Concepts

The explanation pages describe the pipeline as a set of boundaries and records: which package owns each computation, how artifacts move between stages, and why validation and staging stand between training and publication.

Read these pages when reviewing a proposed workflow, diagnosing a handoff, or deciding which identifier and artifact should be authoritative.

• Pipeline architecture separates computation, scheduling, evidence, and promotion responsibilities.
• Data lineage and run identity traces a network back through MLflow experiments, runs, artifact identifiers, configs, and the lockfile.
• Promotion safety explains staging, gate semantics, destination refusal, and the decisions that remain human-governed.
• Parameter recovery separates a wrong network from a model the design cannot identify, and says which statistic tests which.