A Reader Is Not Its Readout, and a Capability Is Not Its Utility

A time-series reader can hold information that its readout throws away, and a capability that is real can still be useless ifthe world rarely supplies the condition that would make it decisive. We show both, cleanly separated. The WCM occupancyreader records how a rank-normalized series fills the 24 vertices of the 24-cell in delay coordinates. Summarizedisotropically, by the total-variation change of the occupancy histogram, it loses to a simple window-mean detector on everykind of regime change we tried. But the occupancy distribution is a coarse-grained empirical measure on delay-coordinatespace, not on the value marginal, so it carries information that no marginal moment can. We build a controlled regimechange in which the mean, variance and skewness of the window are held fixed and only the temporal organization (lag-1rank autocorrelation) changes. A discriminative readout -- a single direction learned on training seeds and frozen -- letsthe occupancy reader beat mean, variance and skewness detectors on 100% of held-out test seeds, down to autocorrelationshifts as small as 0.20 to 0.30. Decomposing the frozen winning direction under the signed-permutation group B4 shows92% of its energy in one six-dimensional block and 64% in the S4 sign-concordance layer: the winning direction is literallysame-sign minus opposite-sign delay-coordinate occupancy, a relation invisible to any marginal moment. That is thecapability. For the opportunity, we ask how often real earthquake magnitude series present this condition. Acrossearthquake regime boundaries the occupancy reader wins only one of four -- the single quiet-period boundary whose meanbarely moves -- while every mainshock boundary goes to the mean; across thirty unselected boundaries the effect is in thepredicted direction but not significant. Capability is proven; the opportunity, in this data, is rare. ※ The license for codes is AGPL3.0-or-later