02 / METHODS & FALSIFIABILITY

Claims must be traceable, testable, and revisable.

Civilization Science separates definitions, observations, models, interpretations, and normative proposals. No equation or index is treated as valid merely because it is elegant.

METHODOLOGICAL STATUS

The stability equation and associated constructs are working hypotheses within an open research framework. They have not yet been validated as universal scientific laws. The program advances only through transparent evidence, criticism, replication, and revision.

From concept to correction

Every research claim should pass through a documented chain. Completion of the chain does not prove a claim; it makes the claim inspectable, reproducible, and open to rejection.

Concept operationalization

Translate abstract concepts into observable constructs, units, scales, and boundary conditions.

Variable definition

Specify symbols, direction, level of analysis, time window, and plausible alternative explanations.

Data sourcing

Record origin, collection method, license, coverage, exclusions, bias, and revision history.

Indicator construction

Publish transformations, normalization, weighting, aggregation, missing-data rules, and uncertainty.

Multi-agent simulation

Make agent rules, network topology, parameters, random seeds, and output metrics reproducible.

Historical backtesting

Freeze the model at a past cut-off and test it only against information available at that time.

Comparative cases

Use explicit case-selection rules and include difficult, negative, and disconfirming cases.

Sensitivity analysis

Vary assumptions, parameters, weights, lags, thresholds, and data treatments across plausible ranges.

Model validation

Assess construct, measurement, internal, external, predictive, and decision validity separately.

Peer review

Invite disciplinary, statistical, computational, historical, regional, and ethical review.

Corrections & versioning

Publish corrections, preserve earlier versions, and document material changes and their consequences.

Falsifiability statement

State in advance which observations would weaken, bound, reject, or replace a proposition.

Working constructs and variables

The stability expression is a research architecture. Each variable must be estimated through multiple observable indicators; no single proxy should be treated as the construct itself.

RESEARCH PROTOCOLWorking hypothesis: S = (C × Lₑ × Eₙ) / E

Multiplicative form, direction of effects, functional form, lags, interactions, and scale dependence are all testable assumptions—not settled facts.

Symbol	Working construct	Possible observations	Required caution
S	Civilizational stability	Continuity of essential functions; recovery time; institutional reliability; violence and displacement; adaptive capacity	Define system, population, horizon, and acceptable trade-offs; stability is not moral legitimacy.
C	Cooperation structure	Network connectivity; collective-action completion; institutional coordination; reciprocity; conflict-resolution performance	Distinguish voluntary cooperation from coercive compliance and unequal extraction.
Lₑ	Trust field	Interpersonal and institutional trust; expectation reliability; verification cost; information credibility	Survey trust, behavioral trust, and institutional reliability are related but not interchangeable.
Eₙ	Effective energy	Usable material, financial, informational, organizational, and technological capacity available for adaptation	Measure accessible and productive capacity, not gross resource stocks alone; account for distribution and conversion losses.
E	Civilizational entropy	Coordination loss; institutional friction; corruption; information disorder; fragmentation; irreversible waste	Entropy is an operational analogy unless a formally defined measure and conservation logic are specified.

Data sources and provenance

A result is only as auditable as its evidence trail. Each observation should retain a machine-readable source record and a human-readable limitation note.

Official and administrative data

Censuses, budgets, health, education, justice, infrastructure, trade, conflict, and institutional records.

Surveys and panels

Repeated measures of trust, expectations, behavior, well-being, legitimacy, and cooperation.

Historical and event data

Archives, contemporaneous records, chronologies, coded events, institutional histories, and case reconstructions.

Environmental and digital traces

Remote sensing, climate and ecological observations, mobility, network, communication, and open-platform data where ethically and legally permitted.

Indicator construction

Composite indicators are useful only when their construction remains transparent and when conclusions survive reasonable alternatives.

Define

Link every indicator to a stated construct and causal rationale.

Inspect

Report distribution, coverage, missingness, outliers, breaks, and measurement invariance.

Normalize

Justify scale direction, baseline, winsorization, standardization, and treatment of bounded values.

Weight

Prefer theory- or evidence-based weights; publish equal-weight and alternative-weight results.

Aggregate

Test additive, multiplicative, threshold, and non-compensatory forms where relevant.

Uncertainty

Propagate sampling, measurement, imputation, parameter, and model uncertainty into reported results.

No ranking, warning score, or composite index should be published without component values, uncertainty, sensitivity results, and a clear statement of non-comparability where data quality differs.

Simulation, backtesting, and comparative cases

Multi-agent simulation

Define agent types, states, incentives, information, learning, interaction rules, network topology, shocks, institutional constraints, stopping rules, and outcome metrics. Publish code, configuration, random seeds, calibration targets, and repeated-run distributions whenever lawful and practical.

Historical backtesting

Use rolling or fixed historical cut-offs; prevent future-data leakage; define events and horizons before testing; compare against simple baselines; report false positives, false negatives, calibration, lead time, and performance decay.

Comparative cases

Select cases by explicit rules rather than desired outcomes. Combine most-similar, most-different, typical, deviant, negative, and boundary cases. Preserve historical context and do not treat coded observations as substitutes for primary-source interpretation.

Sensitivity analysis and model validation

Validation is multidimensional. A model may measure a construct reasonably yet fail to predict, generalize, or support decisions.

Construct validity

Do indicators represent the intended concept rather than convenience, visibility, or institutional bias?

Measurement validity

Are observations reliable across sources, languages, regions, periods, and subpopulations?

Internal validity

Are associations robust to confounding, reverse causality, selection, and alternative specifications?

External validity

Do findings generalize beyond the calibration cases and conditions?

Predictive validity

Does preregistered out-of-sample performance exceed transparent baselines with useful calibration?

Decision validity

Do proposed uses improve outcomes without unacceptable harm, rights violations, or distributional failure?

Required sensitivity tests

Parameter and weight sweepsAlternative operational definitionsLag and time-window variationMissing-data and outlier treatmentsNetwork topology and agent-rule variationPlacebo outcomes and negative controlsLeave-one-case-out and subgroup analysisMonte Carlo and uncertainty propagation

Peer review, corrections, and version management

Academic credibility depends on visible correction, not the appearance of never being wrong.

Review before authority

Preprints, protocols, datasets, code, and major claims should be reviewed by relevant specialists. Conflicts of interest and reviewer scope should be disclosed.

Reproducible record

Archive data dictionaries, code, environments, seeds, model cards, study protocols, analysis plans, and output checksums where permitted.

Correction classes

Label minor corrections, analytical corrections, material revisions, retractions, and superseded models differently. Never silently replace a consequential result.

Semantic versioning

Major: construct or conclusion changes. Minor: compatible method or dataset additions. Patch: factual, typographic, or implementation corrections that do not alter conclusions.

Falsifiability statement

The central equation and related indicators are working hypotheses. They should be weakened, bounded, rejected, or replaced when transparent tests fail.

Directional failure

If preregistered, adequately measured studies repeatedly find no expected association—or a robust opposite association—between C, Lₑ, Eₙ, E, and S under stated conditions, the relevant directional claim must be revised or rejected.

Predictive failure

If the model does not outperform simple, transparent baselines in independent out-of-sample tests, it should not be presented as a forecasting or early-warning improvement.

Non-robustness

If conclusions depend on one proxy, weight, case, threshold, time window, or undocumented data treatment and disappear across plausible alternatives, the claim is not robust.

Failure to generalize

If effects do not transfer across preregistered populations, periods, or system types, the scope must be narrowed rather than declared universal.

Causal failure

Associations must not be described as causal when experimental, quasi-experimental, longitudinal, or process evidence supports credible rival explanations.

Decision failure

If model-guided interventions produce no benefit or create unacceptable harm, inequality, coercion, or rights violations, the decision claim must be withdrawn even if descriptive fit remains.

No book, website, simulation, equation, author, or institution can exempt a proposition from these conditions. Evidence has priority over the framework.