State Erosion: Testing Collapse Without Precursors

Can collapse be predicted? We built the instruments to kill our own theory before asking whether it was true. Here is what survived. Slowly stressed systems (forests, fault networks, financial markets) erode through discrete release events. This study formalises a state-erosion hypothesis: event sizes follow a truncated power law whose cutoff s* defines a measurable containment range, and collapse requires a prior drift of that range toward the system scale. Instead of seeking confirmation, we did the opposite. Every claim was given a preregistered falsifier with frozen, hash-stamped decision thresholds. Every statistical test was certified on synthetic ground truth before touching real data, and tests that lacked the power to refute were annulled, not reported. Then we attacked the theory with about 165,000 public events across five domains: US wildfires, Southern California earthquakes, a century of S&P 500 drawdowns, US grid disturbances, and interstate wars. What survived. The descriptive core is remarkably solid where it can be tested. A set of 109,722 wildfires follows a truncated power law (tau = 1.63; s* of about 264,600 acres, 95% CI from 230,465 to 308,399) that decisively beats the lognormal, while 46 years of seismicity show pure Gutenberg-Richter scaling with no detectable cutoff. Containment, it turns out, is a joint property of the system and of the observation window. What died. Every predictive claim. Of the three great crashes with usable pre-history (1987, 2008, 2020), two arrived with no prior drift of the containment range; in 2008 the range was contracting. No crash is a dragon-king outlier: the giants sit inside the tail fitted on their own past. A preregistered drift alarm fired once in 99 years, outside its own success window. A second preregistered study on the ten largest out-of-sample earthquakes and wildfires found the drift phase before giant events to be statistically indifferent: collapse is blind to the observable state of erosion. The result, in one line: giant events are ordinary members of the tail, quantifiable in possibility, unheralded in actuality. Within the statistical power certified by the protocol, this is more consistent with Bak's self-organised criticality than with the dragon-king programme. The oracle retires and the insurer takes over: like building codes dimensioned on Gutenberg-Richter, risk can be priced on the measured tail without pretending to time the disaster. Why this may interest you even if you care about none of these domains: the paper's primary contribution is the protocol itself, a demonstration that a broad theory can be preregistered, severely tested, partially falsified in public, and survive only as its data-resistant core. All frozen hashes, annulled verdicts, amendments, code and data sources are part of the record. In fields where extreme events are routinely explained after the fact, this inversion is the transferable result. All data are public (USGS FDSN, Yahoo Finance, USDA FPA-FOD, DOE OE-417, Correlates of War); the full pipeline is reproducible from the archived code.