Hi everyone, I have finished a pre-registered, frequentist test of a quadratic-in-frequency group-velocity modification (the alpha = 4 case of the LVK parametrisation) using public strain and the released posterior samples (GWTC-2.1/3 for O1–O3, GWTC-4.0 for O4a). The method works on residuals against the posterior samples, projects out a posterior-derived nuisance subspace, treats calibration as a known covariance and builds Neyman sets from profile scans. Stopping rules were fixed in advance and checked on off-source data before any on-source estimate. Short version of what happened: - With the SEOBNRv4PHM labels, injections into real off-source noise were covered by the 90% sets only 300/360 times (v1) and 299/352 (v2), below the pre-registered threshold (~0.859); IMRPhenomXPHM passed (~0.89). Both versions stopped there. - A pre-registered diagnostic did not find the mechanism, but one thing stood out: the time shifts needed to align individual posterior samples with our reference waveform scatter by 2–18 ms for the SEOBNRv4PHM releases (all from LALInference in O1–O3), versus 0.1–0.6 ms for bilby releases. Model and sampler are fully confounded in that set, so we cannot tell which one causes it. - A third version builds the belts from the full analysis chain (including our time/phase refinement) and mixes IMRPhenomXPHM and SEOBNRv5PHM. Coverage in new real-noise segments is now fine (0.96 at Λ = ±2σ, 0.89 under the model mismatch), but the honest intervals become very wide and hit the edge of the pre-registered grid, so it also stopped before any on-source estimate. Preprint: Zenodo, DOI 10.5281/zenodo.23324211 Code, registrations and full audit trail: GitHub, Kretski / residual-first-framework (folder scanner/gw) Two questions where I would value the community’s view: 1. Is a ms-level spread in the effective merger time of individual LALInference SEOBNRv4PHM samples (relative to a reference waveform generated from the same parameters with current LALSuite) expected, for example from time or reference-frequency conventions of that release, or does it point to a problem on my side? 2. Our coverage check at a non-zero true value suggests that refining time and phase per sample absorbs part of a large dispersive phase. Is this a known issue in dispersion tests that marginalise over time and phase, and is there a standard way to account for it? No dispersion is detected and no validated bound is claimed; this is a methods result. Any feedback, including on the setup, is welcome. Thanks, Dimitar