Covariance audit of the live registry: of 56 live claims, exactly ONE (USR-2026-0003's local-structure half) uses diagonal errors on a dataset that ships a full covariance — and the registry had already flagged it. Re-running the flagship both ways quantifies the exposure: full covariance gives Δχ² = 6.973 (2.16σ), while dropping only the Pantheon+ off-diagonals inflates it to 17.934 (3.83σ) — a 1.77× significance inflation, the same mechanism that produced this registry's 6.8σ quarantine.

, VIPERS-v7(3), WiggleZ(3)
       rho    chi2_M1    chi2_M2    gamma_a    sigma8    sigma    min eig
      0.00    15.0211    13.9577    +0.6277    0.8322     1.03   4.84e-04
      0.20    16.0138    14.9689    +0.6279    0.8290     1.02   4.84e-04
      0.40    18.1189    17.1654    +0.5964    0.8216     0.98   3.92e-04
      0.60    22.6476    21.8869    +0.5163    0.8077     0.87   2.64e-04
      0.80    36.4375    36.0621    +0.3252    0.7780     0.61   1.32e-04
      0.90    63.6288    63.5466    +0.1276    0.7463     0.29   6.63e-05

  CURRENT catalogs.py (post z=1.48 correction)
    28 points; correlation groups with >1 member: 6dFGS(2), BOSS-DR12-RSD(3), DESI-LRG-RSD(2), GAMA(2), VIPERS-v7(3), WiggleZ(3)
       rho    chi2_M1    chi2_M2    gamma_a    sigma8    sigma    min eig
      0.00    19.5742    14.4433    +1.1854    0.9098     2.27   4.84e-04
      0.20    20.6101    15.6003    +1.1902    0.9072     2.24   4.84e-04
      0.40    22.8618    18.1209    +1.1634    0.9000     2.18   3.92e-04
      0.60    27.7193    23.5022    +1.0823    0.8848     2.05   2.64e-04
      0.80    42.3730    39.3701    +0.8486    0.8470     1.73   1.32e-04
      0.90    70.7866    69.1414    +0.5444    0.8003     1.28   6.63e-05

  CROSS-CHECK against the live registry.  USR-2026-0044 currently reads:
    correlated-systematics scan:
      assuming within-survey correlation ρ = 0 → 0.9 drives the
      significance 1.03σ → 0.29σ; no value of ρ rescues the claim
    data-fix robustness check (2026-07-26):
      the reproduction surfaced a wrong fσ₈ value in our own catalog (z
      = 1.48: 0.315 ± 0.095, a number absent from its own cited source;
      Hou et al. 2020 reports 0.464 ± 0.045). Corrected and re-fit: the
      evolution signal STRENGTHENS to 2.27σ but γ_a goes MORE positive
      (+0.628 → +1.185), so the kill survives and the corrected data
      argue harder against the claim's direction
  reproduced, PRE fsigma8-fix: 1.03 sigma at rho=0 -> 0.29 sigma at rho=0.9
  reproduced, CURRENT catalogs.py: 2.27 sigma at rho=0 -> 1.28 sigma at rho=0.9
  The registry's recorded 1.03 -> 0.29 scan was run on the PRE-fix
  catalog; the CURRENT catalog starts higher (the z = 1.48 fsigma8 value
  was corrected the same day) and still loses most of its significance
  to rho.  Both scans are printed above so neither is quoted out of its
  data version.  rho = 0 is an ASSUMPTION in every row of this table.

  WHO INHERITS THE UNQUANTIFIED FSIGMA8 CORRELATION?
    USR-2026-0044 [killed      ] The growth index γ evolves from 0.41 (z < 0.
    USR-2026-0069 [provisional ] Covariance audit of the live registry: of 56
    USR-2026-0042 / USR-2026-0043 [quarantined]: the 'multi-probe
      compilation' behind the ghosts includes fsigma8 (the surviving
      sibling run_gen237_bootstrap.py sums (fs8-model)^2/err^2 in BOTH
      its likelihoods) -- so the ghosts inherit it too, on top of their
      Pantheon+ diagonal shortcut.
    USR-2026-0045 [not_rerunnable]: gen1-31 era, cannot be checked.
    NO LIVE (provisional) CLAIM quotes an fsigma8-derived number.
    The one that did -- USR-2026-0044 -- is already KILLED.  That is
    the cleanest part of this audit: the dataset with NO covariance at
    all has no surviving live claim resting on it.

==============================================================================
VERDICT -- how much of the live registry rests on diagonal errors?
------------------------------------------------------------------------------
  62 claims are LIVE (provisional or contested).

  RANKED BY EXPOSURE:
   1. USR-2026-0049 (SPARC a0)  -- UNQUANTIFIABLE, most fragile.
      sigma_median = sd/sqrt(34) assumes 34 galaxies are independent
      draws.  No covariance exists.  An assumed rho = 0.05 shared
      systematic takes the canonical a0 from +2.6 to +1.6 sigma; a
      coherent 5% distance-scale error alone moves it 0.47 sem.  The
      claim is not refuted, it is underdetermined.
   2. USR-2026-0003 (PULL 2, local-structure test) -- the only LIVE
      claim using diagonal errors on a dataset that SHIPS a covariance.
      Re-run here both ways.  Shell test: 'all shells equal' chi^2 18.95
      -> 7.53 (full covariance makes it MORE uniform).  Dipole: 0.92 ->
      1.88 sigma (full covariance makes it LESS uniform).  Both stay
      below 2 sigma; the claim's conclusion survives, its error bars
      were never right, and the registry already flagged it.
   3. USR-2026-0048 (GD-1) -- Poisson counting errors, no covariance
      exists for the estimator.  UNQUANTIFIABLE, low stakes (p = 0.587).
   4. Everything else in the expansion / black-hole seams (USR-2026-0001,
      0002, 0004, 0005, 0008, 0009, and 0064 which inherits 0001) uses
      the FULL Pantheon+ and DESI DR2 covariances, verified by reading
      and by reproducing 0001's Delta chi^2 = 6.973 -> 2.16 sigma exactly.
   5. The CMB-map claims (boundary_seam, tilt_seam, lattice_cosmology)
      and the information_seam analytics take their significance from
      simulation nulls or closed form.  Not exposed.

  THE AUDIT IS SUBSTANTIALLY CLEAN -- and that is a real outcome.  ONE
  live claim uses diagonal errors where a full covariance is on disk,
  and the registry had already flagged it before this audit ran.  The
  quarantine of 2026-07-09 did its job.

  WHAT THIS AUDIT FOUND THAT WAS NOT ALREADY KNOWN:
   (i)  The project's own covariance re-run script,
        audit_bootstrap_cov.py, is BROKEN: it compares the STAT+SYS
        covariance against the MU_SH0ES_ERR_DIAG column, two different
        error definitions differing by a factor 1.40, and prints a
        'deflation' to 9.2 sigma while its prose says 4.  Anyone
        checking the quarantine with it would be misled.
   (ii) The exposure of the SN diagonal shortcut in this registry's
        FLAGSHIP direction is now measured, not assumed: 2.16 -> 3.83
        sigma, i.e. the shortcut would have inflated the live headline
        by 1.8x in sigma.  That is the quarantine's mechanism,
        reproduced on live code.
   (iii) The sign of the diagonal bias is NOT universal.  On redshift-
        shell contrasts, ignoring the Pantheon+ off-diagonals FABRICATES
        structure (chi^2 x2.52 too large); on a sky dipole it HIDES
        structure.  'Diagonal is conservative' is false in both
        directions and should not be used as a defence anywhere.
   (iv) The err_DIAG columns Pantheon+ ships are not the diagonal of
        the covariance it ships, and the release says in capitals not
        to fit with them.  Any project comparison of 'diagonal vs full'
        must say WHICH diagonal.  This one does.

  NOT A PHYSICS RESULT.  Nothing here measures the universe.  GLS is
  Aitken (1935); every failure mode named is textbook; all of it is
  REPRODUCED, NOT DISCOVERED.  No claim's status is changed by this
  script -- it reports, the founder decides.
==============================================================================

This script needs data files, a heavy dependency,
or more time than a browser tab should take.
Download it from /data and run it locally.