Observer121: A Matched-Control Framework for Attributing Productivity in Adaptive Observation

An observation strategy can look dramatically more efficient than another while leaving the reason for that advantage unresolved. Fewer measurements may result from better selection, adaptive use of feedback, a richer query language, a more compact representation, or simply a different accounting rule. We introduce Observer121, a diagnostic framework for separating these sources before assigning credit. An experiment is represented by a configuration of worlds, admissible queries, selection policy, representation, and resource accounting; matched interventions then classify a claimed gain as isolated, confounded, equivalent-control, interaction-dependent, or unresolved. Calibration experiments recover known effects when the relevant factor is controlled: adaptive threshold search reaches the binary information bound while the matched non-adaptive policy requires all thresholds, and fixed bit queries expose query-language rather than adaptivity gains. A compilation control shows that online generation and an equivalent procedural realization can produce the same query sequence, thereby separating query efficiency from representation cost. In an 11-by-11 system–observer laboratory, 21 observations determine an additive model, yet exact certification against a single unrestricted hidden interaction requires all 121 cells. A finite-population audit connects these extremes under risk assumptions. Finally, a matched real-data experiment on breast-cancer diagnosis shows how the same attribution logic localizes performance differences under matched experimental conditions.