The End of Fixed Grouping: Grouping as an operation on data, not a property of it

Data analysis commonly begins by partitioning observations into fixed groups. We examine the assumption that such a partition belongs to the data, without proposing that any other organisation does. A partition is an equivalence relation — binary and transitive — and group-mean estimation is the Nadaraya–Watson estimator with that relation. Removing the two constraints recovers established nonparametric methods; no new mathematics is introduced. The reorganisation has consequences. A partition can be derived from any symmetric graded relation by thresholding, and graded relations can be derived from partitions by averaging; we argue that neither should therefore be treated as prior. A group estimator is an orthogonal projection, so its stability under re-estimation from its own outputs is guaranteed by construction and cannot count as validation; graded estimators need not be idempotent, and can make such self-reinforcement observable. The authority of a categorical boundary can be estimated, through a categorical smoothing parameter, rather than assumed. In synthetic experiments a graded relation reduced error against the truth by 78% relative to administrative groups, whose errors concentrated along their own boundaries; cross-validation treated a real boundary as nearly hard (λ̂ = 0.05) and largely discounted an irrelevant one (λ̂ = 0.7); and under feedback, self-disagreement fell monotonically while external error rose after the first pass. Selection, representation and objective remain. What changes is that they become examinable. Files. The paper (PDF); experiments.py, which reproduces every figure in Section 7 with a fixed seed (Python 3 and NumPy only); and experiments_output.txt, the expected output. Preparation. This paper was developed by the author in sustained collaboration with AI systems: Claude (Anthropic) and ChatGPT (OpenAI). They were used for drafting, mathematical development, implementing and running the experiments, checking references and critical review. The author set the question and direction of the work and takes responsibility for its content. The experiment script was run in two separate environments, with identical results. Status. This is a preprint and has not been peer reviewed.