Consequence entanglement, measured: a falsifiable operationalization of the polycrisis severity claim

Abstract

The polycrisis field asserts that crises are causally entangled across domains but has never written down a quantitative claim that could be shown to be false. This paper supplies the missing object for the half of the claim the field's own mechanisms make most plausible — consequence propagation: that crises make each other worse, not merely more frequent — and then tests it. A crisis stream is formalised as a marked point process: per month a count and, per event, a mark equal to log severity. An intensity channel (a latent regime modulating the event rate) and a mark channel (a latent regime modulating event severity) are kept structurally separate, and each is given the explicit null the field never wrote down — within a domain, random labelling, which holds the counts and the marginal severity distribution exactly fixed; across domains, a circular-shift null on each domain's detrended monthly mean log severity. Both tests are gated on decidability atlases computed on known-truth simulation.

The result is a split verdict. Within conflict (UCDP GED fatalities, 347,109 events over 420 months) the excess over the random-labelling null survives decisively and seed-robustly in all five pre-declared cuts: the monthly allocation of conflict severity is not exchangeable given the counts. Within disasters the same excess is decidable by rank but near-negligible. Across domains there is no shared severity regime: the detrended contemporaneous correlation of the two domains' monthly mean log severity is −0.1147 against a circular-shift null (one-sided p = 0.9492), and no lag in a ±6-month scan turns positive. So on these data the most plausible form of polycrisis fails in the same direction as the least plausible one: whatever severity structure exists is domain-local. The negative is published with the effect size below which it says nothing, and with its coordinates.

Re-scoped register (2026-08-30), stated up front. This version of the paper narrows its positive result in three specific ways, all of them consequences of an adversarial audit and of the calibration record deposited beside it. (i) The magnitude is not diagnostic. The 7,654 BIC excess establishes that the monthly allocation of conflict severity is not exchangeable given the counts, and nothing more: an exogenous monthly-mean log-severity drift of about 0.13 log-units, with no severity dynamics of any kind, reproduces an excess of the same size, as does a literal two-population composition model. Every effect-size comparison built on that magnitude is withdrawn. (ii) The confirmatory decidability certificate is withdrawn. The recovery bootstrap an earlier draft promoted to "what carries the decidability of this read" scored full-joint BIC arm selection, which the count channel decides; it returns a winner-recovery of 1.000 on streams with the mark coefficient set to zero, i.e. with the severity channel switched off. The criterion is fixed in code, but has not been re-run on the real configuration, so decidability of the severity channel there is uncertified. (iii) A quoted null-calibration ruler is withdrawn as unregenerable. The "0 of 24 false decisives, false excesses topping out at 23.5 BIC" figure had no script, seed, or record anywhere; it is replaced by a measured ceiling of +7.19 BIC at the conflict configuration and +4.02 BIC at the disaster configuration. The negative half of the paper — no shared severity regime across domains — is unaffected and stands as published.