Scale-Indexed Drift Geometry in Loophole-Free Bell Experiments

We ask whether order-dependent changes in the four correlators of loophole-free Bell experiments can be separated into reproducible symmetry-adapted drift modes without mistaking large trial counts or apparatus structure for new physics. A retrospective analysis of 1,048,576 ETH Zurich 2023 trials found that a half-file change placed 97.557% of its Euclidean energy in a common mode and was compatible with a one-factor multiplicative visibility model. Before archive access for a 1,342,177,280-trial ETH Zurich 2026 study, we checksum-locked the training procedure, held-out split, transport tests, and decision thresholds. The direction estimated on training blocks did not transport to held-out blocks under those criteria. Exploratory Walsh-Hadamard decomposition instead found a checkerboard-aligned validation geometry. Clean-room Node.js and independent Python recomputations reproduced all 21 locked quantities. However, an exhaustive five-versus-five native-block permutation test gave p = 0.531746 under block exchangeability, and a block-resampling sensitivity interval crossed zero. We therefore propose a scale-indexed diagnostic workflow that separates arithmetic reproducibility, geometric structure, and native-block evidence. A hierarchical mode-routing hypothesis is retained only as a prospective apparatus hypothesis requiring preregistered telemetry and independent-platform replication.