quorum is a security-sensitive eval runner. It can launch real agent CLIs in permissive modes, collect transcripts, read local session logs, and send run artifacts to an LLM verifier.

Do not file public issues or pull requests containing exploitable details, tokens, session logs with secrets, or private filesystem paths.

Report security issues privately through GitHub private vulnerability reporting when available. If that is not available, contact a Prime Radiant maintainer privately and share only the minimum detail needed to reproduce the issue.

Never commit:

• .env files or API keys
• results/ output
• raw session.log, trajectory.json, filesystem.json, verdict.json, or meta.json from real runs
• local agent config directories such as .claude/, .codex/, or .gemini/

If a real run prints or captures a secret, rotate the secret before sharing any artifact derived from that run.

Public CI must run only static/unit checks. Live quorum run ... evals require credentials and permissive agent CLIs, so they belong in trusted local or maintainer-controlled environments only.

Kimi live evals launch kimi --yolo inside a Quorum-prepared workdir. Quorum uses a fresh KIMI_CODE_HOME, installs only the local Superpowers plugin via plugins/installed.json, and supplies model auth through a temporary runtime env file that is deleted by the launcher and runner cleanup paths.

Treat Kimi results/ artifacts as sensitive. Raw wire.jsonl logs may contain model outputs, tool arguments, and provider env until Kimi tool-subprocess env scrubbing is verified. Do not run live Kimi evals in public CI or against untrusted PR scenarios.