The short version

Three hypotheses from the 31 October 2026 sets were selected because each made a quantitative prediction that could be checked against public data without a new instrument. Each test was pre-registered before the data were opened. The headline is not that the hypotheses were right. It is that two of the three mechanisms left no measurable signature at the stated scale, and the third survived only in a narrow, class-specific form.

The rest of this report gives the claim, the data, the method, the numbers, and the limits for each. Every figure is regenerated by the code in code/; every input is listed with a URL and a checksum in DATA.md.

Why these three

The 31 October sets contain 29, 40 and 99 hypotheses, most of them requiring a laboratory, a beamline, a cryostat or a spacecraft. Three were chosen against a simple filter: the claim had to name a measurable quantity, the data had to exist in a public archive, and the test had to be able to fail for a reason other than “the archive was empty.”

• H12 is a local, high-rate claim. A swarm of small earthquakes and a pressure record are enough to test the predicted semidiurnal pacing.
• H21 is a claim about a systematic in the supernova distance ladder. Pantheon+ is the largest public compilation with a published covariance matrix, so the test can be done without re-reducing light curves.
• S07 is a claim about a correlation between two large public archives. Gravity Spy classifies glitches; the IRIS/EarthScope archive has seismometers at both LIGO sites.

The three also span three failure modes: a null in a point-process test, a parameter-degeneracy problem in a cosmological fit, and a correlation that survives only in a subclass. That spread is deliberate.

1 · The San Ramon swarm is a barometric valve

Data

The USGS/ANSS FDSN catalogue was queried for the box 37.6–37.9°N, 122.1–121.8°W, 1 January 2018 to 31 October 2026, magnitude ≥ 0, event type earthquake. It returned 2,439 events; a conservative Gardner–Knopoff-style declustering (5 km/1 day, 10 km/2 days for M ≥ 3) left 1,120 background events. Hourly mean-sea-level pressure came from the KCCR ASOS station at Concord, about 15 km from the swarm, and was gap-filled only for runs of six hours or less.

Method

A Poisson harmonic regression was fitted to the hourly event count:

with terms at M2 (12.42 h), S2 (12.00 h), K1 (23.93 h), O1 (25.82 h), the 24 h solar day and the annual cycle. The same design was fitted to pressure by least squares. The pressure–rate cross-correlation was computed on 3 h and 24 h anomalies, with a phase-randomised null (300 surrogates) for the maximum correlation over lags −24…+24 h. The storm composite used 103 pressure falls greater than 8 hPa in 24 h, matched to 20 control windows per storm.

Result

The pressure tide at KCCR is dominated by the solar semidiurnal component: S2 amplitude 0.61 hPa, M2 only 0.02 hPa, the diurnal 0.80 hPa. The declustered seismicity rate shows no comparable semidiurnal term:

The pressure–rate cross-correlation peaks at r = 0.021 at a lag of +18 h; the phase-randomised probability of a maximum that large is p = 0.79. At zero lag, r = 0.007 (p = 0.44). The storm composite is slightly higher before the pressure minimum than after it: 0.106 versus 0.088 events per 6 h, a ratio of 1.20 with a 95% bootstrap interval of 0.77–1.76 and a random-start p = 0.21. That is the wrong direction for a valve that opens when pressure falls.

Limits

KCCR is 15 km away and sits in a different boundary-layer regime from the fault; a null here is a null for this pressure proxy. The catalogue is a regional network product with variable completeness, and declustering removes much of the swarm. A borehole pressure sensor in the fault zone, or pore-pressure data, would be a stronger test. Nothing in this result says that rainfall, creep or fluid migration is the cause; it says the barometric-valve version of the claim is not visible in the public record.

2 · The supernova phantom crossing is a zero-point drift

Data

Pantheon+ SH0ES: 1,701 supernovae, corrected m_b,corr, z_HD, and the full STAT+SYS covariance matrix (33 MB, 1,701 × 1,701). The diagonal errors in the public table are explicitly labelled “do not fit cosmological parameters with these”; the full covariance was used. The fit marginalises the absolute magnitude M analytically and fixes H₀=70 km s⁻¹ Mpc⁻¹.

Method

Flat ΛCDM, flat wCDM, and flat w₀wₐCDM were fitted by a global bounded search followed by a bounded polish. A redshift-dependent offset δ m=A(z-0.3) and a step at z=0.5 were added to the baseline models. The drift parameter was also estimated directly by generalised least squares with M marginalised, which is the cleanest single-number test of the 0.01 mag claim. Single-survey subsets were fitted separately as a check on the “single instrument shows no crossing” prediction.

Result

The constant-w fit gives w=-0.706±0.091: a mild preference for w > −1, not a crossing. The w₀wₐ fit runs to the wₐ=-3 prior bound and is not a physical fit; its formal crossing at z≃0.10 is a boundary artefact, not evidence. Adding the drift to w₀wₐ improves χ² by only 0.35 for one extra parameter.

The direct drift estimate is the decisive number:

At z=1 that is 0.0016±0.016 mag. A 0.01 mag drift is therefore allowed but not required; this sample cannot exclude it, and cannot detect it either. The 95% interval on A is roughly [−0.042, +0.047] mag per unit z.

Limits

Pantheon+ is a compilation, not the DESI DR2 supernova sample that motivated the claim. Its calibration is already partly harmonised, which works against the hypothesis. The single-survey subsets (SDSS 321, PS1 269, DES 203, FOUND 179, SNLS 160, LOSS2 105) are too small to constrain wₐ; their fits hit the prior bounds. A proper test needs the DESI SN sample and a joint calibration model, not a compilation. The full covariance was used, but the intrinsic scatter and selection model are the ones published with Pantheon+.

3 · Gravitational-wave glitches are ocean weather

Data

Gravity Spy machine-learning classifications for O3a: 63,917 H1 glitches and 90,738 L1 glitches after requiring ml_confidence ≥ 0.80 and excluding No_Glitch and None_of_the_Above. Co-located seismometers: US.HAWA LHZ (28 km from H1) and N4.545B LHZ (30 km from L1), 1 Hz, reduced to hourly RMS in the microseism band 0.05–0.30 Hz and a higher-frequency control band 0.30–0.45 Hz. The LIGO-site station LI.LTL had no LHZ/BHZ data for O3a, so the nearby N4 station was used; this is a proxy, not the site instrument.

Method

Glitches were counted per day and per hour. The primary test is daily, because the glitch catalogue does not include the detector duty cycle and daily aggregation is less sensitive to short lock losses. Correlations are reported raw and after subtracting a 15-day running median, which removes the seasonal trend and tests the short-term association. A Poisson GLM with log microseism, log high-frequency noise, a linear trend and weekday dummies was fitted to the daily counts. Class-resolved correlations use the Gravity Spy labels; the light-scattering family combines Scattered_Light and Fast_Scattering, the two labels used for the same physical family.

Result

The total glitch rate is not correlated with microseism:

The class-resolved picture is different. The light-scattering family — Scattered_Light plus Fast_Scattering — is positively correlated with microseism at both detectors:

This is a modest effect — ρ≈0.2, roughly 4% of the variance — but it is in the predicted direction, it is present at both sites, and it is absent or negative in most other classes. A cross-detector control weakens the word “local”: H1 light-scattering glitches correlate with the L1-region microseism at ρ = 0.13 (p = 0.08), and L1 glitches correlate with the H1-region microseism at ρ = 0.25 (p = 0.001). The L1 association is therefore not site-specific; the effect is at least regional, and this test cannot separate regional ocean forcing from local coupling. L1’s Whistle class is strongly anti-correlated (ρ=-0.46), and L1’s Tomte, Blip and Koi_Fish classes are negative, so the effect is not a detector-wide artefact.

Limits

The detector duty cycle is not in the glitch catalogue, so the daily rate is not exposure-corrected. The L1 seismic proxy is 30 km from the site, not at it. Daily aggregation cannot resolve the minutes-to-hours lag the hypothesis predicts; an hourly analysis is possible but is dominated by zero-inflation and lock-state changes. The classifier labels are not identical across detectors, and the light-scattering family is defined post hoc from the two labels, so the replication is suggestive rather than pre-registered. No veto efficiency was measured.

What the three tests show together

Two of the three mechanisms are not visible in the public data at the stated scale. The third is visible only in a subclass and only at ρ≈0.2. That is a useful outcome, but it is not the outcome the hypotheses wanted.

• H12 fails cleanly: the predicted semidiurnal pacing is absent, the pressure–rate correlation is zero, and the storm composite is in the wrong direction. The most likely reading is that the San Ramon swarm is paced by something other than the local barometric tide — rainfall, creep, or a fault-valve process with a different pressure coupling.
• H21 is not settled. Pantheon+ does not require the drift, but its sensitivity to a 0.01 mag drift is poor, and the single-survey test cannot discriminate. The honest verdict is “not demonstrated,” not “disproved.”
• S07 survives narrowly. The light-scattering family is the physically expected channel — scattered light is sensitive to test-mass and beam-tube motion — and it replicates at both detectors. The total rate does not, which is also expected if only a fraction of glitches are seismically coupled.

The common thread is that the hypotheses that named a specific channel (light scattering) did better than the ones that named a bulk rate (all glitches, all earthquakes). That is a useful design lesson for the next set: predict the subclass, not the total.

Reproduction

The environment is Python 3.14 with NumPy 2.5.4, SciPy 1.18.1, pandas 3.1.0 and Matplotlib 3.11.2. Each test is a standalone script:

python code/t1_san_ramon.py          # writes data/t1_*.csv, results/t1_results.json, figures/t1_*.png
python code/t2_pantheon_drift.py     # writes data/t2_*.csv, results/t2_results.json, figures/t2_*.png
python code/t3_fetch_seismic.py      # downloads and reduces the O3a seismic data
python code/t3_glitch_seismicity.py  # writes data/t3_*.csv, results/t3_results.json, figures/t3_*.png

The raw downloads are in data/ (see DATA.md for URLs and checksums). The original monthly seismic files are not kept; the hourly reduced series is. The raw Gravity Spy CSVs are 90 MB and 142 MB and are kept outside the deliverable, with checksums in DATA.md.

Pre-registration and changes

The pre-registration is in PRE-REGISTRATION.md, written before the data were opened. Three changes were made and are recorded here:

1. T1: the storm-composite permutation test was added after the first pass; it does not change the verdict.
2. T2: the full STAT+SYS covariance was used instead of the diagonal errors named in the first draft, because the Pantheon+ documentation explicitly warns against the diagonal. The diagonal fit was not used for any verdict.
3. T3: the light-scattering family was defined after the class-resolved table was seen. It is reported as a post-hoc family, not as a pre-registered class. The total-rate test remains the pre-registered primary test, and it is null.

Sources

The hypotheses are quoted from the 31 October 2026 sets: Forty Hypotheses from the Side Channels (H12, H21) and 99 Original Hypotheses (S07). The data sources are:

• USGS/ANSS FDSN event service, earthquake.usgs.gov/fdsnws/event/1/.
• Iowa Environmental Mesonet ASOS archive, mesonet.agron.iastate.edu.
• Pantheon+ SH0ES data release, github.com/PantheonPlusSH0ES/DataRelease.
• Gravity Spy machine-learning classifications, Zenodo record 5649212.
• IRIS/EarthScope FDSN web services, stations US.HAWA, N4.545B.

Full citations, URLs and checksums are in DATA.md.