Model Consistency

One of the main failure modes this repo is designed to prevent is inconsistency between:

• transform output
• model object used for predict
• model/decomp logic used for contribution outputs

Rule

For a given market/KPI build, transform, predict, and decomp must refer to the same intended model representation.

What This Means In Practice

If a model is rewritten for forecast/upload purposes, for example by replacing an inline combined term with an explicit single variable, then:

• transform must produce that rewritten variable
• the API model must be the updated model that expects it
• decomp must use that same updated model representation

Not Acceptable

Examples of inconsistency:

• transform produces updated variables but the API still uses the old stored model
• predict uses one model object while decomp closes over another
• decomp assumes a mapping or formula representation different from the build model

Why This Exists

This exact class of inconsistency was one of the initial problems surfaced by the China store visits investigation.