Independent re-computation of published quantitative claims, from raw or released data. Where I could, the test was fixed in writing before it ran; not every entry had that, and the entries make no stronger claim than their own test. When a claim holds, that is recorded with the same care as when it does not: an auditor who only ever says no isn't checking anything. Author and institution responses are described only with permission.
The work is done with an AI model doing much of the coding, computation and drafting. I choose what to test, check every number on this page against its evidence file, and answer for all of it; own mistakes are corrected in public.
Rebuilt the full search from public GBM photon data; the three peaks reproduce. Against a null that keeps each burst's own light-curve shape (instead of a constant rate), the same search matches or beats the observed statistic in 89%, 40% and 4.9% of simulations.
not significant under a burst-shaped nullCorrespondence with the authors open; methods paper in preparation (A&A).
The published bootstrap tests whether burst times are random — not whether they are periodic. With the burst storms kept intact, the period can be neither confirmed nor refuted from three cycles: a power test with a truly periodic source shows the folding statistic detects it in only 5–7% of trials.
not decidable from these dataIn correspondence with the author; an out-of-sample test as monitoring grows is the test that matters.
The 42 points contain the same galaxy samples two to four times, and the Eisenstein–Hu sound horizon is converted for some points but not others. With one measurement per sample and a free H0·r_d, ΛCDM fits the paper's own data better than the void profile.
independence assumption failsThe authors' own follow-up paper (Universe 2026) independently reached the free-H0·r_d point. In correspondence with the authors.
The supernova points are calibrated on SH0ES and the BAO points on the Planck sound horizon, so the "trend" restates the H0 tension between the two anchors. With calibrations included as uncertainties: 2.3σ; with them free: 0σ.
trend = tension, restatedManuscript submitted to A&A (Sep 2026); code and data: doi:10.5281/zenodo.23002508.
The signal reproduces only with one shape parameter frozen; profiled over it the significance is 2.6σ / 2.4σ, and the binning choice carries a global p of 0.20. (An earlier note here, that newly catalogued objects go against the warp, was withdrawn: their orbits are too short to apply the paper's resonance filter.)
below discovery-grade once profiledShared with the lead author; a research note is planned.
With the flat WW* background treated consistently and the search's own look-elsewhere effect (~13 effective trials), the combination drops to roughly 4.4σ, and one channel's status changes between the 2023 and 2025 datasets.
overstated; parked until journal publicationThree duplicate galaxies and a σ-convention reduce it to 3.05σ; parity and spatial checks pass. Within its own framing the claim stands — no comment was warranted.
held (3.4σ → 3.05σ)Autocorrelation-corrected: N_eff ≈ 15, p = 0.002–0.005. The correlation survives the same test that kills most smoothed-series claims.
heldRe-measured with the paper's own sample size: 1.45σ.
consistent with noiseIndependent re-analysis of a published cluster line candidate; the significance depends on one cluster (Ophiuchus) and a background model choice (ρ², 3.7σ).
under review (minor revision), Physics of the Dark UniverseBoth series are 11-year running means; the effective sample size is 6–8, and phase-randomised surrogates give p = 0.33 and 0.42. The unsmoothed annual correlations are ±0.3 and not significant. Other results in the paper held in the same tests: the change in the ENSO and PDO relationships around 1950 (surrogate p = 0.002 and 0.01) and excess multidecadal power over red noise (p = 0.02).
smoothing artefactShared with the authors; no comment will be submitted.
Autocorrelation-corrected: N_eff ≈ 80, p ≈ 0.003. It holds; no letter was needed.
heldThe main correlation reproduces. The "forecast error" signal, however, equals the forecast's own +2σ cloud excess before the event — it is not additional evidence.
partly reproduces; letter with the authorsThe protein-affinity result is genuine (p = 0.005 with leakage rules enforced). The peptide ordering is not (p = 0.56).
split: half held, half didn'tLetter with the corresponding author.
Finding sent to Kela's statistics unit. Kela reviewed it, had no objections, suggested one methodological merge, and gave permission to cite their response. Submitted to a Finnish economics journal (Sep 2026).
acknowledged by the institution99.3% of the index's impossible years are Buddhist-era dates: the offset is exactly +543 in 600 of 600 sampled records. Reported to the registering agency.
systematic metadata defect, root cause identifiedNot an error: the rate is an artefact of the export level. The data itself is sound — but totals across cities are not comparable. The ledger cuts both ways: this one defends the data.
data defended…with sample size 0 and no confidence intervals anywhere in the document.
performance claims without measurementsEight of eight submitted comments accepted so far (four more under moderation), including independence requirements and LLM-judge validation. Separately: the 10²⁵ FLOP threshold lacks a tolerance statement — noted to the AI Office (doi:10.5281/zenodo.22966931).
accepted into the draftThe ministry opened its draft tender for an evaluation contract to comment. Of the ten points raised, nine changed in that direction in the published version and one was clarified. The data section now requires bidders to state how sources are chosen, how they are made comparable, and how uncertain they are. One change traces to these comments directly: the added sentence that lets a new firm without a confirmed financial year bid is the wording proposed. Others commented too, so the other changes are consistent with the comments, not proven to follow from them.
adopted into the tenderEleven audits of AI benchmarks and leaderboards from released data — ten overstated, one certified as holding. Now with a deposited rank-stability study (doi:10.5281/zenodo.22541065).
see the case files →Kaggle scores every competition on two disjoint test slices, so rank noise is observed, not modelled. The August measurement over 19 competitions replicated on an independently built dataset of 1,157 finished boards, method file unchanged and pre-registered: pairs closer than one local movement keep their order in 55.1% of 2.2M pairs, pairs further apart in 90.8% of 7.3M; a shuffled control stays at 50%. 58 boards invert and are kept in every figure.
replicated, 19 → 653 boardsReport, pre-registration, code and per-board results: doi:10.5281/zenodo.23017379 · data: Kaggle Shakeup Atlas (G. Mamarin).
Deposited (doi:10.5281/zenodo.22825533, code 10.5281/zenodo.22834627); under review at JDMDH. A companion measurement: the same proto-Elamite corpus yields 1,738 / 680 / 468 signs under three published counting conventions — sign-count disagreements are notation, not discovery.
under reviewMethod paper deposited (doi:10.5281/zenodo.22162050). The periodicity work above runs on an open null-model toolkit (burst-shaped nulls, selection-aware calibration, trials-factor verification) that ships with each paper.
publishedEvery entry above was computed from public or released data, with the test pre-registered in a version-controlled file before it ran. Where an author or institution is quoted, it is with permission. If you rely on a number — a benchmark, a procurement criterion, a published effect — and want it re-measured the same way: ilpo@ipezygj.com.
How to verify, mechanically. The pre-registration claim is checkable, not a promise: deposits are moving to include a PREREG_PROOF.txt — the git first-add commit and UTC timestamp of the pre-registration and of every result file, generated by an open script, so a reader can confirm the test was locked before any result existed (first deposit with one: doi:10.5281/zenodo.23017379, the Kaggle threshold above). The counts on this page are re-derived from the entries by a checker that runs on every update.