tribunallisted

# tribunal — external coding agents as your QA panel You wrote the code, so you share every assumption that produced its bugs. An auditor from a different model family (xAI's Grok, OpenAI's Codex, a local Qwen) breaks that correlation — that is the entire value, and it survives only if the auditor is genuinely independent and the audit is genuinely proven. Roles are fixed: **external agents audit, read-only; you (the session) verify their findings, apply fixes, and own the report.** An auditor's word is never quoted as a conclusion, and an auditor never touches the tree. Your panel roster — who's installed, how they're invoked, what they cost — lives in `roster.md` alongside this skill. Read it; verify a member is reachable before you claim it ran or that it's down. If `roster.md` is missing, build it from `roster.example.md` before running any auditor. ## The four receipts — PER AUDITOR, every one, no exceptions These apply to **each panel member separately**. Receipting your headline auditor and narrating the second one from memory is a real observed failure: a test agent proved Grok rigorously and then wrote that its local-endpoint member "Returned 7 findings" — with no call, no output, no file. Half a receipted panel is a half-faked panel. 1. **It ran.** The exact invocation and its raw or structured output appear in the report, or in a saved file whose path is given (one file per auditor is fine). An audit narrated from memory scores as **not-run — even when i