neuralnations · CAMS · PR-GAP-2

Does CAMS See a Coup Coming?

CAMS structural stress as a leading indicator of societal rupture — 44 societies, 1900–2025, 126 rupture events

Study · Kari Freyr McKern · July 2026 · 4,541 society-years, blinded five-scorer ensemble runs, cross-referenced against the Cline Center Coup d'État Project, V-Dem ERT, the Reinhart/Rogoff crisis database, and the historical record · Companions: Study 1 · Study 2

Take 44 countries, 125 years, and every coup, revolution, economic crisis, and regime collapse in between. Now ask: in the one to three years before each of those events, did our structural stress measurement look any different from an ordinary calm year? Tested naively — treating every rupture the same — the answer is a shrug. Tested properly, once two structurally different kinds of "rupture" are told apart, the answer is a clear yes.

This is the largest and most rigorously tested cross-national study CAMS has been put through. The first two studies looked at a single country (the United States) across two different mechanisms — how it absorbs fiscal pressure, and how structural stress erodes public trust. This one asks a blunter question across the whole world: does the same structural stress signal show up before any kind of societal rupture — coups, revolutions, civil wars, economic collapses, regime transitions — regardless of country or era? And it holds itself to a harder standard than the first pass: leave-one-society-out cross-validation, where the model is trained on 43 countries and tested, blind, on the one it's never seen.

The rupture calendar came from established academic sources, not from CAMS itself, so the test can't cheat: 126 events, drawn from the Cline Center's global coup database, the V-Dem project's regime-transition data, the Reinhart-Rogoff financial crisis database, and the historical record for wars, occupations, and revolutions. For each of the 4,541 society-year data points, we asked whether the structural readings one to three years before a coded rupture differ from readings in ordinary calm years.

They do, descriptively — every core structural variable is statistically different in the run-up to a rupture. But when that's turned into a genuine out-of-sample test (hold out one country, train on the rest, repeat 37 times), pooling all 126 events into one outcome barely clears chance: AUC 0.563, not statistically significant (p = .087). That's the honest result of asking the wrong question. Not every "rupture" in the historical record is the same kind of event.

Splitting them by what CAMS itself measures — whether the society's expressed institutional cognition had inverted (stress overtaking capacity) in the three years beforehand — separates 76 deficit crises, where the system's coordination capacity had already failed before the visible break, from 41 coordination transitions, where institutions were still functioning and the "rupture" is a negotiated handover, an independence declaration, or a postwar reconstitution rather than a collapse. Once the deficit crises are evaluated on their own, the same cross-validation reaches AUC = 0.740 (p < .001), with 26 of 29 held-out countries scoring above chance. The model wasn't failing on the whole panel — it was correctly telling two different processes apart.

The mechanism matches what the first two US studies already suggested, and it now has direct panel-wide confirmation: across the full 44-society, 125-year panel, rising stress suppresses a society's expressed institutional cognition in 41 of 42 evaluable societies, and suppresses coordination strength between institutions in all 42. This isn't just a crisis-window pattern — it's a structural regularity that holds almost everywhere, all the time. The signal doesn't come from fast-moving noise — day-to-day political drama, protest cycles, news-cycle chaos. It comes from slow institutional decay: coordination strength between institutions weakening, the system's capacity to retain and apply institutional memory eroding, governing capacity declining.

And there's a live reading attached to this. As of 2025, three societies — South Africa, Venezuela, and the United States — are at or above the highest stress level any of them has recorded in their own 125-year history. All three show the same institutional-memory erosion that the model flags as the leading warning sign. For the US specifically, the 2025 stress reading is now higher than the peak of the Great Depression in 1932 — the previous high point in over a century of data. The US and Germany also show the two steepest stress increases of any society over the past decade, which is notable precisely because neither is a fragile state on a classic crisis trajectory — both are large, mature democracies experiencing a newer, less familiar pattern of institutional erosion happening in the open, not during a war or an economic emergency.

Headline

A naive model that treats all 126 coded ruptures as one outcome barely beats chance in honest hold-one-country-out testing (AUC 0.563, p = .087). Once ruptures are correctly split into deficit crises (institutional coordination capacity already failed beforehand, n = 76) and coordination transitions (institutions still functioning, n = 41), the same test reaches AUC = 0.740 for deficit crises (p < .001, 26/29 countries above chance) — comparable to published political-science early-warning models that use far richer covariate sets. Panel-wide, rising stress suppresses expressed institutional cognition in 41/42 societies and coordination strength in 42/42: the mechanism holds almost everywhere, not just around crises.

1. Panel and rupture chronology

44
Societies
4,541
Society-years, 1900–2025
126
Rupture events coded
7.5%
Base rupture-incoming rate

CAMS scores come from blinded five-scorer ensemble runs (the CAMNATIONS5 protocol). Each society-year carries 8 nodes (Helm, Shield, Lore, Stewards, Archive, Craft, Hands, Flow) × 4 dimensions (Coherence, Capacity, Stress, Abstraction), from which Node Value, Bond Strength, Stress/Capacity Ratio, Reactivity Ratio, and expressed institutional cognition are derived under the JUNO v1.2-Final canonical operators. Ruptures — coups, regime transitions, financial crises, civil wars, occupations, revolutions — were coded independently from the Cline Center Coup d'État Project, V-Dem's Episodes of Regime Transformation v15, the Reinhart/Rogoff crisis database, and expert annotation for major events the standard sources miss, across 40 of the 44 societies (4 had no major rupture in the study window).

2. Two structural regularities, confirmed at panel scale

Before testing rupture prediction, two basic structural claims were checked across the full panel: does rising stress suppress a society's expressed institutional cognition (its capacity to translate accumulated knowledge into coordinated action), and does it suppress coordination strength between institutions? Both held almost everywhere.

ClaimSupported inMean correlation
Stress suppresses expressed cognition41 / 42 societiesρ̄ = −0.860
Stress suppresses coordination strength42 / 42 societiesρ̄ = −0.910

This is not a crisis-window result — it holds across each society's entire observed history, spanning every world region and political system in the panel. The one exception, India, turns out to be a data-quality artifact rather than a real counter-example (Section 9).

3. Not every rupture is the same kind of event

The 126 coded events lump together things that look similar from the outside but are structurally different underneath. Splitting each rupture by whether the society's expressed cognition had already inverted (stress overtaking capacity) in the three years beforehand separates two populations:

TypenShareWhat it looks like structurally
Deficit crises7660.3%Coordination capacity had already failed before the visible break
Coordination transitions4132.5%Institutions still functioning — negotiated handovers, independence declarations, postwar reconstitution, some coups in states with intact capacity
Insufficient data97.1%Excluded

This split was derived from the CAMS data alone, before looking at how well any model predicted either group — the point isn't to discard inconvenient cases, it's that a model built to detect structural degradation shouldn't be expected to "predict" a rupture where no degradation occurred.

Leave-one-society-out results. Trained on 43 countries, tested blind on the 44th, repeated for every country. Pooling all ruptures together: AUC 0.563 (p = .087) — not significant. Restricting to deficit crises: AUC 0.740 (p < .001), 26 of 29 evaluable countries above chance. Adding Dream-composition variables (how concentrated a society's expressed cognition is across its eight nodes) lifts deficit-crisis AUC to 0.763.
ModelOutcomeCountriesMean LOSO AUCp
Structural baselineAll 126 ruptures (pooled)370.563.087
Structural + compositionAll 126 ruptures (pooled)370.577.025
Structural baselineDeficit crises only290.740<.001
Structural + compositionDeficit crises only290.763<.001

For context, published political-science early-warning models that combine economic indicators, regime-type scores, ethnic composition, and conflict history typically land at AUC 0.68–0.75. The 0.740 figure above uses none of that — only internal structural coordination metrics.

4. Descriptive check: pre-rupture years vs calm years

Before cross-validating anything, the simplest possible check: do the raw readings one to three years before any coded rupture differ from ordinary calm years? Yes, and by a wide margin — this is the pattern that motivated the more rigorous test above.

VariablePre-Rupture MeanCalm MeanDifferenceSignificance
Mean Stress6.275.36+0.92p<0.001
Stress/Capacity Ratio1.391.06+0.33p<0.001
Mean Node Value7.169.45−2.29p<0.001
Mean Bond Strength0.2370.297−0.060p<0.001
Archive Node Value8.1510.44−2.29p<0.001
Helm Node Value7.069.69−2.64p<0.001

And single variables, taken individually, separate pre-rupture from calm years above chance across the board:

VariableAUCDirectionReads as
Mean Stress0.648Higher → ruptureBest single variable
Bond Strength0.641Lower → ruptureCoupling collapse
Stress/Capacity Ratio0.637Higher → ruptureStress-capacity gap
Archive Stress0.628Higher → ruptureMemory/legal strain
Helm Node Value0.628Lower → ruptureGovernance capacity
Archive Node Value0.604Lower → ruptureFixation capacity
Reactivity Ratio0.515Near-chance: not a rupture signal

These descriptive numbers are useful for intuition but overstate confidence on their own — they're in-sample, not held-out-country tests. Section 3's LOSO results are the load-bearing evidence.

5. Where the model doesn't work

Three countries scored below chance on the deficit-crisis LOSO test: Russia (0.366), Nigeria (0.420), Indonesia (0.481). All three share a common thread — the structural-degradation signal needs a few years to accumulate, and these countries' ruptures didn't give it that time.

Nigeria and Indonesia are event-compression cases: military coups and rapid authoritarian collapses where structural deterioration happened over weeks to months, not the two-to-three-year window the model scans. Russia combines classification ambiguity across very different eras (1905, 1917, 1991) with elevated scorer uncertainty for the Soviet period, where institutional opacity was often the point — the 1991 dissolution reads as a deficit crisis federally while simultaneously looking like a coordination transition for individual successor states. This marks a real scope limit: the model detects gradual structural decay, not fast-onset collapse or historically ambiguous, multi-era rupture sequences.

6. The mechanism holds across all three studies

Archive Node Value — institutional memory, legal precedent, the system's capacity to learn and retain coordination knowledge across administrations — is among the strongest CAMS predictors of both legitimacy collapse (Study 2) and rupture incidence (this study). South Africa's Archive Node collapsed from +7.5 in 2020 to −2.5 in 2025, a 10-point drop in five years — the most severe Archive degradation in the panel outside active state-collapse events like Afghanistan 2001 or Iraq 2003. The USA's Archive Node Value fell from 9.5 to 6.5 over the same window, continuing the erosion flagged in Study 2.

The Reactivity Ratio — a fast-node measure of operational churn — sits at 0.515, statistically indistinguishable from chance. Rupture risk is not written in short-term political noise; it's written in the slow decay of coordination and memory.

Current alarm.

Three societies are at or above their own all-time CAMS stress readings within the 1900–2025 panel: South Africa (105%), Venezuela (104%), and the USA (104%). For South Africa and Venezuela this matches or exceeds their historical late-apartheid and petroleum-crisis stress peaks. The USA's 2025 reading (7.33) exceeds its 1932 Great Depression reading (7.075) — the highest stress level in 125 years of US CAMS data. All three show measurable Archive degradation, and the model assigns each of them elevated deficit-crisis probability within a 3-year window.

7. A scoring artifact: the India exception

India is the one society where Section 2's stress–cognition claim fails, and it's worth explaining why rather than burying it. India's stress readings only take 9 distinct values across 73 scored years and rise in an almost perfectly straight line from 1950 to 2010; capacity rises in parallel, consistently a few points above stress. Because expressed cognition depends on capacity minus stress, both climb together for a mechanical reason — a scoring-resolution artifact — not because stress was somehow generating more cognition. Tellingly, the scores register no change at all during the 1975–77 Emergency, India's most acute authoritarian episode in the panel window: there isn't enough resolution in the India data to test the claim properly. Excluding India, the stress–cognition result is 41 of 41 evaluable societies, pending re-scoring at annual resolution.

8. Interpretation and limits

An AUC of 0.740 across 44 societies and 125 years, on a correctly-specified outcome, is a meaningful and reasonably robust result — not a definitive one. It does not mean CAMS can say which stressed country will rupture, or when. The model's scope is specific: it detects gradual structural degradation over a two-to-three-year horizon. It is not expected to, and does not, predict fast-onset coups, externally-imposed ruptures, or events in societies where historical scoring uncertainty is high. Restricting the outcome to deficit crises also shrinks the evaluable sample from 37 to 29 countries, so the 0.740 figure should be treated as a strong first result pending replication on a held-out panel, not a locked-in number.

9. Verdict

PR-GAP-2 establishes three things with reasonable confidence. First, at panel scale, rising CAMS stress consistently suppresses both expressed institutional cognition (41/42 societies, excluding the India scoring artifact) and coordination strength between institutions (42/42 societies) — a structural regularity holding across every region and political system in the panel, not just around crises. Second, once ruptures are correctly split by whether coordination capacity had already failed beforehand, the structural framework discriminates deficit crises from calm periods at AUC = 0.740 in honest leave-one-country-out testing (26/29 countries above chance) — comparable to published early-warning models that use far richer covariates, using only internal structural metrics. Third, this discrimination does not extend to coordination transitions, exactly as the theory predicts: coordinated institutional reorganisation and structural collapse are different processes, and a model built to detect the second should not be expected to "predict" the first.

South Africa, Venezuela, and the USA are currently at or above their own all-time stress readings, all three showing measurable Archive degradation, all three carrying elevated deficit-crisis probability within a 3-year window.

10. Files & downloads

Open science — the full interactive report (six Chart.js charts), the rupture chronology, and the underlying panels are all available below.

Full academic paper (with Robson Gauge & cross-system parsimony, Aug 2026) Full interactive report (HTML) Society-year panel CSV (4,541 × 19) Rupture chronology CSV (126 events) Event-study frame CSV (4,451 rows)
Companion: Study 1 — Interest Burden Companion: Study 2 — Legitimacy