PR-GAP-2-NCT-v1: Dream-Composition LOSO Results

Does Dream novelty, concentration, and weighted JUNO support add signal beyond the structural baseline?

Locked 2026-07-28 · 43 societies · 37 with valid ΔAUC · Baseline: Stress + Bond + Helm/Archive · Augmented: + N + C_d + T

1. Operationalization (PR-GAP-2-NCT-v1)

Three variables were locked before any outcome inspection. The naming convention below resolves the notational collision between node-level Coherence (C_ist, used in q_ist) and aggregate Dream Concentration (C_st, labelled Dream_Concentration throughout):

Dream composition D_ist = A_ist · C_ist (node Coherence, not Dream_Concentration)
p_ist = D_ist / Σ_j D_jst

N — Dream path deviation (Aitchison distance from 5-yr trailing centroid) clr(p_ist) = log p_ist − (1/8)Σ_j log p_jst
μ_is,t⁽⁵⁾ = (1/5) Σ_k=1..5 clr(p_is,t−k)
N_st = √[ (1/8) Σ_i (clr(p_ist) − μ_is,t⁽⁵⁾)² ]

Dream_Concentration — Pielou-evenness complement C_d,st = 1 − [ −Σ p_ist log p_ist ] / log 8

Dream_Weighted_JUNO_Support q_ist = (0.6·C_ist + 0.4·A_ist) / 10 [C_ist = node Coherence]
B_ijst = √(q_i·q_j) · 2^{−(S_i+S_j)/10}
T_st = Σ_i<j p_ip_jB_ij / Σ_i<j p_ip_j

Data note

Two society-years (Ukraine 1933, 1941) had D=0 on one node (Stewards: A=0, C=0), violating the p>0 guarantee. Both were excluded from NCT computation. N requires five consecutive preceding years with complete node data; this removed approximately 5 rows per society (first 5 years of each series), reducing the augmented test frame from 4,096 to 3,911 rows across the 37 evaluable societies.

2. LOSO Results: All Societies

Society	N base	N aug	Ruptures	AUC Base	AUC Aug	ΔAUC
Russia	126	121	17	0.3335	0.7296	+0.3961
South Africa	126	121	9	0.6842	0.7153	+0.0310
Colombia	126	121	6	0.8757	0.9029	+0.0272
Argentina	76	71	12	0.4596	0.4839	+0.0242
Syria	112	107	9	0.7292	0.7494	+0.0202
Poland	126	121	9	0.6287	0.6488	+0.0201
China	126	121	15	0.6793	0.6956	+0.0163
Finland	122	117	3	0.6947	0.7105	+0.0158
Iran	126	121	9	0.6163	0.6303	+0.0140
Iraq	124	119	24	0.7031	0.7129	+0.0098
Estonia	123	118	6	0.6211	0.6220	+0.0009
Philippines	117	112	6	0.8604	0.8616	+0.0013
Pakistan	78	73	12	0.4426	0.4447	+0.0021
Afghanistan	126	121	18	0.5247	0.5165	−0.0082
Thailand	124	119	12	0.3092	0.3006	−0.0085
Germany	126	121	9	0.9031	0.8968	−0.0063
USA	126	121	12	0.3874	0.3761	−0.0113
Norway	126	121	3	0.7005	0.6893	−0.0113
India	73	68	3	0.1429	0.1308	−0.0121
Egypt	126	121	11	0.7652	0.7529	−0.0123
Venezuela	56	51	9	0.6738	0.5397	−0.1341
Brazil	126	121	12	0.7156	0.6995	−0.0161
Nigeria	75	70	7	0.3550	0.3379	−0.0172
New Zealand	126	121	3	0.5881	0.5678	−0.0203
Japan	126	121	6	0.6889	0.6659	−0.0229
France	126	121	6	0.9056	0.8826	−0.0229
Ukraine	84	79	7	0.3748	0.3447	−0.0300
United Kingdom	126	121	12	0.4331	0.3998	−0.0333
Chile	126	121	6	0.6194	0.5855	−0.0339
Italy	124	119	9	0.7005	0.6626	−0.0379
Saudi Arabia	125	120	3	0.0410	0.0000	−0.0410
Indonesia	85	80	7	0.1905	0.1351	−0.0553
Hong Kong	116	111	6	0.3515	0.2849	−0.0666
Lebanon	83	78	13	0.3643	0.2953	−0.0690
Israel	80	75	11	0.4170	0.3224	−0.0946
Palestine	81	76	9	0.5154	0.3714	−0.1440
Australia	126	121	0	—	—	—
Canada / Sweden / Türkiye / UAE	—	—	0	—	—	—
Denmark	3	3	0	—	—	—

6 societies excluded from evaluation (no rupture events in test set or n<5). Italicised rows have AUC undefined.

3. Summary Statistics

Fig 1. ΔAUC distribution — all 37 evaluable societies

Fig 2. Baseline vs Augmented AUC per society

Statistic	Value
Evaluable societies (valid ΔAUC)	37
Mean ΔAUC (all 37)	−0.0094
Median ΔAUC	−0.0121
SD of ΔAUC	0.0783
ΔAUC > 0 (augmented better)	13 / 37
ΔAUC < 0 (baseline better)	24 / 37
One-sample t-test (H₀: mean=0)	t = −0.722, p = 0.475
Sign test (H₀: P[δ>0]=0.5)	p = 0.099
Russia ΔAUC	+0.396
Russia: outlier? (>2 SD from mean)	Yes (2 SD = ±0.157)
Mean ΔAUC excluding Russia	−0.0207
t-test excl. Russia	t = −3.054, p = 0.004
Sign test excl. Russia	p = 0.065

4. The Russia Outlier

Russia produces the only large positive ΔAUC (+0.396): the baseline model achieves AUC 0.334 (below chance) on Russia when held out, while the augmented model reaches 0.730. This is the dominant feature in the overall distribution. It almost certainly reflects a Russia-specific scorer pattern where Dream-composition variables carry a sharp structural signal not captured by the baseline — but it may equally reflect scorer drift or a peculiarity in how Russia's node scores shift across the Soviet/post-Soviet transition, amplified by the small number of Russia-specific folds contributing to the augmented model's training distribution.

Including Russia: mean ΔAUC = −0.009, p = 0.475 — no significant effect in either direction.
Excluding Russia: mean ΔAUC = −0.021, p = 0.004 — the augmented model is significantly worse than the baseline across the remaining 36 societies.

Critical finding

The Russia case must be treated as a leverage point, not as evidence for the augmented model. A single held-out society swings the aggregate ΔAUC from −0.021 to −0.009. The sign test (13/37 positive, p = 0.099) shows that even including Russia, the augmented model is not significantly better by sign — and excluding it the test approaches significance in the negative direction (p = 0.065).

5. Effective Sample Note

As flagged in the pre-execution review: the 5-year lookback requirement for N removed approximately 5 rows per society, reducing the augmented test frame from 4,096 to 3,911 rows. The missing rows are the first 5 years of each society's series — a non-random subset (earliest observations, often covering pre-modern or early-state periods). The baseline and augmented models were therefore evaluated on different (though nearly identical) subsets. The effect is small but worth noting: any society where the earliest years carry strong rupture signal could show augmented AUC artificially elevated or depressed relative to the baseline run on those rows.

6. Verdict

PR-GAP-2-NCT-v1 Verdict

The Dream-composition layer (N, Dream_Concentration, Dream_Weighted_JUNO_Support) does not improve rupture discrimination over the structural baseline under leave-one-society-out cross-validation.

Across 37 evaluable societies: mean ΔAUC = −0.009 (p = 0.475, t-test; 13/37 positive by sign, p = 0.099). The only society with a large positive ΔAUC is Russia (+0.396), which is a statistical outlier more than 2 SD from the mean and likely reflects a leverage artefact rather than generalizable Dream-composition signal. Excluding Russia: mean ΔAUC = −0.021, p = 0.004 — the augmented model is significantly worse.

The most defensible interpretation is that N, C_d, and T as computed from this scored panel do not carry rupture-prediction information beyond what is already encoded in Stress, Bond Strength, and slow-node health. This does not falsify the Dream-composition hypothesis in general: it could reflect scorer limitations (Dream-composition depends on A×C products, and if scorers treat A and C as roughly proportional the composition is nearly uniform and uninformative), insufficient variation in p_ist across years within a society, or genuine redundancy with existing structural variables.

The result does, however, meet the preregistered test. The ΔAUC criterion was LOSO; the direction of the result is clear for 36 of 37 societies. The narrative-layer hypothesis requires either stronger prior evidence for heterogeneous Dream configurations in the scored data, or a reconceptualization of how occupational-narrative deviation should be operationalized before a retest.

Files: pr_gap2_loso_results.csv · pr_gap2_panel_derived.csv · pr_gap2_event_study.csv · pr_gap2_ruptures.csv — all in wintermute/analysis/