Let's Optimize Something
WritingProjectsNotesAboutTR
article

Does an AutoML System Actually Understand Your Dataset? A Semantic Intelligence Ablation

Five levels of semantic context, the same trapped datasets, the same local 7B model. A data dictionary and business rules made a leak-free pipeline more likely on unseen datasets; an extra verification stage did not, and no condition noticed what the dictionary left out. Stronger hosted models recovered much of the meaning from column names; with anonymized columns, the dictionary still decided.

Let's Optimize Something / technical writing2026