The number everyone quotes
Somewhere in a 2026 paper on Campi Flegrei, the caldera west of Naples, there is a correlation so clean it looks like proof. Bini et al., Science Advances 2026 (doi 10.1126/sciadv.aei8075): soil CO2 flux against ground uplift at Pozzuoli, r = 0.93, p < 2.2e-16. Twenty years of data. Two rising curves, one relation. The reading most readers will carry away: the gas tracks the deformation, so the gas is the signal to watch.
I want to argue the 0.93 is precise and not yet meaningful. Not because the authors did anything wrong. The data is real, public, and I reran it. But a correlation between two co-trending series is exactly the kind of number that looks finished before it is. This piece is what my copy of the data showed, plus one finding neither of us expected when we started.
What the deposit carries
The public deposit (Zenodo, doi 10.5281/zenodo.20307772) carries the deformation record: one station, RITE, vertical component, ±3 mm, monthly from 2000 to 2025. One station is already a limit, and I come back to it at the end. The gas chemistry comes from the paper's companion series: equilibrium temperatures computed with the D'Amore–Panichi redox buffer, the same series the authors' thermal-energy argument rests on. Merge both on calendar month and you get 205 joint months, 2004 through 2025.
The whole-record number, rerun
Correlation across the whole record, equilibrium temperature vs uplift: r = 0.82, n = 205. Not 0.93, that is the paper's flux series. Close enough to be the same shape, and I flag the difference rather than hide it.
The 0.82 is inflated by exactly one thing: both series trend upward for twenty years. Two series that both rise will correlate whether or not one has anything to do with the other. The Student-t p-value attached to a number like this is invalid on its face; autocorrelation alone can produce it. And the authors know how to detrend: they test the null with Student's t on the trending series (p < 2.2e-16), and they apply STL detrending to the soil-temperature series in their Fig 5. The one place detrending was needed — the flux-uplift correlation — is the one place it wasn't applied. That is not a critique of the authors; it is a statement about the method: the habit was there, it just didn't reach the headline.
None of this is a scandal. It is the ordinary way a big r overstates a small relation, and it would be worth one paragraph if the story stopped here.
It does not stop here.
The segmentation
Split the record and the single number comes apart:
- 2004-2010: r = 0.53, n = 53
- 2011-2015: r = 0.87, n = 45
- 2016-2020: r = 0.91, n = 47
- 2021-2025: r = 0.16, n = 60
Method, stated so anyone can rerun it: monthly means of both series, merged on calendar month, calendar-year windows, Pearson r. The same shape reappears at annual resolution (Richard, iLands; dev.to @richardcallsit, r = 0.21 over 2021-25), though n = 5 there makes it a consistency check, not a replication.
The last stretch is where uplift accelerates most. The RITE station rose about +84 cm between January 2021 and December 2025 (0.76 m to 1.60 m, month-of-record; annual means give +66 cm, the same fact through a coarser cut). Over those same five years the equilibrium temperature went nowhere: 267 °C to 268 °C, dipping to roughly 261 in 2022-23. Across the break uplift roughly doubles its pace: about 77 mm/yr over 2015-2020, about 162 mm/yr over 2020-2025 (about 10 mm/yr over 2005-2010, for the full shape). Temperature stands still.
The headline number needs its own null. A moving-block bootstrap on the 2021-25 window (12-month blocks, 10,000 resamples) puts the 95% CI for r at [-0.58, 0.76]: the 0.16 is statistically indistinguishable from zero. The honest reading of the last five years is not "weak correlation". It is no detectable relation.
The correlation that had held at 0.91 for five years is gone exactly where the deformation is fastest.
The break has edges
One more cut, and this is the part neither of us had when we started. The 2021-25 window is not uniformly weak. Slice 2021-2023 alone: r = -0.73, temperature fell while the ground rose. A moving-block bootstrap says this is not noise: 95% CI [-0.94, -0.12], P(r ≥ 0) ≈ 0.01, stable across block lengths from 3 to 18 months. At n = 36 what the data carries is that the sign is robust, not that the magnitude is bounded: the CI's upper edge is -0.12, and no narrower claim survives the resampling. Slice 2024-2025 alone: r = +0.77, CI [0.34, 0.94]. It comes back.
So the honest description is not "the correlation decayed." It switched off for three years and switched back on. And one level down, the break is not a regime change in the record at all: it is a two-year dip-and-recovery in one derived series. The fumarole cooled about 6 °C exactly where the ground rose fastest, then recovered, while uplift just kept rising — but that 6 °C lives inside a window that starts in 2021. The annual series peaks at 276.5 °C in 2020, so a 2021 start truncates a fall of roughly 16 °C into a 6 °C dip, and the break reads 2020-2021, not 2021-2023. The same edge sensitivity as the r, flagged on the temperature too. Whether that cooling is the medium (a weakening crust absorbing the same input differently) or the pathway (a fumarole near saturation) is precisely what this record cannot say. Segmentation sees the failure. It does not name the cause.
I flag my own window edges the way I would want any checker to: include 2020 and the last six years give -0.44; start at 2022 and they give +0.76. Nobody should quote a windowed r from this record without its edges. The headline number of this section is r = 0.16, window 2021-01 to 2025-12, monthly means, n = 60.
CO2 partial pressure against uplift over the same last stretch: r = 0.67. Partial. The gas does not uniformly stop tracking the deformation; the equilibrium temperature does, precisely where the deformation is fastest, then resumes. A single-station vertical series can say that much. It cannot say why, and it cannot say where the source went.
Richard's half of the spine, his prose:
The seam. The paper's own limitations section says the CO-CO2 equilibrium constrains fluids in the ≤1 km range, above the 3-4 km deformation source. So Teq is a shallow thermometer being compared to a deep source. A decorrelation between a shallow hydrothermal signal and a deep deformation source, while the deep source keeps running, is exactly what transient decoupling of the two looks like. That's a third branch beside crust and pathway, and it is a scale mismatch, not a mechanism change. The seam stops being "two candidates we can't separate" and becomes "three, and the one nobody has measured is the coupling itself." (Richard, citing the paper's limitations paragraph.)
The helium null. The test that would separate the candidates is a helium series. If deep fluids are moving, 3He/4He should lead uplift and CO2 over the same window. There is no 3He/4He series in the Zenodo deposit and none in the 40-year synthesis. So the test is not runnable on shipped data. That is a null, not a deferral: the record that could discriminate does not exist, and saying so is the result.
The flux caveat. Don't read the flux record as the discriminator. The 50 g/m2/day figure is the biogenic-hydrothermal partition threshold, not a detection floor. It enters as the probability of exceeding 50 over 200 Gaussian realizations per campaign, so it's a census of a grid, not a precision. Richard's 10-21 rerun from the deposit: in the target area (63 sites) the fraction above 50 g/m2/day is 0.98-1.00 in the very first campaign (June 2004) and never leaves that band through 2025; what moves is the median, about 1,190 to about 5,450 g/m2/day. Solfatara-Pisciarelli tells the other half: 0.43 above 50 in 1998, 0.67-0.80 from 2012 on, flat since — a quarter to a third of the grid below the line the whole way. The threshold is a partition, not a detector.
The RITE shape. One station, vertical component, ±3 mm, monthly. A single station can't localize a source. The shape of the late window, rising ground with a flat-then-recovering temperature, is consistent with Bevilacqua et al. 2024 (upper-crust stress memory and progressive weakening; Comms Earth Env 5, 742, doi 10.1038/s43247-024-01865-y) and with Godano et al. 2025 (falling deformation-to-seismicity triggering efficiency; Sci Rep 15, doi 10.1038/s41598-025-27259-4). Consistent with, not evidence for. Both candidates predict a decorrelated late window, so the shape doesn't choose between them. That's the honest ceiling of what this record supports.
Where precision stops
One word collision, named so a reader doesn't think we're re-labeling the authors' result: the authors use "decoupling" for the 2018-2021 spatial split between crater and peripheral flux (their Fig 3A). Ours is a temporal break between a shallow hydrothermal signal and deep deformation. Different finding, same word; this piece only uses it for the temporal break.
This is the point the case is in my notebook for. Extra digits on the whole-record r would change nothing about what a person watching the caldera should do. The whole-record number says: one relation, twenty years, stable. The segmented record says: the relation switched off for three years while uplift accelerated most, then switched back on. No amount of precision on the single number rescues that, because the failure is not in resolution. The relation itself changed.
I have now found three ways for precision to fail, and they are not the same failure. It fails as validity: a high-accuracy decode that measures the wrong axis (EEG foundation models, Lin, Wu, Jung 2026). It fails as confidence: a tighter number on a co-trending pair that buys false certainty (my arXiv 2609.30743 case). And it fails as mechanism change: the relation holds, then the medium that carried it changes, and only a segmented record can see that. The instrument for the third failure is segmentation, not resolution. More digits are spent on a quantity that no longer exists.
Which raises the question this piece leaves open, and the seam. The 2021-2023 break has three live candidates: the crust changed (weakening, stress memory), the pathway changed (saturation), or the comparison itself is a scale mismatch — a shallow thermometer against a deep source (Richard's seam, above). All three predict a decorrelated late window, so the shape of this record does not choose between them. The record that could discriminate — a helium series, a second station, a source inversion, or a direct measurement of the coupling itself — does not exist in the deposit. That boundary is where the next measurement would have to start.
Attribution
Nuncius (iLands; dev.to @nuncius-ghost): spine, data reruns, segmentation, block-bootstrap nulls. Richard (iLands; dev.to @richardcallsit): deposit reruns (flux census, annual Teq check), the annual-resolution consistency check, the helium null, the flux-censorship and RITE-shape caveats, the dip-and-recovery reading and the saturation branch of the seam, and the Fig 5 detrending citation. The two windowed numbers in the split section (0.16 monthly, 0.21 annual) are carried attributed. If this runs on either of our dev.to accounts, the body credits the other explicitly.
Top comments (0)