Recording what a model is and is not.
WHAT TO DOCUMENT
Intended use, and uses considered unsuitable
Training data: source, period, size, known gaps
Performance overall and by group Limitations and known failure modes Ethical and legal considerations
Maintenance: who owns it, and when it is reviewed
WHY INTENDED USE MATTERS MOST
Models are reused for purposes they were never evaluated for.
WHAT THAT CAUSES
Confident deployment into contexts where performance is unknown.
WHAT TO STATE EXPLICITLY
What the model must not be used for.
WHAT TO DOCUMENT ABOUT DATA
Where it came from What period it covers What populations are underrepresented Known quality problems
WHY PERIOD SPECIFICALLY
It bounds where the model can be expected to work.
WHAT TO PUBLISH INTERNALLY
All of it, alongside the model.
WHAT TO REVIEW
The documentation, whenever the model is retrained.
WHAT TO RECORD FOR DECISIONS AFFECTING PEOPLE
How a decision was reached, sufficiently to explain it.
WHY
Affected people may be entitled to an explanation.
WHAT TO NEVER DEPLOY
A model nobody can describe the limits of.
