Papers  /  What Holds When the Money Moves: tax, trust and community in…

What Holds When the Money Moves: tax, trust and community in the Americas

InstanceSomaSoft
Published2026-10-12
SAGL-1.0 preprint open access
The licence All papers
Cite this paper 
SomaSoft. (2026-10-12). "What Holds When the Money Moves: tax, trust and community in the Americas". SomaSoft Research. https://somasoft.ai/papers/what-holds-when-the-money-moves. Licensed under SAGL-1.0.

How to read this. Written for a general reader, not for economists. [measured] marks a figure from a named primary source with its date. [I] marks our own argument — not a finding, offered to be argued with. Where we could not verify something, we say so. There is a glossary at the end for the technical words we could not avoid.

What we are not doing. We are not proposing an optimal tax system. We do not think there is one. People want different things from a tax system — security, growth, fairness, simplicity — and those cannot all be maximised at once or ranked objectively. So this paper does what analysis can honestly do: it surfaces trade-offs and names which ones are being made silently. The choice stays with the people who live with it.

The question is not whether it is fair. It is whether it holds.

Arguments about tax are almost always arguments about fairness — who pays, who should, who gets away with it. Those arguments matter. But there is a prior question, and in the Americas it may be the more urgent one: is the money there in the year it is needed?

Government revenue is not a steady stream. It rises and falls with the economy, with commodity prices, with the weather. A system is stable if those movements partly cancel out — when one source falls, another holds. It is unstable if everything moves together.

[I] That distinction is the whole paper. It is not really an economic idea. It is a structural one, and it appears identically in systems that have nothing to do with money.

The same failure, in three different systems

In ecology, what protects a community from a shock is not how many species it contains. It is whether they respond differently to the same disturbance. [measured] The term is response diversity: it is response diversity, rather than the number of members, that buffers a system, because components that fail together provide no buffering however numerous they are (Loreau and de Mazancourt, 2013). A forest of forty species that all die in the same drought is, for surviving that drought, one species.

The same idea has a name in computing. [measured] Algorithmic monoculture is the condition in which the same automated system makes a decision across a whole domain, so that someone rejected by it is rejected everywhere rather than by one evaluator (Kleinberg and Raghavan, 2021). And [measured] correlated failure is failure of many components at once because they share a vulnerability or a decision rule — converting what would have been independent random error into systematic error across the whole system.

[I] Hold onto or a decision rule. A shared vulnerability is a physical fact: one port, one harvest. A shared decision rule is a design choice. Somebody made it, usually for a good reason, usually in the name of efficiency.

A tax base is a portfolio, and most of them are badly diversified

Governments raise money from what people earn, what companies profit, what people buy, what people own, and — in much of the Americas — what comes out of the ground.

[I] These are not independent. Consumption taxes and corporate profits both fall when the economy slows. Commodity royalties fall when prices fall, and in several economies a commodity price fall is the slowdown. A system leaning on those three holds three assets that drop together. The two that move differently — taxes on personal income at the top, and on property — are typically the smallest part of the base.

This produces a specific, cruel property. [I] Revenue falls hardest in the years when need rises hardest. The budget shrinks during the recession, the drought, the hurricane season — exactly when unemployment support, food assistance and reconstruction are most required. It is not that the government chose austerity. The revenue system was built with its failures correlated, and a shock arrives as one event on both sides of the ledger.

[I] So the measurable question for a finance ministry is not "how many taxes do we have?" It is: do our revenue streams move together? That can be computed from data every treasury already holds. [I] A system with five taxes whose yields rise and fall almost in lockstep has, for stability purposes, roughly one tax. We know of no finance ministry that publishes this number, and we think it belongs in budget documentation next to the debt ratio.

The gap is one tax, and it is not the one people argue about

Here the evidence is unusually clean, and it is not what the public debate assumes.

[measured] From ECLAC's Fiscal Panorama of Latin America and the Caribbean 2026 (published 4 May 2026, using OECD data for 2023):

Latin America & Caribbean OECD
Total tax take, % of GDP 21.3% 33.7%
Income tax, % of GDP 6.2% 12.2%
Corporate income tax, % of GDP 3.8% 3.8%
Personal income tax, % of GDP 2.0% 8.2%

Read the third row again. Corporate income tax collects an identical share of GDP in Latin America and in the rich world. The entire income-tax gap is personal. And Latin America's personal income tax take is lower not only than the OECD's but than Africa's (2.8%) and developing Asia-Pacific's (2.7%). Personal income tax is 9.2% of the regional tax take against 24.4% in the OECD.

[I] This single table reframes most of the argument. The regional shortfall is not companies paying too little relative to their peers elsewhere — on this measure they pay the same. It is that the tax which falls on individual income, and which is the main counter-cyclical and progressive instrument available to any state, barely functions. That is a question about who the system can reach, not about rates.

And reaching them is exactly where the system fails. [measured] ECLAC's new harmonised estimates of personal income tax non-compliance — the only Latin American figures of their kind published with uncertainty intervals — give:

Country Share of potential revenue not collected Gap, % of GDP
Honduras 51% 0.93 [0.86–1.00]
Peru 34% 0.70 [0.62–0.78]
Costa Rica 18% 0.33 [0.31–0.35]
Panama 17% 0.43 [0.36–0.50]
Uruguay 13% 0.39 [0.31–0.52]

ECLAC's own summary is that losses "may represent up to half of estimated receipts." And [measured] its six-year panel finds the gap "has remained fairly stable in most of the countries analysed, with no clear downward trend." That period covers the region's largest wave of investment in tax technology.

Chile is the control case, because it is the region's best-administered system and the one that measures itself most carefully. [measured] Chile's own revenue service (SII, April 2025) reports VAT non-compliance of 18.2% in 2023 — and an income tax gap of roughly 45%, with total revenue loss from non-compliance at 6.0–7.6% of GDP. Its VAT gap moved 13.0% (2021) → 17.8% (2022) → 18.2% (2023), back to where it started.

What the technology actually did

Latin America is widely described as the world leader in electronic invoicing, and the claim is directionally right but smaller than the rhetoric. [measured] Mandatory e-invoicing covers 44.1% of Latin American and Caribbean countries against 48.1% of upper-middle income countries, 40.0% of low-income countries — and 25.5% of high-income countries (CIAT, from the ISORA 2023 survey). The real contrast is rich-world laggardness. Chile began in 2003 and was first.

Now the part that matters. Three separate studies, in three countries, using three different research designs, find the same thing: these systems raise reported income without raising collected revenue, because enforcement visible on one margin is displaced to another.

[measured] Brazil. São Paulo's Nota Fiscal Paulista rewarded consumers for making firms report sales. Over four years firms' reported revenue rose at least 21% — and net tax revenue rose 9.3%. Roughly half the gain was absorbed by firms increasing their reported costs (Naritomi, American Economic Review, 2019).

[measured] Ecuador. When the revenue service told firms it had detected a discrepancy between their declared revenue and third-party records, most simply did not respond. Those that did corrected their revenue to match — and also increased reported costs by 96 cents for every dollar of revenue adjustment, producing only minor increases in tax collected (Carrillo, Pomeranz and Singhal, AEJ: Applied Economics, 2017).

[measured] Peru. Mandatory e-invoicing increased reported sales, purchases and VAT liabilities by over 5% in the first year. But "existing stocks of credits were used to offset the reform's positive effects on collection" (Bellon and colleagues, Journal of Public Economics, 2022).

[I] Three countries, three designs, one mechanism. Making a transaction visible does not make the money appear; it moves the problem to whichever margin is not being watched. This is the single most important practical finding in this paper, and it is the opposite of what is usually claimed for tax technology.

The broader evidence agrees. [measured] The World Bank's Raising Revenue Right (2026) records consumer card-payment rebates that raised card transactions 50% and "ultimately did not affect total tax revenues"; a Rwandan mobile-money effect that "lasts only a few months"; a Pakistani land-record digitisation that reduced agricultural income tax revenue by removing an informal enforcement lever; and a Senegalese trial in which inspector-selected audits generated 23% more revenue than algorithm-selected audits. Its conclusion: "when tasks are complex and data are weak, the state's most-valuable resource may not be digital infrastructure but its well-trained, well-embedded civil servants."

[I] None of this says administration does not matter. It matters more than anything else here — [measured] compliance gaps dominate in developing countries while policy gaps (exemptions and reduced rates) dominate in rich ones, and the numbers confirm it: the EU's VAT compliance gap is 9.5% against a policy gap of 50.5%, while the Latin American compliance residual is 28.4% against a policy gap of 14.6%. Administration is the binding constraint. What the evidence denies is that technology by itself has been shown to relieve it.

How much is lost to corruption and illicit flows? Nobody knows, and the honest range is embarrassing

This section exists because the numbers in public circulation are not measurements, and a paper that repeated them would be doing the thing it criticises.

[measured] Published annual estimates of illicit financial flows from Latin America run from about US$0.3 billion to about US$322 billion — a factor of roughly a thousand. The low end is the sum of all outward flows any Latin American government has officially reported to the United Nations. The high end is Global Financial Integrity's trade "value gaps" for eight economies. Even comparing like with like — region-wide trade misinvoicing by the same method — the range runs from US$55.5 billion to US$322 billion.

The method underneath almost all of it is a single statistical artefact: the gap between what two countries' customs services write down for the same shipment. The objections are not fringe.

[measured] The IMF states that "official estimates of trade misinvoicing cannot be derived by transforming trade data from the IMF Trade Statistics and/or UN COMTRADE." The World Customs Organization classifies mirror analysis as a risk-screening tool requiring field verification before any inference of fraud. And GFI's own 2026 report renamed the output "Value Gaps" rather than illicit financial flows, and says results "should be interpreted as potential risk rather than definitive measures of illicit activity."

[measured] The internal evidence is worse. Reviewing GFI's own published tables, Nitsch (2016) found that the method's "inflows" exceed its "outflows" by roughly a factor of two — which would mean developing countries received a vast unrecorded net capital inflow. Switching between GFI's own methods moved Brazil's estimate by +743% and Chile's by +126%. Forstater (2018) documents that three quarters of a headline South African gold gap disappeared once someone looked up the official statistics, and that COMTRADE simply holds no record of South Africa's platinum exports for two years.

[measured] The United Nations, which put illicit flows into the Sustainable Development Goals, has kept the indicator at Tier II since 2019 and still as of 30 March 2026 — meaning "data are not regularly produced by countries." It has endorsed no global or regional total. Across the whole of Latin America, exactly four countries have produced official estimates, all for specific crimes, and no Latin American country has tested the tax-and-commercial methodology that corresponds to trade misinvoicing.

[I] One number deserves singling out because it will be quoted for years. ECLAC's 2026 Fiscal Panorama states that tax non-compliance represented 6.7% of regional GDP in 2023, or US$433 billion — with no footnote, no source line and no stated method. We are flagging it rather than using it. A figure published by a United Nations regional commission without a traceable method is exactly what this section exists to catch, and we would rather be wrong in public about ECLAC than repeat an unreplicable number.

What survives

Three claims do, and they are smaller and more useful.

[measured] The ECLAC personal income tax gaps above — the only Latin American figures published with uncertainty intervals.

[measured] The Chilean and Peruvian copper study. Built from exporters' own customs declarations benchmarked against actual contract terms rather than from partner-country mirror data, it found under-invoicing of 2.2% of Chilean copper-concentrate export value and 3.2% of refined-cathode value over 2006–2016 — falling from averages of 3.1% to 0.9% in Chile and 2.4% to 0.7% in Peru (Hanni and Podestá, ECLAC Review, 2019). Note what the authors refuse to say: that it was illicit. They say it "warrants scrutiny… to determine if [it] involved fraudulent misinvoicing or not." [I] This is the only series in the entire literature we reviewed that shows a policy visibly working.

[measured] Offshore wealth, estimated from central-bank-reported bilateral banking statistics rather than from two customs services disagreeing. The world average is about 10% of GDP, with "about 60% in Gulf countries and some Latin American economies."

Measuring corruption, and why the league tables mislead

The two indices that drive nearly all public discussion — Transparency International's Corruption Perceptions Index and the World Bank's Control of Corruption indicator — are, by their publishers' explicit statements, measures of perception held by a narrow population.

[measured] The CPI draws on 13 sources, every one an expert assessment, an executive opinion survey or a commercial risk rating. TI's own list of what the index does not capture includes "citizens' perceptions or experience of corruption; tax fraud; illicit financial flows; enablers of corruption; money-laundering; private sector corruption." The World Bank subtitles its 2026 methodology "Measuring Governance Using Perception Data," warns that scores are "not intended to serve as definitive criteria for use in credit assessments, sovereign credit ratings, investment risk," and notes that "perceptions may be slower to change than actual governance reforms."

[measured] The precision is also thinner than the league tables suggest. Cuba is scored on five sources with a confidence interval spanning 35.7 to 44.3. On the World Bank measure, nearly 40% of all country pairs globally cannot be statistically separated; in the Americas, Brazil (40.4) and Argentina (40.3) are indistinguishable, as are Haiti (17.0) and Venezuela (16.9). In its 2026 edition the Bank stopped publishing percentile ranks entirely — the word "rank" does not appear in its 74-page methodology.

Now the comparison that matters. [measured] Olken (2009) dug core samples out of Indonesian village roads to measure what had physically been stolen, then asked the villagers living beside those roads what they thought. The elasticity of perceived corruption with respect to actually-missing expenditure was about 0.08 — and instrumenting away measurement error moved it only to 0.085. Across 26 countries, the individual-level correlation between perceiving corruption and having experienced it is 0.09, with 84% of people thinking corruption likely against an 18% victimisation rate (Donchev and Ujhelyi, 2014). And when 348 experts of exactly the type these indices recruit were placed beside 35,000 households in eight African countries, the experts overestimated victimisation by a factor of four, with 94% erring in the same direction (Razafindrakoto and Roubaud, 2010).

[I] We report the counter-evidence too, because this is contested: Charron (2016), with 85,000 respondents across 24 European countries, found consistency between reported corruption and both citizen and expert perceptions "remarkably high." The disagreement tracks the sample — perception indices may work far better in high-capacity European states than in the settings where the sceptical results were produced. That is itself a finding, and it does not favour using them to triage policy in Latin America.

What measurement actually finds

[measured] Brazil runs the closest thing to a corruption census anywhere: the federal comptroller selects municipalities by public lottery and audits their use of federal transfers. Those audits found corruption in 73–78% of audited municipalities, averaging R$125,000 per violation — 8.7% of audited funds in the early lotteries and 30% in a later random sample.

And the audits work. [measured] Mayors who could stand for re-election misappropriated 27% fewer resources than those who could not (Ferraz and Finan, AER, 2011). Having been audited reduced future corruption by 8% and raised the likelihood of subsequent legal action by 20% — achieved with an audit probability of only 1.3% per lottery (Avis, Ferraz and Finan, JPE, 2018). A randomised increase in audit risk of about 20 percentage points cut corruption in procurement by about 15 percentage points (Zamboni and Litschig, 2018). And in Buenos Aires, when auditors began checking what public hospitals paid for standardised supplies, the prices paid fell 15% during the crackdown and stayed 10% below the old level afterwards (Di Tella and Schargrodsky, 2003).

[I] One warning that cuts against the enthusiasm, and we think it is the most sophisticated finding in this literature. [measured] In Chile, audits reduced the use of competitive auctions, because auctions "underwent more than twice as many checks and led to twice as many detected infractions" than comparable direct contracts (Gerardino, Litschig and Pomeranz, AEJ: Applied, 2024). Detected infractions scale with how auditable a procedure is, not with how much is being stolen. Audit counts reward opacity. Any system that measures corruption by counting findings will push officials toward the procedures where findings are harder to make.

Elite tax behaviour: the denominator is the argument

Headlines report that billionaires pay almost nothing. The underlying research is better than the headlines and says something more precise.

[measured] From the same administrative dataset — US tax records matched to the Forbes 400, 367 tax units, 2018–2020 — the effective tax rate of the wealthiest Americans is:

Measured against Top 400 All taxpayers
Economic income 23.8% 30.2%
Wealth 1.3% 5.6%
Haig-Simons income (incl. unrealised gains) 21.2% 21.5%
Individual fiscal income 57.1% 43.9%

[I] Note that the sign flips. On economic income the top 400 are below average. On wealth, far below. On Haig-Simons income they are indistinguishable from the population. On individual fiscal income they are far above it. The famous "0.3% of wealth" figure is personal taxes divided by wealth; it is a real number answering a specific question, and it is not the same question as any of the others. Anyone quoting one of these four without the denominator is making a choice and presenting it as a fact. There is also an active critic who, with four stated adjustments, moves the 23.8% figure to 38% — above the population average.

[I] Which is why we think the strongest evidence here is not the billionaire rate at all. It is the evasion gradient. [measured] Combining random audits, leaked banking data and population-wide wealth registers in Scandinavia, Alstadsæter, Johannesen and Zucman (2019) found that about 3% of personal taxes are evaded on average — against roughly 30% among the top 0.01% of the wealth distribution. That result needs no contested denominator, rests on administrative records rather than estimates, and has not been seriously rebutted.

The Latin American evidence points the same way. [measured] In Colombia, the wealthiest 0.01% conceal about one-third of their wealth offshore; the top 1% faces an effective rate of 7–8% and the top 0.01% under 4%. In Ecuador the top 1% pays about 7% and the richest 0.001% under 2%. And across the region, wealth taxes rarely exceed 0.1% of GDP.

[measured] Colombia also supplies the most useful warning about wealth taxes as an instrument: taxpayers "instantly lower their reported wealth to fall below the tax bracket cut-offs," and the effects "persist for years — even after the wealth tax is no longer in force," through overstated debt, understated business assets and offshore entities (Londoño-Vélez and Ávila-Mahecha, 2024). [I] A tax that is easy to announce and hard to assess teaches avoidance that outlives it.

Why this matters structurally, not just morally

[I] Return to the portfolio argument. Elite wealth and top-end income are precisely the part of the base least correlated with the commodity and consumption cycle. A fortune in diversified global assets does not fall when the local harvest fails. It is the hedge. Removing it from the base does not merely shrink the base — it removes the part that was holding steady while the rest fell.

And there is a second mechanism, about belief rather than arithmetic. [I] No revenue agency anywhere can audit most of its taxpayers. Compliance is overwhelmingly voluntary, and voluntary compliance rests on a belief that others are also paying. Visible elite escape does not only remove that person's contribution; it tells everyone else that compliance is for the unsophisticated.

Our own knowledge base holds a note that we did not expect to be the most useful line in a paper about tax. It concerns Marcus Aurelius:

[measured] The condition the Meditations was written against: its author held more unchecked power than almost any person before or since, and the text is a private record of a man arguing himself out of using it. The relevant observation is that the restraint is entirely internal, with no institution compelling it, which is precisely why it cannot be relied upon as a governance mechanism.

[I] That is the strongest available argument against voluntary codes and philanthropic substitutes for taxation. The Meditations is the best-documented case of elite self-restraint in history — a man with absolute power, writing privately, in good faith, talking himself out of using it. It did not outlive him by a generation. If that is the ceiling on internal restraint, no weaker version of it is a policy.

The counterpart is institutional. [measured] Separation of powers is dividing the authority to propose, decide and carry out action among independent bodies, so that no single one can act unchecked (Montesquieu, 1748; Madison, Federalist 51). [I] Applied to revenue: the body that writes tax law, the body that assesses liability and the body that audits collection should not answer to the same people.

The leaks, and what they actually recovered

[measured] Across the Panama Papers, Pandora Papers, Swiss Leaks and FinCEN Files — roughly 23.4 million documents and more than US$2 trillion in flagged transactions — the total publicly tallied recovery across all ICIJ investigations was US$1.3 billion as of April 2025. In June 2024, 28 defendants were acquitted in Panama, including the founder of the firm at the centre of the largest leak. The Pandora Papers have produced no published recovery total and no verified conviction. In the FinCEN Files, the only person convicted is the person who leaked the documents, sentenced to six months.

[measured] And the transparency reforms the leaks were meant to catalyse have been narrowed since: the United States exempted 99% of companies from its beneficial-ownership regime in March 2025, made permanent that August; the EU dropped its promised shell-company crackdown in 2025; Chile's public-registry bill has been stalled since 2023.

[I] We report this not to argue that transparency fails, but because the gap between exposure and recovery is the single most consistent finding in the field, and any roadmap that assumes revelation produces recovery is assuming something the record does not support.

What does work

[measured] The mechanism with reliably large measured numbers is the dull one: voluntary disclosure under credible threat of detection.

[I] Note that last distinction, because it is where this field most often misleads. Assets declared and revenue collected differ by an order of magnitude, and headline figures routinely quote the first while implying the second. [measured] The same applies to the OECD's regional aggregate: Latin American members have identified at least €27.8 billion since 2009 — identified, not collected — of which only €118.9 million came from automatic information exchange, and 63% of the information-request total came from a two-year spike in four countries. It is not a steady yield. It is one episode.

Brazil is running the experiment

[measured] Brazil is mid-way through the largest consumption-tax reform in the hemisphere. A constitutional amendment of December 2023 and a complementary law of January 2025 replace five taxes with a dual value-added tax. 2026 is a test year at token rates — and the levy is waived: assessment runs "for informational purposes only," and no line for either new tax appears in the federal collections report through August 2026. Full collection begins in 2027; the old state and municipal taxes are extinguished in 2033. As of October 2026, 86% of fiscal documents from obliged firms carry the new tax highlight.

[I] Two honest notes. The headline rate everyone quotes does not yet legally exist: the reference rate is set by Senate resolution on a federal audit-court calculation, and no such resolution had been filed as of October 2026. We looked for the widely-circulated figure in the 252-page implementing regulation and it is not there, so we are not printing it. And because the reform is designed to be revenue-neutral by construction, it is a test of administration and correlation, not of revenue level — which, if the argument in this paper is right, is the more important test anyway.

The layer that catches people when the fiscal layer fails

Suppose revenue falls anyway. Something has to absorb it, and in practice that something is local: a clinic, a school meal, a neighbour, a transfer payment, a cooperative.

The mechanisms are well established. [measured] Social support — emotional, practical and informational help from relationships — protects against the harmful effects of stress (Cohen and Wills, 1985). [measured] Loneliness, the distressing sense that one's connections are fewer or less close than wanted, is distinct from being alone (Cacioppo and Hawkley, 2009). [measured] Job control — the decision latitude a worker has over how and when tasks are done — combined with high demand produces job strain, associated with mental strain and cardiovascular risk; the Whitehall II cohort found job control contributing to social variation in coronary heart disease (Karasek, 1979; Marmot and colleagues, 1997).

[I] Those three findings together say something uncomfortable about how emergency support is delivered. A transfer that arrives reliably, that the recipient can plan around, and that does not require repeatedly proving desperation, is doing something different from a transfer of the same value delivered unpredictably through a process that strips decision latitude from the person receiving it. The second is a smaller intervention wearing the same price tag.

We have to be straight about a gap here. Our research on the hemispheric evidence — long-run follow-up on conditional cash transfers, the mortality effects of Brazil's family health programme, what happened when pandemic-era emergency transfers were withdrawn, whether community institutions measurably improve outcomes after disasters — did not complete. We are not going to characterise that literature from memory. The mechanisms above are well-sourced; their regional application here is [I], and the evidence that would test it is the most important thing missing from this paper.

[measured] We also keep a curated caution against our own enthusiasm: fungal connections between tree roots exist and some carbon moves through them, but popular claims that forests cooperate through them rest on few field studies and are inflated by citation bias (Karst and colleagues, 2023). [I] We keep that entry because borrowing an admired mechanism from a field without checking whether it is real in that field is the standing failure mode of papers like this one.

The governance shape that fits

[measured] Ostrom's work on commons that lasted identifies eight conditions: clear boundaries, locally fitted rules, participation by those affected, accountable monitoring, graduated sanctions, cheap conflict resolution, recognised rights to organise, and nested layers of governance (Ostrom, 1990). [measured] Graduated sanctions means responses that start small and escalate, rather than a single harsh penalty.

[I] Three do specific work here.

Graduated sanctions describes what revenue enforcement already looks like at the bottom and conspicuously does not at the top: a late filing by a small trader attracts escalating small corrections, while large-scale avoidance attracts either nothing or a prosecution a decade later that may end in acquittal. The ladder is missing in the middle, exactly where the money is. The Brazilian audit evidence suggests why that matters: what worked was not severity but probability — a 1.3% chance of audit, applied consistently and published.

Nested layers answers a real objection: that local services cannot be funded locally, because the places with greatest need have the smallest base. Money raised where the base is broad, spent where the knowledge is local.

Accountable monitoring is what corruption attacks first, and the cheapest to restore.

A roadmap, ordered by what binds

[I] Everything here is our argument. Each step names what would show it worked and what would show it wrong, because a roadmap whose steps cannot fail is a wish.

First: publish the correlation. Every treasury should compute and publish how closely its major revenue streams move together, and what share of revenue is counter-cyclical. It costs nothing and uses data already held. Wrong if correlation turns out to be low in systems that nonetheless collapse during shocks — then the instability is elsewhere, most likely debt service.

Second: fix who the system reaches, not what the rates are. The regional gap is personal income tax at 2.0% of GDP against 8.2%. Corporate rates are not the shortfall. Wrong if countries that broadened personal income tax coverage show no revenue gain relative to those that changed rates.

Third: measure collected, never reported. Every technology programme should be evaluated on revenue collected, with the offsetting-margin check built in from the start — because in Brazil, Ecuador and Peru the reported-revenue gains were absorbed by costs and credits. Wrong if a programme shows a durable collection gain with no offsetting adjustment; we would want to know what it did differently.

Fourth: buy audit probability, not audit severity. Brazil bought an 8% reduction in corruption with a 1.3% audit probability, published. Buenos Aires cut hospital supply prices 15% by checking. Wrong if raising audit probability stops producing effects above some threshold, which nobody has mapped.

Fifth: use disclosure programmes, and report both numbers. They have the largest measured yields in this entire literature. They should always publish assets declared and revenue collected, separately, as Ecuador did. Wrong if repeated programmes stop producing yield — the obvious risk, since each one teaches that another may come.

Sixth: fund the community layer counter-cyclically, and design it to preserve agency. Pre-commit funding so it does not depend on a decision taken during the shock. Wrong if outcomes under predictable transfers are no better than under discretionary ones of equal value. [I] This is the step we are least able to support with evidence, and we are ranking it on mechanism rather than measurement.

Seventh: separate the three revenue powers. Writing, assessing and auditing should not answer to one authority. Wrong if jurisdictions with formally separated revenue powers show no better collection or lower capture.

What we will not claim

We have not modelled anything. There is no simulation, no projected revenue, no costing. [I] The central argument — that correlation between revenue streams is a first-order stability problem and is not routinely measured — is an argument from structure, transferred from ecology and computing. Transferred arguments should be checked in the destination field before they are believed, and we would like someone with treasury data to try to falsify it.

We hold no proprietary fiscal data and have ingested none. The community-services evidence is missing, as stated. And we have deliberately not used the headline illicit-flow aggregates, including one published by a UN regional commission, because we could not replicate them.

[I] The honest summary is that the things which demonstrably work in this field are unglamorous and already known: audit people at random and tell them you did; make the ladder of consequences continuous all the way to the top; offer a way back in while the threat of detection is credible; and measure what you collected rather than what was reported. The Chilean copper programme cut under-invoicing from 3.1% of export value to 0.9% by comparing declarations against actual contract terms. That is the whole trick, and it is not a trick.

Glossary

Base — the total pool of income, spending or property a tax can reach. Compliance gap — the difference between tax legally owed and tax actually collected. Correlation — how closely two things move together. Two revenue streams correlated at 1.0 are, for planning purposes, one stream. Counter-cyclical — moving opposite to the economy; a counter-cyclical revenue source holds up when the economy falls. Effective tax rate — tax actually paid divided by some measure of income or wealth. Which measure you choose changes the answer enormously; see the table above. Informality — economic activity outside official registration, and therefore mostly outside the tax base. Mirror statistics — comparing what one country records as an export against what its partner records as the matching import. The gap is widely treated as evidence of fraud; the IMF says it cannot be used that way. Policy gap — revenue forgone through deliberate exemptions and reduced rates, as opposed to through non-compliance. Progressive / regressive — a progressive tax takes a larger share from larger incomes; a regressive one takes a larger share from smaller ones. Consumption taxes are usually regressive unless deliberately offset. Tier II indicator — a UN classification meaning the method is agreed but countries do not regularly produce the data.