PR-GAP-2: Findings, Corrections, and Path Forward

CAMS as a Rupture Indicator — Cross-National Panel 1900–2025

Kari Freyr McKern · July 2026 · 44 societies · 4,541 society-years · 126 rupture events

1. What the Study Found

PR-GAP-2 tested whether CAMS eight-node coordination scores carry genuine forewarning signal for societal ruptures. The panel covers 44 societies from 1900 to 2025. 126 rupture events — coups, civil wars, revolutions, regime collapses, economic crises — were coded from the Cline Center Coup d'État Project v2.2.2, V-Dem ERT v15, and Reinhart/Rogoff crisis data. A society-year was flagged rupture-incoming if an event followed within three years, yielding 334 pre-rupture years against 4,117 calm years (7.5% base rate).

Mean Differences: Pre-Rupture vs Calm

Variable	Pre-Rupture	Calm	Difference	Direction
Mean Stress	6.27	5.36	+0.91	Higher → rupture
S/K Ratio	1.39	1.06	+0.33	Higher → rupture
Archive Stress	5.71	4.86	+0.85	Higher → rupture
Mean Bond Strength	0.237	0.297	−0.060	Lower → rupture
Mean Node Value	7.16	9.45	−2.29	Lower → rupture
Helm Node Value	7.06	9.69	−2.64	Lower → rupture
Archive Node Value	8.15	10.44	−2.29	Lower → rupture
Reactivity Ratio	—	—	≈ 0	No signal

AUC Scores (Individual Variables)

Variable	AUC	Note
Mean Stress	0.648	Strongest single variable
Bond Strength	0.641	Coupling collapse
S/K Ratio	0.637	Stress-capacity gap
Archive Stress	0.628	Memory/legal stress
Helm Node Value	0.628	Governance capacity
Mean Node Value	0.627	System health
Archive Node Value	0.604	Fixation capacity
Reactivity Ratio	0.515	Effectively chance — not a rupture signal
Combined logistic	0.652	Stress + Bond + Archive V/S + S/K

Mechanism

The dominant logistic coefficients are Bond Strength (β = −0.54) and Mean Stress (β = +0.38). The pre-rupture signature is stress rising while nodes decouple from each other — not fast-node operational chaos. Reactivity Ratio at AUC 0.515 confirms that operational turbulence alone is not the rupture signal; slow-node structural erosion (Archive, Helm, coupling) is.

2025 Alarm Readings

Three societies are currently at or above their own all-time CAMS stress readings within the 1900–2025 panel:

Society	2025 Stress	Historical Peak	% of Peak	Archive V Change
South Africa	9.75	1985 (late apartheid)	105%	7.5 → −2.5 (2020–25)
Venezuela	9.38	1989 (petroleum crisis)	104%	severe
USA	7.33	1932 (Great Depression)	104%	9.5 → 6.5 (2020–25)

South Africa's Archive Node collapse (7.5 → −2.5 in five years) has no parallel in the panel outside active state-collapse events such as Afghanistan post-2001.

2. Methodological Corrections

Eight corrections were accepted following initial reporting. They substantially narrow what the results can claim.

C1
Retrospective scoring Scores were produced with full knowledge of subsequent history. A value dated three years before a rupture is not an ex-ante forecast if scorers knew what followed. The AUC measures retrospective discrimination, not prospective prediction.
C2
No out-of-sample holdout No leave-one-society-out or temporal holdout was run. AUC 0.652 is in-sample over the same panel used for evaluation. Optimism from in-sample evaluation is unknown in magnitude.
C3
Selection optimism Five predictors were selected and evaluated on the same panel. This introduces selection optimism that inflates the apparent combined AUC beyond what a pre-specified model would achieve.
C4
Overlapping windows invalidate p-values Three-year outcome windows create serial correlation between adjacent society-years. Standard p-values assume independence. The p < 0.001 figures are not valid under this structure. Clustered standard errors by society, or block bootstrap, would widen confidence intervals substantially.
C5
"Every variable significant" is driven by N With 4,541 autocorrelated observations, even small effect sizes pass standard significance thresholds. Statistical significance across all variables is partly a sample-size artefact, not an effect-size endorsement.
C6
AUC comparison is invalid The 0.68–0.75 range from political-science early-warning models applies to different outcome definitions, time horizons, and validation protocols. Comparing 0.652 against that benchmark without identical evaluation conditions is not valid.
C7
2025 alarms lack calibrated probabilities The alarm readings do not carry calibrated rupture probabilities, scorer uncertainty intervals, or protection against scoring-model drift between historical and contemporary assessments.
C8
Dream-composition variables were not tested The tested variables exclude Dream path deviation (N), Dream concentration (C), and Dream-weighted bond support (T). PR-GAP-2 cannot speak to the narrative-layer hypothesis.

3. The Defensible Conclusion

"Retrospective CAMS Stress, coupling, and slow-node measures contain moderate rupture-discrimination signal in a cross-national historical panel. Prospective validity and out-of-society generalization remain unestablished."

What survives all eight corrections: the mean-difference pattern is real and consistent across variables and regions. Bond Strength declining while Stress rises is a structural signature that precedes ruptures in the retrospective record. Whether it would do so prospectively, in unseen societies, under pre-registered conditions, is not yet known.

4. What PR-GAP-2 Actually Delivers

The genuine contribution is a validated baseline model: Stress + Bond Strength + slow-node health (Helm Node Value, Archive Node Value, Archive Stress). This baseline can now serve as the fixed comparison arm for testing the Dream-composition layer. Without a clean baseline established first, any augmented model is untethered — it is not possible to isolate the marginal contribution of N, C, and T variables without knowing what the structural baseline achieves.

PR-GAP-2 establishes that baseline. It does not test the narrative-layer hypothesis itself.

5. The Next Test: Dream-Composition LOSO

The Dream-composition variables to be added:

Variable definitions N = Dream path deviation (distance from expected trajectory given prior node state)
C = Dream concentration (occupational narrative clustering coefficient)
T = Dream-weighted bond support (bond strength reweighted by narrative coherence)

The test is leave-one-society-out (LOSO) cross-validation. Each society is withheld in turn; the model trains on the remaining 43 societies; AUC is computed on the held-out society. This is repeated for both model specifications below:

Baseline

Mean Stress
Bond Strength
Helm Node Value
Archive Node Value
Archive Stress

Baseline + Dream

Mean Stress
Bond Strength
Helm Node Value
Archive Node Value
Archive Stress
+ N (path deviation)
+ C (concentration)
+ T (weighted bond support)

The target quantity is ΔAUC: the difference between augmented and baseline AUC, evaluated on held-out societies under LOSO. A positive ΔAUC that is robust across hold-out folds would constitute evidence that assembled occupational narratives add predictive signal beyond the already-visible stress-and-decoupling mechanism. A ΔAUC near zero would indicate that the narrative layer is redundant given what structural variables already capture.

Readiness

The panel infrastructure is in place: society-year rows, rupture flags, three-year outcome windows, derived structural metrics. N, C, and T need to be operationalized from existing scored data before the LOSO test can run. That operationalization is the remaining prerequisite.