What problem does it solve? ML pipelines that version only code produce irreproducible models: the same git sha trained on different data snapshots yields different models, monolithic training scripts make every experiment pay full compute cost, and ungated promotion lets bad runs reach production. ## Core Features & Use Cases - Run Fingerprinting: Records git sha, dataset URI, dataset manifest hash, params hash, and environment hash so any past model can be rebuilt from its fingerprint alone. - Stage Decomposition and Caching: Splits pipelines into ingest, validate, build_features, train, evaluate, and register stages with content-hash cache keys so unchanged stages are skipped. - Blocking Data Validation: Raises on row-count, category-cardinality, and label-base-rate failures instead of logging warnings, stopping bad data before compute is spent. - Champion Comparison and Gated Promotion: Evaluates candidates against the live champion on the same deterministic holdout, including per-segment regression checks, and separates registration from promotion. - Use Case: A team retraining a churn model uses this guidance to snapshot source tables immutably, split holdouts by deterministic hash buckets, and block automatic promotion when a segment-level AUC regression is detected. ## Quick Start Ask the AI to review your ML training pipeline design against reproducibility, validation, and promotion-gating best practices.
