Ten pages, one mental model. Read a slow training step all the way down to a single marginal optic — then build the pipeline, SLOs and alerting that make it repeatable. Start anywhere; the arc below is the recommended path.