An analytics protocol for calibrating benchmarking-community claims in data envelopment analysis

Reading a data envelopment analysis intensity matrix as a reference network invites three claims that a partition cannot separate: that the structure is imposed by the envelopment programme, that the panel is extreme relative to a declared reference technology, or that more than one technology generated the units. We audit a pipeline on one confidential retail panel and six public panels and ask which claim each component identifies. One result is exact: a basic feasible solution of the variable returns-to-scale programme has at most m + s positive intensities per unit, and 63 solver-tolerance cells show no violation. A degree-preserving rewiring null fires in every replicate of both structurally valid single-regime arms, so its size is invalid. A reference ensemble generated from a declared technology does not identify latent regimes: it misses regimes not visible as network blocks, fires on a single-technology panel with concentrated benchmark mass, its false-fire rate moves from 0.025 to 0.125 between two plausible technologies, and at 200 draws it reproduces its verdict about half the time near its threshold. On the application panel and five public panels the statistic falls below its reference and the verdicts are stable; on the application panel the object classification is unassessed, so no community-informed decision rule is constructed as a confirmatory analysis. The deliverable is a calibration record rather than a validation certificate, stating, for each claim, the inferential object, the reference model, the observed rank with its uncertainty, the model’s own error rate, and where the analysis stops.