What problem does it solve? Verifying that a multi-step composition of AI skills actually performs as predicted is hard: individual step scores do not guarantee the wired pipeline behaves as the product-of-means model forecasts. This Skill runs a typed composition end-to-end over eval cases and seeds, scores the final artifact deterministically, and reports the delta between measured and predicted performance. ## Core Features & Use Cases - End-to-end composition execution: Resolves each step's frozen winner variant, builds the prompt/schema map, and invokes the emitted patterns/<name>.js workflow over eval cases x seeds. - Prediction validation: Computes mu_pred as the product of step winner means, compares it to the measured mean, and flags composite variance violations and biased prediction models. - Calibration gating: Refuses to run when any step winner lacks a calibrated grade, since saturated steps make the prediction vacuous. - Use Case: After freezing winners for a classify-then-synthesize pipeline, ask to measure the composition and receive a report with measured mu/sigma, delta_predicted, and a running-bias check on the prediction model. ## Quick Start Run the composition named in my request end-to-end and report the measured versus predicted performance delta.
