Dark-Energy Stress Lab

Project B v2 — an independent compressed-likelihood implementation and stress test of the DESI evolving-dark-energy claim

de-stress-lab is now also an installable Python package for selection-aware influence diagnostics, null calibration, and frozen scientific prediction ledgers. The original scripts remain frozen as the flagship validation case.

python -m pip install -e ".[test]"
destress demo
pytest
destress verify-ledger predictions/2027-ledger.json

• Package quick start
• Frozen 2027 prediction ledger
• Reproducibility and statistical red-team services
• JOSS paper and eligibility status
• Architecture decision
• Contributing and support
• Release clarifications and errata

Scope statement (audit-reconciled): this lab independently reimplements the published compressed likelihoods (DESI DR2 BAO summaries, Pantheon+/DES5Y distances, the DR2 compressed early-CMB prior) and reproduces their maximum-likelihood improvements and compressed-branch posterior structure. It is NOT a raw-data reanalysis and does not reproduce DESI's exact PR4/ACT full-CMB or 2024 DES5Y/Union3 headline rows. Four AI systems performed software, statistics, or manuscript checks; these are not independent scientific review. A clean-room human rerun is the outstanding verification step.

Status (2026-07-27): reproduction and diagnostics complete, including the 5,000-mock selection calibration, direct time-variation calibration, and held-out LRG2 prediction. The canonical DOCX/PDF has been regenerated from whitepaper.md, validated, and inspected page by page. A human clean-room rerun is prepared and pending. Canonical completion states, hashes, and numbers are in RESULTS_MANIFEST.json.

Frontier result: what advanced and what did not

The project now tests the scientific question more directly than the original ΛCDM-versus-w₀wₐ headline:

• Direct time variation: DESI+cCMB gives T = χ²(wCDM) − χ²(w₀wₐCDM) = 7.8506. Only 30 of 5,000 mocks generated at the best-fit constant-w null are as large: p = 0.0060 [0.0041, 0.0086], or 2.75σ.
• Without LRG2: T falls to 3.4263; 349/5,000 mocks are as large: p = 0.0698 [0.0629, 0.0772], or 1.81σ.
• Held-out LRG2: fitting ΛCDM without LRG2 and predicting it gives a joint p = 0.0408 (2.05σ). The isotropic distance is low (D_V/r_d p = 0.0204), but the AP ratio D_M/D_H is ordinary (p = 0.374).

This is stronger evidence that the compressed CPL likelihood genuinely rewards time variation, but weaker evidence for a revolutionary physical interpretation: most of the direct significance remains LRG2-sensitive, and the targeted LRG2 p-values are not independent of selecting that tracer after inspection.

The physics already visible: BAO prefers Ωm ≈ 0.297; supernovae prefer Ωm ≈ 0.333. Within ΛCDM these should agree. That ~2σ mismatch between expansion-history probes is a seed of the w₀wₐ preference — the "evolving dark energy" fit is, in part, the model bending to reconcile these two numbers. Whether that reconciliation reflects physics or calibration is what this lab exists to test.

Provenance rules (non-negotiable)

1. No number enters the pipeline from memory. BAO values were machine-extracted from the arXiv HTML of the DR2 paper; SN data downloaded from the official PantheonPlusSH0ES release.
2. No interpretation until reproduction matches published results (PASSED for BAO-alone, SN-alone, and the w₀wₐ ladder rows −4.7/−4.9/−8.0; NOT attempted for full-CMB rows — out of scope).
3. Both the baseline analysis and published critiques get implemented as switchable pipeline options.
4. Results publish regardless of which way they point.

Layout

src/       reusable installable `destress` package
paper/     JOSS paper, bibliography, and readiness gate
predictions/ frozen 2027 ledger plus SHA-256 digest
data/      pantheon_plus.dat + cov, des_dovekie_hd.csv + npz (official releases), desi_dr2.html (BAO/CMB source)
scripts/   fit_lcdm.py (wk1 reproductions) · fit_w0wa.py (w0wa + calibrated-surrogate CMB, validation gates)
           influence.py (LOO/pulls/SN cuts) · fit_des5y.py (DES5Y + intercept test)
           ccmb_ext.py (anchored intercept + w(z) swap) · null_mocks.py (parametric bootstrap)
           max_influence_mocks.py (all 7 deletions inside each null mock; selection calibration)
           time_variation_mocks.py (wCDM-vs-CPL direct test, full and no-LRG2)
           lrg2_posterior_predictive.py (held-out joint/AP/isotropic test)
           camb_check.py (full-Boltzmann swap — REQUIRES local CAMB)
results/   *.json per analysis
cleanroom/ blind independent-verifier protocol, worksheet, pinned environment, one-command runner
docs/      pinned canonical-whitepaper build dependency and instructions
tests/     numerical equivalence, observed-result, and determinism checks

Documented simplifications (v0)

• Gaussian likelihood on published BAO summary values (this mirrors the official compressed likelihood; full-shape comes later).
• (superseded — later scripts include radiation in E(z); early-universe module is a calibrated surrogate, see fit_w0wa.py header)
• SN target value (0.334 ± 0.018) taken from Brout et al. abstract; re-verify exact cut definitions against their §4 before treating small differences as meaningful.

Week 2–4 results: the w₀wₐ preference REPRODUCED (2026-07-19)

Compressed-CMB likelihood implemented from the DR2 paper's own eqs. 35–36 (Gaussian in θ*, ωb, ωbc; CamSpec-based). Validation gates: G1 — pivot-calibrated sound-horizon integral matches the paper's rd formula to 0.17% across a parameter grid; G2 — ΛCDM DESI+BBN+θ* reproduces the published Ωm=0.2967±0.0045, H0=68.45±0.47 (we get 0.2976, 68.37). Two bugs were caught by the gates before any dark-energy fit ran (an ωbc bookkeeping error and a recombination-calibration offset) — the stop/go protocol working as designed.

Best-fit parameters also land on the published values: DESI+SN w₀=−0.886 (pub −0.888), DESI+cCMB w₀=−0.44, Ωm=0.352, H0=63.6 (pub −0.42/−0.43, 0.353, 63.6). Run: python3 scripts/fit_w0wa.py data (~7 s).

What this means: the DR2 evolving-dark-energy preference is independently reproduced from primary-source inputs on a laptop — for the BAO, BAO+SN, and BAO+compressed-CMB legs. The paper's own table shows the climb to 3.1σ requires full CMB information (mostly lensing), and to 3.8–4.2σ requires Union3/DESY5 supernovae — neither yet in this pipeline.

Week 4–6 results: INFLUENCE DIAGNOSTICS (2026-07-19) — the novel analysis

Probe 1 — leave-one-tracer-out (DESI+cCMB, 2.4σ baseline):

Probe 2 — χ² improvement decomposition (w₀wₐ vs ΛCDM at respective best fits): LRG2 −4.1, CMB anchor −3.9, LRG1 −1.2, all others ≈ 0, Lyα +1.0 (fits worse under evolving dark energy).

Probe 3 — SN low-z cut sweep (DESI+Pantheon+): σ falls monotonically 1.8 → 1.6 → 1.4 → 1.3 as z_min goes 0.01 → 0.025 → 0.05 → 0.10, while w₀ stays pinned at ≈ −0.89 throughout. The crossing component (wₐ) is what the nearby supernovae supply.

Probe 1b — selection-calibrated maximum influence (completed 2026-07-26): define I_j = Δχ²_−j − Δχ²_full and select M = max_j I_j inside each dataset. Observed M = 4.1249 at LRG2. Repeating the full fit and all seven deletion fits inside each of 5,000 ΛCDM mocks gives 78/5,000 equal-or-larger selected maxima: p = 0.0156 [0.0124, 0.0194]. However, among the 70 null mocks whose global preference is at least as strong as observed, 34 have an equal-or-larger maximum: p_cond = 0.486 [0.364, 0.608].

Reading (selection-calibrated): the observed fit is genuinely sensitive to LRG2, so “load-bearing” remains an accurate descriptive label. But this concentration is not an additional anomaly once one conditions on a global w₀wₐ fluctuation of the observed strength. The localization map directs follow-up; it must not be counted as independent evidence against ΛCDM.

Week 6 results: DES5Y + THE EFSTATHIOU TEST (2026-07-19)

Data: official DES-SN5YR repo, 2025 Dovekie-recalibrated release (N=1820; inverse covariance unpacked per the repo's own likelihood code; external low-z sample identified by IDSURVEY≠10, N=197, z≤0.087).

Published no-CMB row (2024 release): Δχ² = −13.6, 3.3σ. Two findings:

F1 — The Dovekie recalibration itself appears to have weakened the DES5Y evidence (−13.6 → −7.3 in this pipeline; machinery validated on Pantheon+ at −5.1 vs published −4.9). Part of the gap could be pipeline simplifications, but the SN-leg machinery is the same one that reproduced Pantheon+.

F2 — The intercept test reproduces Efstathiou's critique on the newer data: freeing ONE magnitude offset on the 197 external low-z SNe returns a fitted D ≈ +0.037 mag under ΛCDM — nearly identical to the ~0.04 mag Efstathiou inferred — and the evolving-dark-energy preference collapses to 0.2σ with w₀ = −0.99: ΛCDM restored to within noise. One calibration nuisance parameter explains the DES5Y contribution as effectively as two dark-energy parameters.

Counter-arguments, documented: the DES team maintains cross-survey calibration systematics are already in the covariance (freeing D may double-count); the offset test is a post-hoc choice motivated by a critic; and this combo omits the CMB leg (published 4.2σ row includes full CMB).

Week 7 results: CMB-ANCHORED INTERCEPT + PARAMETERIZATION SWAP (2026-07-19)

Anchored intercept test (DESI + cCMB + DES5Y): baseline Δχ² = −10.9 (2.9σ — our compressed-CMB analog of the published 4.2σ row). Adding the single low-z intercept: Δχ² = −3.2 (1.3σ), fitted D = +0.0355 mag. Even with the CMB anchor, the calibration nuisance strips out the supernova contribution; the surviving ~1.3σ is the BAO+CMB geometry piece (which our LOO showed is itself concentrated in LRG2).

w(z) parameterization swap (all families reduce to Λ at w₀=−1, wₐ=0):

Two conclusions cut opposite ways. For the DESI claim: the preference is not a CPL artifact — all two-parameter evolving families find it at nearly equal strength. Against over-interpretation: constant-w captures almost none of it — the evidence lives entirely in the second degree of freedom (time variation/crossing), best-fit parameters are family-unstable (JBP wants w₀=+0.09, wₐ=−5.4), and that crossing is precisely the component our diagnostics localized to LRG2 and the low-z supernova calibration seam.

Consolidated lab verdict (honest form): a real, parameterization-robust ~2–3σ preference for late-time evolution exists in the anchored data; its observational support is concentrated in one BAO tracer and the low-z SN calibration; one ~0.036 mag calibration offset — matching both Efstathiou's inference and the direction of the Dovekie recalibration — reduces it to noise-compatible levels. Probability mass should sit between "small real dynamical effect" and "two stacked systematics," with the 2027 DESI final likelihood and Rubin-calibrated SNe as the decider.

Next

Public write-up (notebook + report), per the publish-regardless rule.

FINAL SECTION — External audits, bootstrap, and consolidated verdict (2026-07-19)

AI-assisted software review. Four AI systems reviewed software, statistics, or manuscript text over the life of the project; these are not independent scientific replication. One re-executed the pipeline and confirmed every χ² to 4 decimals, checked minima against differential evolution and varied starts, confirmed distance-grid accuracy at 3×10⁻⁷, and verified covariance positive-definiteness. Its corrections are adopted: the early-universe module is labeled a calibrated surrogate (rd correction 2.32%, θ* anchor 0.125%; G2 partially circular), the "independent reproduction" claim is narrowed to "independent compressed-likelihood implementation," and DESI's early-matter condition (w₀+wₐ<0; satisfied by all reported best fits) is documented.

Anchor-uncertainty propagation: varying the surrogate's θ* anchor by ±1σ of the published anchor uncertainty (H₀ = 67.14 ± 0.47) moves the baseline between Δχ² = −6.1 and −11.1 → the surrogate pins significance to 2.4σ ± ~0.5σ. Exact CLASS/CAMB required to do better (camb_check.py ready; blocked in this sandbox by CPU instruction set — run locally).

Historical 75-mock checkpoint (superseded): this early consistency check was too small to calibrate a percent-level tail. The completed unrounded 5,000-mock result appears below.

Cross-AI extensions adopted into the record: LRG2's leverage is primarily its radial D_H/r_d measurement (Δχ² −6.5 with radial only vs −4.9 transverse only vs −4.3 neither); LRG2 sits ~2.36σ from the joint ΛCDM prediction (unusual, not outrageous); shifting LRG2 halfway to ΛCDM (a 1.18σ move in its own covariance) drops the preference below 2σ; error-inflation degrades the signal smoothly (not a numerical pathology).

Literature balance (added per audit): DES's rebuttal (Vincenzi et al., arXiv:2501.06664) reproduces the Efstathiou offset but attributes much of it to improved intrinsic-scatter/host modeling — reverting assumptions lowers their result 3.9σ→3.3σ rather than eliminating it. Cortês & Liddle (arXiv:2504.15336): overlapping SN significances must not be averaged; DESI+CMB 3.1σ is the safest single summary. Afroz & Mukherjee (PRD, arXiv:2504.16868): a redshift-dependent Pantheon+–DESI inconsistency; ΛCDM-compatible once modeled.

Consolidated verdict at that checkpoint: a real, reproducible, parameterization-robust likelihood-level tension exists (~2.4σ compressed / 3.1σ full-CMB per DESI); it is localized (LRG2 radial BAO + CMB anchor + low-z SN calibration); the Efstathiou intercept (D≈+0.036 mag) neutralizes the SN contribution even CMB-anchored, while DES's rebuttal keeps that contested; and constant-w captures little of it. The honest status is a reproducible, localized 2–3σ hint — not proof. No numerical probability of “new physics” is assigned because this analysis does not derive one.

Remaining evidentiary gates (current): independent human clean-room rerun → official LRG2 likelihood-surface audit → official Union3/DES5Y likelihoods → growth/lensing consistency (RSD, fσ8, E_G) → DESI final likelihood and independently calibrated supernovae.

GATE 1 RESULT — THE LRG2 AUDIT, public-data version (2026-07-19)

The z≈0.7 BAO measurement across three analyses, tested in the DESI+cCMB frame (all values primary-sourced: eBOSS from the official SDSS DR16 consensus files in the cobaya bao_data release; DR1 from arXiv:2404.03002; DR2 from arXiv:2503.14738):

T1 — distance from the ΛCDM best-fit prediction:

The z≈0.7 deviation has changed observable between DESI's own releases — DR1's pull was transverse (D_M low), DR2's is radial (D_H low) — and the independent eBOSS measurement shows neither (slightly ΛCDM-high in D_M, if anything).

T2/T3 — swap test (DESI+cCMB, the z≈0.7 slot filled by each):

Finding: substituting the independent eBOSS measurement yields exactly the same preference as having no z≈0.7 measurement at all — eBOSS's deviation pattern has essentially zero projection onto the w₀wₐ direction. The entire lift from 1.6σ to 2.4σ is supplied by DR2's LRG2 specifically.

Balanced reading: DR2's LRG2 is by far the most precise of the three (±0.177 vs ±0.30/±0.32 on D_M) — more data can legitimately resolve what noisier data could not, DESI's blinded internal-consistency checks passed, and eBOSS's weaker precision means it could not strongly confirm the signal even if real. But the audit establishes two facts the interpretation must carry: the z≈0.7 signature changed observable across DESI releases (transverse → radial), and the earlier, partially sample- and footprint-correlated eBOSS comparator is ΛCDM-consistent. The official-likelihood-level LRG2 reproduction remains the decisive test.

GATE 2 RESULT — EXACT BOLTZMANN PHYSICS (CAMB, run on Bobby's Mac, 2026-07-20)

The calibrated surrogate's two empirical constants — the audits' primary criticism — are now replaced with CAMB (full RECFAST recombination, exact r_drag/θ*, w0wa-PPF background), executed on local hardware (homebrew Python 3.13 venv; sandbox CPU could not run CAMB's binary).

Best-fit parameters agree to the third decimal (exact: w₀=−0.447, wₐ=−1.64, Ωm=0.351, H0=63.7; surrogate: −0.44, −1.68, 0.352, 63.6). CAMB-vs-surrogate at the best fits: rd differs by 0.12%, θ* by 0.035% — and the fits absorb these into slightly shifted (h, ωb), leaving Δχ² essentially unchanged.

Conclusions. (1) The calibrated surrogate was accurate: every headline number of this lab survives exact early-universe physics to within ~0.2 in Δχ² (~0.03σ). (2) The LRG2 localization is confirmed at the Boltzmann level: removing LRG2 still collapses the exact-physics preference from 2.4σ to 1.5σ. (3) The audits' surrogate objection is resolved empirically, not rhetorically. Remaining ladder: posterior sampling with DESI priors (+DIC), full-CMB lensing leg, official likelihoods — then the write-up.

GATE 3 RESULT — POSTERIOR INFERENCE + DIC (emcee on Bobby's Mac, 2026-07-20)

Full MCMC under DESI's official priors (w₀∈U[−3,1], wₐ∈U[−3,2], w₀+wₐ<0, verbatim from the paper), 80k samples, CAMB-validated surrogate likelihood.

DIC convention note: naive plug-in at the posterior mean gives ΔDIC = −8.99 — an artifact of the strongly curved w₀–wₐ ridge (mean sits off-ridge). With the standard MAP plug-in, effective parameter counts come out exactly right (4.75 for 5 params, 3.0 for 3) and ΔDIC = −4.96, matching the published −4.4 to the same offset as our Δχ². The audit's "best fits, not posteriors" objection is now closed: intervals, contours (chain released), correlation structure, and information criteria all reproduce.

Ladder status: reproduction ✓ · diagnostics ✓ · calibration tests ✓ · parameterization ✓ · bootstrap ✓ · LRG2 audit ✓ · exact physics ✓ · posteriors+DIC ✓. Remaining beyond current resources: full-CMB lensing leg and official uncompressed likelihoods. → REPORT-draft.md is the write-up.

AUDIT 4 ADOPTED (2026-07-20) — eight corrections, all accepted

A fourth external review (13-page document) was received and adopted in full. At that historical checkpoint it required: (1) eBOSS reframed as a partially correlated comparator, not an independent non-confirmation; (2) the 75-mock bootstrap relabeled as a consistency check and a 5,000-mock run queued (now completed); (3) the two LRG2 sigmas explicitly distinguished; (4) "exact CAMB"→"full CAMB Boltzmann," "independent hardware"→"separate execution environment," and AI reviews relabeled as software/manuscript checks; (5) Dovekie framed as a multi-element release change; (6) the low-z intercept labeled a sensitivity test; (7) full-CMB claims held until the lensing-leg run completed; and (8) a frozen, reproducible release. The bottom-line framing remains: lead with the influence map, not with "evolving dark energy."

EXPEDITION 2 RESULT — THE FULL-CMB LENSING LEG (2026-07-20)

Cobaya + CAMB on local hardware: Planck 2018 native likelihoods (low-ℓ TT, low-ℓ EE, plik-lite TTTEEE, lensing) + DESI DR2 BAO, MAP minimization (BOBYQA, dual runs converged), DESI official w(z) priors. Comparison caveat pre-registered: DESI DR2 used PR4 CamSpec + PR4/ACT lensing; this uses the 2018 (PR3) native equivalents — expect the same qualitative lensing lift, not identical numbers. Prior-volume constants (2·log 20) removed to convert posterior minima to likelihood Δχ².

Published full-CMB row (PR4): −12.5, 3.1σ — our −10.4 / 2.8σ with the PR3-native stack is in the expected range given the likelihood-generation difference.

THE ANSWER TO THE PROJECT'S CENTRAL OPEN QUESTION: the CMB-lensing information does not rescue the preference from LRG2 removal. With full CMB including lensing, dropping the single z=0.706 measurement collapses the evolving-dark-energy preference from ~2.8σ to ~1.4σ — the sharpest form of the localization result, now demonstrated at the full-likelihood level rather than the compressed level. Caveats: PR3-native vs PR4 likelihoods; MAP (not posterior) comparison; single-pipeline result awaiting the clean-room rerun; LRG2 removal carries the usual look-elsewhere qualifier.

H-X CROSS-PROBE TEST RESULT + 5000-MOCK CALIBRATION (2026-07-20)

Mock calibration (audit item 2 CLOSED; reconciled 2026-07-26): the fresh unrounded least-squares runner gives 70/5,000 at Δχ² ≤ −8.4575, p = 0.0140 [0.0109, 0.0177], vs Wilks 0.0146. This supersedes the earlier 1.50% number taken from the rounded legacy JSON; the scientific conclusion is unchanged. The same deterministic run performs all seven deletion fits per mock: selected maximum-influence p = 0.0156 [0.0124, 0.0194], but p = 0.486 [0.364, 0.608] conditional on global preference at least as strong as observed. Localization is not independent evidence beyond the global fluctuation.

H-X (pre-registered cross-probe w(z) consistency): verdict INTERMEDIATE, per kill conditions. BAO–Pantheon+ shift 1.44σ, BAO–DES5Y 1.35σ, Pantheon+–DES5Y 0.24σ. Crossing redshifts: BAO z× = 0.49 ± 0.07; both SN legs z× ≈ 0.29–0.30. Neither the consistency nor the inconsistency threshold was met; reported without promotion. Notable texture (not promoted): the two supernova datasets agree with each other almost exactly and both point at a different turning point than the BAO leg, despite a shared CMB anchor pulling all legs together. Registered follow-up: P1-without-LRG2 z× test. Full record: HYPOTHESIS-cross-probe.md, results/hx_tension.json, results/probe_posteriors.json.

H-X registered follow-up (2026-07-20): P1 without LRG2 gives z× = 0.478 [0.384–0.564] — the BAO leg's crossing-location preference does NOT dissolve when LRG2 is removed (registered condition not met). LRG2 carries the significance of the BAO preference, but the crossing location is a distributed property of BAO+CMB geometry. Reported per registration; overall H-X verdict remains INTERMEDIATE.

H-Y RESULT (2026-07-20): registered kill condition MET. DES5Y+cCMB with the low-z intercept marginalized: wₐ = +0.50 [−0.33, +1.03] (includes 0, sign flips), w₀ → −1.17 (straddles Λ). The SN crossing feature is fully degenerate with the calibration seam. Registered caveat carried: degeneracy established, cause not — the intercept can absorb real cosmology. Chain complete: HYPOTHESIS-cross-probe.md holds the full registered record.