Independent Cross-Temporal Replication of Coalition Phase Transition in Cognitive Resource Allocation: A Preregistered Validation Study

This preregistration specifies confirmatory hypotheses and a complete analysis plan for Study 2, an independent- sample replication of findings from Study 1. BACKGROUND AND RATIONALE Study 1 (completed prior to this preregistration) was a cross-sectional exploratory study (N = 1,457) examining how five cognitive-affective variables — Learning Belief Strength (LBS), Self-Monitoring Capacity (SMC), Perceived Survival Concern (PSC, operationalized via public self- consciousness), Brief Fear of Negative Evaluation (BFNE), and Attentional Control Scale (ACS) — interact to predict dual-channel adaptive outcomes: Learning Adaptation (G_Adapt) and Social Adaptation (S_Adapt) among Chinese university students. Study 1 employed a multi-method analytic framework including EBICglasso network analysis, information- geometric diagnostics (Total Correlation, O-information), segmented threshold regression, Transformer-based deep learning with SHAP (SHapley Additive exPlanations) decomposition, and hypergraph-based coalition strength analysis. Study 1 data collection, analysis, and manuscript preparation were fully completed before this preregistration. STUDY 1 KEY FINDINGS (basis for all predictions): (a) Phase transition in the Learning channel: Segmented regression identified a statistically significant threshold at LBS γ = 4.18 [95% CI: 3.80, 6.00] (Davies test p = .0055, ΔAIC = −11.28). The LBS → G_Adapt slope approximately doubled from β₁ = 0.122 (SE = 0.021) below the threshold to β₂ = 0.228 (SE = 0.014) above it, representing a shift from medium (β* = 0.31) to large (β* = 0.58) standardized effect. No significant threshold was detected in the Social channel (S_Adapt ~ LBS; Davies p = .386). (b) SHAP sign-flip of the Growth Coalition: Transformer- derived SHAP analysis revealed that the Growth Coalition (LBS + SMC) contribution to Learning Adaptation reverses sign across LBS regimes: suppressive (mean SHAP = −0.238) in the Survival- dominant regime (LBS < 3.75) and facilitative (mean SHAP = +0.223) in the Growth-dominant regime (LBS ≥ 4.75). This sign-flip was replicated across 20 independent random seeds (100% sign consistency), with regime membership accounting for 42% of SHAP variance (η² = .420, F(2, 1454) = 526.44, p < .001). (c) Channel asymmetry: Learning Adaptation showed 2.83× larger regime sensitivity than Social Adaptation (η² = .167 vs. .059; F ratio = 145.44 vs. 45.41). (d) Coalition coherence U-shape: System-wide coalition coherence (5-Way metric) exhibited a U-shaped trajectory across LBS intervals — high coherence at LBS extremes (0.961 at LBS < 2.81; 0.757 at LBS > 5.23) and fragmentation during the transition zone (0.554 at LBS 4.02–4.63). (e) Prediction–Explanation Dissociation: The optimized Transformer achieved slightly lower global predictive accuracy than linear models (R² = .280 vs. .309) but revealed qualitatively richer process information (the sign-flip) invisible to constant-coefficient models. STUDY 2 PURPOSE Study 2 collects an independent sample (target N = 500) from the same institutional population approximately one month after Study 1, using identical measures. Its primary purpose is to test whether the four core findings above replicate in a new sample at a different time point. All hypotheses below are derived exclusively from Study 1 results and are registered before any Study 2 data are collected or inspected. PROSPECTIVE DESIGN NOTE Study 2 also collects the last four digits of each participant's student ID number. This identifier is not used in any Study 2 analysis but is collected prospectively to enable individual-level longitudinal linkage should a future Study 3 (within-person follow-up) be conducted. The decision to conduct Study 3 will depend on peer review feedback; this contingency is documented transparently.