critic

# Critic skill (tool-grounded) **TL;DR**: a subagent reads the draft, runs 2 grounding checks (top-3 claim sourcing + one blind-spot probe) via read-only tools, and returns a JSON verdict (SHIP / SHIP_WITH_FIXES / BLOCK). Target 3-5 tool calls, ceiling 8. No tool call = auto-BLOCK. Verdict ≤200 tokens to the caller. Used by content drafts, regulatory-style takes, partner outreach, and proposals. Layer 2 of the layered eval architecture. Verifies high-stakes drafts by fetching evidence, not by re-reading the draft. Self-critique without tools is noise. Tool-grounded critique is the only lift that survives evals (per arXiv 2510.14942 GroundedPRM, ICLR 2024 CRITIC). A prior inline self-critique missed the ~20% lift by skipping the fetch step. This skill enforces it. ## When to fire - **Auto-fire on high-stakes turns**: a PostToolUse hook reads the `stake_classified` event from `logs/session.jsonl`. If `stake: high`, it dispatches this skill as a Task subagent. - **Manual invocation**: any skill that produces external-bound output (content, regulatory-take, partner-outreach, proposal, content-publish) calls this Critic before locking the deliverable. - **Operator-driven**: the operator invokes `/critic` on a draft to get a verdict before posting. ## Why a subagent (not inline) Per the architecture's Layer-2 substrate-virtue: the Critic's reasoning stays in the subagent context, only the verdict returns to the main context. The lowest-pollution zoom-out primitive. The Critic...