Methodology specification
Stratos governance diagnostic. Version 0.1 (draft for public review).
This document specifies the analytical methodology underlying Stratos. It is intended for technical reviewers, contributors implementing scoring logic, and academic users who need to assess the validity of the dashboard's outputs.
The methodology has five components: regime classification, structural-constraint conditioning, dimensional scoring, uncertainty quantification, and diagnostic narrative generation. Each is specified in detail below.
The methodology is open to revision. Substantive changes are tracked in METHODOLOGY_CHANGELOG.md and discussed in public issues before adoption.
1. The conceptual frame
Stratos asks one core question for each country-year:
Given this state's regime classification, its structural constraints, and the validity of available data, how is it performing on five governance dimensions relative to its actual peers, and what is the direction and rate of change?
The answer is not a number. The answer is a five-dimensional profile with confidence intervals, peer comparisons, trajectory information, and a diagnostic narrative.
Three theoretical commitments shape the methodology.
First, governance is uncertainty navigation under constraint. States face exploration-exploitation trade-offs at every level of operation. The five dimensions are chosen to capture both sides of that trade-off and the distributional question that exploitation alone cannot answer.
Second, regime type is the operating system. Liberal democracies, electoral autocracies, and closed autocracies are not differently-performing instances of the same governance model. They are different governance models. Evaluating them on a single rubric mislabels both successes and failures.
Third, structural inheritance is causal. Geography, demography, resource endowment, and historical institutional legacy explain most of the cross-national variance in governance outcomes. A methodology that does not condition on these inheritances will repeatedly mistake luck for performance.
2. The five dimensions
Stratos collapses the conceptual space into five dimensions, each captured by a basket of indicators rather than a single proxy.
2.1 Execution capacity
Can the state do things.
This dimension captures the operational backbone of governance: whether the state can collect revenue, deliver services to its territory, regulate its economy, maintain order, and produce reliable statistics about its own activities.
Core indicator families:
- Fiscal capacity: tax revenue as percentage of GDP, government revenue as percentage of GDP, tax buoyancy
- Service delivery reach: electricity access, water access, sanitation access, internet access, geographic coverage of basic services
- Statistical capacity: World Bank Statistical Capacity Indicator, completeness of vital registration, frequency and revision history of national accounts
- Bureaucratic quality: World Bank Worldwide Bureaucracy Indicators (public employment, wage premium, wage compression)
- Regulatory infrastructure: Worldwide Governance Indicators regulatory quality, OECD product market regulation indicators where available
- Public order: homicide rate, conflict deaths, political violence events
This is the most regime-neutral of the five dimensions. Execution capacity matters in every regime type, and the indicators are comparable across regimes with appropriate care.
2.2 Conversion efficiency
Does spending become outcomes.
This dimension captures the relationship between resource inputs and human outcomes. It is the core efficiency question: per dollar of revenue, per percentage of GDP spent on health or education, what outcomes does the state actually produce.
Core indicator families:
- Health conversion: life expectancy at birth, infant and maternal mortality, universal health coverage index, avoidable mortality, per public health expenditure per capita
- Education conversion: learning-adjusted years of schooling, completion rates, learning poverty, per public education expenditure per capita
- Infrastructure conversion: logistics performance, road quality, transport access, per public infrastructure investment
The dimension uses paired input-output ratios with confidence bands, not raw outcomes. A country with high life expectancy might be producing that outcome efficiently or inefficiently depending on its spending baseline.
This is the dimension where Probability Cloud Regression provides the most value, because both inputs and outputs are measured with substantial error and the relationship between them is the quantity of interest.
2.3 Distributional reach
Who benefits.
This dimension captures whether governance outcomes are concentrated among elites and dominant regions or distributed across the population. A high average outcome with severe distributional concentration is a different result from the same average with broad distribution.
Core indicator families:
- Income inequality: Gini coefficient (post-tax post-transfer where available), Palma ratio, income share of bottom 40 percent
- Multidimensional inequality: Inequality-Adjusted HDI, UNDP Gender Inequality Index, regional HDI variance
- Mobility: where measured, intergenerational mobility indicators
- Poverty: poverty headcount at international and national poverty lines, multidimensional poverty index
- Demographic gaps: gender labour force participation gap, youth unemployment, minority and indigenous outcome gaps where consistently measured
- Spatial inequality: subnational HDI variance, urban-rural service access gaps
The dimension uses the loss between HDI and IHDI as one anchor and supplements with more granular indicators where available. Data quality on this dimension is uneven across regime types and a confidence indicator accompanies each country's score.
2.4 Accountability infrastructure
Is power answerable.
This dimension captures the institutional channels through which power can be checked, criticised, contested, and replaced. It is the most politically loaded of the five dimensions and the most methodologically careful in how it handles regime difference.
Core indicator families:
- Electoral integrity: V-Dem electoral democracy index, fairness of elections, suffrage extent
- Civil liberties: V-Dem civil liberties index, freedom of expression, freedom of association, freedom of religion
- Judicial independence: V-Dem judicial constraints on the executive, independent judicial appointments, judicial corruption
- Legislative constraints: V-Dem legislative constraints on the executive, oversight bodies, audit capacity
- Press freedom: Reporters Without Borders index, V-Dem media freedom indicators
- Civil society space: CIVICUS civic space rating, V-Dem civil society repression
- Anti-corruption mechanisms: Transparency International CPI, V-Dem accountability indices
This dimension is scored differently for democracies and autocracies. For democracies, the question is the strength of accountability infrastructure. For autocracies, the question is partly about formal infrastructure (often weak) and partly about within-regime accountability mechanisms (party discipline, succession norms, etc.). The dashboard does not pretend these are equivalent. It scores both honestly within their regime context and flags the comparison limits.
This is a political choice. We name it explicitly: Stratos treats accountability infrastructure as a positive governance feature. Users who hold a different view about the trade-off between accountability and other governance virtues can adjust dimensional weights in the comparison view.
2.5 Adaptive capacity
Can the state respond to novelty and shock.
This dimension captures the state's capacity to learn, experiment, update, and respond to changing conditions. It is the exploration-side complement to execution capacity's exploitation focus.
Core indicator families:
- Fiscal space: debt-to-GDP, interest-to-revenue, currency stability, foreign reserves adequacy
- Climate adaptation: ND-GAIN readiness, climate adaptation planning, disaster preparedness, early warning systems
- Innovation infrastructure: R&D as percentage of GDP, patents per capita, scientific publications, tertiary education completion
- Digital state capacity: UN E-Government Development Index, digital ID adoption, paperless bureaucracy indicators
- Regulatory updating: OECD regulatory governance indicators, policy experimentation indicators where measured
- Food and energy security: food import dependence, energy import dependence, agricultural productivity
Adaptive capacity is harder to measure than execution capacity because the relevant phenomena (learning, updating, experimentation) are often invisible in standard statistics. The dimension currently relies on proxies. Improvement is a known priority.
3. Regime classification
Every country-year is assigned to a regime cluster before any cross-country comparison. The classification scheme uses V-Dem regime types as the primary input and supplements with capacity indicators.
3.1 Cluster definitions
Cluster A: high-capacity liberal democracies. V-Dem liberal democracy index above 0.7, government effectiveness above 1.0, GDP per capita above $25,000 PPP. Approximately 25-30 countries.
Cluster B: established electoral democracies. V-Dem liberal democracy index between 0.5 and 0.7, free and fair elections regularly held, varied capacity levels. Approximately 30-40 countries.
Cluster C: emerging or constrained electoral democracies. V-Dem electoral democracy index above 0.5 but with significant capacity limitations or accountability weaknesses. Approximately 30-35 countries.
Cluster D: electoral autocracies. Elections held but not free or fair. V-Dem regimes of the world classification of electoral autocracy. Approximately 30-35 countries.
Cluster E: closed autocracies. No meaningful elections. V-Dem closed autocracy classification. Approximately 20-25 countries.
Cluster F: fragile or contested authority. States with limited control over territory, post-conflict transitions, or contested sovereignty. Approximately 15-20 countries.
Classifications are revisited annually based on updated V-Dem data. Transitions are flagged on the country profile.
3.2 Comparison rules
Within-cluster comparison is the default view on every country profile. Peer countries are drawn from the same cluster, with secondary filtering on structural-similarity variables (see section 4).
Cross-cluster comparison is available but framed as comparing different operating models. The dashboard does not display cross-cluster percentile ranks without an explicit user action and a contextual note about what such a comparison can and cannot tell you.
3.3 Edge cases and contested classifications
Some countries sit on cluster boundaries or have contested classifications. The methodology handles these in three ways:
- Soft assignment. Where a country sits near a cluster boundary, it appears in peer comparisons for both clusters with appropriate weighting.
- Transition flagging. Countries that have moved between clusters in the past five years carry a transition flag with the direction and confidence of the move.
- External classification disagreement. Where V-Dem classification disagrees with other major democracy classifications (Freedom House, Economist Intelligence Unit Democracy Index, etc.), the country profile shows the disagreement explicitly.
4. Structural-constraint conditioning
Before scoring, each country is conditioned on structural variables it did not choose. The conditioning is implemented through hierarchical modelling and peer matching.
4.1 Conditioning variables
The variables used for structural conditioning are selected on three criteria: substantial cross-national variance, evidence of causal influence on governance outcomes, and reasonable data availability across all regime clusters.
The current variable set:
- Geography: landlocked status, small island state status, climate band (tropical, temperate, polar)
- Population: total population (log), population density
- Demographic structure: median age, dependency ratio, urbanisation rate
- Resource endowment: natural resource rents as percentage of GDP, fossil fuel export dependence
- Historical institutional legacy: colonial origin (proxy for institutional inheritance, coded carefully), post-Soviet status, decades since independence
- Conflict exposure: ten-year history of major armed conflict, ongoing conflict status
- Baseline disease burden: malaria endemic status, HIV prevalence baseline
The conditioning variables are explicit. Users can see exactly what is being controlled for in any peer comparison. We resist the temptation to over-condition, which would let bad governments off the hook by absorbing all variance into structural inheritance.
4.2 Implementation
Conditioning is implemented through Bayesian hierarchical modelling in the scoring layer. For each indicator, the model estimates expected performance given the structural variables, and the country's score is the deviation from that expectation with appropriate uncertainty bands.
The dashboard shows both adjusted and unadjusted views. The adjusted view answers "how is this government performing given what it inherited." The unadjusted view answers "how is this country doing in absolute terms." Both are legitimate and the user chooses which to read.
The conditioning model is published alongside the dashboard. Specification details, prior choices, and diagnostic plots are in the methodology repository.
5. Indicator selection and weighting
Within each dimension, multiple indicators contribute to the dimensional score. Selection and weighting follow four principles.
5.1 Principles
Construct validity. Each indicator must plausibly measure the construct it is assigned to. The dimensional scoring is not exploratory factor analysis. It is theory-driven measurement.
Independence. Indicators within a dimension should capture different aspects of the construct, not redundant measurements of the same aspect. We use correlation diagnostics to identify and remove redundancy.
Coverage. Indicators with poor coverage across regime clusters are deprioritised. Where an indicator is only available for high-capacity democracies, it cannot drive cross-regime comparison.
Source diversity. Each dimension draws on indicators from multiple sources to reduce single-source bias. No dimension is more than 50 percent dependent on any single data provider.
5.2 Weighting
Within-dimension weighting starts at equal weights and adjusts based on three factors:
- Reliability: indicators with higher measurement reliability receive higher weight
- Coverage: indicators with poorer cross-country coverage receive lower weight to avoid biasing comparisons
- Theory: where strong theoretical or empirical evidence supports a particular indicator's primacy, weights can be adjusted
The weights are explicit, documented, and adjustable in the user interface. Users can see how the score changes under different weighting schemes. This is a form of robustness analysis built into the product.
Cross-dimensional aggregation is not done. There is no single "Stratos score." The dashboard shows the five-dimensional profile. Users who want a composite can compute one from the published dimensional scores, but the product does not encourage this and does not provide a default aggregation.
6. Uncertainty quantification
This is the methodologically distinctive component of Stratos. Most governance dashboards report point estimates as if the underlying indicators were measured precisely. They are not.
6.1 Sources of uncertainty
Four sources of uncertainty are quantified:
- Measurement error in indicators. Survey-based indicators have sampling error. Administrative indicators have revision and reporting error. Both are quantified where the data provider supplies confidence information, and estimated where they do not.
- Model uncertainty in conditioning. The hierarchical conditioning model has parameter uncertainty that propagates to country-level estimates.
- Indicator-to-construct uncertainty. Each indicator measures its target construct imperfectly. Multi-indicator dimensional scoring partially addresses this through aggregation, but residual construct-validity uncertainty remains.
- Cross-source disagreement. Where multiple sources provide overlapping indicators, disagreement between sources is treated as an uncertainty signal.
6.2 Probability Cloud Regression on conversion efficiency
The conversion-efficiency dimension is the natural fit for Probability Cloud Regression because both inputs (spending, resources) and outputs (outcomes) are measured with substantial error, and the relationship between them is the quantity of interest.
PCR is applied to the input-output relationships within this dimension, producing posterior distributions over the conversion-efficiency estimates rather than point values. The dashboard displays these as confidence bands on the dimensional score.
PCR remains the intellectual property of its developer (Soliman, 2026). It is integrated into Stratos as a methodology dependency. The Stratos codebase calls a PCR implementation maintained in a separate repository. See OPEN_SOURCE_GOVERNANCE.md for the IP boundary specification.
6.3 Display
Every score on the dashboard is displayed with its confidence band, not as a point estimate. The band is wider for countries with poorer data quality and narrower for countries with better data. This is visible in the user interface, not hidden behind a click.
Confidence intervals are 80 percent by default, with 95 percent available as a user toggle. The choice of 80 percent as default reflects the practical reality that most governance comparisons are decision-relevant at this confidence level and that wider intervals risk being read as licence to ignore the data.
7. Data quality and validity flagging
Each country profile carries an explicit data-quality assessment alongside its dimensional scores.
7.1 The data quality score
A composite data-quality indicator is computed for each country-year based on:
- Coverage: percentage of indicators available
- Recency: average age of the most recent data points
- Source diversity: number of independent sources contributing
- Cross-source agreement: variance between sources reporting overlapping indicators
- Statistical capacity: World Bank Statistical Capacity Indicator
- Independent verification: whether key indicators are corroborated by non-state sources (satellite imagery, third-party surveys, civil society monitoring)
The score is displayed prominently on the country profile and warnings are raised when the score falls below a threshold.
7.2 Known-falsification flagging
Where credible evidence exists that a country has falsified statistical reports (Argentina inflation pre-2015, persistent China provincial GDP overstatement, fictional electoral turnout in specific autocracies), the affected indicators are flagged and either excluded or replaced with alternative-source estimates.
The flagging logic is explicit and documented. Flag decisions are public and contestable through the issue tracker.
7.3 Missingness handling
Missing data is not imputed silently. Where an indicator is missing for a country, the dimensional score is computed from available indicators with explicit notation that the score is partial.
Where missingness is severe (more than 40 percent of indicators in a dimension missing), the dimensional score is suppressed and the user sees a missingness warning instead of a number.
8. The diagnostic engine
The diagnostic narrative is the killer feature. This section specifies how it is produced.
8.1 Inputs
For each country, the diagnostic engine takes:
- Five-dimensional scores with confidence bands
- Regime classification and any transition flags
- Structural-constraint profile
- Within-cluster peer rankings on each dimension
- Five-year and ten-year trajectory data
- Data quality score and known-issue flags
- Cross-source disagreement signals
8.2 Output structure
The diagnostic produces a structured narrative covering:
- Headline assessment: one-paragraph summary of the country's governance profile, written in plain language
- Strengths: dimensions where the country outperforms its regime cluster and structural peers, with reasoning
- Gaps: dimensions where the country underperforms, with reasoning
- Trajectory: direction and rate of change with confidence assessment
- Constraints: structural factors that materially shape the score
- Data limits: what the available data can and cannot tell us about this country
8.3 Generation
The diagnostic engine uses templated narrative generation with parameterised content. Templates are written by the project team, parameterised on the inputs above. The dashboard does not use free-form LLM generation for the public diagnostic, because such generation is unauditable and inconsistent across countries.
LLM use in the diagnostic engine is limited to two roles:
- Pre-publication review: an LLM reads the generated diagnostic and flags any internal inconsistencies, factual errors, or tone issues for human review
- Translation: diagnostic narratives are translated into multiple languages with LLM assistance, with human-language-expert review for high-traffic languages
The templating approach is auditable, consistent, and reproducible. Country profiles can be regenerated identically from the same inputs.
9. Trajectory analysis
Each country's profile includes a trajectory module showing the direction and rate of change on each dimension over five and ten year windows.
9.1 What we compute
For each dimension, three trajectory indicators:
- Direction: improvement, decline, or stability
- Rate: magnitude of change relative to within-cluster variance
- Consistency: monotonic trend or volatile
Trajectory analysis uses the same uncertainty quantification as point-in-time scoring. A country with a noisy improvement signal is distinguished from one with a clean improvement signal.
9.2 Early-warning signals
For Cluster A through C countries, an early-warning module flags trajectories consistent with democratic backsliding patterns documented in the comparative politics literature. The module is conservative: it flags concerns for human review rather than autonomously labelling backsliding.
This is the component most analogous to trajectory-geometry methodology in adjacent research. Detailed specification is deferred to a planned methodology extension.
10. Methodology governance
The methodology will evolve. Changes are governed by three rules.
First, substantive methodology changes require an open issue, a public discussion window of at least 14 days, and a recorded decision. The decision can be made by the maintainer or maintainer team, but the discussion must be public and the reasoning recorded.
Second, all methodology changes are versioned. Dashboard outputs are stamped with the methodology version that produced them. Historical outputs remain reproducible.
Third, methodology critiques receive responses. Issues raising substantive methodological concerns are triaged within seven days. The response may be agreement, disagreement with reasoning, or deferral for further analysis. The response is public.
Appendix: known limitations and open questions
Three known limitations are flagged explicitly here.
Subnational governance is invisible. Stratos operates at the country level. Federal countries with significant subnational variance (United States, India, Brazil, etc.) are inadequately represented by national-level data. A subnational extension is on the long-term roadmap but is not in the MVP scope.
Time-lag attribution is approximate. Current outcomes reflect past policies, often by decades. The methodology partially addresses this through trajectory analysis and structural conditioning but cannot fully attribute outcomes to current administrations. The diagnostic narratives flag this where it is materially relevant.
Authoritarian state capacity is partly invisible. The accountability dimension explicitly handles regime difference. The capacity dimensions face a subtler problem: state capacity in autocracies may be high but directed at coercion rather than service delivery. The methodology partially addresses this through the coercive-stability indicators but acknowledges remaining limits.
Open methodological questions are tracked in the issue tracker under the methodology label.