The corrected NANOGrav 15-year free-spectrum release weakens the preference for curved background spectra

Many refits of the nanohertz gravitational-wave background use NANOGrav's public 15-year free-spectrum kernel density estimates, release v1.1.0. In August 2026 NANOGrav released v2.0.0, rebuilt from chains rerun after a parallel-tempering swap bug in PTMCMCSampler was fixed. Starting from the model comparison of Xu, Zhang and Guo, whose numbers we reproduce, and adding seven NANOGrav new-physics templates, we rerun the comparison on v2.0.0 under a protocol that was written, reviewed by a separate AI instance and hashed before any evidence was computed on the new files (the time stamp is our own; the 14-bin analysis was made co-primary after the v2.0.0 densities had been inspected). With the 14 signal-carrying bins, every preference with ln B > 1 for a curved or new-physics spectrum over the gamma = 13/3 SMBHB power law shrinks by 0.7 to 1.6; none now exceeds 2.5. Against NANOGrav's astrophysical SMBHB prior the largest falls from 4.4 to 3.3. With all 30 bins the ranking changes: Xu et al.'s phase transition falls from fourth to eighth, its Bayes factor from about 140 to 2.7, but that fall depends on the noise-dominated bins 15 to 30. Our errors cover refit Monte Carlo noise only. We also show that Xu et al.'s out-of-grid convention produces spurious ln B of 40 to 50 on the v2.0.0 grid; the density should be extended flat below it. Code, protocol, hashes and all results are included. Note on versions. Version 2 corrects version 1 after independent review; no result file changes. Version 1 wrongly said no refit had used v2.0.0; an earlier one (Liu 2026, Zenodo 22733734) with different templates is now cited. "Pre-registered" is removed from the title: the time stamp is self-asserted and the 14-bin analysis was promoted after the new densities were inspected; both are now stated. The protocol's mixed-dataset check is now reported. The error bars are stated to cover refit Monte Carlo noise only. Misquoted numbers are corrected (0.7 to 1.6; 4.4 to 3.3; 43; 0.32). The erratum's and the release notes' statements are both quoted. Full list in the paper's Note on versions.