Timescape versus ΛCDM with the Seifert et al. supernova method: a replication, an application to DES-Dovekie and a measurement-error test

Seifert et al. (2025) reported very strong evidence for the timescape cosmology over flat ΛCDM from Pantheon+ supernovae (ln B > 5 on the full sample) and moderate evidence (ln B > 1) when every supernova below z = 0.075 is removed. An independent implementation reproduces their Bayes factors at all 21 redshift cuts to within 0.08 (never below theirs; 0.03 higher on average). With the same likelihood and priors, and our own contamination cut and covariance construction, the recalibrated DES-Dovekie sample gives ln B < 0 at every cut (−2.70 to −0.71; −1.08 to −0.71 above z = 0.075). The sign at low redshift depends on the redshift treatment, and above z = 0.075 the result depends on the contamination cut: our PROBCC_BEAMS cut uses Hubble residuals, and a classifier-only cut gives an inconclusive −0.49. The method fixes stretch and colour at their fitted values, so their measurement errors never widen the distance uncertainty. We simulate data carrying those errors (no selection effects), with truths calibrated so that the simulations reproduce the real-data fits (added in version 2 after a post-publication review), and compare the models by the likelihood ratio of the observed ln B. This is not a Bayes factor and is graded on the Jeffreys scale only by analogy. Pantheon+, z_min = 0.075: 1.42 for timescape (5.3 without the errors; 1.21 to 2.95 across all 28 pairings of truths tried): no preference by analogy. Pantheon+, full sample: 5.2 for timescape (5.2 to 7.0 across pairings), moderate by analogy; the method omits peculiar-velocity corrections there. DES-Dovekie, z_min = 0.075, PROBCC_BEAMS cut: 8.1 for ΛCDM (4.6 to 8.9 over four independent simulation sets; 4.8 to 10.0 for other kernel widths; 3.0 to 26.7 across 32 pairings), moderate by analogy; only 9 of 1200 timescape simulations reach the observed value. DES-Dovekie, full sample: 1.7 for ΛCDM, but 1.0 to 5.1 across pairings; the grade depends on the truths. Applying the collaboration's bias corrections changes the Dovekie ln B at a given cut by up to 2.6 with the widest priors (6.3 with the 2σ priors), so these numbers are provisional until selection is simulated under each model. Note on versions. Version 2 supersedes version 1, which stays available. It adds simulations with calibrated truths (version 1's "matched" truths did not reproduce the real-data fits), which move the headline likelihood ratios (Pantheon+ at z_min = 0.075: 2.08 → 1.42; DES-Dovekie at z_min = 0.075: "about 3 to 5" → 8.1, partly from estimator and Monte Carlo noise), and it withdraws version 1's comparison of Seifert et al.'s Bayes factor with a likelihood ratio, its description of z_min ≥ 0.075 as the regime on which their claim rests, and its statement that Leith et al. (2008) and Smale & Wiltshire (2011) made the first supernova tests. It corrects several descriptions of other work and adds Kwan, Francis & Lewis (2009). Full list in the paper's Note on versions. The deposit contains all scripts, per-cut and simulation outputs, the calibration runs (including a superseded one), the aggregation that produces every number, and a checker that tests the result-dependent statements it lists and the quotations of other work against their sources. Computations and drafting were assisted by an AI system (Claude).