Mara, the capacity planner at fictional Harbor Parcel, needs a reviewable model experiment. The sandbox table practice.capacity_samples contains one row per original fictional observation: sample_id identifies it, x is a planning input available in advance, and y is the observed capacity outcome. Train on sample_id <= 40 and reserve the later identifiers for validation. Deliver a prediction register and a one-row report of holdout size, absolute error and signed bias. The exercise data is synthetic and does not establish a production forecast.
A model review needs more than the statement that training finished. You will inspect the split, fit a numeric line, evaluate a disjoint holdout and preserve a row-level prediction register. That register makes an aggregate metric auditable: another analyst can trace each residual back to an observation.
Each step starts with your previous SQL. Keep the model definition in the script as the final report develops, so every result can be rebuilt from the same stated training split. The setup recreates the fictional samples for each step, including the hidden copy. Don't replace the query with numbers copied from the grid. A signed average can be close to zero while individual misses are large, so the final report includes absolute error as well as signed bias. No future scheduling or production deployment is implied.