Two GJ 876 orbit libraries leave 3.2 m/s median scatter on 39 later CARMENES nights after one offset

We test whether later GJ 876 velocities separate two groups of published four-planet solutions. The source paper retained rows with stellar-noise coherence of at least 20 days and shape w at most 0.7. We draw 2,048 unique orbit rows from each group and predict 39 public CARMENES nights after the source time range. The two median residual scatters are 3.16 and 3.20 m/s after one offset. With formal errors and 3 m/s of added independent noise, the retained-minus-other log score is +0.020. Its approximate 95% row-bootstrap range is [-0.099, +0.150]. Transferring each row's source noise gives +0.406, under a stronger assumption. A simulation under the primary noise selects the generating group in 59.8% of draws at equal group frequency. Thus, the tested libraries give similar orbital forecasts. This test does not establish the source noise assumption. Original chain weights were lost during source thinning, so these are finite-library scores, not posterior odds. We preserve the full source archive, observations, fixed predictions, code, provenance, figures, numerical checks, and replay instructions. No new planet is claimed.

GJ 876 · CARMENES DR1 · radial velocity · orbit forecasts · stellar noise · finite orbit library · reproducible benchmark