The Acquisition Clock in Bell Records: Exact Fibers and Schedule-Dependent Temporal Inference

Bell statistics characterize setting-conditioned one-trial behavior. They do not generally determine the joint law of an ordered record. We define exact one-trial fibers that contain processes with identical marginal Bell behavior and different cross-trial dependence. We then separate three components of the acquisition architecture. These components are sampling geometry, trial admission, and programmed measurement context. The distinction determines which mathematical and causal assumptions apply. For equatorial GHZ measurements, we construct explicit temporally distinct models in the same one-trial fiber and derive the sharp one-direction hiding cap. A second-order schedule map shows how periodic sampling can suppress informative lags or create exact aliases. Outcome involutions yield an anytime-valid e-process under predictable conditional symmetry. Source-gated analyses of NIST, Delft, Weihs, AlmostDI, and ETH records illustrate exclusion, anomaly routing, event-ready admission, pairing sensitivity, cross-talk localization, and model comparison. Controlled experiments on one IBM Marrakesh QPU test the analysis under internally committed ground truth. IBM-1 recovers all 180 injected telegraph blocks up to a global label and recovers an injected line frequency to 1.13 × 10−5 cycles per tick. IBM-2 recovers twelve programmed phases with mean circular error 0.043 rad. A training phase-orbit model predicts held-out co-located CHSH witnesses. The aligned witness is 2.8086, while a full phase sweep gives 0.0306. Matched separable controls remain below a simultaneous upper bound of 1.113. These experiments demonstrate identifiability and aggregation procedures under the programmed models. They do not establish a spontaneous hardware clock or a spacelike Bell test on the QPU.