The Two Anchors of a Measure

A stored analytical result is conventionally described by one grain: the grain at which the result is reported. That is often sufficient for additive quantities such as total revenue. It is not sufficient in general.

Many common analytical operations involve two distinct analytical locations. The input anchor identifies the points over which an operation is formed or reduced. The output anchor identifies the points at which the resulting member is reported. Average order value over orders and average revenue over customers can occupy the same output anchor and use the same underlying revenue while denoting different quantities because their input anchors differ. Weighted averages, rates, ratios, counts, extrema, order statistics, and multi-stage reductions expose the same issue in different ways.

Within the Theory of Data framework, a measure is a stable governed family and a member is one anchored realization of that family. “Two anchors of a measure” is therefore compact language for a more precise claim: an anchor-sensitive operator instance or derivation may require both an input member at an input anchor and an output member at an output anchor, and the input anchor may be identity-bearing.

Two questions must be kept separate. Sufficient-state composability asks whether staged reductions preserve the information required for exact continuation. Input-anchor sensitivity asks whether changing the input anchor changes the value or interpretation of the requested quantity. COUNT, for example, composes through additive count state even though reapplying COUNT to displayed counts is wrong; MIN and MAX can compose exactly while still being uninterpretable without knowing what entities they ranged over. This distinction yields a general materialization rule: values, sufficient state, and analytical identity must not be collapsed.

The practical importance is not hypothetical. The companion silent-failure taxonomy classifies five of fourteen planted benchmark defect classes as anchor freedom failures, including re-aggregation, semi-additive summation, empty-bucket behavior, and definition ambiguity. In a nine-model text-to-SQL study, the structural failure families persisted despite strong schema documentation. The lesson is architectural: analytical systems need an explicit, structured, verifiable account of anchor-sensitive meaning. Query languages and AI interfaces that can name only a metric, output dimensions, and filters are therefore not expressive enough for every analytical request. A language must either derive the missing input anchor uniquely, ask for it, or represent it explicitly.