Three 2017–2018 intervals reverse the 2023 ranking of two fitted torque models for Swift J0243.6+6124

The benefit of a constant torque term depends on the earlier fit interval for Swift J0243.6+6124. We fit an integrated BAT count-rate power law to public GBM frequencies from 2017–2018. We compare B = 0 with fitted constant B. We keep the supplied orbit correction fixed. Only a frequency offset is set from the first three 2023 measurements. The other 40 measurements give prediction error. With 139 earlier points, fitted B raises the 2023 root mean square error from 6.41 to 27.51 microHz. A faint tail selected from count rates has 45 points. There, fitted B lowers the error from 15.92 to 2.83 microHz. Removing the last 15 days leaves 37 points and raises the error from 16.36 to 31.15 microHz. Fitted B improves the earlier fit in all three cases. The added term therefore has no stable prediction advantage across these intervals. The estimator and the 2023 spin increase are established in prior work. This contribution supplies a fixed test for two events. It shows that the model ranking changes with the earlier interval. It includes all 64 control fits, 2400 conditional residual resamples, independent numerical checks, raw source files, provenance, code, and figures. The test uses observed later count rates and is not a real-time forecast. One source and one event pair cannot identify a universal physical torque law.

Swift J0243.6+6124 · accreting pulsar · torque · prediction · Fermi GBM · Swift BAT · fit interval · reproducible benchmark