The saturation curve was bent by the distance between the sensor and the storm

Where the error enters

The driver in this relationship is never measured where it acts. WIND, ACE and DSCOVR sit at the L1 point, 1.5 million kilometres sunward of Earth, and the solar wind they sample has to cross that gap before it leaves any mark on the ground. Sivadas and colleagues modelled what happens in between: a propagation delay to the bow shock with a scale of about 8 minutes, a further delay from the nose to the polar cap averaging 17 minutes with a standard deviation of 25 minutes, and shock fronts that add noise growing with the size of the event. The response side is cleaner, taken from ground magnetometers through the polar cap index and the westward auroral electrojet index. What sits on the table is twenty-five years of one-minute averages, 1995 to 2019.^[1]

That asymmetry is the whole mechanism. When the measured driver is extreme and the measurement is uncertain, the true value behind it is likelier to sit closer to the mean than the reading suggests. Pair enough such readings with the milder responses they actually produced and the curve bends downward at its top end, leaving a nonlinear bias with no physics in it. The authors argue that the saturation charged to the magnetosphere for decades is that bias, and that the theories built to explain it were fitted to an artefact. The competing possibility has to be written down as well: a real physical limit could exist and be erased by an error model that over-corrects, in which case the recovered linear relation would itself be the artefact.^[1]

What the correction changes

Regression calibration does one thing: it replaces each uncertain measurement with the conditional expectation of the true value given that measurement, and refits. After that step the relation between solar wind driving and geomagnetic response stays linear across the range the data cover, with no saturation. Carried to about 25 mV/m, the response comes out roughly twice what the saturating fits implied. The error model was not assumed: its output was checked against the data's own statistics, the probability density of the driver, its standard deviation, its conditional distributions and the regression bias it predicts.^[1]

The paper marks its own edge, and that edge matters more than the headline figure. Beyond 15 mV/m, the authors write, the data are insufficient to conclude anything about the shape of the relation. The doubled extreme quoted at 25 mV/m therefore sits above the range where the corrected curve was measured; it is an extension from a linear fit rather than an observation. The confidence intervals carry the familiar caveat too: they describe how often a repeated procedure would contain the regression curve, not a 95 per cent probability that the true value lies inside them. For a grid operator sizing a worst case, that distinction separates a measured quantity from a planning assumption.^[1]

How far the claim travels

The authors carry the argument past space weather. The same bias, they write, may have a role in climate models that underestimate extreme events such as heatwaves, in the assessed impact of strong earthquakes away from their epicentre, and in severe medical symptoms. The condition is narrow and checkable: a driver measured with an error that grows with its own magnitude, set against a response measured more precisely. Where both hold, an apparent ceiling should be doubted before it is explained. Models trained on such inputs will learn that ceiling as though it were a property of the world.^[1]

The test is repetition rather than persuasion. The upstream monitors are interchangeable in principle: WIND supplied the driver data here, while ACE and DSCOVR watch the same wind from the same point with different instruments and different error characteristics. A reanalysis that starts from one of those and applies the same calibration should recover a linear relation up to 15 mV/m, or fail to. If such a reanalysis on an independent upstream monitor is published before 30 June 2027, that comparison is the first thing to read. A corrected curve that repeats on other instruments is the result itself; a curve that does not repeat belongs to WIND's error model.^[1]