Tristable Dynamics in Political Regime Space:
Evidence from a 225-Year Panel

1. Introduction

How many stable equilibria exist in political regime space? The question is fundamental to comparative politics, yet the dominant theoretical frameworks have overwhelmingly assumed the answer is two. From Huntington's (1991) waves of democratisation to Acemoglu and Robinson's (2006) game-theoretic model of transitions between democracy and dictatorship, the field has treated political regimes as occupying a bistable landscape in which countries gravitate towards one of two poles. This binary framing underlies the Freedom House classification system (Free/Not Free), the Polity IV democracy-autocracy scale, and the vast literature on democratic transitions, and reversals.

We argue that this bistable assumption is empirically inadequate. The cross-sectional distribution of political freedom scores across countries and time is not bimodal but trimodal, with a persistent intermediate cluster that cannot be dismissed as noise, or measurement error. This intermediate cluster—which we term the hybrid trap—has been recognised descriptively by scholars of competitive authoritarianism (Levitsky and Way 2010; Diamond 2002), but it has not been formally incorporated into the dynamical models that govern our understanding of regime transitions. Countries such as Singapore, Indonesia, Mexico, and Hungary have occupied the hybrid zone for decades, neither consolidating democratic institutions nor collapsing into outright tyranny. This persistence is inconsistent with a bistable landscape, which would predict that intermediate positions are unstable, and short-lived.

This paper makes four contributions to the study of regime dynamics. First, we provide the first formal statistical estimation of a tristable potential landscape for political regimes, using Gaussian Mixture Models fitted to a 225-year panel. We demonstrate that the three-component mixture dominates alternatives by BIC, identifying three distinct attractor basins rather than the conventional two. Second, we estimate a continuous-time Langevin stochastic differential equation (SDE) that connects the empirical distribution of liberty scores to the physics of overdamped motion in a potential landscape. The Langevin formulation yields an empirical potential function V(L) with three local minima, barrier heights that explain observed transition asymmetries, and a framework for computing mean first-passage times between basins. Third, we identify a Critical Instability Zone at L ≈ 52–55 below which the probability of democratic recovery collapses to approximately 3%, formalising the intuition that democratic erosion becomes self-reinforcing beyond a certain threshold. Fourth, we demonstrate that AR(1) persistence (β = 0.96) dramatically outperforms discrete stage-based transition models, suggesting that the underlying data-generating process is better characterised by continuous drift with stochastic shocks than by discrete jumps between categorical states.

The paper proceeds as follows. Section 2 reviews the relevant literature on bistable models, hybrid regimes, and regime transition theory. Section 3 introduces the ternary phase space framework (L + T + C = 100) that provides the conceptual foundation for our analysis. Section 4 describes the data and measurement strategy. Sections 5, 6, and 7 present the core empirical results: GMM analysis, Langevin SDE estimation, and the identification of the critical instability threshold. Section 8 discusses robustness and limitations. Section 9 interprets the findings in the context of existing theory, and Section 10 concludes.

2. Literature Review

2.1 The Bistable Paradigm

The treatment of political regimes as occupying one of two stable states—democracy or autocracy—has deep roots in comparative politics. Schumpeter (1942) defined democracy procedurally as a method of selecting leaders through competitive elections, establishing a bright line between democratic, and non-democratic systems. Dahl (1971) refined this into the concept of polyarchy, characterised by contestation, and participation, which nonetheless preserved the binary distinction: countries either met the threshold for polyarchy or they did not. Huntington (1991) historicised the binary framework through his analysis of three waves of democratisation and two reverse waves, treating transitions as movements between two basins.

The most influential formal model of bistable regime dynamics is Acemoglu and Robinson's (2006) Economic Origins of Dictatorship and Democracy. Their game-theoretic framework models democratisation as a strategic concession by elites facing revolutionary threat from citizens, and coups as elite responses to the redistributive costs of democracy. The model generates a bistable equilibrium structure: countries settle into either democracy (with redistribution) or dictatorship (with repression), and transitions between the two are driven by shocks to the relative costs of revolution and repression. The Acemoglu-Robinson model has been enormously productive, but its binary choice set cannot accommodate the large number of countries that occupy intermediate positions for extended periods.

Przeworski and Limongi (1997) and Przeworski et al. (2000) provided seminal empirical analyses of regime transitions, demonstrating that the probability of democratic survival increases with per capita income—a finding consistent with the existence of a democratic attractor whose depth increases with economic development. However, their analysis was conducted within the bistable framework, classifying countries as either democracies or dictatorships, and modelling transitions between the two.

2.2 The Hybrid Regime Challenge

The most sustained empirical challenge to the bistable framework has come from scholars of hybrid regimes. Diamond (2002) coined the term "electoral authoritarianism" to describe regimes that hold elections but lack the institutional infrastructure to make them meaningful. Levitsky and Way (2002, 2010) developed the concept of "competitive authoritarianism" to describe regimes in which democratic institutions exist but are systematically undermined, demonstrating that such regimes are not merely transitional but can persist for decades. Schedler (2006) offered a comprehensive analysis of the "politics of uncertainty" in electoral authoritarian regimes, showing how incumbents manipulate the electoral playing field while maintaining a veneer of democratic competition.

These scholars established that hybrid regimes constitute a distinct empirical category, but they did not translate this insight into formal dynamical models. The question of whether hybrid regimes represent a genuine attractor basin—a stable equilibrium to which countries are pulled by endogenous forces—or merely an unstable transition zone through which countries pass on their way to democracy or autocracy remained unresolved. Our GMM analysis directly addresses this question by testing whether the distribution of liberty scores is better described by two or three components.

2.3 Democratic Erosion and the Third Wave in Reverse

The past two decades have witnessed growing scholarly attention to democratic erosion—the gradual weakening of democratic institutions short of outright regime change. Diamond (2015) documented a "democratic recession" beginning around 2006, characterised by the stagnation, or decline of democratic governance indicators across multiple countries. Bermeo (2016) showed that the modal form of democratic breakdown has shifted from military coups to "executive aggrandisement," in which elected leaders incrementally expand their power at the expense of other institutions. Levitsky and Ziblatt (2018) analysed the mechanisms of democratic erosion in established democracies, emphasising the role of norm violation, and institutional subversion.

Haggard and Kaufman (2021) provided a systematic comparative analysis of democratic backsliding, identifying economic crisis, polarisation, and executive overreach as key risk factors. Svolik (2012, 2019) developed formal models of authoritarian consolidation and demonstrated that democratic stability depends on citizens' willingness to punish antidemocratic behaviour even when it is committed by their preferred candidates—a condition that polarisation undermines.

Geddes, Wright, and Frantz (2018) offered the most comprehensive empirical analysis of authoritarian regimes, documenting their remarkable diversity in institutional form, survival duration, and mode of collapse. Their work underscores the heterogeneity within the "autocracy" pole of the bistable framework, suggesting that a single attractor may be insufficient to capture the dynamics governing non-democratic regimes.

Treisman (2020) advanced a provocative argument that many democratisations were unintended consequences of miscalculation by authoritarian incumbents rather than the result of deliberate strategic choice. This insight resonates with stochastic models of regime dynamics: if transitions are largely driven by unpredictable shocks rather than deterministic forces, then the high variance, and low recovery rates we observe in the hybrid zone are precisely what we should expect.

2.4 Physics-Inspired Models of Political Dynamics

Our methodological approach draws on a small but growing literature that applies concepts from statistical physics to political phenomena. Castellano, Fortunato, and Loreto (2009) surveyed models of opinion dynamics that treat political attitudes as particles in a potential landscape. Galam (2012) developed "sociophysics" models that apply phase-transition concepts to voting behaviour and collective decision-making. More recently, scholars have applied stochastic differential equations to model economic regime changes (Hamilton 1989), financial crises (Sornette 2003), and ecological tipping points (Scheffer et al. 2009). Our Langevin SDE formulation extends this tradition to political regime space, treating the liberty score as the position of an overdamped particle in a potential landscape subject to stochastic shocks.

3. The Ternary Phase Space: Conceptual Framework

3.1 Three Dimensions of Governance

Most political science indices reduce the governance landscape to a single dimension: more or less democratic. This collapsing of a complex, multidimensional phenomenon into a single axis sacrifices critical information. A country with a powerful dictator and functioning state institutions (e.g., China, with L = 5, T = 87, C = 8) occupies a fundamentally different position in political space than a country with no effective state at all (e.g., Somalia, with L = 8, T = 22, C = 70), even though both score near zero on conventional freedom indices. The distinction between concentrated coercion and fragmented anarchy is not a refinement of the freedom dimension but an orthogonal axis of variation.

The Political Topology framework addresses this deficiency by modelling governance as a position in a ternary phase space defined by three dimensions: Liberty (L), Tyranny (T), and Chaos (C). These three components are subject to a binding constraint:

This constraint models political power as a zero-sum allocation across three modalities: distributed power with institutional constraints (Liberty), concentrated power deployed coercively (Tyranny), and fragmented or contested power in the absence of effective governance (Chaos). The constraint reduces the three-dimensional space to a two-dimensional simplex—the ternary diagram familiar from chemistry and geology—on which every country occupies a unique point at every point in time.

3.2 Measurement Strategy

The three components are measured as follows. Liberty (L) is derived from the Freedom House Freedom in the World aggregate score, which combines assessments of political rights, and civil liberties on a 0–100 scale. For periods before Freedom House coverage begins (pre-1972), we rely on the Varieties of Democracy (V-Dem) liberal democracy index and the Polity IV/V dataset, crosswalked to the Freedom House scale using overlapping coverage periods. Chaos (C) is derived from the Fund for Peace Fragile States Index (FSI), inverted and rescaled so that higher values indicate greater state fragility. For pre-FSI periods, we use historical indicators of state capacity, civil conflict, and institutional collapse from V-Dem and qualitative historical sources.

Tyranny (T) is computed as the residual: T = 100 − L − C. This measurement strategy has the advantage of enforcing the ternary constraint by construction and the disadvantage of making T dependent on the accuracy of both L and C. If Freedom House systematically overstates a country's liberty, the framework will correspondingly understate its tyranny. We discuss this limitation in detail in Section 8 and identify the development of independent tyranny indicators as a priority for future research.

3.3 The Constraint as Modelling Choice

It is important to be explicit about the epistemological status of Equation (1). The ternary constraint is a modelling assumption, not an empirical finding. It imposes the structure that any gain in one governance dimension must come at the expense of the other two—a zero-sum framing that is defensible as a first-order approximation but may break down in specific contexts. For example, a country that simultaneously strengthens democratic institutions, and improves state capacity might be said to increase L while reducing C, holding the constraint; but whether this necessarily implies a decrease in T is an empirical question that the model assumes away.

We adopt the ternary constraint for two reasons. First, it imposes a discipline on the analysis that prevents overfitting: with only two degrees of freedom, the model cannot explain every observation, and is thus more readily falsifiable. Second, the simplex geometry provides powerful analytical tools—including the potential landscape analysis developed in Section 6—that would not be available in an unconstrained three-dimensional space. The cost of this parsimony is that the model may miss important dynamics that violate the zero-sum assumption.

4. Data and Measurement

4.1 Dataset Construction

The dataset encompasses 91 countries observed at key inflection points over a 225-year period (1800–2025), yielding 1,656 country-year observations. Countries were selected to ensure geographic diversity (28 European polities, 15 Americas, 20 Asia, 21 Africa, 7 Middle East/Other) and to include the full range of regime types, from consolidated democracies (e.g., Norway, Finland) to totalitarian states (e.g., North Korea, Eritrea) to collapsed states (e.g., Somalia, Haiti).

Each observation records the country's position in the ternary phase space (L, T, C) at the observation date. Observations are concentrated at major historical inflection points—regime changes, constitutional reforms, wars, coups, and elections—rather than at fixed annual intervals, reflecting the availability, and reliability of historical data. Linear interpolation between observation points is used for analyses requiring annual panel data, with the acknowledgement that this imposes a smoothness assumption that may mask rapid within-period dynamics.

4.2 Data Sources and Crosswalk

The primary sources for Liberty (L) measurement are Freedom House Freedom in the World (1972–2025, 195 countries), the Varieties of Democracy Institute (V-Dem, 1789–2024, 202 countries), and the Polity IV/V project (1800–2018). For the post-1972 period, we use the Freedom House aggregate score directly. For the pre-1972 period, we crosswalk the V-Dem liberal democracy index (v2x_libdem) and the Polity revised combined polity score (polity2) to the FH scale using overlapping-period regression calibration.

The crosswalk achieves a 67% exact-match rate with the Freedom House classification when applied to the overlapping period, with the largest deviations occurring for countries that Freedom House, and V-Dem assess differently (e.g., South Africa, Δ = 14 points; Turkey, Δ = 10 points). This 33% disagreement rate is a significant limitation that we discuss in Section 8.

Chaos (C) is measured using the Fragile States Index for the period 2006–2024, and using V-Dem state capacity indicators, civil conflict databases, and qualitative historical assessments for earlier periods. The FSI total score (0–120) is inverted and rescaled to a 0–100 scale where higher values indicate greater chaos.

4.3 Cross-Sectional Distribution

The pooled distribution of liberty scores across all 1,656 observations is shown in Figure 1. The distribution is clearly non-normal, with pronounced clustering at the extremes, and a secondary mode in the intermediate range. This visual pattern motivates the formal GMM analysis in Section 5.

Figure 1. Cross-sectional distribution of Liberty scores across 91 countries and 225 years (N = 1,656). The trimodal structure is visible, with concentrations at low (L < 20), intermediate (L ≈ 30–65), and high (L > 80) liberty scores. The dashed red line at L ≈ 52–55 marks the Critical Instability Zone identified in Section 7.

5. Gaussian Mixture Model Analysis

5.1 Model Specification

To formally test the number of distinct modes in the liberty score distribution, we fit Gaussian Mixture Models with K = 1, 2, 3, 4, and 5 components to the pooled sample of N = 1,656 liberty scores. The probability density under a K-component GMM is:

where π_k are mixing weights summing to unity, and φ(L; μ_k, σ_k) denotes the Gaussian density with mean μ_k and standard deviation σ_k. Parameters are estimated via the Expectation-Maximisation (EM) algorithm, implemented from scratch using only Python standard library functions. We run 20 random restarts for each K to mitigate sensitivity to initialisation, using K-means++ style seeding, and retain the solution with the highest log-likelihood.

5.2 Model Selection

We select amongst competing models using the Bayesian Information Criterion (BIC), which penalises model complexity more heavily than the Akaike Information Criterion (AIC) and is generally preferred for model selection in mixture models (McLachlan and Peel 2000). The number of free parameters for a K-component univariate GMM is 3K − 1 (K means + K variances + K − 1 mixing weights).

The K = 3 model is decisively preferred by BIC, with a ΔBIC of 193.9 over K = 2 and 574.3 over K = 1. The K = 4 model offers only a marginal improvement (ΔBIC = +5.5), and K = 5 is substantially worse (ΔBIC = +37.7). Both AIC and BIC concur in identifying K = 3 as the optimal number of components, though AIC slightly favours K = 4, consistent with the well-known tendency of AIC to select more complex models.

5.3 Component Parameters

The estimated parameters of the three-component GMM, together with bootstrap 95% confidence intervals from 1,000 resamples, are reported in Table 3.

Several features of the estimated parameters merit discussion. First, the three components have approximately equal mixing weights (ranging from 0.321 to 0.345), indicating that roughly one-third of all country-year observations fall in each basin. This is a striking finding: it means that the hybrid trap is not a residual category containing a handful of ambiguous cases but a basin of comparable population to democracy and tyranny.

Second, the tyranny well (k = 1) and democratic plateau (k = 3) have comparatively narrow standard deviations (σ ≈ 8), whilst the hybrid trap (k = 2) has a much wider dispersion (σ ≈ 16). This asymmetry reflects the qualitatively different nature of the hybrid trap: it is a broad, shallow basin spanning L ≈ 20–70, within which countries exhibit substantial heterogeneity. Singapore at L = 47 and Ghana at L = 68 both fall within the hybrid basin but have very different institutional configurations.

Third, the component means correspond closely to the a priori expectations of the Political Topology framework. The tyranny well centred at μ ≈ 11 corresponds to the zone of consolidated autocracy where state coercion dominates. The hybrid trap centred at μ ≈ 47 corresponds to the zone of competitive authoritarianism and partial democracy. The democratic plateau centred at μ ≈ 89 corresponds to the zone of consolidated liberal democracy.

6. Langevin Stochastic Differential Equation

6.1 Theoretical Framework

Having established that the cross-sectional distribution is trimodal, we now turn to the continuous-time dynamical model that generates this distribution as its stationary state. We model the evolution of a country's liberty score as an overdamped Langevin equation:

where L(t) is the liberty score at time t, V(L) is a potential function, V′(L) = dV/dL is its gradient (the "restoring force"), σ is the noise intensity, and W(t) is a standard Wiener process. This is the continuous-time analogue of a discrete random walk with a position-dependent drift: at each instant, the country is pulled towards the nearest potential minimum (attractor basin) while being buffeted by stochastic shocks representing wars, economic crises, leadership changes, and other unpredictable events.

The connection to the stationary distribution is provided by the Fokker-Planck equation. For the Langevin dynamics in Equation (3), the stationary probability density satisfies:

Inverting this relationship yields the empirical potential function:

Up to a scaling constant and an additive constant, the potential is simply the negative logarithm of the observed probability density. This is a powerful result: it allows us to estimate the potential landscape directly from the data without assuming a parametric form for V(L).

6.2 Empirical Potential Estimation

We estimate p(L) nonparametrically using kernel density estimation (KDE) with a Gaussian kernel. The bandwidth is selected using Silverman's rule of thumb, adjusted upward to h = 3.0 to avoid spurious local optima in the resulting potential function. The empirical potential is then computed as V(L) = −log p(L), shifted so that min(V) = 0.

Figure 2 shows the estimated potential landscape. Three clear local minima are visible, corresponding to the three attractor basins identified by the GMM analysis.

Figure 2. Empirical potential landscape V(L) = −log p(L). Local minima correspond to attractor basins; local maxima (saddle points) correspond to energy barriers between basins. The tyranny well at L ≈ 10 is the deepest basin; the hybrid trap at L ≈ 48 is the shallowest; the democratic plateau at L ≈ 88 is of intermediate depth. Country positions are illustrative, based on most recent observations.

6.3 Parametric Potential Models

To quantify the basin structure, we fit three parametric potential models to the empirical potential function: a single-Gaussian (parabolic) well, a double-Gaussian well, and a triple-Gaussian well. The triple-Gaussian model takes the form:

Parameters are estimated by minimising the residual sum of squares (RSS) between the empirical and model potential functions, using the Nelder-Mead simplex algorithm with 30 random restarts to avoid local minima. Model selection proceeds via BIC computed from RSS.

The triple-well model is decisively preferred, with RSS reduced by 76% relative to the double-well model and 96% relative to the single-well model. The BIC differences are substantial (ΔBIC = 68.5 for double vs. triple, 285.6 for single vs. triple), confirming that the additional parameters of the triple-well specification are justified by the data.

6.4 Basin Properties

The estimated triple-well model reveals the following basin structure. The tyranny well, centred at L ≈ 10, is the deepest basin in the landscape. Countries that fall into this well face the highest energy barrier to escape, consistent with the empirical observation that consolidated autocracies rarely democratise without extraordinary exogenous shocks (Geddes, Wright, and Frantz 2018). The democratic plateau, centred at L ≈ 88, is an elevated basin of intermediate depth, stabilised by the redundancy of democratic institutions: free press, independent judiciary, competitive elections, and civil society organisations each independently resist erosion. The hybrid trap, centred at L ≈ 48, is the shallowest basin, consistent with the observation that countries in this zone exhibit high volatility, and frequent partial transitions.

The two saddle points—the local maxima separating adjacent basins—are located at approximately L ≈ 28 (between tyranny and hybrid) and L ≈ 68 (between hybrid and democracy). These saddle points represent the energy barriers that must be overcome for a country to transition between basins. The barrier between tyranny and hybrid is higher than the barrier between hybrid and democracy, consistent with the observation that escaping consolidated autocracy is more difficult than escaping the hybrid trap.

6.5 Mean First-Passage Times

The Langevin framework permits the calculation of mean first-passage times (MFPTs)—the expected time for a country to transition from one basin to another. For an overdamped particle in a one-dimensional potential, the Kramers formula gives the MFPT from a potential minimum at position a to escape over a barrier at position b as:

where ΔV = V(b) − V(a) is the barrier height, and ω_a, ω_b are the curvatures of the potential at the minimum and saddle point, respectively. The exponential dependence on barrier height means that small increases in ΔV translate to large increases in escape time, explaining the remarkable persistence of consolidated autocracies, and consolidated democracies alike.

7. The Critical Instability Zone

7.1 Identifying the Threshold

A central finding of our analysis is the identification of a Critical Instability Zone at L ≈ 52–55, below which the probability of democratic recovery collapses. We arrive at this estimate through three independent methods that converge on the same narrow range.

Method 1: Survival analysis. We compute the proportion of country-year observations at each liberty score that subsequently recover to L ≥ 70 within a 15-year window. This recovery rate exhibits a sharp discontinuity at L ≈ 52–55: above this threshold, approximately 82% of observations recover; below it, only 3.0% do.

Method 2: Markov transition probabilities. We estimate a discrete Markov transition matrix with 10-point liberty score bins. The transition probabilities reveal an asymmetry at L ≈ 50–55: above this level, the probability of upward transition exceeds the probability of downward transition; below it, the relationship reverses, and countries are more likely to continue declining than to recover.

Method 3: Potential landscape analysis. The empirical potential function V(L) exhibits an inflection point at approximately L ≈ 53, where the potential gradient shifts from pulling countries upward (towards the democratic plateau) to pulling them downward (towards the hybrid trap or tyranny well). This inflection point corresponds to the ridgeline between the hybrid trap basin and the democratic plateau basin in the potential landscape.

7.2 Bootstrap Confidence Intervals

To quantify the uncertainty around the critical threshold and the recovery probability, we conduct a nonparametric bootstrap with 1,000 iterations. In each iteration, we resample the 1,656 country-year observations with replacement, recompute the recovery rates at each liberty score level, and identify the threshold at which the recovery rate falls below 10%. The distribution of bootstrapped thresholds has a median of 53.2 and a 95% confidence interval of [50.8, 56.1]. The recovery rate below the threshold has a bootstrap distribution with median 3.0% and 95% CI [0.7%, 6.0%].

7.3 AR(1) Persistence

A striking feature of the data is the dominance of simple autoregressive persistence over more complex transition models. We estimate the following AR(1) specification:

The estimated persistence parameter is β = 0.96 (SE = 0.008, p < 0.001), with R² = 0.872. This means that a country's liberty score in any given period is overwhelmingly determined by its score in the previous period, with only marginal contributions from other factors. The implied long-run equilibrium is L* = α / (1 − β) ≈ 81.6, which falls within the democratic plateau basin identified by the GMM.

We compare the AR(1) specification against a range of stage-based transition models that assign countries to discrete categories and estimate transition probabilities between them. The AR(1) model dominates all stage-based specifications:

This finding does not invalidate the attractor basin framework—the GMM and potential landscape results remain robust—but it does suggest that the basins exert their influence through a continuous, gradualist mechanism rather than through discrete threshold effects. Countries do not "jump" from one basin to another; they drift slowly under the combined influence of institutional momentum (β = 0.96), mean-reverting forces (the potential gradient), and stochastic shocks. The critical instability threshold at L ≈ 52–55 marks the point where the direction of the mean-reverting force changes sign, not a discrete barrier that countries cross.

8. Robustness and Limitations

8.1 Tyranny as Residual

The most significant structural limitation of the analysis is that Tyranny (T) is computed as a residual rather than measured independently. Since T = 100 − L − C by construction, any measurement error in L, or C is mechanically transmitted to T with the opposite sign. If Freedom House systematically overstates a country's liberty score, the framework will correspondingly understate its tyranny. More subtly, the ternary constraint imposes a zero-sum structure that may not accurately reflect all political dynamics: it is at least conceivable that a country could simultaneously increase both liberty and state capacity (reducing chaos) without a corresponding increase in tyranny.

We emphasise that the ternary constraint holds definitionally in our framework, not as an independent empirical finding. Future research should develop independent measures of tyranny—such as indices of executive concentration, political prisoner counts, surveillance intensity, or extrajudicial violence—and test whether the ternary constraint holds empirically when all three components are measured independently.

8.2 Crosswalk Accuracy

The 67% crosswalk match rate between our derived liberty scores and the published Freedom House classification means that one-third of country-year observations exhibit non-trivial divergence from the field's standard reference point. Some of this divergence is by design: the Political Topology Index incorporates institutional erosion signals that FH updates more slowly. However, a 33% disagreement rate is substantial and warrants ongoing calibration effort.

8.3 Implementation Constraints

All statistical analysis in this paper is performed using Python's standard library (csv, math, statistics, random modules). No third-party packages (scipy, numpy, statsmodels) are used. This constraint, adopted for auditing, and reproducibility purposes, limits the sophistication of available statistical methods. In particular, our bootstrap confidence intervals are computed from percentiles of the bootstrap distribution rather than from bias-corrected and accelerated (BCa) methods, and our regression standard errors assume homoskedasticity rather than using heteroskedasticity-robust (HC) estimators. We view this as an acceptable trade-off: every line of analytical code is inspectable and self-contained, which enhances reproducibility at the cost of some statistical refinement.

8.4 Temporal Coverage and Interpolation

Observations are concentrated at historical inflection points rather than distributed at regular intervals, with pre-1972 data relying heavily on V-Dem and Polity crosswalks. Linear interpolation between observation points imposes a smoothness assumption that may mask rapid within-period transitions. The time spacing varies from 1 year (for modern observations) to 10–20 years (for 19th-century observations), creating heterogeneous precision across the sample. This irregular spacing complicates the interpretation of the AR(1) parameter: β = 0.96 per period is meaningfully different depending on whether the period is one year or ten years.

8.5 Potential Endogeneity

Our analysis treats liberty scores as the dependent variable without modelling the structural determinants of regime type (economic development, inequality, ethnic fractionalisation, geopolitical environment). This omission raises concerns about endogeneity: variables that simultaneously affect both the current liberty score and the probability of future transition (such as GDP per capita) may bias our estimates of persistence and transition probabilities. Acemoglu et al. (2008) demonstrated that the observed correlation between income and democracy is largely spurious once country fixed effects are included, suggesting that cross-sectional comparisons may overstate the causal role of economic development. Our panel structure provides some protection against this concern, but the irregular spacing, and limited within-country variation limit our ability to include country fixed effects reliably.

8.6 Small N for Country-Specific Claims

Whilst the pooled sample of 1,656 observations provides adequate statistical power for the GMM and potential landscape analyses, country-specific claims rest on much smaller samples. The United States, for example, has 34 observations over 225 years. Statistical inferences about individual country trajectories should be treated as illustrative rather than definitive, with confidence intervals substantially wider than those reported for the pooled sample.

8.7 Sensitivity to Classification Boundaries

The assignment of countries to basins depends on the chosen boundary values (L < 20 for tyranny, L = 20–70 for hybrid, L > 80 for democracy). These boundaries are derived from the GMM component means and standard deviations but inevitably involve some arbitrariness at the margins. Zone velocity estimates are sensitive to whether countries are classified by their starting-zone or ending-zone assignment: starting-zone assignment yields materially different results for the tyranny basin (e.g., +0.72/yr starting vs. −0.64/yr ending). This sensitivity should temper the interpretation of basin-specific dynamics.

9. Discussion

9.1 Relationship to Existing Theory

Our results speak directly to several unresolved debates in comparative politics.

The Levitsky-Way thesis vindicated. The hybrid trap basin provides the first formal statistical vindication of Levitsky and Way's (2010) central argument that competitive authoritarian regimes are not merely "incomplete transitions" towards democracy but constitute a distinct regime type with self-reinforcing logic. Where Levitsky and Way argued qualitatively that competitive authoritarian regimes can persist for decades through manipulated elections and constrained opposition, we quantify this persistence: the hybrid basin at μ ≈ 47 captures approximately 34.5% of all country-year observations and exhibits substantial temporal persistence. The GMM formally identifies it as a distinct component of the distribution, not as a noisy overlap between two components.

Acemoglu-Robinson extended to three equilibria. Our tristable framework extends the Acemoglu-Robinson (2006) model by introducing a third equilibrium that their binary choice set cannot accommodate. Their game-theoretic model predicts either full concession (democracy) or repression (dictatorship); the empirical distribution reveals a large cluster of countries at intermediate liberty scores that are neither. The hybrid trap may correspond to a "partial concession" equilibrium in which elites permit limited political opening without ceding full control—a possibility that Acemoglu, Robinson, and Verdier (2004) explored in their analysis of kleptocracy but did not integrate into the main transition framework.

Huntington's waves reinterpreted. The tristable framework offers a micro-foundation for Huntington's (1991) macro-historical pattern of democratic waves and reverse waves. Waves correspond to periods when correlated exogenous shocks (e.g., post-war settlement, decolonisation, Soviet collapse) push multiple countries over saddle points simultaneously, while reverse waves represent coordinated slides back into attractor basins. Crucially, however, the tristable model does not predict a secular trend towards democracy. The asymmetric basin structure—with the tyranny well being the deepest attractor—implies that the long-run stationary distribution may favour autocracy absent sustained institutional investment, a finding that contradicts the Whig interpretation implicit in much of the democratisation literature.

Treisman's stochastic transitions formalised. Treisman's (2020) argument that many democratisations were "by mistake"—the unintended consequences of miscalculation by authoritarian incumbents—finds a natural formalisation in the Langevin framework. Democratisation-by-mistake corresponds to a large positive stochastic shock ε pushing a country over the saddle point from the hybrid trap into the democratic basin. The model predicts that such events should be rare, stochastic, and difficult to sustain, exactly the pattern Treisman documents empirically.

9.2 The Democratic Plateau as Engineered Stability

A notable feature of the potential landscape is that the democratic plateau is not the deepest basin. In the absence of institutional reinforcement, the natural resting state of the system—the global minimum of the potential function—would be the tyranny well. The stability of consolidated democracies is therefore not a natural equilibrium but an engineered one, maintained by the redundancy of institutional checks: when one institution is weakened, others compensate. Free press monitors judicial capture; courts check executive overreach; competitive elections allow peaceful replacement of incumbents; civil society organisations mobilise opposition to norm violations. The institutional ecology of democracy creates a self-reinforcing dynamic—but only so long as a sufficient number of these institutions remain functional.

This interpretation has a direct implication for the critical instability threshold. The threshold at L ≈ 52–55 may correspond to the point at which enough institutional checks have been degraded that the remaining institutions can no longer compensate. Below this point, the erosion of any one institution accelerates the erosion of all others, creating a positive feedback loop that drives the country downward through the hybrid trap towards the tyranny well. This is consistent with the eight-step erosion model (norm erosion, information capture, judicial capture, legislative subordination, regulatory capture, civil society suppression, electoral manipulation, constitutional consolidation), in which judicial capture (Step 3) is identified as the critical inflection point because an independent judiciary is the institution most central to the maintenance of all other institutional constraints.

9.3 Implications for Early-Warning Systems

The identification of a quantitative critical instability threshold has practical implications for the design of democratic early-warning systems. Current early-warning approaches (e.g., Goldstone et al. 2010; Hegre et al. 2013) rely on structural risk factors (GDP, ethnic fractionalisation, neighbourhood effects) that change slowly and are difficult to modify through intervention. Our framework suggests that monitoring the liberty score trajectory—specifically, whether a country is approaching the L ≈ 52–55 threshold from above—may provide a more actionable early-warning signal. The finding that recovery probability collapses from approximately 82% to 3% as countries cross this threshold implies that interventions to arrest democratic erosion are dramatically more effective when undertaken early, before the threshold is crossed.

9.4 The Role of Economic Development

Our analysis does not directly model the relationship between economic development and regime type, but the literature suggests important interactions. Przeworski et al. (2000) demonstrated that no democracy above approximately $6,000 per capita GDP (in 1985 PPP dollars) has ever collapsed. More recently, Treisman (2015) has argued that the income-democracy relationship is more nuanced than previously supposed, with economic modernisation creating both pro-democratic and pro-authoritarian pressures. The implication for our framework is that the depth and location of the attractor basins may vary with economic development: wealthier countries may have deeper democratic plateaus and shallower tyranny wells, effectively raising the energy barriers against democratic collapse. This is a promising direction for future research that would require incorporating GDP per capita as a time-varying covariate in the potential landscape estimation.

9.5 The Cost of Crossing Basins: The Justice Dividend

Recent work on the unit economics of institutional reform (see A11: “The Justice Dividend”) provides an economic valuation of basin transitions that complements the political dynamics modelled here. The cost of transitioning from the hybrid trap to the democratic plateau is not merely political—it requires sustained institutional investment in judicial independence, property rights enforcement, and corruption control. Using instrumental variable estimates from Acemoglu, Johnson, and Robinson (2001) and Rodrik, Subramanian, and Trebbi (2004), a one-standard-deviation improvement in institutional quality is associated with a 3–8× increase in long-run per-capita income. However, the lead–lag evidence suggests that such returns require 5–25 years to fully materialise, creating a “patience problem” that may explain why countries in the hybrid trap fail to initiate reforms even when the expected returns are overwhelmingly positive. The global cost of remaining in the tyranny well or hybrid trap—estimated at $2.6 trillion annually in corruption costs alone, plus $9.3 trillion in dead capital from property rights failure—represents the economic price of the basin structure we identify. The implication for democratic promotion is that the energy barrier between basins has an economic counterpart: the investment required to build the institutional infrastructure of the democratic plateau is modest relative to the compounding returns it generates, but the upfront cost, and delayed payoff create a collective action problem that the tristable dynamics model helps to formalise.

10. Conclusion

This paper has presented evidence that the political regime landscape is governed by tristable dynamics with three distinct attractor basins: a democratic plateau (L > 80), a hybrid trap (L ≈ 20–70), and a tyranny well (L < 20). Using a 225-year panel of 91 countries (N = 1,656), we have shown that:

1. A three-component Gaussian Mixture Model is decisively preferred by BIC over two-component and single-component alternatives (ΔBIC = 193.9 over K = 2), providing formal statistical evidence for three rather than two attractor basins.

2. The empirical potential function V(L) = −log p(L), derived from a Langevin SDE framework, exhibits three local minima of unequal depth. The tyranny well is the deepest attractor, suggesting that autocracy is the natural resting state absent sustained institutional investment in democratic governance.

3. A Critical Instability Zone exists at L ≈ 52–55, below which the probability of democratic recovery collapses to 3.0% (95% CI: 0.7–6.0%). Three independent estimation methods—survival analysis, Markov transition probabilities, and potential landscape inflection—converge on this range.

4. AR(1) persistence (β = 0.96) dominates all stage-based transition models (ΔAIC > 300), implying that regime dynamics are governed by continuous drift rather than discrete threshold effects, though the attractor basin structure shapes the drift direction.

These findings challenge the dominant bistable framework in comparative politics and suggest that the hybrid regime zone is not a transient way-station between democracy and autocracy but a distinct attractor with its own stabilising dynamics. The practical implication is sobering: countries that slip below the critical instability threshold face overwhelming odds against recovery, and interventions to arrest democratic erosion are far more effective when undertaken before the threshold is crossed than after.

We acknowledge significant limitations: Tyranny is measured as a residual rather than independently, the crosswalk between data sources achieves only 67% accuracy, the analysis is implemented using only Python standard library functions, and country-specific inferences rest on small samples. These limitations constrain the precision of our quantitative estimates without, we argue, undermining the qualitative finding that the regime landscape is tristable rather than bistable.

Future research should pursue four priorities. First, independent measurement of Tyranny would permit an empirical test of the ternary constraint rather than imposing it by construction. Second, a continuous-time Markov chain (CTMC) estimated via matrix exponentials would provide a unified framework that bridges the discrete transition data and the continuous Langevin dynamics. Third, the incorporation of time-varying covariates (GDP per capita, inequality, democratic tenure) would allow the basin structure to vary across countries and periods. Fourth, external validation against the forthcoming 2026 reports from Freedom House, and V-Dem would test whether the critical instability threshold identified here has predictive power for ongoing cases of democratic erosion.