https://www.kelytix.com/en/column/ Notes on the data layer of AI in vertical domains: concrete scenario difficulties, engineering methods and trend judgement. en Thu, 01 Oct 2026 09:00:00 +0800 https://www.kelytix.com/en/column/pen-dataset.html https://www.kelytix.com/en/column/pen-dataset.html Thu, 01 Oct 2026 09:00:00 +0800 Annotation accuracy measures the samples that were labelled, not model behaviour. Drawing on public research, this article explains why a dataset of positives alone cannot define a model's boundary, why annotator disagreement is signal rather than noise, why evaluation sets saturate with age, and what each of the three PEN data categories — Positive, Edge, Negative — solves.