Build the baseline before the extension.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Start with the question

A colleague supplies a better score but no split or seed. Can you tell whether the algorithm changed or the evaluation changed?

Mechanism and worked comparison

Reproduction begins with an executable contract. Name the data generator, split rule, model, initialization, update rule, number of updates and scoring metric. This study uses three authored tasks. Row indices modulo three assign training, development and final membership; no row is assigned twice within a task. Task formulas and partitions are public, so disjoint membership is not a claim of secret or independently collected data.

Every fit starts from three small deterministic coefficients derived from seed 7, 19 or 41. The linear baseline keeps its quadratic coefficient at zero. A full-batch gradient uses all twelve training rows at one checkpoint, then updates active coefficients simultaneously at learning rate 0.15. Forty or eighty updates cost 480 or 960 training-example visits per model, excluding scoring and overhead. Prediction never updates weights.

The downloadable Python reference reconstructs the same rows and outputs all runs. Compare numerical values with a declared floating-point tolerance, not a screenshot. The existing dependency-free two-game starter provides a second branch with exact task contracts and programmed search; it is not a neural implementation. Keep each branch’s manifest separate. A passing replay establishes agreement with the supplied software recipe, not independent scientific validity or reviewed implementation proficiency.

Try and explain

Inspect training and development row IDs. Export a recipe, run the independent Python studio and compare each run before adding your own intervention.

Open the reproducible study project or use the implementation studio.