eval-newlisted

# Eval New Arguments: `$ARGUMENTS` If the case is missing, stop and show: ```text /eval-new <case> <model-label> [--suite <path>] ``` `--suite <path>` takes the case and its evaluator reference from an external suite directory laid out like this repository (`cases/<id>/case.yaml`, `evals/reference/<id>.yaml`) — a private holdout that must not be committed here. Only the suite's directory name is recorded in the run. ## Procedure 1. Run the deterministic primitive with the supplied arguments: ```bash npm run eval:new -- $ARGUMENTS ``` 2. Capture the generated run ID from the command output or the newly created directory under `evals/runs/`. 3. Do not inspect `evals/reference/` (nor a suite's), other runs, other Case Library entries, or git history for expected answers. 4. Do **not** start the benchmark in this session. A fresh session is part of the contamination boundary. ## Completion Tell the user to open a fresh Claude Code session and run exactly: ```text /eval-run <generated-run-id> ``` Do not reproduce the run procedure. `eval-run` owns it.