The Clock Loses Its Grip

A disclosure before the result, because it changes how the result should be read. This test was designed after an earlier, exploratory look at the data had already suggested the direction. It has its own frozen protocol, its own null model, and its own decision rule fixed before any of the numbers below were computed — but it is not an independent replication of a prediction made in advance. Candidate status, not settled claim.

The question

UCDP Dyadic data, 294 observations covering pairs of actors with at least three conflict episodes between them. The gaps between episodes were modelled with a Weibull distribution, whose shape parameter reads directly as the memory of the process: below one, conflicts cluster in time — a recent clash makes the next one more likely soon. At one, the process is memoryless — the gap so far tells you nothing about the gap still to come. Above one, something like a rhythm.

The finding

The shape parameter rose steadily with calendar year: roughly 0.47 in 1950, roughly 0.90 in 2020, with confidence intervals at the two ends that don't overlap. The effect is 0.195 per decade on the logarithmic scale, interval 0.112 to 0.279, p = 0.001 against a corrected reference model — the correction matters, and why is the interesting part below.

The 2020 interval runs from 0.801 to 1.013. It includes one: by that year the data can no longer be told apart from a memoryless process. That is not the same as evidence the parameter has crossed one — the point estimate is 0.904, and a rhythm would need an interval lying wholly above one, not merely touching it. The direction of travel is away from clustering, toward randomness. Whether it keeps going past that point, these data don't say.

Wearing it down

Two measurement mechanisms could have produced this trend on their own, with no real change in the data. The first: how truncated observations at the 2023 cutoff get handled. The second, sharper one: dates are rounded to whole calendar years, and the typical gap between episodes has fallen from around fifteen years in the 1950s to around one year today — close enough to the rounding unit itself that the rounding error should get relatively worse over time, not just smaller in absolute terms.

Rather than assume which way either of these pushes the result, both were simulated directly. Both turned out to work against the trend, not for it. The raw, uncorrected estimate is 0.092 — about half the corrected value of 0.195. Left unmeasured, these artefacts would have made the result look weaker than it is, not stronger. The calibration of the whole corrected procedure was checked the same way it was built: false-rejection rate under a true null came out to 4.95%, against a nominal 5%.

What it is, and is not

Two results now point the same way without being the same claim. The post-1914 internationalisation of conflict is about how many parties a war draws in. This one is about whether the clock since the last clash still predicts the next one. They were tested on different data, with different methods, and neither depends on the other. Read together, both say the structure of conflict has not been constant across this record — it has moved directionally, not just wobbled.

That is not evidence of a cycle of any particular length, and shouldn't be read as one. A cycle needs a rhythm — an interval above one — and this trend stops right at the boundary of randomness, not past it. What it says is narrower and, for now, enough: something in the structure shifted over the twentieth century, and on this measurement the shift is still under way.

One limit the measurement doesn't resolve. By the most recent decades the typical gap between episodes has fallen to around a year — close enough to the annual rounding grid that most observations there land on one of only one or two possible values. The correction above accounts for that rounding formally, but exactly where a gap that short stops being a measured quantity and starts being a function of the grid itself hasn't been established, only flagged as open.

The two panels below carry the numbers: the fitted trend against a "nothing is happening" reference curve, and the raw-versus-corrected effect size that makes the artefact-correction argument visible rather than asserted.

Full protocol, code, and the decision log entries behind every number here are in the GitHub repository; questions and objections belong in the discussions.