Material-Dependent Clustering in Measurements of the Newtonian Gravitational Constant: Pattern, Controls, and Falsifiable Predictions

This preprint serves as a timestamped prediction record (de facto preregistration) for the outcomes described in Part II, filed before publication of the NIST copper–sapphire comparison result. Part I (Postdiction) reports a previously unidentified correlation in published Big G data: when thirteen precision measurements of the Newtonian gravitational constant are grouped by source mass material, inverse-variance weighted means follow a monotonic ordering Cu > SS > W, with a weighted copper–tungsten separation of 387 ± 39 ppm. The copper (n=2) and tungsten (n=2) groups are each internally consistent, while the stainless steel group (n=7) shows the expected substantial internal scatter. The pattern is explicitly identified as a postdiction: material is perfectly confounded with laboratory and method for the extreme groups, and the sample sizes are small. No physical conclusion is drawn from Part I alone. Part II (Prediction) converts this postdiction into five falsifiable predictions for experiments not yet published: P1 (Tier 1, flagship): G(copper) ≠ G(sapphire) at NIST, using the BIPM Mark-2 balance with both source masses on the same apparatus — the first experiment to break the material–lab confounding. Pre-analysis thresholds: confirmed if z > 2, falsified if z < 0.5. P2 (Tier 3): Sign of the Cu–sapphire difference discriminates electronic vs lattice coherence. P3 (Tier 2): Single-crystal > polycrystalline for any material. P4 (Tier 3): Residual G increase at cryogenic temperatures after conventional thermal corrections. P5 (Tier 2): Material corrections collapse the historical scatter (Birge ratio < 2.0). A speculative phase-coherence mechanism is proposed (Section 6) but is not required for the model-independent (Tier 1) and pattern-extrapolation (Tier 2) predictions. Supplementary file: big_g_verification.py independently reproduces every numerical claim in the paper (group statistics, chi-squared values, combinatoric null, leave-one-out analysis, threshold calculations).