A controlled posterior-mean and best-fit diagnostics of w₀CDM support conditioning using the Planck 2018 baseline likelihood combination (high-ℓ TTTEEEₗite, low-ℓ TT, low-ℓ EE, and CMB lensing) combined with DESI DR2 BAO galaxy and QSO tracers. Three inference branches are compared under identical datasets, likelihoods, and solvers: (i) a neutral control run (CTRL2) with broad prior support in w₀, (ii) a non-phantom restricted-support run enforcing w₀ > −1, and (iii) a phantom restricted-support diagnostic run enforcing w₀ < −1. The branches are treated as truncations of a common posterior manifold rather than distinct physical models. Posterior-mean block-wise χ² decomposition reveals a structured early–late redistribution: the non-phantom branch improves the lowest-redshift BAO component (LRG1, z ≈ 0. 14) while incurring a modest penalty in the high-ℓ CMB; the phantom branch exhibits the opposite behavior. These component-level contributions largely cancel in the total goodness-of-fit, indicating redistribution along the (w₀, H₀) degeneracy direction rather than decisive model preference. Early-universe parameters remain stable across all branches, with shifts remaining below the 1σ level. A BOBYQA multi-start minimizer comparison at the likelihood maximum reveals a qualitative change: the non-phantom branch is statistically indistinguishable from CTRL2, while the phantom branch is mildly disfavored. A pre-committed curvature scan along the validated (w₀, H₀) degeneracy direction additionally served as a stress test of the minimization procedure. The scan identified a lower-χ² region not reached by the initial five-start optimization; the minimizer was subsequently re-run with scan-seeded starting points, and the corrected best-fit preserved the qualitative result while reducing the magnitude of the phantom disfavor. The novelty is therefore not the existence of the (w₀, H₀) degeneracy itself, but the diagnosis of how posterior conditioning redistributes likelihood contributions and how this redistribution changes at the likelihood maximum. All analyses use Cobaya with the CLASS Boltzmann solver in a four-chain MPI configuration. The accompanying reproduction kit (DOI: 10. 5281/zenodo. 20122148) provides archived MCMC chains, Cobaya configuration files, post-processing scripts, posterior summaries, χ² block decompositions, convergence diagnostics, and best-fit minimizer outputs. Chain-derived results are reproducible from the archived chains using the included scripts. Best-fit minimizer outputs and curvature-scan products are archived in the kit; their regeneration requires local installation of the Planck 2018 likelihood data.
Erik Tobias Hummel (2026) studied this question.