You already know: how to train and evaluate a classifier, choose thresholds and avoid leakage (D3 Supervised learning, ML evaluation); features and their pitfalls (D3 Feature engineering); release gates and online monitoring for LLM features (D10); model registries as database tables (D1 Databases); the training-serving skew from the D10 bridge.
Why this matters for Atlas. A model is a snapshot of the world at training time, and the world keeps moving. In production, the hard part of machine learning is rarely the algorithm; it's everything around it: defining the problem so the model optimizes what the business needs, reproducing how a model was made, deciding when to retrain, and noticing when the model changes the very data it learns from. Google's well-known paper on the subject put it bluntly: the ML code is a small box in the middle of a large system. This lesson maps the whole lifecycle, measures the failure modes that make "88% in March" useless by September, and sets up the rest of the domain.
1. The lifecycle, stage by stage
The question. What has to exist, apart from the model, for an ML feature to keep working?
Start from nothing. Every production model passes through the same stages, and every stage feeds back into earlier ones: