What problem does it solve? Agent-completed work and proposed contracts can be accepted on weak evidence, hidden assumptions, or unchallenged reasoning. This Skill provides an independent, read-only adversarial review that challenges decisions before they are accepted, without making the decision itself. ## Core Features & Use Cases - Decision-boundary review: Challenge a proposed contract before acceptance or a pending human gate after process evidence exists, using loopbench challenge packet and loopbench challenge record. - Commit-sized checkpoints: Review coherent, non-trivial work units before they become a commit, storing bounded memos under work-unit-reviews with source-backed task notes. - Reviewer independence ordering: Prefer a different model family in a fresh/read-only context, fall back to same-family fresh context with explicit labeling, and always ask the user before sending material to an uncleared external service. - Structured rubric and simplicity assessment: Grade declared rubric criteria with traceable evidence and map new long-lived mechanisms to contract goals. - Use Case: Before accepting a contract that changes project scope, run a contract-target challenge review to surface unsupported claims, omitted options, and the strongest counter-hypothesis, then record a verdict of clear, concerns, blocker, or unknown. ## Quick Start Ask the assistant to run a challenge review on the pending human gate for the current Loopbench task and record the verdict memo.
