What problem does it solve? Determining whether one prompt variant of a skill actually outperforms another requires more than a few ad-hoc runs; this Skill runs a rigorous Layer-1 variant sweep that scores variants across cases and seed batches until a statistical rule decides a real winner, tie, or inconclusive outcome. ## Core Features & Use Cases - Adaptive seed rounds: Automatically adds seed batches until a precision-width rule decides, capped by n_max, so you stop spending budget once the answer is statistically clear. - Four-state verdicts: Produces winner, tie, inconclusive_saturated, or underpowered outcomes with bootstrap confidence intervals, so you know whether to ship, harden the eval, or raise the budget. - Provenance-gated scoring: Reconciles relayed scores against disk via a reconcile CLI before writing any winner.json or ledger rows, preventing corrupted results from being recorded. - Use Case: You have three prompt variants of a claim-verification skill and want to know which to ship. Run the sweep to get a leaderboard, a blinded LLM-judge comparison, and a durable winner.json only if the difference is statistically real. ## Quick Start Run an eval sweep on the claim-verify skill with default settings and show me the verdict.
