Do not publish a live credential, private dataset sample, device identifier or unredacted diagnostic archive in a GitHub issue. Revoke an exposed credential first. Use GitHub's private vulnerability reporting/security-advisory channel when available; otherwise open a minimal issue that contains no exploit detail or secret so a private contact path can be established.

Include the affected revision, operating system, whether physical adapters were enabled and a redacted reproduction. Never attach .env, .local-data, robot calibration files or raw terminal history.

• The optional dashboard is a separate local-first project and should remain bound to loopback unless it is placed behind an independently reviewed access boundary.
• LLM output is advisory until deterministic schemas, roles, leases, limits and approval gates accept it.
• Policy paths are not accepted from model output; the runtime resolves only revision-pinned local Hugging Face snapshots.
• Secrets belong in an OS secret store or process environment, never source, JSON profiles, notebooks or logs.
• HASHTAG_ENABLE_PHYSICAL=true is only one gate. It does not replace per-run opt-in, calibration checks, operator approval or a physical E-STOP.

If motion is unsafe, use the physical power cut/E-STOP first. Software stop, Ctrl-C and optional dashboard controls are secondary. Do not reach into the arm sweep. After an incident, keep logs and the exact artifact revisions, inspect mounting, gearing, calibration and cables with torque disabled, and do not resume from the failed state without a fresh preflight.