Distinguishing persistence from deviation in composite policy indices: A diagnostic framework for AI readiness trajectory screening and stratification

Composite indices of national capability reduce dynamic trajectories to static rankings, offering no mechanism to distinguish expected persistence from departures that warrant diagnostic attention. This study asks when a year-to-year change in a national AI-readiness score warrants additional analytical attention. We develop a five-step diagnostic framework (Expect, Detect, Explain, Contextualise, Interpret & Investigate) and apply it to the Oxford Government AI Readiness Index (118 countries, 2021–2024) across two transition windows. An autoregressive baseline explains 97.9% of the variation in next-year scores (W1 in-sample R² = 0.9790, MAE = 1.6506), establishing persistence as the dominant structure and residuals as the primary analytical unit; under a common held-out protocol no flexible specification improves on that baseline, so pillar disaggregation is used for attribution rather than prediction. Government is the pillar most frequently selected by the attribution rule among flagged deviations, with wide uncertainty and a threshold-sensitive ranking (50.0% in W1, 45.8% in W2 after normalisation; Wilson 95% CI 31–69% in W1). Regime analysis defines three Level-dominated readiness strata (k-means, k = 3; W1 trajectory stability ratio, TSR = 0.9407); the W1 directional differentiation in ΔLevel (p = 0.0020) is consistent with mean reversion under the below-unity AR(1) slope (0.954), with the regime term non-significant once continuous prior score is controlled (ANCOVA p = 0.389), and the W2 non-recurrence (p = 0.3643) coincides with a slope near unity (1.007). The regime typology therefore serves as an interpretive stratification device rather than an independent source of trajectory differentiation. Zero skip-a-level movements occur in both windows, consistent with strong persistence rather than independent evidence of ordered mobility. By screening departures from expected persistence using a signed-decile rule with window-specific cut-points (24 of 118 observations flagged per window by construction), the framework reconceptualises composite indices as state-dependent diagnostic systems and supports targeted follow-up monitoring.