jury

What problem does it solve? Plausible-but-wrong findings survive when nobody's job is to refute them. This Skill stress-tests one specific claim by spawning multiple independent skeptic subagents whose explicit mandate is to break it, then applies a refute-vote so you only act on claims that withstand adversarial attack. ## Core Features & Use Cases - Blind parallel skeptics: Dispatches N subagents in one batch, each isolated from the others and each instructed to refute the claim, defaulting to "refuted" when uncertain. - Adjustable kill threshold: Choose lenient (majority-refute kills), strict (zero refutes allowed), or supermajority bars, with jury sizes of 3, 5, or 7 matched to the stakes. - Distinct attack lenses: Optionally assign each juror a different angle — correctness, reproducibility, counterexample, security — so diverse failure modes get caught. - Use Case: Before gating an expensive training run on the premise "the data pipeline deduplicates correctly," run 5 jurors at a strict bar; if even one produces a counterexample, you fix the pipeline first instead of wasting the run. ## Quick Start Ask the AI to jury the claim "the login fix resolves the timeout bug" with 5 jurors at a strict bar before deploying it.