Fits with two free periods give larger error scores for 38 later GJ 1002 velocities in seven fixed noise cases

We test whether an early velocity fit predicts later variation at GJ 1002. We fit two circular signals to 86 raw CARMENES velocities and predict 38 later ESPRESSO velocities. The first test row follows the last training row by 420.143 days. The period bands are 9.5 to 11.2 days and 19 to 23 days. All orbit parameters remain fixed during testing. Only a test instrument constant is removed. The error score is a generalized squared residual under a fixed covariance. The free-period fit gives a larger later score than a flat prediction in all seven noise cases. Score ratios are 1.030 to 1.544. In the primary case the early free-period score is 77.883, versus 78.699 when both periods have their published values. The later scores reverse this small early advantage: 43.213 for free periods, 22.100 for published periods, and 41.953 for flat. The published periods use the full source record and do not form an independent early-data forecast. A profile finds early fits within one score unit of the minimum whose later scores range from 26.821 to 85.713. Other selected data splits improve on flat predictions. The result is specific to the late chronological test, the stated models, and point predictions. It is not evidence of a physical period change. The fits omit the joint activity model used in the discovery work and do not integrate parameter uncertainty. The contribution is a measured forecast limit and a reproducible benchmark, with raw data, provenance, source code, full fit records, figures, and numerical checks.

GJ 1002 · radial velocity · time prediction · period uncertainty · CARMENES · ESPRESSO · stellar activity · reproducible astronomy