eval-firelisted

# /eval-fire — single-class judge dispatch Fires one eval class against one or more curriculum files. The class determines which compendiums load (all `memory/check_*.md` whose `eval_classes:` includes the class), which judge prompt template is used (`curriculum/evals/judges/<class>.md`), and which model the subagent runs on. ## When to invoke - **Iterating on a single file** during authoring — fastest signal, cheapest token cost. - **Triaging what still owes a judge** — `npm run evals:queue` (`curriculum/evals/scripts/eval-queue.js`) walks the universe and derives every (file, class) pair still owing, from each file's own Quality pins. - **High-stakes story sweep:** `/eval-fire story --personas 3` runs the audience triangle (mid-layer / opinionated senior / fast operator). Default is single canonical persona; the flag is opt-in for cohort-imminent or sponsor-critical files. - **NOT** for a full ship-time audit — that's `/curriculum-pre-ship-audit`, which dispatches all four classes in parallel and adds source-verify + capability-check + Quality-tag check on top. - **NOT for clearing the queue.** More than a handful of (file, class) pairs → `.claude/workflows/eval-sweep.js`, below. ## Clearing a queue: use the workflow, do not hand-write the dispatch `Workflow({scriptPath: '.claude/workflows/eval-sweep.js', args: {items, confirm, sets}})`, where `items` is `eval-queue.js --training <t> --json` with a per-class `pins` map attached. It fires one judge per pair, adversarial