tooling locally. Brand generation uses its Python source. Neither generated SVG should be edited manually; these are different toolchains.

Aborted run records (schema 5) store the stopping exception's type and message, which can include local paths or device details. Review them before sharing.

Report vulnerabilities through a private GitHub security advisory rather than a public issue. Maintainer response is best-effort.