manager-auditlisted

# /manager-audit ## Goal One interpreted verdict over a supervise run, answering four questions the supervisor cannot be trusted to answer about itself: 1. **Did it restate the ask?** — lane `restatement-quality`: the intake restatement exists, all six sections are filled, success criteria are measurable, no wave drifts outside the interpreted goal. 2. **Did it route models per policy?** — lane `routing-waste`: no silent TOP-inherit leaks, verify/judge never below the MID floor, fan-out at LOW (conventions from `docs/model-routing.md`, measured by `core/hooks/model-routing-observer.py`). 3. **Where did the tokens go?** — lane `token-spend`: relative dispatch cost (tokens × tier multiplier — relative ranges from docs/model-routing.md, never prices), top waste sources ranked. 4. **Did it follow its own loop?** — lane `role-compliance`: every wave audited, never-auto-retry honored, RECORD.md written, review lane dispatched after code waves. The machine layer *detects*; this skill *interprets and proposes*. The split mirrors harness-audit: deleting this skill weakens no gate. ## Steps ### 1. Run the machine layer ```bash bash core/infra/manager-audit.sh <slug> --json ``` Optional: `--session <id>` to scope routing records to one session. Render the findings as a per-lane table (lane / check / severity / evidence). The script always exits 0 — severity lives in the findings, not the exit code. ### 2. Interpret the semantic halves The script flag