REVIEW 0.2.0-rc3
Interpreting DESI's evidence for evolving dark energy
2404.08056v3 · M. Cortês, A. R. Liddle
Reproduced in scope
The pivot transformation and main interpretation reproduce, with qualifications to the precision and direction of prior effects.
Limit The Dark Energy Survey five-year supernova sample does not meet the stated pivot precision; effects depend on the chosen prior.
Full report Checking method and assumptions Code and evidence package
What these results mean
- Reproduced in scope: the named checks support the reported result, with the stated exceptions; not a certification of the whole paper.
- Partly reproduced: some named results reproduce while important parts remain unverified.
- Confirmed finding: the specific discrepancy has author or independent researcher confirmation; downstream consequences may remain unknown.
- Technical finding — not externally confirmed: a supported, explicitly limited discrepancy without external confirmation.
- Attempt documented: the completed work and reasons it stopped are recorded. Failure to finish does not establish a paper error.
What we checked
- Published DESI numbers, the pivot-redshift transformation and selected summaries of the released parameter samples.
- How the stated conclusions depend on prior choices, confidence contours and the assumed dark-energy parametrization.
| Check | How it was checked | Result | Evidence assessment | Evidence |
|---|---|---|---|---|
| The quoted detection significances agree with conversions of the printed fit improvements using two additional parameters. Report reference: C1 | Checked against source material | Reproduced in this check | Supports this assessment | Evidence 1 |
| The stated priors match the DESI table. The sampled high-redshift equation-of-state combination lies below the tested boundary, although that boundary is not explicitly listed in the configuration file. Report reference: C2 | Checked against source material | Reproduced in this check | Supports this assessment | Evidence 1 |
| The pivot redshifts and scale factors reproduce from the released samples within the declared criteria. Report reference: C3 | Checked against source material | Reproduced in this check | Supports this assessment | Evidence 1 |
| Figure 2 uses marginal means, not best-fitting points. The claimed ±0.02 proximity holds for Pantheon+ only; the broader 'few hundredths' description fits all three samples. Report reference: C4 | Checked against source material | Difference identified | Supports this assessment | Evidence 1 |
| The pivot transformation is algebraically correct. The induced prior on the pivot equation of state is slightly tilted rather than flat, with a small effect on its mean. Report reference: C5 | Checked against source material | Difference identified | Supports this assessment | Evidence 1 |
| The confidence regions reach the cosmological-constant line, but the Dark Energy Survey five-year sample's central pivot value does not lie on that line. Report reference: C6 | Checked against source material | Difference identified | Supports this assessment | Evidence 1 |
| The parameter-region argument holds in the checked samples. One plotted boundary differs from the corresponding calculation. Report reference: C7 | Checked against source material | Difference identified | Supports this assessment | Evidence 1 |
| The maximum-density derivation holds. For the Dark Energy Survey sample, the quoted pivot lies outside the central 68% interval of the crossing redshift, so the interpretation is only partly reproduced. Report reference: C8 | Checked against source material | Difference identified | Supports this assessment | Evidence 1 |
| Prior dependence was already disclosed by DESI. The tested prior changes do not all reduce the Bayesian evidence in the same direction; an additional later test was exploratory. Report reference: C9 | Checked against source material | Difference identified | Supports this assessment | Evidence 1 |
| The selected constant-equation-of-state, survey-swap and curvature comparisons reproduce. One sentence describes a mild discrepancy more strongly than DESI did. Report reference: C10 | Checked against source material | Difference identified | Supports this assessment | Evidence 1 |
“Supports this assessment” means the evidence supports the stated result, including a discrepancy or limitation. It does not mean the paper's claim passed.
Run the packaged code
mkdir replay-work && python3 -m zipfile -e legacy-original.zip replay-work && cd replay-work && python3 reproduce.py --allUse the package instructions for dependencies, data and limits. This command repeats the packaged checks; it does not necessarily reproduce the whole paper.
AI used and recorded resources
Partial records · 6/6 sessions with complete token counts · 6/6 with activity time recorded
OpenAI Codex · gpt-6-astra · reasoning: high
Anthropic Claude · claude-opus-5-5 · reasoning: high · settings reported by the project owner
Token breakdown and what the time measures
- New input
- 487.1K
- Input read from cache
- 71.9M
- Input written to cache
- 1.7M
- Output, including reasoning
- 291.8K
- Recorded agent activity
- 1.9 h
- Scientific computing time
- Not established by these records
- Human time
- Not recorded
Recorded case work before 2026-11-10; later shared collection preparation is excluded. Tokens count processed traffic, including repeated cached input, not unique text or a monetary cost. Reasoning tokens are already part of output.
Session time sums the recorded session windows, including waits and parallel sessions. Activity time sums only explicitly logged activity durations and may include tool waits. Neither is elapsed project time, scientific CPU/GPU time, or human labour. Missing, estimated or overlapping records are not silently counted as zero or added to measured totals. 0 records are excluded from sums; remaining records can still lack particular counters. This is a partial accounting snapshot, not a full cost estimate.
Claude contributed the first pass; Codex contributed separate checks and criticism. Later packaging and informed corrections are distinguished in the audit record. A separate AI check is not external scientific confirmation.
Technical scope, original wording and review history
Reference labels such as C1 or N3 identify checks and notes in the report, not physical quantities. Workflow letters denote the first pass (A), separate checking (B) and criticism (C). P2 means package validation; G2 means the rule to commit a plan before scientific execution. V-numbers identify internal workflow steps. S0–S3 are internal issue-severity labels, not statistical significances.
Original scope and exclusions
- The DESI significances, priors and pivot redshifts the paper quotes, against DESI 2024 VI v1, v2 and v3 (C1–C3, C10)
- The pivot values of w(a) behind Fig. 2 and the ± 0.02 / 'a few hundredths' statements, from DESI's chains (C4)
- The algebra of the pivot transformation, and the claim that the induced prior on w_p is flat (C5)
- The orientation of the w0–w_a contours, their offset from w_p = −1, and the pivot-shifted contours the paper could not draw (C6)
- The four zones of Fig. 1, where the posterior lies among them, and the drawn lines (C7)
- The phantom crossing, the maximum of the dark-energy density, and its position against the pivot (C8)
- The prior argument: what DESI's significances and Bayes factors depend on, and what five prior variants fixed before the run do to the Bayes factor and to the credible level of ΛCDM (C9)
Not checked: A new fit or optimisation. All numbers come from reweighting DESI's released chains, so DESI's own maximum-posterior statistic was not recomputed under a changed prior; New nested-sampling evidences. Absolute Bayes factors are DESI's printed ln B (0.65, 2.4, 2.8) plus the change computed here. If DESI normalised its prior over the full box rather than the w0 + w_a < 0 wedge, every absolute value would move by 0.255 in magnitude; the changes would not; The never-phantom prior (T5) beyond a rough estimate: only 2 to 31 effective samples survive it, so it needs its own sampler; How wide the observational window behind the pivot is, and how likely an alignment of the contours with w_p = −1 would be under ΛCDM; Whether the CPL family describes the dark energy beyond the observed range; the 'maximum it will ever have' statement holds only for CPL extrapolated to all times
Original claim wording
C1 · C1 in fixed reviewed version: 2.5/3.5/3.9 sigma printed in DESI VI v1-v3; Delta chi2_MAP -8.7/-15.2/-18.1 with k=2 give 2.486/3.481/3.852 (A and B)
C2 · C2 in fixed reviewed version: Priors match DESI Table 2; w0+wa 99.9th percentiles -0.98/-1.17/-1.14, below -0.3; no explicit wedge entry in the yaml (Addendum A1)
C3 · C3 in fixed reviewed version: z_p 0.26/0.33/0.26 and a_p 0.79/0.75 exact; chain DETF pivots 0.2615/0.3260/0.2628; fidelity gate passed for all four runs
C4 · C4 in fixed reviewed version: Fig. 2 values are DESI's marginal means, not best fits (N1); '+-0.02' holds for PP only, 'a few hundredths' for all three (|E[w_p]+1| 0.018/0.040/0.055); DY5 carried by N2 at C6
C5 · C5 in fixed reviewed version: Algebra holds (sympy, corr ~1e-16, sigma ratios in [0.8, 1.25]); marginal prior on w_p tilts 11.8-12.3% over [-1.2,-0.8], not flat; effect on E[w_p] ~2e-4 (N3, S1 after V6)
C6 · C6 in fixed reviewed version: Axes within 0.18 deg of w_p=-1; all 95% contours reach the line; d at DESI pivots 0.66/0.97/2.15: DY5 does not lie on the line (N2)
C7 · C7 in fixed reviewed version: Zones hold; mass outside Z4 <= 0.003, no 95% cell outside Z4; Fig. 1 magenta line off by up to 0.068 in w0 (N4)
C8 · C8 in fixed reviewed version: Maximum-density derivation holds, P(wa<0) >= 0.9989; DY5 pivot 0.26 outside z_x 68% [0.307, 0.418] (N5, same fact as N2); partially_reproduced
C9 · C9 in fixed reviewed version: (a) DESI stated prior dependence of its evidence from v1 (N6, S1 after V6); (b) T1-T4 lower the LCDM level, ln B falls for all three only under T2, rises under T3/T4; post_hoc T6 lowers both (N7); partially_reproduced
C10 · C10 in fixed reviewed version: wCDM rows within 2 sigma of -1, SDSS swap 0.3/0.4/0.4 sigma, curvature |Omega_k|/sigma <= 0.21; 'significantly' for DESI's 'mildly' at z=0.51 (N6 wording list)
Scientific abbreviations
DESI: Dark Energy Spectroscopic Instrument; BAO: baryon acoustic oscillations; SN: supernovae; CMB: cosmic microwave background; DESY5/DES5Y/DY5: Dark Energy Survey five-year supernova sample; PP: Pantheon+; PD: Planck + DESI; PDPP: Planck + DESI + Pantheon+. CPL, JBP and BA name the Chevallier–Polarski–Linder, Jassal–Bagla–Padmanabhan and Barboza–Alcaniz dark-energy parametrizations. EXP and LOG mean exponential and logarithmic parametrizations. Full-plik refers to the full Planck high-multipole likelihood; lite is its compressed approximation. ESS is effective sample size; IC means initial conditions; MG means modified gravity. NEC is the null energy condition; GR is general relativity; FLRW denotes Friedmann–Lemaître–Robertson–Walker cosmology. GB means Gauss–Bonnet. AIC and DIC are model-comparison information criteria. Further symbols are defined in the linked paper and protocol.
See original source-payload README and recorded P2; not rerun during assembly. Unpacks the unchanged original candidate into a new work directory and runs its documented scoped replay. Package validation checked the artifact and inherited the earlier replay after verifying the relevant files were unchanged; it did not rerun the scientific calculation.
Original provenance disclosure
Prepared with AI assistance. Historical Claude first pass and Codex independent review/critique remain distinct; this assembly reads saved evidence and is neither blind B nor new P2. No scientific code was run in this assembly. Before the replay command below, extract legacy-original.zip into a new directory and read its README.md and requirements.txt for installation, dependency, data and scope instructions.
Tools: Claude (historical A); Codex (historical B/C and current informed assembly)
Human involvement: Owner selected the preparation scope and approved the predecessor release published on 10 November 2026; the owner approved publication of this editorial revision on 11 November 2026