연결

Critique Gate

Ink01101011byInk01101011
0•
생산성 및 워크플로

Enforces strict quality standards on AI-generated content through automated, multi-mode critic evaluation loops before shipping.

Critique Gate separates generation from evaluation to ensure drafts and code are objectively judged before being finalized. Operating in tandem with loop-ledger MCP tools, it establishes enforceable feedback cycles across four specialized modes: single-critic score retries, multi-critic reviews gated by the weakest score, adversarial red-teaming that demands every objection be resolved, and multi-judge ensemble scoring that surfaces rating divergence. It is essential when outputs must pass strict rubrics, withstand active counterexamples, or receive independent multi-perspective validation without manual intervention.

주요 기능

01Uses weakest-link gating in multi-critic mode to prevent high scores from hiding critical flaws
02Integrates seamlessly with loop-ledger MCP tools for automated stall detection and session gating
03Separates generator and critic roles to eliminate self-review bias in model outputs
04Provides four evaluation modes: score-retry, multi-critic, adversarial, and judge-ensemble
05Surfaces judge variance in ensemble mode to highlight disagreement rather than silently averaging
060 GitHub stars

사용 사례

01Red-teaming critical security configurations or sensitive documentation before merging
02Evaluating code drafts independently across correctness, style, and performance criteria
03Iteratively refining complex technical explanations until all rubrics meet quality thresholds