Does Big Government Kill Growth? The Armey Curve Tested on 151 Countries

Show Real Country Data:

Real-world data from World Bank API (Government expenditure & GDP growth)

Time Period for Averaging: Longer periods smooth volatility but may include outdated regimes

Loading...

Chart displaying the Armey Curve with government spending on x-axis and GDP growth rate on y-axis

Model Fit Ranking

Each model is auto-fitted to find its best-case parameters before computing AIC/R² — a fair competition at peak performance. Bold row = currently selected model. 95% CI computed via Fisher’s Z transformation on the sample correlation (N = 113, standard exclusions applied).

R² (R-squared)

Measures how well the curve explains the variation in growth rates across countries. A perfect fit = 1.0. A score of 0 means the model is no better than just predicting the average growth rate for every country. Negative values mean the model is worse than that baseline. When models are auto-fitted across 113 countries (resource-dependent, externally-funded, conflict/fragile, and GDP-distorted countries excluded), the top models reach R² ≈ 0.42 — meaning government spending % explains roughly 42% of the variation in growth. Note that Power Law and Inverse are tied on combined R² (0.4913); Power Law edges ahead on AIC (53.29 vs 53.82) — it fits the spending–growth relationship well across diverse economies. Each exclusion category has a specific theoretical justification: resource-dependent economies grow via commodity windfalls unrelated to government size; externally-funded states have aid-distorted budgets; conflict/fragile states have suppressed growth from instability; GDP-distorted economies have artificially inflated or deflated GDP figures.

AIC (Akaike Information Criterion)

Ranks competing models by balancing how well they fit the data against how many parameters they use (simpler models are rewarded). Lower AIC = better model. Unlike R², AIC is useful for comparing models even when none of them fits well — it tells you which is the least bad option. With 113 countries (standard exclusions), Power Law (AIC 53.29) and Inverse (AIC 53.82) are effectively tied, with Log-Linear (54.31) close behind. Exponential (AIC 55.48) and Quadratic (62.17) trail — the constrained Armey Curve shape finds no empirical support in this dataset.

p-value

The probability of observing a fit this strong (or stronger) by chance alone, assuming government spending has no real effect on growth. Lower p-value = stronger evidence against the null hypothesis. A threshold of 0.05 is conventional; all models except Quadratic show p < 0.001, meaning there is less than a 1-in-1000 chance the observed relationship is a statistical accident. The p-value is derived from an F-test on the regression: F = (R² / k) / ((1 − R²) / (N − k − 1)), where k is the number of fitted parameters.

N (Sample size)

The number of countries included in the regression for the current filter and time-period settings. Larger N increases statistical power and makes both R² and p-value estimates more reliable. Excluding resource-dependent or externally-funded countries reduces N and may change all fit metrics — a drop in R² after exclusion means those countries were pulling the curve in a predictable direction.

What else explains growth beyond government spending?

Greedy stepwise regression across 27 candidate variables (macro, governance, human capital, finance, structural) added one at a time to the power-law spending baseline. Each row shows the single best remaining variable at that step. The ceiling — stacking all 17 variables that clear the contribution threshold — is R² ≈ 0.680. The remaining ~32% is not recoverable from standard World Bank data. Notably, all 6 WGI governance indicators dropped out again — no marginal gain after income, investment, education and population are controlled. Military spending, R&D, and remittances all entered — for structural reasons explained in the interpretation column.

113 countries, 2005–2023, power-law baseline, standard exclusions. Greedy stepwise over 27 candidate World Bank indicators. Threshold: marginal R² > 0.003. Source: scripts/ceiling-r2.mjs. Updated by scripts/update-static-tables.mjs.

Stage 2: Spending + R&D + Military + Capital Formation + Population Growth + Domestic Credit + Tertiary Enrollment + Initial Income — combined-R² ranked model comparison

The spending curve above predicts a growth rate for each country. The residual is how far off that prediction is (actual − fitted). These charts ask: do R&D, military expenditure, capital formation, population growth, domestic credit, tertiary enrollment, initial income, and terms-of-trade volatility systematically predict the errors? All eight are controlled jointly via eight-variable OLS. Each chart shows the partial slope — the effect of one variable holding the other seven at their means. A positive capital formation slope reflects the investment-growth channel. A positive domestic credit slope would indicate financial depth matters beyond the spending level. A negative income slope (poorer countries grow faster conditional on spending) is expected textbook beta-convergence. Note: military expenditure, R&D, and tertiary enrollment are all components of total government spending (the Stage 1 x-axis). Their Stage 2 coefficients should be read as composition effects — among countries with the same total spending share, how does allocating more of it toward military (vs. other uses) correlate with growth? This is a guns-vs-butter trade-off coefficient, not an additive one. The table below ranks all models by combined R² on the joint subset.

The theory seemed reasonable: The Armey Curve suggested an inverted U-shaped relationship between government spending and economic growth. Named after economist Richard Armey, this curve claimed there exists an optimal level of government spending that maximizes economic growth.

But here's the problem: When you actually look at real-world data from dozens of countries over multiple decades, the theory doesn't hold up. Countries with lower government spending consistently achieve higher growth rates, while high-spending countries cluster in the low-growth zone.

What the data actually shows: Instead of a neat U-shaped curve with an "optimal" government size around 20-30% of GDP, we see patterns that better fit power law (s⁻ᵅ) or inverse (1/x) models - suggesting that any government spending beyond the absolute minimum reduces economic growth.

What the Traditional Theory Claimed

The Armey Curve theory proposed three distinct phases:

• Rising Phase (0-20%): Government spending supposedly provides essential infrastructure, legal framework, and public goods that enhance productivity and growth
• Peak (20-30%): The mythical "optimal" government size where growth is supposedly maximized
• Declining Phase (30%+): Excessive spending creates inefficiencies, crowds out private investment, and reduces growth through higher taxes and regulatory burden

What the Data Actually Shows

There is no "rising phase." Countries with minimal government spending (Singapore ~17%, historically Hong Kong ~15%) consistently achieve solid growth. Meanwhile, countries that spend 30-45% of GDP (most of Europe) cluster in the low-growth zone (0.5-1.5%).

There is no clear "optimal" zone. The data doesn't show clustering around 20-30% spending. Instead, we see a consistent negative relationship: lower spending = higher growth.

The relationship is better described by power law or inverse decline, not a quadratic curve. The data shows government spending is harmful to GDP growth from the first dollar — but that's precisely the point. The defensible use of government is as a deliberate brake on economic activity we don't want: pollution, overfishing, systemic financial risk. Slowing the economy is the feature, not the bug, in those applications.

The Historical Arc Confirms the Pattern

The simulator window (2005–2023) captures only a narrow slice of history in which all major economies already operate above 25% of GDP. But the pre-WWII record is instructive: in 1913, government spending averaged roughly 10–15% of GDP across Western Europe, and annual per-capita growth ran at ~2–3% — consistent with where the power law curve projects at those spending levels. The post-war expansion of the state shifted every major Western economy rightward along the curve into the low-growth zone. Where high-spending economies have sustained rapid growth, compositional factors — high investment shares, catch-up convergence, or off-budget financing — tend to account for the exception.

Traditional Theory vs. Empirical Reality

The traditional quadratic theory claimed:

Growth Rate = β₀ + β₁ × Government Spending + β₂ × (Government Spending)²

Where β₀ represents baseline growth, β₁ captures supposed initial positive effects, and β₂ (negative) represents diminishing returns.

But the data actually fits these patterns much better:

Power Law: Growth Rate = β₀ × (Government Spending)⁻ᵅ
Inverse Model: Growth Rate = β₀ / (Government Spending + 1)
Exponential Decay: Growth Rate = β₀ × e^(-decay × Government Spending)

The power law model achieves the highest R² of any model tested, explaining ~42% of the variation in growth rates among the 113 countries that pass the standard quality filters (excluding resource-dependent, externally-funded, conflict-fragile, and GDP-distorted economies). Without those filters the figure is ~24%, because the excluded groups add noise without adding signal. In cross-country macroeconomics, 42% is an exceptionally strong result for a single explanatory variable. For comparison, Robert Barro's landmark 1991 growth study — one of the most cited papers in economics — achieved R² ≈ 0.35–0.50 using ten or more variables simultaneously. Most single-variable growth regressions explain only 5–15% of variation. Government spending alone explaining 42% (or even 24% on the full unfiltered sample) means it is, by a wide margin, the single most important measurable determinant of cross-country growth differences.

The power law generalizes the inverse model (which is just power law with α=1) and captures the steep initial harm of government spending that gradually flattens at higher levels. These models all suggest there's no "beneficial phase" of government spending - it crowds out private investment from day one.

Understanding the Intercept (β₀)

The intercept represents the natural economic growth rate in the absence of government intervention. This baseline reflects:

• Entrepreneurial Innovation: Natural human creativity and problem-solving driving new products and services
• Voluntary Exchange: Wealth creation through mutually beneficial trade
• Capital Accumulation: Private savings and investment in productive assets
• Knowledge Spillovers: Information sharing and learning between economic actors
• Competition: Market pressure driving efficiency improvements
• Specialization: Gains from division of labor and comparative advantage

Historical evidence suggests this baseline ranges from 2-4% annually in developed economies, representing the economy's natural tendency toward improvement when people are free to innovate, trade, and invest.

Real-World Policy Implications

If the data is right and the traditional theory is wrong, the policy implications are dramatic:

• No "Optimal Size" to Target: There's no sweet spot to fine-tune toward - just minimize government and maximize growth
• Every Program Has a Cost: Each government program, no matter how well-intentioned, reduces overall economic growth
• Composition Matters, But Level Dominates: This data treats all spending identically — it cannot distinguish productive from wasteful programs. But the cross-country pattern holds at the aggregate level: countries with smaller governments consistently outgrow those with larger ones, suggesting the total level is the primary driver
• The Public Goods Question: Standard economic theory identifies a narrow category of goods (defense, basic rule of law, core infrastructure) where market provision may be insufficient. This data cannot resolve those debates — it only shows that countries with very low aggregate spending still achieve strong growth, suggesting private and market alternatives can substitute for more services than conventional theory predicts
• Government as a Selective Brake: If government spending reliably slows economic activity, the implication is not only to minimize it — but to aim it deliberately at the parts of the economy we want to slow down. River pollution, overfishing, antibiotic overuse, and systemic financial risk are all cases where unchecked private activity grows at society's expense. A government that acts as a targeted brake on these harms can improve welfare while keeping aggregate spending — and its drag on growth — small
• Maximum Reduction Strategy: The best policy is to cut government to the absolute minimum needed for basic rule of law

Why High-Spending Countries Struggle to Cut

If smaller government means faster growth, why don't high-spending countries simply cut their way to prosperity? The power law curve itself supplies the most under-appreciated answer: the marginal gain from cutting is tiny when you are already on the flat part of the curve.

The mathematics are exact. For the power law model $g = \beta_0 \cdot s^{-\alpha}$ (growth = scale × spending^−exponent), the marginal growth gain from cutting spending by a small amount $\Delta s$ (a tiny change in spending) at level $s$ is:

$\Delta g \approx \alpha \cdot \beta_0 \cdot s^{-(\alpha+1)} \cdot \Delta s$

In plain English: the growth boost from cutting spending equals the exponent ($\alpha$) times the baseline scale ($\beta_0$) times how flat the curve is at that spending level ($s^{-(\alpha+1)}$) times the size of the cut ($\Delta s$). At high spending the $s^{-(\alpha+1)}$ term is tiny, so the boost is tiny.

The ratio of that gain at a low-spending country ($s_1 = 20\%$) versus a high-spending one ($s_2 = 50\%$) is:

$\frac{\Delta g(s_1)}{\Delta g(s_2)} = \left(\frac{s_2}{s_1}\right)^{\alpha+1} = \left(\frac{50}{20}\right)^{\alpha+1}$

In plain English: the ratio of how much a cut helps at 20% spending versus 50% spending equals $(50/20)$ raised to the power of $(\alpha+1)$. With $\alpha \approx 0.45$, that is $2.5^{1.45} \approx 3.5$ — meaning the same cut is about 3.5× more valuable if you start from a lean government.

At 20% of GDP, the curve is steep. Cutting from 20% to 15% produces a meaningful jump in predicted growth. At 50% of GDP, the curve has flattened dramatically. Cutting from 50% to 45% produces a barely perceptible improvement. With the empirically fitted exponent α ≈ 0.447, the same five-percentage-point reduction delivers roughly 3.5× less growth dividend when you start from 50% versus 20% — and the ratio grows with the spending gap. If α were closer to 1.5 (the simulator default), that multiplier would reach ~10×. Either way, the directional logic is the same: the flatter the curve, the harder it is to make reform politically legible.

This asymmetry creates a political trap. The costs of cutting are immediate and concentrated — public sector jobs lost, entitlement recipients mobilized, contractors lobbying for reinstatement. The growth benefits are diffuse, delayed, and, crucially, small enough to be statistically invisible in the first few years of data. A government that cuts 5 points of GDP from a 50% baseline cannot credibly promise its population a visible boom; the power law says it will get perhaps 0.2–0.3 additional percentage points of annual growth. That is real wealth compounded over decades, but it is not a headline.

Contrast that with the experience of Ireland in the 1980s–90s or Sweden in the early 1990s: both cut from high bases, but their recoveries were amplified by other tailwinds (EU single market access, currency devaluation, rapid catch-up from deep recessions) that made the growth payoff visible and large. The power law contribution was real, but it was packaged with other forces. Without those tailwinds, a country cutting from 55% to 50% in a stable mature economy will see a reform dividend the curve predicts to be almost invisible over a parliamentary term.

There is also a ratchet effect: spending programmes create constituencies. Each percentage point of GDP spent builds a group of beneficiaries who will resist reversal more fiercely than the diffuse taxpayer will reward it. The political equilibrium drifts rightward along the curve — toward higher spending, lower growth, and ever-smaller marginal returns to cutting — until a fiscal crisis forces discontinuous adjustment. The power law explains not only the economic cost of big government, but the political economy of why large states tend to stay large.

The Selective Brake: What Government Is Actually Good For

The data in this simulator shows that government spending reliably slows economic activity. The conventional policy response is to minimize it. But there is a more precise implication: if government is a brake, use it on the parts of the economy you want to slow down.

The immediate objection is that this is not what actually happens. Most government spending is not a targeted instrument aimed at pollution or systemic risk — it is broad wages, transfers, subsidies, and procurement that slow the whole economy at once. A carbon tax that raises the cost of emissions is a brake applied to one specific activity. A ministry that grows its headcount, a pension system that expands its eligibility, a procurement rule that reserves contracts for incumbents — these apply the same braking force to productive and unproductive activity alike. The aggregate data captures this reality: the countries with large governments are not primarily running large externality-correction programs. They are running large redistribution and public employment programs, and the drag on growth reflects that indiscriminate scope. The selective brake is a theoretical possibility that real-world fiscal expansion rarely achieves.

Many harmful economic activities are fast precisely because their costs are externalized. A factory grows by dumping into a river. A fishing fleet expands by depleting a commons. A bank profits by taking on systemic risk that others will absorb. These activities are not growing because they are productive — they are growing because the people driving them don't bear the full cost. Left alone, they will outcompete cleaner, more careful alternatives and drag down the overall quality of the economy.

This is exactly where government's braking force belongs. A liability rule that makes polluters pay, a quota that limits overfishing, a capital requirement that forces banks to internalize their own risk — these interventions slow down the harmful activity without requiring large budgets or broad redistribution. The government is not producing anything; it is correcting the price signal so the market slows down the right things.

The corollary is equally important: government should not be used to slow down productive activity. Licensing barriers that limit entry into competitive professions, procurement rules that protect incumbent suppliers, tariffs that insulate domestic industries from competition — these apply the brake where it destroys rather than creates value. The data captures this aggregate: the countries that use government broadly slow their entire economy. The countries that keep government narrow preserve the growth engine while, in principle, still being able to target genuine harms.

A Single Objective Test: The Inclusive Wealth Criterion

A note on the simulator above. The simulator scores spending levels by their fit to GDP growth. The criterion developed below argues GDP is the wrong target — it can be inflated by depletion, transfers, and coercion. The two answer different questions: the simulator asks how much government spending maximises measured growth; the criterion asks where that spending should be aimed and which private activities deserve a brake. Both can be right at once: the empirical curve disciplines size, the criterion disciplines target.

If government should brake some activities and not others, we need a non-arbitrary way to decide which. "Things I dislike" is not a criterion. "Things that fail a cost–benefit test" is closer, but GDP-based cost–benefit can be gamed by transfers, depletion, and coercion — the very pathologies a brake should target. We need a metric that cannot be inflated by the activity it is meant to evaluate.

The whole rule, in one 2×2

Before the formula, the verdict it produces. Two binary questions decide every case: does the activity impose a net wealth loss on parties outside the transaction, and is the brake cost-effective?

Three of the four cells tell government to do nothing. Only the bottom-left — an external wealth loss that a cost-effective brake can address — justifies action. That is a much narrower licence than “whatever the median voter wants” — and a much wider one than “government should never act.” The rest of this section defines $\Delta W_{\text{ext}}$ precisely.

The formula behind the 2×2

The cleanest formulation comes from the inclusive wealth tradition. The pedigree:

Sources, in order: Pigou (1920), Coase (1960), Buchanan & Tullock (1962), Solow (1974), Dasgupta & Heal (1974), Weitzman (1976), Hartwick (1977), Arrow, Dasgupta, Mäler et al. (2004), Dasgupta (2021). each ingredient has a published source, but the unified rule is not stated in any single paper.

Define a country's wealth as the present value of every productive capital stock — produced, human, natural, knowledge, institutional — each weighted by its shadow price:

$$\frac{dW}{dt} = \sum_k p_k \frac{dK_k}{dt}$$

In plain English: the rate at which a country's wealth grows ($dW/dt$) equals the sum, across every kind of capital $k$ (produced, human, natural, knowledge, institutional), of how fast that capital is changing ($dK_k/dt$) multiplied by what that capital is worth at the margin ($p_k$, its shadow price). "$\sum_k$" just means "add up over every kind of capital."

Long-term growth, honestly measured, is $dW/dt > 0$. An activity is objectively brake-worthy if and only if it reduces this quantity for parties outside the transaction (those who bear cost or benefit without choosing to engage), and only when the brake itself is cost-effective:

$$\text{Brake}(a) \iff \mathbb{E}\!\left[\sum_k p_k\,\Delta K_k^{(a)}\right] < 0$$

(summed over external parties — those outside the transaction — and only when brake cost ≤ |ΔW_ext|)

In plain English: apply the brake to activity $a$ if and only if ("$\iff$") its expected ("$\mathbb{E}$") effect on total wealth, summed over every kind of capital, is negative — counting only parties outside the transaction (those who bear cost or benefit without choosing to engage). $\Delta K_k^{(a)}$ means "the change in capital of type $k$ caused by activity $a$." A second check (not shown in the formula) requires the brake cost to be no greater than this loss.

// Capital kinds — module-level constant, not per-call state.
const CAPITAL_KINDS = [
  "produced",      // factories, infrastructure, durable goods
  "human",         // health, skills, life expectancy
  "natural",       // air, water, soil, biodiversity, reserves
  "knowledge",     // R&D stock, know-how, useful information
  "institutional", // rule of law, trust, contract enforcement
];

// ── Three functions, one concern each ───────────────────────────────────────
//
//   shouldBrake(a, brake)     : pure decision, trusts its inputs
//   inputsAreTrustworthy(a)   : optional upstream gate over activity data
//   brakeIsWellFormed(brake)  : optional upstream gate over the brake descriptor
//
// Calling code composes them:
//   if (!inputsAreTrustworthy(a).ok || !brakeIsWellFormed(brake).ok)
//     refuseToDecide();
//   else act(shouldBrake(a, brake));
//
// The criterion itself is small and stable. The epistemic stack
// (measurement → vetting → pricing) lives upstream as a separate concern.
//
// a = {
//   affectedParties: [{ id, external, shadowPrices: { [kind]: price },
//                       measurementIntegrity?: true,  // explicit opt-in; absence is flagged
//                       priceVetted?: true,           // explicit opt-in; absence is flagged
//                       jurisdiction?: string }],     // for coordinationFloor
//   capitalDeltas:      { [kind]: { [partyId]: number } },
//   decidingJurisdiction?: string,                    // for coordinationFloor
// }
//
// brake = {
//   deadweightLoss:  number,   // Harberger triangle estimate
//   enforcementCost: number,   // budget cost of the brake
//   captureRisk:     number,   // expected rent-seeking cost (instrument-specific)
// }
//
// captureRisk belongs on the brake descriptor, not the activity, because it
// depends on the instrument chosen: Pigouvian tax → high (rate-setting is
// capturable); strict liability + competitive insurance → low (insurer pays,
// so they audit). Same activity, very different captureRisk.
//
// "external" means the party bears a cost or receives a benefit from the
// activity without choosing to participate in it (standard economic definition
// of an externality). Internal parties — those who voluntarily engaged — are
// excluded from ΔW_ext regardless of whether they gain or lose.
//
// A `price` is either a scalar (point estimate) or an object:
//   { mean, sd?, irreversible?, confidence? }
//
// The certainty-equivalent rule unifies three flavors of epistemic risk into
// one formula: p* = max(0, mean × confidence − λ × sd)
//
//   • sd            : statistical uncertainty (Arrow–Fisher–Hanemann option value)
//   • irreversible  : λ = 2 if true, 0.5 otherwise (Dasgupta Review)
//   • confidence    : strategic + adversarial uncertainty in [0, 1]
//                     1.0  measurement is tamper-evident AND price is independently vetted
//                     0.5  one of the two is weak
//                     0.0  measurement is compromised OR price is self-reported by an
//                          interested party — collapses p* to 0, so the channel
//                          contributes nothing to ΔW_ext (no brake on garbage data,
//                          but no false reassurance either)
//
// confidence = 0 is the "I cannot trust this" degenerate case, expressed as a
// limit of the same pricing rule rather than a separate code path.
//
// ── Out of scope (deliberate, not oversights) ────────────────────────────────
//
// These are scope decisions, named here so they are easy to find if scope
// expands. The criterion is intentionally narrow; the items below are upstream
// or orthogonal concerns that would muddy it if folded in.
//
//   • Internal-party schema errors. inputsAreTrustworthy does not check
//     internal parties' shadowPrices keys against CAPITAL_KINDS, and
//     capitalDeltas entries referencing internal parties pass the gate
//     unflagged. Both behaviors are consistent with the stated invariant
//     (internal parties may appear in capitalDeltas for bookkeeping;
//     shouldBrake filters them out before summing). If the invariant ever
//     changes, the two places to update are the gate's first for-loop and
//     the construction of knownPartyIds.
//
//   • Price-descriptor validation. brakeIsWellFormed sets the precedent that
//     descriptor fields get range-checked (negative cost, NaN, Infinity).
//     There is no analogous priceIsWellFormed for shadow-price descriptors,
//     so confidence: 1.5 or sd: -2 produces a mathematically defined but
//     economically meaningless result. certaintyEquivalentPrice is
//     deliberately a pure formula; if untrusted price objects ever reach
//     this code, add the sibling validator next to brakeIsWellFormed.
//
//   • Temporal structure. All capital deltas are treated as contemporaneous.
//     Discounting, phasing, and path-dependence (e.g. damage that grows if
//     not arrested early) are upstream concerns: they must be priced into the
//     shadow prices before inputs reach this code.
//
//   • Distributional weights. Welfare changes are aggregated by summing,
//     so a party losing $10 of natural capital exactly offsets a party
//     gaining $10 of natural capital at shadow price 1. Whether that
//     aggregation should be weighted by income, vulnerability, or political
//     representation is a prior question the pricing step must resolve;
//     the criterion takes the resulting prices as given.
function certaintyEquivalentPrice(p) {
  if (typeof p === "number") return p;                  // scalar point estimate
  const mean = p?.mean ?? 0;  // ?? 0: a price object with no mean is no signal
  const sd = p?.sd ?? 0;
  const confidence = p?.confidence ?? 1;                // default: fully trusted
  const lambda = p?.irreversible ? 2.0 : 0.5;       // risk aversion weight
  return Math.max(0, mean * confidence - lambda * sd);
}

// Pure decision: external parties lose inclusive wealth on net, AND
// the brake itself is cost-effective.
// Missing shadow prices contribute 0 (no signal), not 1 (arbitrary assumption).
// Typo detection lives in inputsAreTrustworthy; shouldBrake treats unknown kinds
// as zero by the ?? 0 convention, so the gate must be in the call chain for the
// guard to fire.
//
// brake defaults to {} so Q2 always passes (cost = 0). This is intentional for
// Q1-only testing ("is there external harm at all?"), but a caller who forgets
// the brake descriptor in production gets an optimistically cheap answer.
// Convention: always pass an explicit brake when the instrument is known.
function shouldBrake(a, brake = {}) {
  const externalParties = a.affectedParties.filter(p => p.external);
  const deltaW_ext = externalParties.reduce((total, p) =>
    total + CAPITAL_KINDS.reduce((s, k) =>
      s + certaintyEquivalentPrice(p.shadowPrices?.[k] ?? 0)
      * (a.capitalDeltas?.[k]?.[p.id] ?? 0), 0), 0);
  if (deltaW_ext >= 0) return false;                    // Q1: no net external harm

  const brakeCost = (brake.deadweightLoss ?? 0) + (brake.enforcementCost ?? 0) + (brake.captureRisk ?? 0);
  if (brakeCost > -deltaW_ext) return false;            // Q2: cure worse than disease

  return true;
}

// Optional upstream gate. Returns { ok, reasons }. Callers use it to refuse
// to decide when the data pipeline is compromised — separate from the
// criterion itself, so bad-faith "we can't trust the inputs" arguments have
// a single, explicit place to live where they can be debated on merits.
//
// Design note — two concerns in one gate:
//   The epistemic checks below (measurementIntegrity, priceVetted) answer
//   "can we trust the measurement and pricing pipeline?". The schema checks
//   (CAPITAL_KINDS keys, party IDs in capitalDeltas) answer "is the data
//   structure well-formed?". Keeping both here means there is still one
//   place where "do not proceed" lives. If they are ever split — e.g.
//   because schema errors should throw rather than accumulate reasons, or
//   because a caller wants schema validation without epistemic gating — the
//   seam is between the two for-loops below.
function inputsAreTrustworthy(a) {
  const reasons = [];
  for (const p of a.affectedParties.filter(x => x.external)) {
    if (p.measurementIntegrity !== true)
      reasons.push(`measurementIntegrity not confirmed for "${p.id}" — unknown is not safe`);
    if (p.priceVetted !== true)
      reasons.push(`priceVetted not confirmed for "${p.id}" — unknown is not safe`);
    const unknownKinds = Object.keys(p.shadowPrices ?? {}).filter(k => !CAPITAL_KINDS.includes(k));
    if (unknownKinds.length > 0)
      reasons.push(`unknown capital kinds for "${p.id}": ${unknownKinds.join(", ")} — not in CAPITAL_KINDS (typo?)`);
  }
  // capitalDeltas schema: kind keys must be in CAPITAL_KINDS; party IDs must
  // match IDs declared in affectedParties. Both classes of typo produce silent
  // zeros in every computation, so they are detectable only here.
  //
  // Invariant — internal parties in capitalDeltas:
  //   knownPartyIds includes ALL parties (external and internal). A delta
  //   entry for an internal party therefore passes this check rather than
  //   being flagged as an unknown ID. This is intentional: capitalDeltas
  //   may legitimately record internal-party changes for bookkeeping, and
  //   shouldBrake silently ignores them (it filters to external parties
  //   before summing). The silence in both places is consistent design, not
  //   an oversight. If the intended invariant were "capitalDeltas must only
  //   reference external parties", this check would use a filtered set and
  //   flag internal-party entries explicitly.
  const knownPartyIds = new Set(a.affectedParties.map(p => p.id));
  for (const kind of Object.keys(a.capitalDeltas ?? {})) {
    if (!CAPITAL_KINDS.includes(kind)) {
      reasons.push(`unknown capital kind in capitalDeltas: "${kind}" — not in CAPITAL_KINDS (typo?)`);
    } else {
      for (const partyId of Object.keys(a.capitalDeltas[kind])) {
        if (!knownPartyIds.has(partyId))
          reasons.push(`unknown party ID in capitalDeltas["${kind}"]: "${partyId}" — not in affectedParties (typo?)`);
      }
    }
  }
  return { ok: reasons.length === 0, reasons };
}

// Sibling gate over the brake descriptor. Same shape as inputsAreTrustworthy
// — { ok, reasons } — but a separate function because brake is optional and
// has a different signature than the activity. Callers compose them:
//   if (!inputsAreTrustworthy(a).ok || !brakeIsWellFormed(brake).ok)
//     refuseToDecide();
//
// Catches the failure modes that brake = {} cannot defend against:
//   • non-numeric or non-finite values (NaN, Infinity, strings)
//   • negative values (a "negative deadweight loss" silently flips Q2)
// A missing brake (undefined / null) is treated as the documented Q1-only
// case and passes — that's the brake = {} default of shouldBrake, made
// explicit here.
function brakeIsWellFormed(brake) {
  const reasons = [];
  if (brake == null) return { ok: true, reasons }; // Q1-only mode; documented default
  for (const field of ["deadweightLoss", "enforcementCost", "captureRisk"]) {
    const v = brake[field];
    if (v === undefined) continue; // missing → 0 by ?? convention; not an error
    if (typeof v !== "number" || !Number.isFinite(v))
      reasons.push(`brake.${field} is not a finite number: ${String(v)}`);
    else if (v < 0)
      reasons.push(`brake.${field} is negative: ${v} — costs cannot be negative`);
  }
  return { ok: reasons.length === 0, reasons };
}

// ── How to apply the brake ───────────────────────────────────────────────────
//
// shouldBrake() answers a yes/no question. The execution side — what
// instrument actually closes the externality gap — is a separate problem
// with its own capture surface (Pigouvian rates set by lobbyists, quotas
// allocated by political weight, voluntary codes written by industry, etc.).
//
// One instrument satisfies all the structural requirements at once:
//
//   strict liability + mandatory insurance covering the maximum credible loss,
//   with no liability cap, in a competitive insurance market,
//   not subsidised by the state.
//
// Why this single instrument generalises:
//   • Prices the externality, not the transaction (premiums track expected
//     damage, not gross sales)
//   • Self-enforcing — the operator either holds a policy or they don't;
//     no regulator decides case-by-case
//   • Adversarial verification built in — the insurer pays out on harm, so
//     they price, audit, and refuse uninsurable risks. They are the vetter
//     you do not have to appoint, with skin in the game by construction.
//   • State has no fiscal interest in the activity continuing (premiums go
//     to insurers, not the Treasury — breaks the dependency that turns
//     "green taxes" into structural rent extraction)
//   • Tail risk handled automatically: if maximum credible loss exceeds
//     insurance market capacity, no policy exists and the activity is
//     braked by the absence of coverage
//   • Measurement integrity becomes the insurer's problem — and they have
//     skin in the game to solve it
//
// requiredInsuranceCoverage(a) returns the maximum credible external loss
// the operator must insure against. This is the brake in operational form:
//   coverage > 0  → activity must hold a policy of at least this size
//   coverage = 0  → no required coverage (no net external harm priced in)
function requiredInsuranceCoverage(a) {
  // Losses aggregate across all capital kinds and all external parties —
  // the insurer covers the full scope of harm, not a net. Missing prices
  // contribute 0 (no signal) rather than 1 (arbitrary assumption).
  // Note: the coverage amount depends on the same shadow prices the insurer
  // will correct via underwriting — this is an input to that negotiation,
  // not its output. The circularity is resolved iteratively, not analytically.
  const externalParties = a.affectedParties.filter(p => p.external);
  const externalLoss = externalParties.reduce((total, p) =>
    total + CAPITAL_KINDS.reduce((s, k) => {
      const ce = certaintyEquivalentPrice(p.shadowPrices?.[k] ?? 0);
      const dK = a.capitalDeltas?.[k]?.[p.id] ?? 0;
      return s + Math.min(0, ce * dK); // only losses count, gains do not net out
    }, 0), 0);
  return Math.max(0, -externalLoss);
}

// ── Coordination floor: where the externality lives vs. who is braking ───────
//
// The criterion's natural unit is the externality's footprint, but our
// institutions are geographic. coordinationFloor(a) reads the spatial
// structure already encoded in affectedParties to answer: what fraction of
// external loss falls on parties OUTSIDE the deciding jurisdiction?
//
// Three regimes fall out of the same number:
//   floor ≈ 0   Case 1 — externality footprint ⊆ jurisdiction. Local courts
//               and a unilateral insurance regime are fully sufficient.
//   0 < f < 1   Case 2 — externality crosses borders via trade flows.
//               Border adjustments (CBAM-style) or reciprocal liability
//               recognition carry the brake to where the demand originates.
//   floor → 1   Case 3 — externality is global (CO₂, ozone, antibiotic
//               resistance, novel pathogens). No single country can brake
//               unilaterally without leakage proportional to the floor.
//               Coordination (climate clubs, treaty trusts, MFN reciprocity)
//               is constitutive of the brake, not an enhancement.
//
// floor is also the leakage rate of unilateral action: 1 − floor of the
// nominal damage is captured by acting alone; the rest escapes. This is
// not a flaw in the criterion; it is information the criterion produces.
//
// a.decidingJurisdiction is the jurisdiction whose institutions are
// applying the brake. Each affected party may carry a `jurisdiction` field.
// Parties with no jurisdiction declared are treated as being inside the
// deciding one (the conservative default — counts toward sufficient
// coverage rather than toward leakage). The returned { floor, undeclaredCount }
// makes this assumption visible: undeclaredCount > 0 means floor is a lower
// bound, not a precise estimate.
function coordinationFloor(a) {
  const j = a.decidingJurisdiction;
  if (!j) return { floor: 0, undeclaredCount: 0 }; // cannot compute leakage

  const externalParties = a.affectedParties.filter(p => p.external);
  let totalLoss = 0;
  let outsideLoss = 0;
  let undeclaredCount = 0; // external parties with losses but no jurisdiction field
  // Note: parties with no shadow prices have zero computed loss (via ?? 0) and
  // therefore don't appear in totalLoss OR in undeclaredCount. A party that is
  // both jurisdiction-undeclared and price-undeclared is entirely invisible to
  // this function — not flagged, not counted. undeclaredCount is a lower bound
  // on how many unknown parties are being treated as inside the jurisdiction.
  for (const p of externalParties) {
    const partyLoss = CAPITAL_KINDS.reduce((s, k) => {
      const ce = certaintyEquivalentPrice(p.shadowPrices?.[k] ?? 0);
      const dK = a.capitalDeltas?.[k]?.[p.id] ?? 0;
      return s + Math.min(0, ce * dK);
    }, 0);
    if (partyLoss >= 0) continue; // gains do not enter leakage analysis
    totalLoss += -partyLoss;
    if (p.jurisdiction === undefined) {
      undeclaredCount++; // treated as inside j — biases floor toward 0
    } else if (p.jurisdiction !== j) {
      outsideLoss += -partyLoss;
    }
  }
  const floor = totalLoss === 0 ? 0 : outsideLoss / totalLoss;
  return { floor, undeclaredCount };
}

// ── Assertions (truth-table style; throw on failure) ─────────────────────────
function assert(ok, msg) { if (!ok) throw new Error("✗ " + msg); }

// Factory: returns [a, brake] — spread into shouldBrake(...activity(...)).
// Parties carry measurementIntegrity: true and priceVetted: true (clean baseline);
// the gate treats absent fields as unconfirmed, so explicit true is required.
function activity(external, naturalDelta, deadweightLoss, enforcementCost, captureRisk) {
  return [
    {
      affectedParties: [{
        id: "p", external, shadowPrices: { natural: 1 },
        measurementIntegrity: true, priceVetted: true
      }],
      capitalDeltas: { natural: { p: naturalDelta } },
    },
    { deadweightLoss, enforcementCost, captureRisk },
  ];
}

// Core Q1/Q2 logic.
assert(shouldBrake(...activity(false, -10, 0, 0, 0)) === false, "Q1: internal party only → leave alone");
assert(shouldBrake(...activity(true, +10, 0, 0, 0)) === false, "Q1: external party gains → leave alone");
assert(shouldBrake(...activity(true, -5, 3, 2, 1)) === false, "Q2: brake costs 6 > damage 5 → leave alone");
assert(shouldBrake(...activity(true, -10, 1, 1, 1)) === true, "both filters passed → BRAKE");

// ── certaintyEquivalentPrice: direct unit tests ───────────────────────────────
// These test the formula in isolation. If a regression lands here the failure
// message names the function and the inputs, not a downstream Q1/Q2 label.
assert(certaintyEquivalentPrice(7) === 7,
  "cep: scalar passthrough");
assert(certaintyEquivalentPrice({}) === 0,
  "cep: no mean → 0 (no signal, not 1)");
assert(certaintyEquivalentPrice({ mean: 1 }) === 1,
  "cep: mean=1, all defaults → p* = 1");
assert(certaintyEquivalentPrice({ mean: 1, sd: 1 }) === 0.5,
  "cep: reversible, sd=1 → p* = 1 − 0.5×1 = 0.5");
assert(certaintyEquivalentPrice({ mean: 1, sd: 1, irreversible: true }) === 0,
  "cep: irreversible, sd=1 → p* = max(0, 1 − 2×1) = 0");
assert(certaintyEquivalentPrice({ mean: 2, confidence: 0.5 }) === 1,
  "cep: confidence=0.5, no sd → p* = 2×0.5 = 1");
assert(Math.abs(certaintyEquivalentPrice({ mean: 2, sd: 1, confidence: 0.5 }) - 0.5) < 1e-9,
  "cep: mean=2, sd=1, confidence=0.5 → p* = 2×0.5 − 0.5×1 = 0.5");

// The next three tests exercise certaintyEquivalentPrice in isolation via shouldBrake.
// They deliberately omit measurementIntegrity and priceVetted — that's inputsAreTrustworthy's
// domain, not shouldBrake's. Running inputsAreTrustworthy on these inputs would flag them;
// that's expected. These are unit tests of the pricing formula, not the full pipeline.

// Statistical uncertainty: irreversible damage with sd ≈ mean collapses p* to 0.
// Research that shrinks sd raises p* and re-engages the brake.
assert(shouldBrake({
  affectedParties: [{
    id: "p", external: true,
    shadowPrices: { natural: { mean: 1, sd: 1, irreversible: true } }
  }],
  capitalDeltas: { natural: { p: -10 } },
}) === false, "uncertainty: irreversible + sd≥mean → p*=0 → leave alone (Arrow 1972)");

// Adversarial uncertainty: confidence = 0 collapses p* to 0 — same effect
// as throwing, but expressed as a degenerate case of the pricing rule rather
// than a special control-flow path. shouldBrake stays pure.
assert(shouldBrake({
  affectedParties: [{
    id: "p", external: true,
    shadowPrices: { natural: { mean: 1, sd: 0, confidence: 0 } }
  }],
  capitalDeltas: { natural: { p: -10 } },
}) === false, "confidence=0 → p*=0 → leave alone (no brake on untrusted data)");

// Positive confidence test: partial confidence (0.8) still engages the brake when
// the adjusted loss (0.8 × 10 = 8) exceeds brake cost (3). Tests the mid-range
// between full-trust (1) and zero-trust (0).
assert(shouldBrake({
  affectedParties: [{
    id: "p", external: true,
    shadowPrices: { natural: { mean: 1, sd: 0, confidence: 0.8 } }
  }],
  capitalDeltas: { natural: { p: -10 } },
}, { deadweightLoss: 3, enforcementCost: 0, captureRisk: 0 }) === true,
  "confidence=0.8 → p*=0.8 → adjusted loss 8 > brake cost 3 → BRAKE");

// Trustworthiness gate is a separate concern. It reports problems; it does
// not decide policy. Callers that want hard-fail behavior compose it.
const tampered = {
  affectedParties: [{
    id: "op", external: true, measurementIntegrity: false, priceVetted: true,
    shadowPrices: { natural: 1 }
  }],
  capitalDeltas: { natural: { op: -10 } },
};
assert(inputsAreTrustworthy(tampered).ok === false, "gate: detects compromised measurement");
assert(inputsAreTrustworthy(tampered).reasons.length === 1, "gate: reports one reason (not two)");

const unvetted = {
  affectedParties: [{
    id: "ind", external: true, priceVetted: false, measurementIntegrity: true,
    shadowPrices: { natural: 1 }
  }],
  capitalDeltas: { natural: { ind: -10 } },
};
assert(inputsAreTrustworthy(unvetted).ok === false, "gate: detects unvetted prices");

// CAPITAL_KINDS typo guard: a misspelled kind (e.g. "naturel") is flagged by
// inputsAreTrustworthy rather than silently contributing 0. This is the only
// enforcement point — CAPITAL_KINDS is the authoritative list and nothing in
// the computation itself can distinguish a typo from a genuinely zero exposure.
const typo = {
  affectedParties: [{
    id: "tp", external: true, measurementIntegrity: true, priceVetted: true,
    shadowPrices: { naturel: 1 },  // deliberate misspelling of "natural"
  }],
  capitalDeltas: { natural: { tp: -10 } },
};
assert(inputsAreTrustworthy(typo).ok === false,
  "gate: unknown capital kind (typo) is flagged");
assert(inputsAreTrustworthy(typo).reasons.length === 1 &&
  inputsAreTrustworthy(typo).reasons[0].includes("naturel"),
  "gate: names the unknown kind in the reason");

// Clean: both fields explicitly true. Absent fields are also flagged — true is
// an explicit opt-in, not a default, so missing = unknown = flagged.
assert(inputsAreTrustworthy(activity(true, -10, 1, 1, 1)[0]).ok === true,
  "gate: clean inputs pass (measurementIntegrity: true, priceVetted: true)");

// Required insurance coverage = the size of the policy the operator must
// hold to proceed. With shadow price 1 and external loss of 10, the
// required cover is 10. Gains for other parties do not net out — the
// insurer covers harms, not aggregates.
assert(requiredInsuranceCoverage(activity(true, -10, 0, 0, 0)[0]) === 10,
  "insurance: external loss of 10 → required cover of 10");
assert(requiredInsuranceCoverage(activity(true, +5, 0, 0, 0)[0]) === 0,
  "insurance: external gain → no coverage required");
assert(requiredInsuranceCoverage(activity(false, -10, 0, 0, 0)[0]) === 0,
  "insurance: internal-only loss → no external coverage required");

// ── Coordination floor: three cases by externality footprint ─────────────────
//
// Case 1 — local externality. All external parties reside in the deciding
// jurisdiction. Unilateral action is fully sufficient; no leakage.
assert(coordinationFloor({
  decidingJurisdiction: "FR",
  affectedParties: [
    { id: "p1", external: true, jurisdiction: "FR", shadowPrices: { natural: 1 } },
    { id: "p2", external: true, jurisdiction: "FR", shadowPrices: { natural: 1 } },
  ],
  capitalDeltas: { natural: { p1: -5, p2: -5 } },
}).floor === 0, "Case 1 (local): floor = 0 → unilateral fully sufficient");

// Case 2 — trade-coupled externality. 70% of external loss falls outside
// the deciding jurisdiction. Border adjustments carry the brake to the
// consumption point; without them, leakage is 70%.
const case2 = {
  decidingJurisdiction: "FR",
  affectedParties: [
    { id: "fr", external: true, jurisdiction: "FR", shadowPrices: { natural: 1 } },
    { id: "br", external: true, jurisdiction: "BR", shadowPrices: { natural: 1 } },
  ],
  capitalDeltas: { natural: { fr: -3, br: -7 } },
};
assert(Math.abs(coordinationFloor(case2).floor - 0.7) < 1e-9,
  "Case 2 (trade): floor = 0.7 → border adjustments needed");

// Case 3 — global externality. Damage is spread over many jurisdictions;
// the deciding country bears a small share. Floor → 1 means almost all of
// the brake authority lives outside the country. Unilateral action leaks
// in proportion to (1 − share).
const case3 = {
  decidingJurisdiction: "FR",
  affectedParties: Array.from({ length: 10 }, (_, i) => ({
    id: `c${i}`, external: true,
    jurisdiction: i === 0 ? "FR" : `X${i}`,
    shadowPrices: { natural: 1 },
  })),
  capitalDeltas: {
    natural: Object.fromEntries(
      Array.from({ length: 10 }, (_, i) => [`c${i}`, -1])
    )
  },
};
assert(Math.abs(coordinationFloor(case3).floor - 0.9) < 1e-9,
  "Case 3 (global): floor = 0.9 → coordination is constitutive of the brake");

// Gains by external parties do not contribute to the leakage calculation:
// the brake exists to address harm, not to net harm against benefit.
assert(coordinationFloor({
  decidingJurisdiction: "FR",
  affectedParties: [
    { id: "fr", external: true, jurisdiction: "FR", shadowPrices: { natural: 1 } },
    { id: "us", external: true, jurisdiction: "US", shadowPrices: { natural: 1 } },
  ],
  capitalDeltas: { natural: { fr: -10, us: +10 } },
}).floor === 0, "coord: external gains elsewhere do not net against losses at home");

// No deciding jurisdiction → { floor: 0, undeclaredCount: 0 }.
assert(coordinationFloor(activity(true, -10, 0, 0, 0)[0]).floor === 0,
  "coord: no decidingJurisdiction → floor = 0 (analysis not applicable)");

// undeclaredCount surfaces the assumption: parties with no jurisdiction field
// are treated as inside the deciding jurisdiction, biasing floor toward 0.
assert(coordinationFloor({
  decidingJurisdiction: "FR",
  affectedParties: [
    { id: "local", external: true, jurisdiction: "FR", shadowPrices: { natural: 1 } },
    { id: "unknown", external: true, shadowPrices: { natural: 1 } },
  ],
  capitalDeltas: { natural: { local: -5, unknown: -5 } },
}).undeclaredCount === 1, "coord: undeclaredCount flags parties with losses but no jurisdiction field");

// ── Multi-party, multi-kind aggregation ───────────────────────────────────────
//
// shouldBrake sums across both external parties AND capital kinds. These tests
// exercise the aggregation paths that single-party, single-kind tests cannot reach.

// Two external parties in different capital kinds. Individual losses (6 each)
// are below the brake cost (8) in isolation, but their sum (12) exceeds it.
// This confirms the reducer sums across both dimensions before comparing.
assert(shouldBrake({
  affectedParties: [
    { id: "a", external: true, shadowPrices: { natural: 1 } },
    { id: "b", external: true, shadowPrices: { human: 1 } },
  ],
  capitalDeltas: { natural: { a: -6 }, human: { b: -6 } },
}, { deadweightLoss: 8, enforcementCost: 0, captureRisk: 0 }) === true,
  "aggregation: two parties, two kinds, combined loss 12 > brake cost 8 → BRAKE");

assert(shouldBrake({
  affectedParties: [
    { id: "a", external: true, shadowPrices: { natural: 1 } },
    { id: "b", external: true, shadowPrices: { human: 1 } },
  ],
  capitalDeltas: { natural: { a: -6 }, human: { b: -6 } },
}, { deadweightLoss: 13, enforcementCost: 0, captureRisk: 0 }) === false,
  "aggregation: combined loss 12 < brake cost 13 → leave alone");

// requiredInsuranceCoverage with one party losing across two capital kinds.
// The insurer covers the full scope of harm; losses in each kind accumulate.
assert(requiredInsuranceCoverage({
  affectedParties: [{
    id: "c", external: true,
    shadowPrices: { natural: 1, human: 2 },
  }],
  capitalDeltas: { natural: { c: -4 }, human: { c: -3 } },
}) === 10,  // 1×4 + 2×3 = 4 + 6 = 10
  "aggregation: single party, two kinds — insurance covers 4 + 6 = 10");

// ── capitalDeltas schema checks in inputsAreTrustworthy ───────────────────────
//
// A misspelled kind key in capitalDeltas (e.g. "natrual") is as silent as a
// misspelled shadowPrices key — the ?? 0 fallback swallows it. Likewise, a
// misspelled party ID (e.g. "pp" instead of "p") lands in no party's row.
// Both are now caught by the gate alongside the shadowPrices typo guard.

// Misspelled kind key in capitalDeltas.
const deltaKindTypo = {
  affectedParties: [{
    id: "p", external: true, measurementIntegrity: true, priceVetted: true,
    shadowPrices: { natural: 1 },
  }],
  capitalDeltas: { natrual: { p: -10 } },  // "natrual" instead of "natural"
};
assert(inputsAreTrustworthy(deltaKindTypo).ok === false,
  "gate: misspelled kind key in capitalDeltas is flagged");
assert(inputsAreTrustworthy(deltaKindTypo).reasons.length === 1 &&
  inputsAreTrustworthy(deltaKindTypo).reasons[0].includes("natrual"),
  "gate: names the misspelled kind from capitalDeltas in the reason");

// Misspelled party ID in capitalDeltas.
const deltaPartyTypo = {
  affectedParties: [{
    id: "p", external: true, measurementIntegrity: true, priceVetted: true,
    shadowPrices: { natural: 1 },
  }],
  capitalDeltas: { natural: { pp: -10 } },  // "pp" instead of "p"
};
assert(inputsAreTrustworthy(deltaPartyTypo).ok === false,
  "gate: misspelled party ID in capitalDeltas is flagged");
assert(inputsAreTrustworthy(deltaPartyTypo).reasons.length === 1 &&
  inputsAreTrustworthy(deltaPartyTypo).reasons[0].includes("pp"),
  "gate: names the misspelled party ID from capitalDeltas in the reason");

// Clean capitalDeltas with correct kind and party ID passes.
assert(inputsAreTrustworthy(activity(true, -10, 1, 1, 1)[0]).ok === true,
  "gate: correct capitalDeltas kind and party ID pass schema check");

// ── brakeIsWellFormed: brake-descriptor validation ───────────────────────────
//
// brake = {} and brake = undefined are valid (Q1-only mode, documented).
// What's caught here is values that look numeric but break Q2: negatives,
// NaN, Infinity, strings.

assert(brakeIsWellFormed(undefined).ok === true,
  "brake: undefined → ok (Q1-only mode)");
assert(brakeIsWellFormed({}).ok === true,
  "brake: empty object → ok (all fields default to 0)");
assert(brakeIsWellFormed({ deadweightLoss: 1, enforcementCost: 2, captureRisk: 3 }).ok === true,
  "brake: clean positive numbers → ok");

assert(brakeIsWellFormed({ deadweightLoss: -1 }).ok === false,
  "brake: negative deadweightLoss flagged");
assert(brakeIsWellFormed({ captureRisk: NaN }).ok === false,
  "brake: NaN captureRisk flagged");
assert(brakeIsWellFormed({ enforcementCost: Infinity }).ok === false,
  "brake: Infinity enforcementCost flagged");
assert(brakeIsWellFormed({ deadweightLoss: "3" }).ok === false,
  "brake: string deadweightLoss flagged (typeof !== number)");

// All three fields malformed → three reasons (not short-circuited).
assert(brakeIsWellFormed({
  deadweightLoss: -1, enforcementCost: NaN, captureRisk: "0"
}).reasons.length === 3, "brake: accumulates all reasons, no short-circuit");

This single inequality subsumes the standard candidates for "what government should slow down":

Conversely, the criterion refuses to condemn activities that are merely unfashionable, inefficient, or disliked by incumbents. Consumption choices of competent adults that affect only themselves produce $\Delta W_{\text{ext}} = 0$ — there are no external parties. The criterion is silent — which is exactly what an objective rule should do where no objective harm exists.

The decision in sequence

The 2×2 maps directly to two sequential checks. Government action has its own deadweight loss, enforcement cost, and capture risk — both must pass:

Q2 is the step most policy analysis skips. It matters because many real-world activities produce negative $\Delta W_{\text{ext}}$ yet would still be made worse by intervention — either because the brake mechanism is captureable (Q2 fail mode 1) or because enforcement costs exceed the damage (Q2 fail mode 2). The criterion licences action only when both filters are passed.

Q2 is the most qualitative of the two and worth flagging as such. Deadweight loss has a textbook estimator (Harberger triangles), enforcement cost has a budget line, but capture risk — the probability that the brake instrument gets bent toward the very interest it was meant to constrain — is genuinely hard to put a number on. The literature on regulatory capture (Stigler 1971; Dal Bó 2006) describes the mechanism but offers no consensus formula. In practice Q2 acts as a circuit-breaker: if a sector has a strong track record of capturing its regulator, raise the burden of proof for new brakes there; otherwise treat the deadweight and enforcement components as the binding part of the test.

Worked examples: applying the test to three policies

The criterion is only useful if it produces unambiguous verdicts on real cases. Three contrasting examples:

A coal subsidy. $\Delta K_{\text{natural}} < 0$ (emissions, depleted reserves); $\Delta K_{\text{human}} < 0$ (respiratory damage in downwind populations, external to the transaction); $\Delta K_{\text{produced}} > 0$ (cheaper electricity, short-term); $\Delta K_{\text{institutional}} < 0$ (the lobbying apparatus that sustains the subsidy itself erodes rule-of-law neutrality). Net: $\Delta W_{\text{ext}} < 0$ (external wealth loss). Verdict: brake.

A grant for basic-science research. The case is harder than a coal subsidy and worth the honesty. $\Delta K_{\text{knowledge}} > 0$ (results enter the public domain with positive spillovers); $\Delta K_{\text{human}} > 0$ (trained researchers); but $\Delta K_{\text{produced}} < 0$ for taxpayers, who are external parties (they fund the grant through coerced transfers without being participants in the research transaction). The verdict therefore turns on Q1: is $\Delta W_{\text{ext}}$ negative on net? For basic research with broad, non-rivalrous spillovers, the knowledge gain that eventually accrues back to the same taxpayers (cheaper medicine, better materials, public-domain methods) typically exceeds the per-capita tax cost — so $\Delta W_{\text{ext}} \geq 0$ and the criterion declines to brake it. Verdict: don't brake the activity, but the funding mechanism remains a real cost the activity has to pay back in spillover terms. Narrowly captured grants that spill over only to the recipient (industry-specific subsidies, company-town infrastructure) do not pay back and would fail Q1.

Early childhood development. Children are categorically external parties: they cannot choose to participate in, exit, or negotiate the conditions of their upbringing. Parental, community, and state decisions — nutrition, environmental exposures (lead, air pollution), stimulation and care quality, violence — impose irreversible costs on the child as an external party. The empirical record is unusually strong: $\Delta K_{\text{human}} \ll 0$ from ACE (adverse childhood experience) dose-response studies; childhood blood-lead at 10 μg/dL reduces IQ by 2–5 points with documented downstream crime and earnings effects (Nevin 2000; Reyes 2007); iodine deficiency during pregnancy costs 10–15 IQ points at a brake cost of ~USD 0.05/person/year for salt iodization. The externality falls on a party who cannot consent on their own behalf, the harm is largely irreversible (raising shadow prices for irreversibility), and cost-effective brakes exist with well-documented returns. Heckman’s estimates of USD 7–13 return per dollar on high-quality early childhood programs clear Q2 comfortably. Verdict: brake environmental and nutritional externalities imposing costs on children; apply precaution proportional to irreversibility.

Notice what the test did not require: a vote, a moral intuition, or an opinion about energy policy. It only required tallying the capital effects on external parties.

One empirical reference point: the planet, 1992–2014. The Dasgupta Review's Headline Message 2 reports that over this period global per-capita produced capital roughly doubled while per-capita natural capital fell by roughly 40% (Dasgupta 2021; corroborated by the World Bank's Changing Wealth of Nations 2021). Run those through the IWC lens: $\Delta K_{\text{produced}} > 0$, but $\Delta K_{\text{natural}} \ll 0$ for parties — future generations, downstream ecosystems — who are external to the trade. Whether $\sum_k p_k\,\Delta K_k$ was net-positive or net-negative overall depends on shadow prices, which is exactly what the Review argues we have been omitting from standard GDP accounting. The criterion doesn't pretend to settle those shadow-price disputes; it forces them into the open. A growth statistic that aggregates +100% produced capital with −40% natural capital into a single cheerful headline number is not measuring wealth — it is measuring one column of a ledger while pretending the others don't exist.

What "external" actually means here

"External parties" means those who bear a cost or receive a benefit from an activity without choosing to participate in it. This is the standard economic definition of an externality. The boundary matters because the criterion is silent on harms that fall only on willing participants — their wealth change is real, but it is their choice to bear it. Three cases worth distinguishing:

• Clearly external: the harmed party is identifiable and did not participate in the transaction at all. Classical externalities (pollution downwind, systemic risk shifted to the deposit-insurance fund), fraud against counterparties, and coerced transfers — taxation, conscription, eminent domain — where the contributing party was not a voluntary party to the particular use of their resources. The criterion treats these as real external costs.
• Clearly internal: the party freely entered an arrangement that priced the cost in — an insurance pool whose premiums fund payouts, a professional code of conduct accepted on entry, a club whose dues fund the clubhouse, a private contract whose terms were negotiated. The defining feature is voluntary participation: a person who could decline and walk away is an internal party. Their wealth changes do not enter $\Delta W_{\text{ext}}$.
• Contested boundary: a voter who loses on a budget line, a future generation affected by today's investment choices, an animal or ecosystem that cannot transact at all. The criterion requires a convention here — the standard one is to treat future generations as external (hence Hartwick) and non-transacting ecosystems as external via their shadow prices on $K_{\text{natural}}$. These are reasonable defaults, not derivations.

The Coase (1960) reading carries through: when transaction costs are low and parties are identifiable, externalities can be resolved privately and $\Delta W_{\text{ext}} = 0$ — the criterion stays silent. Pigouvian intervention is only licensed when private resolution is structurally impossible, not merely unattractive.

The hard cases (where the criterion is genuinely ambiguous)

An honest framework names the cases it doesn't cleanly resolve. Four worth flagging:

• Public-goods underprovision. Defense, basic research, contagious-disease surveillance: there is no activity to brake, only one to start. The criterion tells government what to slow, not what to provide. A complementary rule (Samuelson on public goods) is needed for the production side.
• Voluntary risky behaviour with social spillovers. Smoking, motorcycling without a helmet, recreational drugs. The user is an internal party; the emergency-room budget is external. The criterion says brake the external portion (e.g. insurance pricing that reflects actual risk) not the activity itself.
• High-variance frontier activity. Gain-of-function research, untested geoengineering, novel financial instruments. The expectation operator $\mathbb{E}$ matters: tail risk to external parties can dominate even when the central estimate of $\Delta W_{\text{ext}}$ is positive. The criterion supports precaution proportional to the size of the external tail.
• Effects on the unborn. Every long-horizon decision affects people who don't yet exist. The criterion treats them as external parties, represented by the discount rate $\rho$ and by an obligation to leave $W$ non-decreasing (Hartwick). This is a convention, not a derivation, and people of good faith disagree about $\rho$.

Why this is genuinely objective (within limits)

• Sign objectivity is robust. Even when shadow prices $p_k$ are uncertain, the direction of $\Delta W_{\text{ext}}$ is often unambiguous: a polluting subsidised monopoly reduces nearly every $K_k$ for outsiders at once.
• It does not require value judgements about lifestyles. It only checks the accounting.
• It is partially measurable. The World Bank's Changing Wealth of Nations covers ~150 countries with produced, human, natural, and net-foreign-asset capital and finds that many resource-dependent economies have shrinking per-capita wealth even while their GDP per capita rises — the exact gap GDP-only accounting hides. The UN Inclusive Wealth Report 2023 covers 163 countries. The Adjusted Net Savings series provides the annual flow counterpart for ~140 countries.

Where objectivity leaks — honestly

• Shadow prices for institutional and natural capital are estimated, not observed. Existing accounts cover produced, human, and natural capital well; institutional capital is not yet integrated into any official wealth account.
• Shadow prices are partly endogenous to the policy under evaluation. $p_k$ depends on the policy environment — a carbon tax changes the shadow price of $K_{\text{natural}}$, which is the very price you'd want to use to score the carbon tax. Honest practice picks a baseline (no-policy or counterfactual-status-quo prices), reports a range, and relies on sign-robustness rather than precise magnitudes. See Dasgupta & Mäler (2000) for the formal treatment.
• The discount rate $\rho$ involves an ethical choice (Stern 2008 vs. Nordhaus 2007).
• The boundary between internal and external parties requires a convention for contested cases (future generations, ecosystems, diffuse publics).
• The criterion is silent on distribution within the internal set. Two policies with identical $\Delta W_{\text{ext}}$ can have very different fairness profiles among participants, and the test will rank them as equivalent. That second question needs a different tool.
• Shadow prices carry uncertainty, not just point estimates. The criterion prices in $\sigma_p$ via a certainty-equivalent: $p^* = \max(0,\, \mu - \lambda\sigma)$, with $\lambda = 2$ for irreversible capital (Arrow–Fisher–Hanemann quasi-option value; Dasgupta Review on natural-capital irreversibility) and $\lambda = 0.5$ for reversible capital. This raises the burden of proof when damage is both uncertain and irreversible. The corollary is an endogenous research incentive: any party that funds work shrinking $\sigma_p$ — better measurement, basic science — lowers the precautionary premium and moves the brake decision toward its true expected-value verdict. This is the value of information (Arrow 1972) made operational.

But these are calibration disputes, not criterion disputes. Two honest analysts using inclusive wealth can disagree on magnitudes substantially — the Stern–Nordhaus split on the social cost of carbon spans roughly an order of magnitude — but they will agree on the sign of $\Delta W_{\text{ext}}$ for most cases, which is what the brake test actually requires. Two analysts using "what's good for society" can disagree by infinity and on the sign too.

The policy rule that follows

Government spending should be aimed at activities where $\sum_k p_k \Delta K_k < 0$ (weighted capital shrinks) for external parties — activities that shrink the inclusive capital base of outsiders — and withdrawn from activities where it is positive.
Don't tax production; tax depletion.

Translated into a tax base, the criterion picks out a small, specific list of things to charge for — and a much larger list of things to leave alone. What it would tax:

• Carbon, methane, particulates, congestion (classical Pigouvian)
• Resource extraction above the regeneration rate (severance / depletion)
• Land value — Henry George's tax on unearned location rent, which captures wealth without distorting production
• Financial leverage above systemic-risk thresholds
• Unfunded pension liabilities — force amortization rather than rollover
• Monopoly rents that rest on regulatory capture rather than productivity

What it would not tax: labour income, corporate profits earned competitively, capital gains on productive investment, consumption of competent adults. Almost every public-finance economist from George to Mirrlees to Friedman has recommended some version of this base — "tax bads, not goods." That it remains politically marginal despite near-consensus in the academic literature is itself a data point about which constituencies the existing tax system serves.

This is also why the data pattern documented above (more government spending → slower growth) is expected under the criterion: most government spending in modern economies funds transfers and current consumption, which appears as $\Delta K \leq 0$ (capital change is zero or negative) across the inclusive-wealth ledger. The countries that beat the curve — Singapore, Switzerland, Norway with its sovereign wealth fund — are precisely those that explicitly reinvest government revenue back into $K_k$ (each type of capital) rather than consuming it.

How the brake is applied: one instrument

shouldBrake outputs a verdict. The verdict has to be executed by some instrument — tax, quota, ban, liability, disclosure mandate. Each instrument has its own capture surface, so this looks like a second hard problem stacked on top of the first. It collapses to a much simpler answer once you ask which instrument self-enforces, prices the externality (not the transaction), and gives the state no fiscal interest in the activity continuing. Strict liability with mandatory insurance covering the maximum credible loss — no liability cap, in a competitive insurance market, not subsidised by the state — is the only instrument that satisfies all those constraints simultaneously.

• Prices the externality, not the transaction. The premium tracks expected damage. A clean activity pays nothing; a damaging one pays heavily. VAT on energy fails this test: it taxes nuclear and coal at the same rate.
• Self-enforcing. The operator either holds a policy or they don't. There is no regulator deciding case-by-case, so there is nothing to capture. Verification is binary and contractual.
• Adversarial verification built in. The insurer pays out on harm, so they price it correctly, audit the operator, demand evidence, and refuse coverage on bad risks. They are the vetter you do not have to appoint, with skin in the game by construction. Measurement integrity becomes the insurer's problem — and they have a financial incentive to solve it.
• State has no fiscal stake in the activity continuing. Premiums go to insurers, not the Treasury. This breaks the dependency that turns "green taxes" into structural rent extraction (TICPE in France, ~€33B/year, is paid into the general budget — the Treasury cannot afford for the activity to actually shrink).
• Tail risk handled automatically. If the maximum credible loss exceeds insurance market capacity, no policy is written and the activity is braked by the absence of coverage. This is how nuclear works in jurisdictions without Price–Anderson-style caps, and how gain-of-function research should work.
• Scales automatically with evidence. As harm materialises, premiums rise. As risk-reducing technology improves, premiums fall. No legislative reset required — the system updates as a side-effect of insurers protecting their reserves.

The hard part is not picking the instrument. The hard part is three preconditions that political systems resist: no liability cap (Price–Anderson-style legislation is the standard failure mode), competitive insurance markets (cartelisation lets insurers under-price collusively), and no state subsidy of premiums (which would transfer the externality back to taxpayers and reintroduce the fiscal-capture problem). Where those three preconditions hold, the criterion needs almost no machinery. Where they fail, the criterion produces a verdict that nobody is structurally equipped to honour. The requiredInsuranceCoverage helper makes this concrete: it returns the policy size the operator must hold, derived from exactly the same certainty-equivalent prices that feed shouldBrake.

When does one country suffice? Three cases

The criterion's natural unit is the externality's footprint, but our institutions are geographic. Whether per-country implementation is sufficient depends on a quantity the framework already encodes: the share of external loss that falls on parties outside the deciding jurisdiction. coordinationFloor(a) reads that share directly from affectedParties. The number it returns sorts activities into three structurally different regimes — not a continuum but three different problems.

• Case 1 — local externality (floor ≈ 0). All external parties reside in the deciding jurisdiction: water pollution, regional smog, soil contamination, urban noise, land-use change, most labour conditions. Per-country implementation is not just adequate; it is optimal. There is no leakage problem because the externality cannot leak across a border it does not cross. Most of what governments actually regulate falls here.
• Case 2 — trade-coupled externality (0 < floor < 1). Water embedded in agricultural exports, deforestation behind cattle exports, labour conditions in supply chains. The harm happens in country B but the demand pull comes from country A. The brake belongs at the consumption point, not the production point. Border adjustments (the EU's CBAM is the early version) carry the brake to where the demand actually originates — this is not a workaround, it is the correct location of the brake when the producer-side framing is wrong.
• Case 3 — global externality (floor → 1). Greenhouse gases, ozone-depleting substances, ocean fishery collapse, antibiotic resistance, weapons-grade fissile material, gain-of-function pathogens. The externality is genuinely fungible across all geography. Coordination is constitutive of the brake, not an enhancement. Climate clubs (Nordhaus 2015), MFN reciprocity, and linked liability schemes all converge to the same result: leakage falls toward zero as the coordinated bloc approaches global market share. Below ~50% of global activity in the bloc, leakage is significant; above ~70% it is de minimis.

This makes the "competitiveness" objection precise. The objection has bite for Case 3 specifically — not as a flaw in the criterion, but as a coordination problem the criterion correctly identifies. For Cases 1 and 2 the objection is a category error: country A's residents win on inclusive wealth even when country A's GDP-share-of-cement-production falls. The Changing Wealth of Nations data shows the pattern repeatedly — resource-dependent economies post rising GDP per capita and falling per-capita inclusive wealth. They are "competitive" only on the wrong metric. The criterion's job is to make the right metric legible, including when unilateral action is sufficient (most cases), partial (trade-coupled cases), or genuinely insufficient without coordination (planet-scale cases).

What would falsify this

A criterion that cannot be wrong is not a criterion. The inclusive wealth test makes a specific empirical prediction: countries that tax depletion and protect inclusive wealth should, over decades, grow $W$ faster than countries that tax production and consume the capital base — even when conventional GDP growth diverges from $W$ growth in the short run.

What would refute the criterion:

• If countries with depletion-style tax bases (Norway's sovereign fund, Chilean copper royalties, Singaporean land-value capture, British Columbia's carbon tax) systematically showed worse long-run inclusive-wealth growth than peers that tax labour and capital, the criterion fails.
• If shadow-price uncertainty turned out to flip the sign of $\Delta W_{\text{ext}}$ (change in external wealth) (not just its magnitude) for most policy-relevant activities, the criterion would lose its claim to sign objectivity.

Current evidence runs the other way — the listed countries outperform on inclusive wealth per capita in the World Bank CWON data — but the prediction is testable, and the dataset is public. That distinguishes this framework from the slogans it replaces.

From actions to regimes: legitimate vs. illegitimate government

The criterion gives a verdict per action. Lifting it to whole governments is straightforward: a regime's legitimacy is the share of its coercive acts that pass the two filters. A government that brakes pollution, fraud, depletion, and systemic risk is using its instrument on the cell the criterion licenses. A government that brakes voluntary trade, productive labour, or peaceful consumption is applying the brake where $\Delta W_{\text{ext}} \geq 0$ — foot on the brake pedal where it should be off it.

How the Theory Misled Policymakers

The Armey Curve theory emerged in the 1980s from observations that both very small governments (lacking basic institutions) and very large governments (socialist economies) had slower growth than moderate-sized governments. This seemed to suggest an optimal middle ground.

But this analysis was flawed. Countries with "very small governments" were often failed states or developing nations with poor institutions, while "very large governments" were communist dictatorships. The comparison wasn't between different sizes of functional government - it was between functional and dysfunctional states.

When you compare functional governments of different sizes, the pattern is clear: smaller government = higher growth. Singapore, Switzerland, and Estonia consistently outperform France, Germany, and Sweden on growth despite having much smaller governments.

The theory gave academic cover to politicians who wanted to justify expanding government by claiming they were finding the "optimal" size. In reality, they were just reducing economic growth.

Why the Quadratic Model Produces Nonsensical Results

The quadratic Armey curve is fundamentally broken because it predicts impossible negative growth rates at high government spending levels. This mathematical artifact reveals why the traditional theory is wrong - real economies don't experience negative 5-10% GDP growth just because government spends 50-60% of GDP.

What actually happens in high-spending countries: European countries with 35-45% government spending don't collapse into economic oblivion. They stagnate at low positive growth rates (0.5-1.5%), which is exactly what the power law and inverse models predict.

The math exposes the flaw: When you fit a quadratic curve (y = ax² + bx + c) to real data, it eventually curves downward so sharply that it predicts economic apocalypse. But Sweden at 35% spending doesn't have -8% growth - it has +0.8% growth. The quadratic model fails basic reality checks.

Worse: the unconstrained free fit is U-shaped, not inverted-U. When a quadratic is fitted to the data with no constraints — letting the sign of the curvature be determined by the data itself — the result is a U-shaped curve with a minimum at ~49% spending, not the inverted-U "sweet spot" Armey theory predicts. The data finds growth declining all the way through the observed range, with the theoretical upturn occurring beyond 49% spending where almost no country in the sample exists. This is arguably more damning than a poor R²: the best quadratic the data can produce is the opposite of the Armey curve.

Linear Models Are Equally Flawed

The linear decline model suffers from the exact same mathematical impossibility. With a negative slope (which is required to show government spending reduces growth), the linear model inevitably predicts negative growth rates at high spending levels:

• Mathematical Inevitability: A linear model with negative slope (Growth = β₀ + slope × spending) must eventually cross zero and go negative as spending increases
• Empirical Absurdity: The model would predict that France (35% spending) should have negative GDP growth every year, which clearly doesn't happen
• No Asymptotic Behavior: Unlike inverse or exponential models, linear models can't capture the reality that even heavily regulated economies maintain some positive growth
• Constant Marginal Damage: Linear models unrealistically assume that each additional percentage of government spending causes exactly the same damage regardless of existing spending levels

Both quadratic and linear models fail the basic empirical test: they predict economic outcomes that simply don't exist in the real world. This leaves only the power law, inverse, and exponential models as mathematically viable alternatives to describe the government-growth relationship.

Why Power Law, Inverse, and Exponential Models Make Sense

• Asymptotic Approach to Zero: Both models approach (but never reach) zero growth, which matches reality where even heavily regulated economies still limp along
• No Mathematical Artifacts: As shown above, these models avoid the impossible negative growth predictions that disqualify both quadratic and linear specifications
• Diminishing Returns Without Collapse: They show government spending becomes increasingly harmful without predicting economic Armageddon
• Empirical Fit: They actually match what we observe - stagnation, not collapse, in high-spending economies

But Even Exponential Decay Goes Too Far

While exponential decay avoids the quadratic model's absurd negative growth predictions, it still doesn't fit the real-world data perfectly. The exponential model suggests that each additional percentage point of government spending causes accelerating damage to growth, but empirical evidence shows this is too aggressive:

• European Resilience: Countries like Germany (30% spending, 0.5% growth) and France (35% spending, 0.8% growth) maintain low but positive growth despite massive government sectors. Exponential decay would predict much sharper decline
• Nordic Stability: Denmark (35% spending) and Sweden (35% spending) have sustained their welfare states for decades with consistent low growth (0.4-0.8%), not the accelerating collapse exponential models predict
• Mathematical Overshooting: Exponential decay curves drop too steeply for high-spending economies, underestimating their ability to maintain basic economic function through institutional momentum
• Real-World Stagnation Pattern: What we actually observe is not accelerating decay but persistent low-growth stagnation - exactly what the power law and inverse models predict

Why the Power Law Model Is Empirically Superior

The power law model (Growth = β₀ × s⁻ᵅ) achieves the highest R² of any model tested, explaining ~42% of cross-country growth variation among comparable economies:

• Best Predictive Fit: With R² = 0.42 (on the 113-country filtered sample — excluding resource-dependent, externally-funded, conflict-fragile, and GDP-distorted economies), the power law outperforms the inverse model (R² = 0.42) and all other alternatives. On the full unfiltered sample the figure is ~0.24; on that same unfiltered sample the quadratic model has R² = 0.00. Those excluded groups add noise unrelated to fiscal policy. To put this in perspective: cross-country growth depends on dozens of factors (demographics, institutions, geography, trade, technology, culture) yet government spending alone accounts for nearly half the variation among comparable economies — a signal-to-noise ratio that most macro variables can only dream of. The exponent α lets the model calibrate the steepness of the spending-growth curve to real data rather than fixing it at α=1 like the inverse model
• Harmful from the First Dollar: Government spending consistently reduces GDP growth even at low levels — with diminishing marginal harm as spending rises, matching the power law's steep-then-flattening shape
• Realistic High-Spending Outcomes: Predicts that welfare states stagnate around 0.5-1.5% growth rather than collapsing, which matches Nordic and European performance
• Generalizes the Inverse Model: The inverse model is just a special case of power law with α=1. Auto-fitting reveals the optimal α ≈ 1.5, meaning government spending is even more harmful at low levels than the simple 1/x curve implies
• Economic Intuition: Reflects how crowding-out works in practice - initial government spending displaces the most productive private investments, while later spending displaces progressively less efficient private alternatives. The tunable exponent captures exactly how steep this displacement is
• Institutional Inertia: Accounts for why high-spending countries don't collapse immediately - existing institutions, human capital, and economic structures provide some resilience even under heavy government burden

This isn't a minor technical issue — it is strong evidence that the theoretical framework is misspecified. When an economic model predicts France should be experiencing Great Depression-level contractions year after year, the more plausible conclusion is that the model's functional form is wrong, not that France's economy is. The power law model avoids this failure while still consistently showing a negative spending–growth relationship.

Power Laws Are Everywhere in Economics

The fact that a power law best describes the government spending–growth relationship shouldn't surprise us. Power laws are among the most robust empirical regularities in economics and finance. The pattern y = β₀ × x⁻ᵅ shows up across wildly different domains — wherever a few extreme observations dominate and the relationship between variables is scale-free rather than bell-curve-shaped.

Established Power Laws in Economics

• Pareto's Law of Income Distribution: Vilfredo Pareto discovered in 1896 that wealth follows a power law: the top 20% hold ~80% of wealth, and this holds across countries and centuries. The tail of the income distribution decays as x⁻ᵅ with α ≈ 1.5–3, meaning extreme wealth is far more common than a normal distribution would predict
• Zipf's Law for City Sizes: The population of the nth-largest city in a country is roughly proportional to 1/n. New York is ~2× Los Angeles, ~3× Chicago. This power law holds across countries from the US to China to Brazil, suggesting deep structural forces rather than coincidence
• Firm Size Distribution: The distribution of company sizes (by revenue, employees, or market cap) follows a power law. A handful of giants (Apple, Amazon) coexist with millions of small firms — exactly what a power law predicts, and exactly what a normal distribution cannot explain
• Financial Market Returns: Stock market crashes and booms follow power law tails, not the Gaussian distributions assumed by traditional finance (Black-Scholes, CAPM). Benoît Mandelbrot showed that "impossible" 5–10 sigma events happen far more often than bell curves predict — because returns follow a power law with α ≈ 3
• Trade Gravity Models: International trade volumes between countries decay as a power law of distance: Trade ∝ (GDP₁ × GDP₂) / distance^α. This gravity equation is one of the most successful empirical models in economics, and it's a power law
• Productivity and Innovation: The distribution of patent citations, scientific paper impact, and startup valuations all follow power laws. A tiny fraction of innovations generate most of the economic value — the "hits-driven" nature of technology
• Network Effects and Winner-Take-All Markets: Market share in platform economies (search engines, social networks, app stores) follows power laws. Google doesn't have 30% of search — it has 90%. This concentration is a natural power law outcome, not a market failure

Why Power Laws Dominate Economic Phenomena

Power laws emerge in economics for the same structural reasons they appear in physics, biology, and network science:

• Multiplicative Processes: When growth compounds (wealth begets wealth, success begets success), the resulting distribution is log-normal or power law — never Gaussian. Economic growth across countries is inherently multiplicative
• Preferential Attachment: In networks, nodes that already have many connections attract more. In economics, large cities attract more migrants, dominant firms attract more customers, and productive economies attract more investment
• Scale Invariance: Power laws have no characteristic scale — the same pattern holds whether you look at small governments (10-15% of GDP) or large ones (40-50%). This is exactly what we see in the spending-growth data: the decay pattern doesn't change shape at any threshold
• Heavy Tails: Unlike exponential decay, power laws allow for extreme observations. Countries like Singapore (17% spending, 2.8% growth) aren't "outliers" — they're exactly where the power law curve says they should be

The bottom line: Power laws are the default functional form for economic relationships involving diminishing returns, scale-free distributions, and multiplicative processes. It would be surprising if the government spending–growth relationship didn't follow a power law. The real question is why economists spent decades forcing a quadratic curve onto data that was always screaming "power law."

Data Sources

The real-world country data displayed in this simulator comes from the World Bank's comprehensive database:

• Government Expenditure: Total government expenditure (% of GDP) - Includes all government spending on goods, services, wages, transfers, and subsidies
• Economic Growth: GDP growth (annual %) - Real GDP growth rate in constant local currency
• Data Period: Averages calculated over user-selected time periods (2005-2023 range)
• Methodology: Countries included only if they have at least 3 years of data in the selected period

Note: You can download the raw data directly from the World Bank's DataBank for your own analysis.

But Does Correlation Mean Causation?

The strongest objection to this analysis is reverse causality: maybe rich countries can simply afford more government spending, so it's prosperity causing big government rather than big government killing growth. Or maybe a third variable — aging populations, for example — drives both higher spending (pensions, healthcare) and lower growth simultaneously. These are serious econometric concerns that deserve honest engagement.

The Endogeneity Problem

Cross-country correlations alone cannot prove causation. Three specific threats to causal inference exist here:

• Reverse Causality: Richer countries may choose larger welfare states as a luxury good — spending more because they can afford it, not because it reduces growth
• Omitted Variable Bias: Demographics (aging), culture (work ethic), geography (distance from equator), and institutional history all affect both spending and growth
• Simultaneity: Growth and spending are determined jointly — recessions increase spending (automatic stabilizers) while reducing growth, creating a mechanical negative correlation

Why the Causal Direction Still Holds

Despite these valid concerns, multiple lines of evidence suggest the dominant causal arrow runs from government spending to growth, not the reverse:

• Wagner's Law Reversed: Wagner's Law (1893) predicted that as incomes rise, government share of GDP would increase. But the highest-growth countries (Singapore, Hong Kong, Botswana) have resisted this tendency through deliberate policy choices. If reverse causality dominated, we would never see rich countries with small governments — yet they exist and outperform
• Time-Lagged Analysis: When researchers use initial government spending levels to predict subsequent growth (Barro 1991, Gwartney et al. 1998), the negative relationship persists. Countries that started the decade with lower spending grew faster over the following 10 years, ruling out simple reverse causation
• Instrumental Variable Studies: Afonso & Furceri (2010) and others use political and institutional instruments (electoral systems, constitutional constraints) that affect spending but not growth directly. The negative effect survives these controls
• Cross-Country Panel Data: Bergh & Henrekson (2011) surveyed the literature and found that "an increase in total government size by 10 percentage points is associated with a 0.5 to 1 percentage point lower annual growth rate" — a result robust across different methodologies

The Natural Experiments: Countries That Changed Course

The most compelling evidence comes from countries that dramatically changed their spending levels, creating quasi-natural experiments where we can observe before-and-after effects:

What these cases establish about causation: In each instance, the spending reduction was a deliberate policy choice made during or immediately before a period of weak growth — not a response to an existing boom. Ireland cut spending in 1987 during a fiscal crisis, before foreign investment arrived. Sweden cut in 1993–1995 during and immediately after its worst recession in sixty years. Canada cut in 1994–1995 under deficit pressure, not amid prosperity. The temporal sequence — policy decision, then growth acceleration 1–3 years later — is inconsistent with reverse causality, which would predict that cuts happen after growth improves a nation's fiscal position. These cases don't prove causation definitively, but they shift the burden of proof: a reverse-causality story would need to explain why austerity programs adopted during crises reliably precede growth recoveries.

Remaining Limitations

In the interest of intellectual honesty, this analysis has real limitations that readers should weigh:

• Cross-Sectional Design: The simulator uses country averages, which lose within-country time dynamics. Panel data with country fixed effects would strengthen the analysis
• Small Sample: 113 countries pass the standard quality filters (N=38 without any exclusions on an older version of the dataset). No confidence intervals or p-values are displayed (though R² = 0.42 with N=113 implies p < 0.001 for the power law fit)
• No Controls: The model uses a single variable. A multivariate regression controlling for demographics, trade openness, and institutional quality would be more rigorous
• Spending Composition: Not all government spending is equal. The power law is fitted to countries in the 2005–2023 World Bank sample that have already funded basic governance, so the marginal dollar in the data is overwhelmingly transfers, subsidies, public wages, and debt service — not courts or property-rights enforcement. This actually strengthens the finding: even when the sample skews toward the least-productive categories of spending, the negative relationship holds. Additionally, some spending intentionally slows economic activity that imposes costs on others (pollution, systemic financial risk) — slowing that activity is the goal, not a flaw
• Survivorship Bias and the Data Floor: No country in the World Bank panel has government spending below ~8% of GDP. This floor is not random. Territories with near-zero formal government either collapse and lose World Bank coverage, or informalize so completely that GDP measurement breaks down. The absence of observable 3–5% spending countries with high growth is itself informative: it may reflect selection failure rather than confirming that a state minimum is necessary.

  There is a subtler version of this bias: what we call "state failure" partly reflects the values of the institutions doing the classifying. A territory where drug markets, organ markets, and unregulated finance operate openly may have high welfare by revealed preference — people are voluntarily choosing those transactions — but low measured GDP and a "failed state" classification by international bodies that are themselves products of high-spending governments. Somalia 2000–2010, for example, had a thriving informal telecoms sector, hawala money networks, and livestock trade — almost none of which appears in official GDP figures. Hernando de Soto documented this systematically in The Mystery of Capital: the informal economy of low-state territories is large, real, and invisible to standard national accounts. So the data may simultaneously undercount economic activity in low-spending territories and misclassify them as failures because the transactions they enable are ones high-spending states have decided to prohibit.

  This does not imply zero government is optimal — the Power Law model is fitted to the observed domain (8–67% spending) and its boundary behavior is a mathematical artifact, not a prediction. What it does imply is that the data cannot speak to the sub-8% range, and the apparent evidence against very small states may partly reflect measurement and classification choices rather than economic reality.

None of these invalidate the core finding — that government spending explains a remarkably large share of cross-country growth differences, and the relationship is consistently negative — but they do suggest caution in drawing absolute conclusions. The power law pattern is real. Whether the implied optimal spending level is 10% or 20% of GDP requires more sophisticated analysis than a single-variable regression can provide.

The Nordic Question: Rich Despite High Spending, or Rich Before High Spending?

The most common objection to the power law model is the Nordic countries. Denmark, Sweden, and Finland spend 49–53% of GDP on government — yet they are wealthy, well-functioning societies with high living standards. Doesn't that disprove the thesis?

It doesn't, for a reason that runs through the entire article: wealth (a stock) is not the same as growth (a flow). The power law model predicts growth rates, not GDP levels. A country that built its wealth over a century of relatively lean government and then expanded its welfare state will still be rich — it just won't grow as fast as it otherwise would.

Sweden is in fact the article's strongest case study for causation: it cut spending from 67% to ~49% of GDP between 1993 and 2007, and had its best growth decades since the 1960s immediately after — a sequence documented in the natural experiments section above. The Nordic model is not an exception to the relationship; it is a confirmation.

Norway requires a separate caveat: roughly 15–20 percentage points of its measured "government spending" flows through the Government Pension Fund (sovereign wealth fund) which invests oil revenues abroad rather than directing them into the domestic economy. Its effective domestic public consumption is closer to 40% of mainland GDP — still high, but materially different from the headline figure.

The correct question to ask about the Nordics is not "why are they wealthy?" but "why do they grow slowly compared to their own potential, and compared to peers with lower spending?" On that question, the data gives a consistent answer.

Why There's Still Scatter Even Though Government Size Matters Most

Even though the cross-country data consistently suggests a negative relationship between government size and growth, you'll notice significant scatter around any curve model. This doesn't undermine the empirical pattern — it just shows that government size is a strong but not the only factor associated with economic growth. Understanding these other factors helps explain why some high-spending countries aren't completely collapsed and why some low-spending countries aren't growing even faster.

Key Growth Factors Beyond Government Size

1. Institutional Quality

• Rule of Law: Strong property rights and contract enforcement enable investment and innovation
• Regulatory Efficiency: Streamlined business regulations reduce costs and encourage entrepreneurship
• Government Effectiveness: Quality of public administration and policy implementation matters more than size
• Corruption Control: Clean governance ensures resources go to productive uses

Example: Singapore (17% spending, 2.8% growth) combines efficient government with strong institutions, while some high-spending countries struggle with bureaucratic inefficiency.

2. Development Stage and Catch-Up Potential

• Convergence Effect: Developing countries can grow faster by adopting existing technologies
• Productivity Gaps: Room for improvement varies dramatically across countries
• Infrastructure Needs: Countries with infrastructure deficits can see high returns to investment

Example: Ethiopia (8.5% spending, 6.5% growth) and Rwanda (19.7% spending, 7.0% growth) benefit from catch-up growth potential despite very different government sizes.

3. Demographics and Human Capital

• Age Structure: Young, working-age populations drive growth; aging populations face headwinds
• Education Quality: Skills matching, innovation capacity, and adaptability to change
• Health Outcomes: Healthy populations are more productive and innovative

Example: Japan (19.7% spending, 0.2% growth) faces demographic headwinds with an aging population, while countries with young populations have natural growth advantages.

4. Economic Structure and Openness

• Trade Integration: Access to global markets and value chains
• Sectoral Composition: Manufacturing vs. services vs. agriculture productivity differences
• Export Competitiveness: Currency levels, labor costs, and specialization
• Foreign Investment: Technology transfer and capital inflows

Example: Ireland (22.3% spending, 6.8% growth) benefits from being a hub for multinational corporations and EU market access.

5. Innovation and Technology

• R&D Investment: Both public and private research spending drives long-term growth
• Technology Adoption: Digital infrastructure and automation capabilities
• Knowledge Spillovers: Proximity to innovation centers and universities
• Entrepreneurial Culture: Risk-taking and business creation rates

6. Macroeconomic Stability

• Inflation Control: Price stability enables long-term planning and investment
• Financial Development: Access to credit and efficient capital allocation
• Exchange Rate Policy: Competitiveness and stability for trade
• Debt Sustainability: Manageable debt burdens and fiscal space

7. Natural Resources and Geography

• Resource Endowments: Can be a blessing or curse depending on management
• Geographic Location: Access to markets, climate, and natural advantages
• Energy Costs: Access to affordable energy for industry and consumers

Example: Norway manages oil wealth well through sovereign wealth funds, while some resource-rich countries suffer from "Dutch disease."

Understanding the Outliers

Policy Implications

The scatter in the data teaches us that minimizing government size is necessary and usually sufficient for growth, but other factors can either amplify or diminish the benefits. Note that this analysis does not refute the theoretical case for a narrow set of public goods (defense, rule of law, core infrastructure) where market provision may be genuinely insufficient — it only shows that, empirically, countries with very low aggregate spending still achieve strong growth. The policy lesson is about scale, not about eliminating every government function. The most successful growth strategies:

• Government Reduction First: Cut spending and regulations as the primary growth strategy
• Let Markets Handle the Rest: Most "other factors" (education, infrastructure, innovation) are better provided by private markets than government programs
• Institutional Quality Means Less Government: Strong rule of law protects private property and voluntary exchange, not government programs
• Demographic and Geographic Factors Are Partial Explanations: Population aging and geography explain some cross-country variation, but they don't account for the full magnitude of the spending-growth relationship in the data
• The Singapore Model (With Caveats): Small budgetary government with strong property rights consistently beats the European welfare state model. However, Singapore uses mandatory savings (CPF ~37%) and state housing (HDB ~80%) that function as off-budget quasi-government programs. The lesson is less "no government" and more "government that operates through market-compatible mechanisms rather than taxes and transfers"

Bottom line: The countries that combine small government with strong institutions achieve the highest growth rates. In single-variable cross-country regressions, government size explains more of the variance in growth than most other commonly tested determinants — which gives it first-order policy relevance even accounting for the endogeneity caveats discussed above.

Why the Quadratic Model Persisted

The traditional Armey Curve survived longer than its empirical record warrants. Part of the explanation is that it offered something for multiple constituencies:

• Politicians: An "optimal zone" argument rather than a demand for maximum reduction
• Academics: A tractable, smooth curve amenable to econometric refinement
• Institutions: A framework compatible with incremental adjustment rather than structural overhaul
• Voters: A story where both spending and growth could coexist at the right level

The data is consistent with the crowding-out hypothesis: each additional dollar of government spending displaces private investment and consumption that would, on average, generate higher growth. Whether this reflects the mechanism Austrian economists emphasize or a simpler reallocation away from higher-return private uses, the cross-country pattern points in the same direction.

The Political Economy of Model Selection

One reason the power law model has not entered mainstream policy discourse may be that its implications are more uncomfortable than those of the quadratic alternative. This is not unusual in economics: as James Buchanan argued, economists respond to incentives like everyone else, and theories that are compatible with existing institutional arrangements tend to receive more development, testing, and citation than those that are not.

What Makes the Power Law Politically Awkward

The quadratic model offers politicians a navigable message: find the sweet spot, fine-tune from there. The power law offers a less negotiable one: any expansion of government spending carries a growth cost. Specific tensions include:

• No "Optimal" Compromise: The quadratic model permits arguments about balance and calibration; the power law does not
• Challenges Technocratic Framing: If aggregate spending is monotonically harmful, the policy question shifts from "how much" to "what for" — a less tractable debate
• Difficult to Campaign On: Electoral systems tend to reward promises of new services; a model that treats all spending as costly is harder to translate into platforms
• Institutional Friction: Large portions of the research and policy apparatus are embedded in, or funded by, government — a structural reason to expect slow uptake of frameworks that challenge its scope

Why the Quadratic Model Remained Useful

The traditional Armey Curve remained in circulation not necessarily because researchers found it more convincing, but because it was a workable shared reference point. It gave both sides of fiscal policy debates a common vocabulary while leaving room for disagreement.

Institutional Inertia and Paradigm Change

This pattern is not unique to the Armey Curve. Economic paradigms tend to persist until the empirical anomalies they accumulate become too large to absorb — and paradigm change is faster when it is not also institutionally costly. The power law model may gain broader acceptance as:

• Fiscal stress: Debt trajectories in high-spending economies force reconsideration of what government spending actually buys
• Cross-country divergence: If low-spending economies continue to outperform, the policy relevance of the pattern becomes harder to ignore
• Private substitutes: Technology-driven alternatives to public services gradually shifting the perceived necessity of government provision
• Independent replication: Researchers outside the main academic-policy pipeline re-examining the data and arriving at similar conclusions

The Opportunity Cost of Model Choice

If the power law model describes the data better, then policies calibrated to a quadratic framework have likely underdelivered on growth — not because policymakers were acting in bad faith, but because the model they were optimizing against was wrong. The stakes of that error compound over decades.

This matters beyond academic model comparison. Small systematic underestimates of government spending's growth cost add up: a country that sustains 0.3% lower annual growth for thirty years is roughly 10% poorer in real terms than it would otherwise have been. Choosing the right model is not a technical footnote — it shapes the trajectory of living standards.

There is also a pro-growth case for the political viability of the power law model: countries that reduce spending and achieve higher growth expand the tax base, which can make fiscal consolidation self-reinforcing rather than just contractionary. The short-term transition costs are real, but the long-run trajectory is more sustainable. Whether political systems can credibly commit to that trajectory is a separate question from whether the underlying economics are correct.

If you reference this analysis or simulator, please cite:

APA

Reszka, J. (2024). The Armey Curve: Government Spending vs Economic Growth [Interactive simulator]. Retrieved from https://julienreszka.github.io/economic-simulator/armey-curve.html

BibTeX

@misc{reszka2024armeycurve,
  author    = {Reszka, Julien},
  title     = {The Armey Curve: Government Spending vs Economic Growth},
  year      = {2024},
  url       = {https://julienreszka.github.io/economic-simulator/armey-curve.html},
  note      = {Interactive simulator. Last modified 2026-04-12}
}

Empirical Foundation

Essential Institutional Protections

Current Threats to Economic Growth Rights

Protecting Growth Rights in Practice