Findings log

16 September 2026, within countries: the national average is mostly the right unit, with two exceptions

Ron asked whether we can compare Alberta with Ontario, or New York with California. For Europe we can, and the answer bears on every other figure in this project, because all of them treat a country as one public.

The source. The World Justice Project's EUROVOICES survey: 64,089 respondents across 110 European regions in 27 countries, asked both their trust in judges and their own left-right placement. 54,598 answered both. One wave, 2024. It carries no survey weight, so everything here is unweighted.

The American equivalent is not available. The General Social Survey codes region rather than state and its state identifiers are restricted. For Canada the Democracy Checkup has province across seven annual waves and has not yet been run this way.

Where the variation sits

The standard deviation of the court gap across all 109 regions is 1.11 points; the average standard deviation across regions INSIDE a country is 0.39. So the national average is mostly the right unit, and the project's country-level design is not hiding the action.

Except in two countries

Germany's internal range is twice the standard deviation of the gap across all European regions. A country can be as internally divided about its courts as Europe is between its countries.

Germany, where the pattern has a name

Gap is the right minus the left in trust in judges, so a negative number means the right trusts courts less.

Thuringia and Saxony-Anhalt are the two most divided regions in the whole 109, both at −2.32. Hesse, at −0.14, is among the least divided in Europe.

But it is the eastern RIGHT that is distinctive, not the east. The left trusts courts at 6.95 in the east against 7.17 in the west, which is nearly the same. The right trusts them at 5.14 against 5.99. And the divide is a tendency rather than a line: Schleswig-Holstein in the west sits at −1.81, exactly the eastern average, while Berlin in the east sits at −1.26, near the western one.

One thing worth noticing about direction

The sign of the court gap flips across Europe. In Germany, France and Austria the political right trusts courts LESS than the left. In Greece, Czechia, Poland and Spain it trusts them MORE. So "polarization about courts" does not mean one side is always the sceptical one, and any pooled analysis that ignores the sign is averaging opposite phenomena. Every gap measure in this project uses the absolute value, which sidesteps the problem but also discards it.

Outputs. data/derived/regional_gaps.csv, built by src/regional_gaps.py.

16 September 2026, the placebo on durability: courts are not distinctive, and the earlier claim falls

Ron asked whether disagreement about courts is more durable than disagreement about anything else. It is not. This is the correction that matters most, because the claim it overturns is the one I told him was the most important result of the day.

The test. Each institution is flagged divided in a country-year when its absolute left-right gap exceeds that INSTITUTION'S OWN median, so the comparison is not driven by which institutions divide publics more in the first place. Persistence is measured against a shuffle of each country's own sequence. No smoothing anywhere.

Europe, 11 institutions, European Social Survey

Courts rank 9 of 11.

Latin America, 10 institutions, Latinobarometro

Courts rank 6 of 10, in the middle of a tightly packed field.

Mean spell lengths are 2.1 to 2.9 observations for every institution in both regions. Courts sit at 2.6 and 2.4.

What this overturns

The site has been carrying, since this afternoon, the claim that a public divided about its courts is in the hardest state to leave, replicated in both regions and beating 2,000 shuffles. That analysis compared composite STATES in an alphabet where the court state absorbed any year in which courts divided a public whatever else did, and it used a three-wave rolling mean. The Eurobarometer rerun already showed the smoothing adds about 0.15. This test removes both problems and compares institutions like for like, and the distinctiveness disappears.

What survives. Disagreement about institutions persists above chance. For courts that is +0.069 in Europe at p 0.006 and +0.077 in Latin America at p 0.002, so the persistence itself is real. It is simply not a fact about courts. It is a fact about institutional disagreement, and courts are unremarkable within it, ranking below hospitals and schools in Europe and below the government and the police in Latin America.

Why I should have run this first. Every other claim in this project has been put through a placebo against other institutions measured the same way, and three earlier claims died that way today. This one was allowed to stand for several hours without it because the state-based design made the comparison awkward. That was the wrong reason not to run it.

The durability section of the results page is being corrected.

Outputs. data/derived/durability_placebo.csv, built by src/durability_placebo.py.

16 September 2026, the durability result on three times the European data

The persistence finding, which Ron judged the most important for a judicial audience, rested on eleven European Social Survey rounds. The Eurobarometer series gives a median of 32 waves per country across 29 countries from 1997 to 2026. Rerunning it there is the strongest available test of the claim, and it holds.

What could and could not be rerun. The full sequence analysis needs several institutions per country-year to say whether courts divide before or after governments, and only the justice item has been harvested from Volume C so far. The whole trust battery is in those volumes and a harvest of it is running. What could be done now is the durability claim itself in its two-state form: divided about courts in a wave or not, how long the spells last, and how persistent the divided state is.

The measure. The gap is the percentage of people at 7 to 10 on the left-right scale tending to trust the justice system minus the same among those at 1 to 4. That is a difference in percentage points and NOT the European Social Survey's difference of means, so the threshold is set on this scale, at the median absolute gap across all country-waves, and repeated at the 60th and 70th percentiles.

The result

It is not chance. Shuffling each country's own waves, keeping its exact mix and destroying only the order, gives 0.556 against the observed 0.705. Not one of 1,000 shuffles reached the observed value.

Spells. Without smoothing, 152 spells with a median of 2 waves, a mean of 3.1 and a longest of 18. Eurobarometer runs two or three times a year, so the longest run is roughly seven years. At the median threshold, 19 countries end the period divided.

Two things this settles that the shorter series could not

The smoothing objection is now quantified rather than argued. A three-wave rolling mean makes adjacent flags share two of their three inputs, so it inflates any persistence figure by construction. Here it lifts the figure from 0.705 to 0.869. Every persistence number this project has published used smoothing, and the honest figure is about 0.15 lower than the smoothed one.

The divided state is stickier than the undivided one, but not by much. Excess persistence is 0.239 for divided against 0.188 for united at the median threshold, and the margin widens as the threshold rises, to 0.307 against 0.115. So a public divided about its courts does hold that position more tenaciously than a united one holds its, and the earlier phrasing overstated how much.

What it means. The claim survives a threefold increase in data, a stricter specification and a permutation test. It survives smaller than it was stated. A public that comes to disagree about its courts stays that way for a median of two waves and a mean of three, and is more likely than chance to still be there at the next wave. That is a real finding and a more modest one than "once divided, permanently divided".

Outputs. data/derived/eb_durability.csv, built by src/eb_durability.py.

16 September 2026, fixing the reader rather than the rows

The Eurobarometer 62 rows were repaired earlier by overwriting them with the correct figures. That left the reader that produced them unchanged, so a rebuild would have reintroduced the fault. The cause is now fixed and the patch script is deleted.

Three bugs, each of which alone produced the symptom.

The country mapping held only the old codes. Eurobarometer switched from single letters (B, D, E, F, I, L, P, S, A) to two-letter codes around 2004. The mapping had the newer members (CY, CZ, EE, HU) but not the older ones (BE, DE, EL, ES, FR, IE, IT, LU, AT, PT, FI, SE). So in EB62 the header row for the fifteen older member states was not recognised as a header at all. Only every second country block was detected, pages were sliced across two questions, and both the French count rows and the English percentage rows came through as separate categories. That is where the 0.65 alongside 577.0 came from.

The page boundary included its own header. The slice took the largest boundary at or BEFORE the header row, and since the boundaries are the header rows themselves, it always returned the header row itself. The offset was therefore always zero, the header area always empty, and the question number never read. The code fell back to the sheet name, which is how the justice item came to be labelled "S2". This had been firing on every stacked wave, not just this one.

One question, two country blocks, one kept. EB62 prints the justice item twice, once for the older member states and once for the newer ones. The build step chose the single best table and discarded the other, returning 18 countries where the published tables carry 33. It now merges tables sharing a question code and item.

The cross-check that found the fault, rerun.

EB62 now agrees with the Volume C totals to 0.0 points across 28 cells.

Two further faults found on the way, both in the other direction.

Fixing the reader surfaced four disagreements in recent waves where Volume A put German trust at 67 to 72 per cent and Volume C at 41 to 45. Volume A was right. The Volume C harvest had read the GEORGIA workbook and labelled it Germany, in four waves, which the source_member column proved outright: the file is ..._volume_C_GE.xlsx. Relabelled.

And I introduced a regression myself, mapping Greece's code EL onto GR, which broke the join against Volume C and against the rest of the project. Reverted before it reached anything.

Not done. The faulty mapping means other stacked waves may have lost their older member states too. The cross-check now passes on every comparable cell, but Volume C does not cover every wave, so coverage in the waves it cannot check has not been independently verified.

Nothing was committed.

16 September 2026, Eurobarometer Volume C: the European series triples, and the gap widens almost everywhere

What arrived. The left-right breakdown of trust in the national justice system, from the Commission's published Volume C, for 35 waves between October 1997 and April 2026 across 46 territories. The median country has 32 waves against eleven rounds at most in the European Social Survey. 2.1 GB of source volumes were read through four different workbook layouts, and every one of the 4,950 band rows sums to between 97 and 103 per cent.

Licence is better than assumed: 115 of 130 distributions carry CC BY 4.0 and the rest the Commission reuse decision, both permitting reuse with attribution.

The headline, which eleven rounds could not see

The gap between right and left in the share tending to trust the justice system is wider at the end of the period in 26 of 29 countries with twenty or more waves.

Only Spain, Estonia and Finland narrowed, and each by under 5 points. Germany is the case worth pausing on: a gap of under 4 points in the early period and over 24 by the end, in a country with 35 waves and no gaps.

What polarization does to it, on 1,009 country-waves

The country-average correlation is still a null and the pooled and quantile estimates are clearly positive. Those are not in conflict. Country correlations cancel because countries differ in sign, from North Macedonia at +0.58 to Greece at −0.57, so their average carries little information. The pooled and quantile estimates use all the variation and both find the effect.

The quantile pattern repeats what the shorter series showed, on three times the data and with a far better-powered test: polarization widens the gap at every point of the distribution above the tenth percentile, and most at the top, 4.6 percentage points per unit at the 95th.

What this changes. The project's European nulls were measured as a country-average correlation on eleven observations. On a 29-year series that specification still returns a null, and two better-powered specifications do not. The honest summary is that the effect is present in the European data and is not visible in the statistic the project had been using.

A repair made along the way

The existing Volume A table carried 90 rows for Eurobarometer 62 read from the wrong sheet, with percentages up to 12,037 and French answer labels mixed with English. The Volume C harvest read that wave correctly from Q10.4, so the broken rows were replaced with the correct country totals rather than deleted, and marked in a source_note column. No percentage anywhere in that file is now above 100. The reader that produced the fault is unchanged and is recorded as an open item.

Outputs. data/eurobarometer/derived/eb_justice_by_leftright.csv, data/derived/eb_lr_gaps.csv, built by src/build_eb_gaps.py and src/repair_eb062.py.

16 September 2026, quantile regression: the effect on courts was there, in the tail

Ron pressed the quantile point a third time and it produced the largest correction of the day, this time in the direction of a finding rather than away from one.

First, a distinction that resolves an apparent contradiction. The project's domestic nulls measured the SPREAD of a public's views, using van der Eijk's agreement statistic on the whole distribution. This run measures the left-right GAP, how far apart the two political sides sit. They are different quantities: a public can be equally dispersed while its two sides converge. The null on the spread stands. What follows is about the gap.

Polarization widens the left-right gap in trust in courts, and it does so at every quantile.

Country and year effects absorbed. The effect is significant throughout and largest in the upper tail: a country-year already among the most divided about its courts gets nearly twice the increment that a typical one does.

It is not specific to courts. The same model on every other institution gives positive coefficients almost everywhere, with the armed forces steepest of all, from +0.21 at the tenth percentile to +0.44 at the ninetieth. So polarization widens left-right gaps in trust generally. Courts are neither exempt nor exceptional, which is what this project has found repeatedly by other routes.

Intensity against duration, which is the question Ron has kept returning to

Putting both in one model, with duration as consecutive years above the global median of polarization, the two behave completely differently across the distribution.

Intensity hurts everywhere and about equally. Its coefficient barely moves across the distribution.

Duration changes sign. In a typical country-year, a longer run of polarization slightly NARROWS the court gap, significantly so at the median. In the worst country-years it widens it, and the effect grows toward the extreme. That is precisely why every mean regression in this project reported nothing for duration: the two halves cancel.

What this says to Ron's framing about nipping it in the bud. Duration does nothing to the average case and may even settle it, which fits a society growing used to its divisions. Duration compounds only where things are already going badly. So early intervention has a rationale exactly where it matters most, in the countries heading for the tail, and none at all for the median country. A judge asking whether to worry about a long-running division should ask first whether their own country is near the tail.

Limits. 688 country-years for the court gap with both measures. Quantile regression with absorbed effects rather than dummies, because dummies would dominate at this sample size. Nothing here identifies a cause, and the sign change at the median is the kind of result that needs replication before it carries weight.

Outputs. data/derived/quantile_courts.csv, quantile_duration.csv, built by src/quantile_courts.py.

16 September 2026, civility in polarized societies: it is the change, not the level

Ron asked the obvious follow-up to the global null. Civility is flat worldwide, but what happens in polarized societies? It falls, and the gradient is steep.

By how much a country's polarization CHANGED, 1990-95 against 2019-24, 174 countries:

The correlation between the change in polarization and the change in respect for counterarguments is −0.46. That is among the strongest relationships anywhere in this project.

By the LEVEL of polarization reached, the gradient is much weaker and not monotone:

This is the reverse of what governs courts, and the contrast is the finding.

Everywhere courts appear in this project, the LEVEL of polarization matters and the change does not. No country below the median on average polarization ever saw its public divide about its courts, none of twenty, while the rise in polarization separated nothing at all. The international threshold works the same way: a country in the top fifth behaves differently, and where it came from is irrelevant.

Civility is the opposite. What predicts a loss of civility is the movement, not the position. A country that was always divided and stayed that way did not lose civility. A country that became divided did.

So a society can be permanently polarized and civil, and a society that is polarizing loses civility on the way. Courts respond to where a society has arrived; elite civility responds to where it is going.

In the most polarized quartile the largest falls are Nicaragua, Venezuela, Hong Kong, Thailand, Russia, Poland, Argentina and Burundi, which is a list of consolidations rather than of ordinary democratic argument.

What to tell judges. The fear that a fall in judicial standing produces incivility is not supported and the temporal order runs the other way. But the fear that polarization and incivility travel together is well supported, and the danger sign is a society in motion rather than a society that is already divided.

16 September 2026, civility: the judges' fear runs the other way

Ron asked whether there is cross-national data on civility, because judges fear that a decline in judicial standing will produce incivility before the courts and then a crisis of the rule of law. There is, it is already on this machine, and it points the opposite way.

The measure. V-Dem codes v2dlcountr, respect for counterarguments, which asks how far political elites acknowledge and respect counterarguments when important policy changes are considered. That is political civility, coded annually by country experts. Alongside it sit reasoned justification, appeals to the common good, range of consultation and an overall deliberative index. 9,528 country-years, 182 countries, 1970 to 2025.

Civility is not falling globally. Comparing 1990-95 with 2019-24, respect for counterarguments rose in 82 countries, fell in 78, and barely moved in 15. The average change is +0.09. Every other deliberation indicator also rose on average. The large falls are Nicaragua, Venezuela, Hong Kong, Benin, Thailand and Russia, which is a list of authoritarian consolidations; the large rises are Timor-Leste, Iraq, Tunisia and Libya, which is a list of post-conflict openings.

The two things move together strongly. Civility and high-court independence correlate at +0.66 across country-years, and their five-year changes correlate at +0.45. So the judges are right that these belong to one story.

But the order is the reverse of the one they fear. A cross-lagged panel on five-year changes, 181 countries, 1980 to 2025:

A change in judicial independence does not predict a later change in civility. A change in civility strongly predicts a later change in judicial independence.

And polarization sits upstream of both. Putting all three in one model, a change in civility is predicted by an earlier change in polarization at −0.045, t −3.2, and not by an earlier change in judicial independence at +0.012, t 0.7.

So the sequence the data suggests is polarization first, then the way political elites treat each other, then the standing of the courts. The judges' fear, that losing standing will make their courtrooms less civil, has the arrow backwards. On this evidence incivility among political elites is a leading indicator of judicial independence rather than a consequence of its loss.

The objection that matters most. Both measures are V-Dem expert codings, and the same country experts may code both. A general impression of a country's political health could move both series and produce this pattern without either causing the other. A cross-lagged panel establishes temporal precedence in the data, which is not the same as cause, and shared method variance is a live alternative here rather than a formality.

What it does not measure. Nothing here is civility in a courtroom. There is no cross-national data on how litigants and counsel behave before a judge, and this measures elite political discourse instead. The step from one to the other is an assumption, not a finding.

Outputs. data/vdem/vdem_civility.csv, 17 indicators.

16 September 2026, the threshold test on the survey data: courts do not break, and what does

The review data showed a court-specific break at the top fifth of polarization. The same test on the survey data does not find one, and the pattern it does find is worth more than the null.

Setup identical to the international version. 5,711 country-year-institution observations, 50 countries, 1995 to 2024, from the European Social Survey and Latinobarometro. Outcome: the absolute left-right gap in trust, in points on 0 to 10. Predictor: the country is above the 80th percentile of division into camps. Fixed effects for country and for year, so the comparison is within a country and within a moment.

Courts: nothing.

Testing each of the 19 institutions in turn, courts sit at t +1.1 and the European legal-system item at −0.6. Neither is close to conventional significance, let alone to the size of the breaks reported below for other institutions.

But the test is not empty. It finds a strong and interpretable split.

In the most polarized fifth of country-years, left and right pull further apart about the incumbent government, the presidency and the working of democracy itself, and they pull closer together about supranational bodies, the church and the army. Courts sit in the middle, unmoved in either direction.

So the two halves of the project now disagree, and the disagreement is the finding.

Internationally, at the top of the polarization range, courts are the one subject on which governments stop accepting outside advice, and nothing else breaks.

Domestically, at the same point in the same range, courts are among the institutions the public does NOT divide about further, while division over the government and over democracy itself widens sharply.

Extreme polarization changes what governments will agree to about their courts. It does not change what their publics think about them. That is a sharper statement than either half alone, and it points the concern at the relationship between a government and its judiciary rather than at public confidence.

On the risk of testing a finding to death, which Ron raised while this was running. It is real, and the record today is mixed rather than one-sided. Three claims died: two to genuine artifacts, the "Justice" tag and the International Criminal Court, and one, the checks-against-instruments reading, to a wider and fairer comparison. One claim was recovered by testing rather than destroyed by it, when the linear specification was replaced by a threshold. And this run produced a null on the thing being tested and a significant structure on eight other institutions, which is not what talking yourself out of a finding looks like.

Outputs. data/derived/threshold_surveys.csv, built by src/threshold_surveys.py.

16 September 2026, the threshold: Ron was right that the effect lives at the top

Ron objected that a linear term averages an effect across the whole range of polarization, and that with polarization the interesting behaviour is often at the extreme. He was right, and letting the effect differ above the 80th percentile recovers a court-specific finding that models 1 to 6 had averaged into nothing.

Below the top quintile, a state's own courts are the easiest thing to be told to fix. Court recommendations are accepted about 6 points more than others. Above it, that advantage is wiped out and slightly reversed.

Polarization on its own does nothing to acceptance at the top. It is the combination that matters.

By decile, the break is sharp. The court premium, in points:

Positive in the bottom eight deciles and negative in the top two. The line sits at about +1.16 on V-Dem's division-into-camps scale.

And this time it IS specific to courts. The same test on seven other subjects:

Courts are at 5.3 and the next largest is armed forces at 2.5 on 767 observations. The gap between the court estimate and every other estimate is what carries this, not any single threshold for significance. Women's rights and children's rights, with 16,305 and 13,822 recommendations, show no break at all, which is the strongest evidence available that the test would have found one had it been there.

What this rescues and what it does not. It rescues a court-specific claim, in a better form than the one it replaces. The relationship is a threshold, not a gradient: nothing distinguishes courts across four fifths of the polarization range, and everything does in the top fifth. It does not rescue the checks-against-instruments reading, which remains unsupported, nor the claim that polarization gradually erodes willingness to reform courts.

The claim as it should now be stated. In a state that is not deeply divided, the judiciary is the institution a government is most willing to be told to improve. Once a state passes into the most polarized fifth of the world, that willingness disappears, and it disappears for courts and not for women's rights, children, the press or the human rights institution.

Why this matters for the rest of the project. Every domestic null in this project was estimated linearly too. Whether the same threshold exists in the survey data has not been tested and should be.

Outputs. data/derived/upr_threshold.csv, built by src/upr_polarization.py.

16 September 2026, the checks-against-instruments reading does not survive either

This is the third time this result has been narrowed, each time by a test rather than by second thoughts, and this is where it settles. The sequence is worth keeping visible because each step looked solid until the next one was run.

First version. Judicial recommendations are accepted less than others, and states accept demands about their courts only in the general register. Wrong: the "Justice" tag is 43 per cent about courts, and the International Criminal Court was doing the work.

Second version. Polarized states refuse recommendations about their own courts 2.5 times more than anything else. True as stated, but "anything else" is the wrong comparison.

Third version. Polarization bites on institutions that CHECK a government, the press, courts and human rights institutions, and not on institutions it USES, the police and prisons. That is what the by-institution table showed and it is not what a direct test shows.

The direct test. Both groups widened so the comparison is fair: armed forces and security services join the police and prisons; civil society and anti-corruption bodies join the courts and the press. Then the two interactions estimated in one model, with a second and independent expert measure of polarization as a check.

The difference between checks and instruments is −0.004 with a t of −0.7 on the first measure and −0.005 with a t of −0.9 on the second. It is not there.

A trap that nearly produced a false reversal. V-Dem's two polarization indicators run in OPPOSITE directions and correlate at −0.65. Denmark scores −3.2 on division into camps and +2.3 on societal polarization. Used as supplied, the second measure appeared to reverse the whole finding. Flipped so that higher means more polarized on both, the two agree on the main effect. Every other analysis in this project uses v2cacamps alone, so nothing else is affected, but it is exactly the error that cost this project three surveys' worth of level statements earlier today.

What actually survives, and it is modest.

Polarized governments accept fewer recommendations of any kind. The effect is −0.018 and −0.017 on two independent expert measures, both with t above 5.7. That is robust and it is the whole claim.

Whether institutional recommendations are penalised extra depends on which measure is used, so it should be reported as unsettled rather than as a finding. Courts sit inside whatever is happening rather than apart from it.

What this costs the project. The international half no longer offers a court-specific result. It offers a general one: polarization makes governments less willing to commit to outside advice. That is consistent with the domestic half, which finds polarization does not divide publics about courts in particular either. Two halves, two nulls on court-specificity, and one robust general effect on governments. Saying so is better than three rounds of a claim that shrinks each time someone tests it.

16 September 2026, the placebo on the international result: courts are not singled out

The previous entry said polarization acts on court recommendations in particular. That was tested against "everything that is not about a court", which is too loose a comparison, and the project's own standard everywhere else is to compare courts against other institutions measured the same way. Doing that changes the claim.

Effect of the reviewed state's polarization on accepting a recommendation, by what the recommendation is about. Same model throughout: fixed effects for state and cycle, holding constant specificity, shared organisation and the recommender's democracy score.

Courts are not the outlier. The media is hit harder and national human rights institutions about the same. So the earlier claim, that this is the one place where polarization does something to courts in particular, does not stand.

What stands is better. The institutions where polarization bites are the ones that check a government: the press, the courts, the national human rights institution, and probably elections, though 716 recommendations cannot carry that. The institutions where it does nothing are the ones a government uses: the police, at +0.018 and not significant, and the prison system, at −0.025 and not significant.

A polarized government does not refuse to be told to improve its institutions. It refuses to be told to strengthen the institutions that constrain it, and agrees as readily as anyone to be told about the institutions that serve it.

That is a sharper claim than the one it replaces, and it sits better with the rest of the project, which has found again and again that courts behave like other contested political objects rather than like a category of their own.

The estimate is robust to any single country. Dropping each of the 25 states with the most court recommendations in turn leaves it negative in 25 of 25 cases, ranging from −0.049 to −0.031. Guatemala is the most influential single case and removing it still leaves −0.031.

What this does to the two halves of the project. The domestic work finds polarization does not reliably divide publics about courts. The international work now finds polarization makes governments refuse outside advice about the institutions that check them, courts among them. The common thread is not courts. It is that polarization operates on the constraint relationship, and courts matter here because they are one of the constraints, not because they are courts.

16 September 2026, the sequence analysis finished: Europe types more cleanly than Latin America

Ron asked whether the sequence analysis had been run on the other data. It has now, and the answer on the remaining sources is that they structurally cannot carry it.

Europe, clustered by trajectory

Optimal matching on 21 countries, 2002 to 2024, with substitution costs from observed transition rates. Five clusters, average silhouette 0.523 against 0.243 for Latin America, so European countries fall into types far more cleanly.

The fourth cluster is Spain, Hungary and Poland, which is the European rule-of-law group falling out of the data rather than being put into it. All three entered the state in the mid-2010s and stayed. Nothing in the method knows anything about judicial reform legislation.

The largest cluster is the one where elected institutions divide a public and courts never do, eight countries including all the Nordics and Britain. That is the modal European trajectory and it is the one the project's general null describes.

What the other sources cannot do, and why

So the sequence analysis is complete on both sources that can support it, and the limit is the data rather than the effort.

What would change this

Standard Eurobarometer asks trust in the national justice system two or three times a year and its published Volume C breaks that item down by left-right band, per country per wave. 42 waves carrying the item are already downloaded. Harvesting Volume C would take the European series from 11 time points to about 42, which is exactly what the sequence and persistence analyses most need. That harvest is now running.

Outputs. data/derived/sequence_clusters_ess.csv, sequence_distances_ess.csv, sequence_plot_ess.png, alongside the Latin American equivalents.

16 September 2026, polarization in the review data: Ron's hypothesis holds

Ron proposed scoring every state in the Universal Periodic Review for polarization in the year it was reviewed, from a third-party source, and predicting acceptance as the corridors paper predicted it from relational position. He then asked the sharper question: do more polarized states refuse recommendations when it comes to courts, but not other recommendations.

They do. src/upr_polarization.py.

Setup. V-Dem's v2cacamps attached to both the state under review and the recommending state at the exact year of review, which the session field gives ("16 - May 2013"). 111,656 recommendations carry scores for both states, 2008 to 2024, 172 states under review. Court recommendations are those naming a court, tribunal, judge or the judiciary and NOT naming an international court or the Rome Statute, which is the distinction that reversed the earlier result. 3,451 of them. Linear probability model with fixed effects for the state under review and for the cycle.

The interaction is the answer.

Run separately rather than as an interaction:

The penalty on courts is 2.5 times the penalty on everything else. Across the range from the tenth to the ninetieth percentile of polarization, 3.3 units, that is 15.6 points of acceptance for court recommendations against 6.2 points for the rest.

At average polarization a recommendation about a state's own courts is MORE likely to be accepted than other recommendations, by 3.5 points. That advantage is wiped out and reversed as polarization rises. A calm state is unusually willing to be told to improve its judiciary. A divided state is unusually unwilling.

What identifies this, stated plainly. The fixed effects mean the comparison is within one state's own answers across its reviews. 164 of 172 states have more than one polarization value across their reviews, but the median within-state range is only 0.64 units, so the estimate rests on modest movement in most states and on a handful of large swings: Brazil 3.3 units, Peru 3.2, Ukraine 2.8, Poland 2.6 from an already high base.

The crude comparison is much weaker and should be reported alongside. Splitting states at the median polarization without any controls gives a difference in differences of only −1.7 points. The regression estimate is larger because it compares a state with itself over time and holds constant how specific the demand is, who is asking and how democratic they are. Both are true and they answer different questions: whether polarized COUNTRIES refuse court recommendations, which is weakly, and whether a country refuses them more WHEN it is more polarized, which is clearly.

Why it matters for the project. The domestic half found that polarization does not lower public trust in courts in any general way. This is the first place in the project where polarization does something to courts specifically and survives a placebo. It does not act on what a public thinks of its judiciary. It acts on what a government will agree to do about it.

Outputs. data/derived/upr_polarization.csv, upr_polarization_models.csv.

16 September 2026, justice or courts: the definition decides the finding, and reverses it

Ron asked whether the corridors extension should be about justice or about courts. The answer is courts, and specifically a state's OWN courts, because the other two definitions produce the opposite result. This entry corrects the one below it.

The "Justice" tag is not about courts

Of 8,450 recommendations tagged Justice by UPR Info, only 42.8 per cent name a court, tribunal, judge or the judiciary. The rest are the Rome Statute, criminal justice generally, investigations and prosecutions, transitional institutions. Meanwhile 27 per cent of recommendations that DO name a court are tagged something else: women's rights, children's rights, detention, human rights education.

So the tag is neither necessary nor sufficient, and my earlier run used it.

Separating domestic from international courts reverses the result

A quarter of court-naming recommendations are about the International Criminal Court, the Rome Statute or a regional court rather than the state's own judiciary. Their acceptance rates are nothing alike.

Three definitions, three different findings, from one dataset:

All with fixed effects for the state under review.

What is actually true

States accept recommendations about their own courts at a higher rate than anything else, 80.9 per cent against 74.7. The small negative coefficient in the third row survives only because the fixed-effects model is comparing within states that receive many such recommendations.

They refuse recommendations about international courts almost entirely, 28.5 per cent acceptance. That single category was driving everything I reported earlier.

Specific demands about domestic courts are accepted MORE, not less. The interaction is +0.131 with a t of 9.1. Specificity costs 29 points across the corpus as a whole and domestic court recommendations recover almost half of that penalty.

What I got wrong, and what it means for the frame

The previous entry said states accept demands about their courts in the general register and refuse them in the binding one, and read that as law held above the contest. That was an artifact of the Justice tag, which is mostly not about courts, and of counting the International Criminal Court as a court.

The corrected reading is close to the opposite and fits the project better. A state's own judiciary is the EASIEST thing to be asked to improve: recommendations about it are accepted more than average and specificity does not deter acceptance. An international court is the hardest, because accepting it costs sovereignty rather than administration. The line between the sacred and the profane in the international register does not fall between courts and other institutions. It falls between a state's own courts and courts above the state.

For the method question

The latent class model, if run, should be run on the domestic-court subset of 3,621 recommendations, not on the 8,381 Justice-tagged ones. That is a small subset for a four-class model and the class count would have to be justified on the subset rather than inherited from the published paper.

16 September 2026, extending the corridors argument to courts

Ron asked whether the corridors argument itself can be extended to courts, and what that would involve. It can. The rebuilt review data tags 8,381 recommendations with the issue "Justice", across four cycles and 193 states under review, with an action category recording how specific each demand is and a recorded response. That is enough to ask of judicial recommendations the questions the paper asked of human rights recommendations in general.

The first test, run rather than proposed. A linear probability model of whether a recommendation is supported, with fixed effects for the state under review, so every comparison is within one state's own responses rather than across states with different acceptance habits. 122,199 recommendations.

What this says. Judicial recommendations are accepted about three points less often than others from the same state. Specificity is what really governs refusal, at 29 points, and judicial specificity is refused a further three points beyond that. So states accept demands about their courts in the general register and refuse them in the binding one, and they do this more than for rights in general.

What it kills. The most attractive version of the story, that states accept recommendations about their courts only from their friends, does not survive. Raw figures suggested it: sharing an organisation raises acceptance by 12.7 points for judicial recommendations against 8.7 for others. With state fixed effects the interaction falls to +0.016 and is not significant. The raw difference was about which states receive judicial recommendations, not about how any state treats them.

What survives is the better finding for this project's frame. Susan Silbey's distinction between law held above the contest and law inside it predicts exactly the pattern that remains. A state will affirm the principle of judicial independence and refuse the mechanism. The international behaviour matches what the domestic survey work shows: courts keep their standing as long as nothing specific is asked of them.

What extending it properly would involve

Four further moves, in order of what each would add.

1. The corridors themselves, for judicial recommendations only. The paper's method is latent class regression on recommendation dyads. Running it on the judicial subset asks whether judicial reform travels through the same corridors as other rights or through different ones. This is the direct extension and the one that would carry the paper's own argument forward.
2. Asymmetry, which is the paper's core mechanism. The finding was that asymmetric relationships lower acceptance. Whether that asymmetry bites harder on courts is the natural hypothesis, and the fixed-effects result above suggests testing it on the SPECIFIC subset, where refusal actually happens, rather than on all judicial recommendations.
3. Over time. The paper had two cycles; there are now four. The justice acceptance gap runs +1.6 points in cycle 1, −8.4 in cycle 2, −4.2 in cycle 3 and zero in cycle 4, which is not a trend and needs explaining rather than reporting.
4. The domestic link, already run. A state whose recommendation traffic stays in its bloc has a public with less confidence in its courts, −0.62 net of democracy. Restricting that to judicial recommendations only would sharpen it.

What is not available. Ron's own analytic file from the paper is not on this machine, and the latent class assignments behind the corridors are model output rather than deposited data. Move 1 means re-estimating the classes rather than reusing them, which is a different claim about reproducibility and should be said plainly in anything written.

16 September 2026, the corridors merge: Ron's own data, and it works on the level

The idea. Ron proposed bringing in his and Ioana Sendroiu's "World Society Corridors" (Social Forces 2023, 102:377-401), on the reasoning that states following their friends in human rights is a polarization of law at the international level. The link to this project is not statistical but conceptual: everything here asks whether law is treated as standing above the political contest or as a position inside it. Inside a country that is whether left and right differ about courts. Between countries it is whether human rights recommendations travel across the whole membership or among blocs.

The data had to be rebuilt. The paper's dataset is not on disk. The Universal Periodic Review database is public, and the six-country fetcher in Ron's separate UN project generalised to all 196 state entities. 122,909 recommendations across four cycles, 193 states under review, 184 recommending states. Cycles 1 and 2 give 57,729 against the paper's roughly 50,000, which is growth in the source since the 2020 extract rather than a different filter. Validation: the paper's two opening examples reproduce word for word, with the same responses and action categories, and 9,987 of 10,000 rows in an independent 2024 export appear in the rebuild with responses agreeing on 99.97 per cent.

The measure. Coleman homophily indices comparing each state's share of recommendation traffic staying inside its own regional group or organisation against what its own volume would produce if partners were drawn at random. The raw share is not used, because a large regional group produces a high share by arithmetic.

The result, and it needed a permutation test to be believed. 987 correlations were computed across 21 predictors, 22 institutions and 3 outcomes. At that many tests a correlation of 0.6 appears by chance, so the countries were shuffled 400 times and the whole procedure rerun.

The level result is real and nothing else is. Strongest single predictors of the level of trust in courts, all net of the electoral democracy index:

Every sign runs the way Ron's argument predicts. A state whose human rights traffic is confined to its own bloc has a public with less confidence in its own courts. A state that takes recommendations from many different partners has a public with more.

The honest qualification. The relationship is stronger for courts than for other institutions, 0.348 against 0.256, but it is present for both and both exceed chance. So this is not something happening to courts alone. It is a general association between a state's international legal posture and its public's confidence in domestic institutions, within which courts are the most responsive.

What it cannot do. Fifty-two to fifty-seven countries, cross-sectional, one average per country across all cycles. No causal claim is available and none is made. A state's bloc confinement and its public's low trust could both follow from something not measured here, and the democracy control removes only the most obvious candidate.

Outputs. data/upr/upr_recommendations_all.parquet, upr_dyads.csv, upr_state_measures.csv, data/derived/upr_courts_merged.csv, upr_courts_results.csv. Built by src/fetch_upr.py, src/build_upr.py, src/merge_upr_courts.py.

16 September 2026, Canadian Hansard: the volume of court talk and the vocabulary of overriding courts

Ron's idea, and why it is the other half of the project. Everything else here measures what publics think. If polarization reaches courts it should also show in what politicians say about them, and Canada is the one case where we hold both the survey series and the full parliamentary record. The record was already on this machine, in the antisemitism project, as a 1.4 GB dump of openparliament.ca. It was read strictly read-only and streamed rather than loaded into a database, because PostgreSQL is not installed here.

Coverage. 3.87 million statement rows read, of which 1,092,106 are House debates from 1994 to 2026. 52,215 mention courts, an overall share of 4.8 per cent, running between 2.6 and 6.7 per cent by year. Committee evidence was excluded.

Matching was done carefully and the near misses were counted. Word boundaries keep "court" out of courtesy, courtyard and Courtenay, and 4,226 such near misses were counted to document the size of the risk rather than assert it away. Further exclusions: the verb "to judge" 3,299, sport and venue 130, royal court 13, a riding name and a surname 6. The word "charter" was narrowed to a rights context because charter flights and charter schools are common in this corpus, and bare "justice" was left out.

Parliamentary attention to courts is not rising

The low point of the entire series is 2023.

The vocabulary moved even where the volume did not

2025 is the highest year for the notwithstanding clause in the whole series and 2023 the second highest, while 2023 is the LOWEST year on record for mentioning courts at all. The two series correlate at −0.08 across the full period, which is to say not at all.

So in 2023 the House talked about courts less than in any year since 1994 and talked about overriding them more than in any year except one. Counting how often courts come up would have missed this entirely.

"Judicial activism" is a period term. 117 speeches in total, concentrated in 1998 to 2005 with a peak of 25 in 1999, then effectively absent for two decades, with a small return of 7 in 2026. The language of confronting courts did not disappear; it changed from questioning what judges do to naming the clause that lets a legislature set aside what they decide.

Elite attention converged while mass opinion diverged

Spread between parties in the share of their speeches mentioning courts:

Over the same years the Democracy Checkup shows the Liberal-Conservative gap in confidence in courts widening from 18 points to 25. Elite attention to courts converged across parties while mass opinion about them pulled apart. Those are not the same process and this is the first evidence in the project that they can run in opposite directions.

Confounds, stated rather than buried. The denominator is every speech a party made, so a party that spends more time in Question Period than in bill debate will look different for reasons that have nothing to do with courts. Volume is not tone: nothing here codes whether a mention is supportive or hostile. The series starts in 1994.

Blocked. LiPaD, which would extend the record to 1901, sits behind a Cloudflare bot challenge that returns 403 to any scripted fetch. That is bot detection rather than a login, so nothing was attempted around it. One signed-out browser download into data/hansard/raw/ would open the earlier century.

Terms. Reproduced under Speaker's Permission: accurate but not an official record, non-commercial, no implied endorsement.

Outputs. data/hansard/derived/court_mentions_by_year.csv, court_mentions_by_party_year.csv, court_terms_by_year.csv, examples.csv, built by src/extract_hansard_courts.py.

16 September 2026, Gallup World Poll: confidence in courts is rising nearly everywhere

What arrived. 540 country-years, 72 countries, 2007 to 2024, from Gallup's own item: "In this country, do you have confidence in each of the following, or not? How about the judicial system and courts?" None of it needed a login. Gallup licenses the microdata, but the country figures are published openly inside OECD Government at a Glance, including its Latin American and Southeast Asian editions, which is where most of the non-OECD countries come from. The 2025 edition switches source to the OECD's own Trust Survey, and those rows are kept separate rather than spliced.

What it cannot do. The item is yes or no, so it gives a level and no distribution. Nothing in this source can measure how divided a public is, and the file carries courts alone, so the placebo comparison that governs every other finding here is unavailable. This source answers one question only: is confidence in courts falling, and does polarization explain it.

Polarization does not explain the level, in the largest sample yet.

That is the cleanest null in the project. Seventy-one countries worldwide and the relationship is indistinguishable from nothing.

The bigger finding is the direction of travel, and it runs against the premise.

Among 61 countries observed at least four times across at least eight years, confidence in courts ROSE. The median change is +7 points and only 14 of 61 fell.

The United States ranks second of 61 on the size of its fall. Gallup's own annual American series, which exists for no other country, runs from 59 per cent in 2006 to 35 in 2024, a fall of 24 points.

What this means for the project's framing. The premise that courts everywhere face a crisis of public confidence is not supported. Confidence is rising in most of the world. What exists is a small set of countries moving hard the other way, with the United States second only to Türkiye, and that is consistent with everything else here: the American case is exceptional and the general claim is not.

Caveats. These are published aggregates, not microdata, so nothing can be broken down by party or ideology. Country coverage is uneven, from one observation to ten. Guatemala and El Salvador appear among the largest rises while also being the two Latin American countries whose publics divided about courts, which is a reminder that a rising average and a widening division are different things and this source can only see the first.

Outputs. data/gallup_world_poll/gwp_confidence_courts.csv, gallup_news_confidence_courts.csv, built by src/build_gallup_courts.py and src/build_gallup_news_courts.py. The access route for the licensed microdata, through the Map and Data Library with a UTORid, is a numbered list in that folder's README.

16 September 2026, the durability result, replicated in Europe

Ron judged the persistence finding the most important one for a judicial audience, so it was rerun in Europe on the European Social Survey as well as in Latin America, with the same four stress tests. src/test_persistence.py ess.

Europe replicates it and strengthens one part. Excess persistence, the chance of staying in a state minus that state's own share of all country-years:

First of four in Europe, first of five in Latin America.

Against chance. Shuffling each country's own states two thousand times, keeping its exact mix and destroying only the order, gives 0.436 in Europe against the observed 0.778, and 0.314 in Latin America against 0.700. Not one shuffle in either region reached the observed value.

The European case is cleaner than the Latin American one. Separating courts divided alone from courts divided alongside elected institutions, there is not one country-year in 21 European countries across 23 years in which a public is split about its courts while united about its politicians, parties, parliament and government. Every European instance arrives as part of a wider political division. In Latin America the isolated case occurs in 11 country-years and has a persistence of 0.091, which is to say it does not last.

So the claim to make to judges is about the combination, not about courts. A disagreement about courts on its own resolves within an observation or two. A disagreement about courts that arrives as part of a general political division is what does not readily reverse. That is a different and more useful warning than "courts are losing legitimacy", because it says the danger sign is not the courts becoming unpopular but the courts becoming one more thing the sides disagree about.

What still does not survive. Without smoothing, the court state is no longer the most persistent of all states in Latin America; a united public overtakes it. It remains the most persistent of the states in which something divides a public. In Europe at thresholds of 1.25 and 1.5 it ranks second, behind elected-and-coercive.

Published as its own section, "What happens after a public divides about its courts", at courts-polarization.pages.dev/data#durability, including the two tests that damaged the claim.

16 September 2026, stress-testing the persistence claim

Ron flagged that "once a public splits about its courts it stays split" is the finding judges will care about most. That is the reason to attack it before anyone else does. src/test_persistence.py runs five tests. Two of them damage the claim and it survives in a narrower form.

First, a number I got wrong

The 0.789 I reported came from the sequence matrix, in which a country's missing survey years inherit the previous year's state so the string stays contiguous. Forward-filling a gap creates a repeat, and repeats are exactly what a persistence figure counts. Computed on observed years only, the figure is 0.700. The published claim should use that.

What survives

The threshold does not matter. At 0.75, 1.25 and 1.5 points the court state ranks first of five on excess persistence, the probability of staying minus the state's own share of all country-years. At 1.5 the excess is +0.842.

The order is not chance. Shuffling each country's own states, which keeps its exact mix and destroys only the sequence, gives a persistence of 0.314 against the observed 0.700. None of 2,000 shuffles reached the observed value.

Spells last about twice as long. Among the states in which something divides a public, the one involving courts lasts a mean of 3.1 years against 1.7 for elected-and-coercive, 1.6 for elected only and 1.4 for coercive only. The longest single spell is 13 years.

What does not survive

Smoothing does part of the work. Each series is a three-wave rolling mean before the threshold is applied, and smoothing makes neighbouring values resemble each other by construction. With no smoothing at all, the court state's excess persistence falls from 0.472 to 0.150 and it ranks second behind a united public at 0.198. The claim that it is the most persistent state of all depends on the smoothing. The claim that it is the most persistent of the DIVIDED states holds either way.

It is not about courts alone. Redefining the court state by exclusion, so that it means only the courts divide a public, splits it in two. Courts divided alongside other institutions keeps an excess of 0.470. Courts divided by themselves occurs in 11 country-years and has a persistence of 0.091, which is to say it does not last at all.

The claim as it should now be stated

When a public comes to disagree about its courts, it does so alongside disagreement about its elected institutions, and that combined state lasts about twice as long as any other state of division. It is not chance, it does not depend on where the threshold sits, and it is not a property of courts on their own. A public that divides about courts and nothing else returns to agreement within a year or two.

Outputs. data/derived/persistence_tests.csv.

16 September 2026, sequence analysis, and a label I had wrong

First, the correction. I had been writing "the judiciary divided" and Ron said he did not know what a divided judiciary was. He was right. Nothing in this project observes judges, judgments or benches. What is divided is the PUBLIC: people who place themselves on the right and people who place themselves on the left differ in how much they trust the courts. Every label in the sequence code now says so, and the note explaining why sits in both scripts.

Why a sequence analysis rather than the onset ranking already done. Ron pointed out that sequence analysis is rare in political science and available to us. The onset analysis dated the year each institution began to divide a public and averaged the ranks. It cannot say whether countries follow the same path, whether courts ever divide a public before the elected institutions do, or whether there is more than one route. Those are questions about whole trajectories. Andrew Abbott's method answers them: represent each country as a string of states, measure distances between strings with optimal matching, and cluster. It is standard in the sociology of careers and life courses and almost unused on institutions.

Run with TraMineR, the canonical implementation, on the same principle that the agreement measure uses agrmt rather than a reimplementation. Substitution costs come from the observed transition rates rather than being chosen by hand, because deciding that a split public is "far from" a united one would build the conclusion into the measurement.

The states. Latinobarómetro, 18 countries, 1995 to 2024, one state per country-year:

Finding 1: the split about courts is the stickiest state

Once a public splits about its courts it tends to stay split. The figure above is computed on the forward-filled sequence matrix; on observed years only it is 0.700, and the stress test in the entry above it sets out what survives. The transition from a court split back to coercive-only never happens at all.

Finding 2: the strict ordering fails

Ten of sixteen countries entered the court-split state from one where elected institutions were already divided. Six did not: Bolivia, Chile, Costa Rica, the Dominican Republic, Mexico and Uruguay reached it from a state where nothing or only the coercive institutions were divided. So publics can split about courts first. The claim that courts come last is a tendency and not a rule, and the earlier entry overstated it.

Finding 3: five trajectories, and one of them reverses

The fourth cluster is the one that matters. El Salvador and Uruguay split about their courts early, stayed split for a decade, and returned to a state where only elected institutions divide. A public that has split about its courts can come back.

Reconciling this with the earlier entry. The onset analysis found three countries whose courts divided; this finds sixteen that touch the state at some point. Both are right and they measure different things. The onset rule required the split to last to the end of the series, which is a question about permanence. The sequence states are year by year, which is a question about occurrence. Episodic splitting is common; permanent splitting is rare.

Outputs. data/derived/sequences.csv, sequence_states.csv, sequence_clusters.csv, sequence_distances.csv, sequence_plot.png, built by src/build_sequences.py and src/sequences.R. TraMineR and cluster were installed from conda-forge, because the source builds need a Fortran compiler this machine does not have.

16 September 2026, two follow-ups: not an interaction, and mostly Europe

Ron asked whether the level-and-duration result is an interaction effect, and whether it holds across all the data or only in Europe. Both have clean answers and the second qualifies the finding.

It is not an interaction. It is a threshold on the level

Multiplying level by duration performs WORSE than level alone. Duration adds nothing once the level is known.

The cleanest form of the result is a two-by-two on the medians:

No country below the median on average polarization saw its public divide about its courts. Not one of twenty. Above the median, seven of twenty did. So a high sustained level looks like a condition for the outcome rather than a cause of it, which is the shape of claim qualitative comparative analysis is built for, and it is what Ron suggested when he first raised QCA.

It is mostly Europe

Europe separates the cases cleanly on both families. Latin America points the same way on every measure and reaches significance on none, which three cases cannot. So the honest statement is that this is a European result which Latin America does not contradict, rather than a finding established across both regions.

One thing is consistent across both: the rise in polarization separates nothing anywhere, 0.46 in Europe and 0.54 in Latin America, both indistinguishable from a coin flip.

Published

The whole sequencing analysis is now a section on the results page rather than only in this log, at courts-polarization.pages.dev/data#sequencing, with the regional breakdown on the page so a reader sees which region carries the claim.

16 September 2026, Europe and the pooled test: Ron was right to push

Two things were done. The Latin American sequencing was repeated on the European Social Survey, and then both regions were pooled to test Ron's objection directly. He challenged the claim that the division of a public about its courts is a symptom of an attack on the judiciary rather than a consequence of political temperature, and asked whether years or intensity of polarization would account for it. They largely do, and the earlier claim was too strong.

Europe replicates the ordering

Eleven institutions, 21 countries, 2002 to 2024, the same onset rule. The legal system ends divided in 4 of 21 countries and ranks 4.38 on onset, later than seven of the ten others. Government satisfaction divides in 14 of 21 and ranks first at 1.07. Schools and hospitals are the last of all, which is what a placebo should do.

So courts are late and rare in both regions. That part stands.

The pooled test, 40 countries, 7 where courts divided

Venezuela 2007, Hungary 2015, Israel 2016, Guatemala 2018, Nicaragua 2018, Austria 2023, Slovakia 2023. All seven started with an unusually united public and ended among the most divided. Hungary runs from a gap of 0.59 to 2.13.

Temperature and attack set against each other on the same countries. The area under the curve says how often a country whose courts divided outranks one whose did not; 0.50 is a coin flip.

Average separation by family: temperature 0.170, attack 0.159. They are not distinguishable.

Correction to the earlier entry. Saying the division looks like a symptom of an attack rather than a consequence of temperature was not supported. Sustained polarization separates the cases at least as well as judicial attack does, and slightly better on the single best measure. The defensible statement is that both mark the same small set of countries and 7 cases cannot tell them apart, which is unsurprising given that governments that attack courts are usually governing polarized societies.

What the test does settle

It is the level, not the change. The rise in polarization across a country's series separates nothing at all, at 0.49, which is exactly a coin flip. Average and peak polarization separate well. So a country whose politics grew more polarized is not thereby a country whose public divided about its courts. A country that sat at a high level for a long time is.

That refines Ron's question rather than answering it either way. Between intensity and duration, the measure that works is sustained intensity: average and peak, with years above the median close behind at 0.68. The measure that fails is the rate of increase.

Among the seven that did turn, everything predicts an earlier turn. Onset year against average polarization −0.77, against years above the median −0.81, against attacks on the judiciary −0.89. All eight measures point the same way. With seven cases the sign is worth more than the magnitude, and the measures are collinear with each other.

Outputs. data/derived/ess_lr_gaps_full.csv, sequencing_onsets_ess.csv, courts_turn.csv, built by src/build_ess_gaps_full.py, src/analyse_sequencing.py ess, src/analyse_courts_turn.py.

16 September 2026, sequencing: courts are last, and mostly they never arrive

The question. Ron asked whether courts are the last institution to become politically divided, whether they get there in the end, and whether the moment can be predicted from the intensity or the duration of the polarization.

Why Latinobarómetro is the only source that can answer it. Dating an onset needs a long annual series. Eight survey rounds cannot tell you which of two things happened first. This uses 19 countries, 1995 to 2024, up to 24 waves each, with the institutional trust grid and a left-right scale in every wave. src/build_lb_gaps.py, src/analyse_sequencing.py.

The measure. For each country, year and institution, mean trust among people placing themselves at 7 to 10 on the left-right scale minus the mean among those at 0 to 3, on a 0 to 10 trust scale. The middle of the scale is excluded rather than split. An institution counts as divided from the first year a three-wave rolling mean of the absolute gap sits above one point and stays above for the rest of the series. Requiring it to stay above is what makes an onset different from one noisy wave.

1. Does it happen at all

Courts are the least likely of the eight to end up dividing a public along political lines, over thirty years, in the region's longest survey series.

2. The order

Ranking institutions by onset year within each country, 1 being the earliest to divide. On the relative threshold, which asks which institutions became more divided than they had been, courts rank 3.67 and are later than six of the seven others. The presidency and the government divide first, at 1.60 and 1.82.

3. Where courts did turn

Venezuela in 2007, Guatemala in 2018, Nicaragua in 2018.

Those are the three Latin American cases of open judicial capture in the period. All three started from an unusually UNITED public about their courts and ended among the most divided:

4. Intensity or duration

Where a country STARTED barely separates the two groups. What accumulated does: six more years above the median, and a higher average. On this evidence duration and accumulated intensity matter and the initial level does not.

The limit that governs all of section 4. Three cases. Four numbers computed on three countries against sixteen are a description of those three countries, not an estimate of anything. The finding that can be defended is section 1, which rests on 19 countries.

What it means for the argument. Courts are not merely slower to polarize than other institutions. In most of these countries they do not polarize at all, across three decades in which governments and presidents reliably did. Where courts did divide a public, a judiciary was being captured. That is a much stronger claim than "polarization erodes confidence in courts", and it points the other way: the division of a public about its courts looks like a symptom of an attack on the judiciary rather than a consequence of political temperature.

Outputs. data/derived/lb_lr_gaps.csv, sequencing_onsets.csv, sequencing_summary.csv.

16 September 2026, Arab Barometer: the grouping does not travel, and that is informative

The test. Latinobarómetro found that as a public spreads out ideologically it loses trust in government, parties, parliament and courts, and not in church, army or police. If that grouping reflects something general about which institutions sit inside the political contest, it should reappear elsewhere. If it merely reflects one region's institutional arrangement, it should break where the arrangement differs. In most Arab Barometer countries the army is a political actor rather than a body held above the contest.

Coverage. 16 countries, nine waves, 2007 to 2024, 66 country-waves matched to V-Dem. Wave III carries no courts item at all and is dropped. Wave I has no weight and is counted unweighted. No wave carries a left-right scale, so polarization here can only be the expert coding, which removes the contamination problem and also removes the survey-based comparison. Countries contribute between one and nine waves, far fewer than Latin America, so the analysis is pooled within countries rather than run country by country.

Result, expert polarization against the level of trust, country means and a year trend removed.

The Latin American grouping does not replicate. Three intervals exclude zero and none of them is a court. As polarization rises in these countries, trust in the military and the police rises and trust in political parties falls. Courts do not move.

What that suggests. The sacred and profane line is drawn in a different place in each region rather than sitting in the same institutions everywhere. In Latin America the church and the army stand outside the contest and are untouched. Here the coercive institutions are what a polarized public turns toward, and the parties are what it turns from. Courts are on neither side of that movement.

Limits, which are severe. The military estimate rests on 18 country-waves in 12 countries and the parties estimate on 16 in 10. Most of these countries are not democracies, so expert-coded polarization is measuring something in a different institutional context from the European and Latin American cases. Nothing here is an effect.

Outputs. data/derived/ab_freq_tables.csv, ab_agreement.csv, ab_placebo.csv, built by src/build_arab_barometer.py.

16 September 2026, Latinobarómetro: the null survives real power, and a pattern appears

Why this source settles something. Every earlier country here gave between four and eleven survey years, and removing a year trend costs two of them. A correlation on that many points cannot tell a modest relationship from nothing, so every null so far has been open to the reply that we simply could not see. Latinobarómetro gives 23 years for the median country across 19 countries, 1995 to 2024, with confidence in the judiciary, a left-right scale and a vote item in every wave. src/build_latinobarometro.py, src/analyse_latinobarometro.py.

The headline null holds. Expert-coded polarization against the spread of views about courts, one correlation per country, year trend removed: mean +0.14, positive in 13 of 19, interval −0.00 to +0.29. It touches zero with 23 observations per country. That is no longer a statement about sample size.

What the placebo shows instead, and this is the substantive result. The trust grid carries up to eleven institutions asked in the same interview, so courts can be placed among them rather than compared across surveys. Against the LEVEL of trust, with the survey's own left-right spread as the measure:

Four intervals exclude zero and they are the four political institutions. Church, military, police and the presidency do not move. Where a Latin American public is more ideologically spread, it trusts its government, its parties, its parliament and its courts less, and its church and army no differently.

Courts sit inside the political cluster, at its weak end. They are not with the church and the army, and they are not as exposed as the government and the parties. That is a finding about where courts are placed in a public's mental map of its own institutions, and it is the first result in this project that survives a placebo outside the United States.

It is about the level, not the spread. Against the spread of views, no institution's interval excludes zero on the expert measure, and courts at +0.14 sit just behind parliament at +0.15. Polarization in Latin America goes with lower trust in political institutions rather than with more divided trust.

Why this matters for the frame. Susan Silbey's distinction between the sacred and the profane predicts exactly this grouping. Institutions held apart from ordinary contest, the church and the army, are untouched. Institutions inside the contest move together. Courts are in the second group, which is the empirical content of saying they have been desacralized, and their position at the weak end of it says the movement is partial.

Caveats. The measure using the survey's own left-right spread is the one that produces this result, and that measure is contaminated when both sides of a correlation are dispersion measures from the same sample. Here the outcome is a mean rather than a dispersion, which is not exposed to that artifact in the same way, but the point deserves a direct test rather than an argument. Country-years are not independent. Nineteen countries is nineteen cases whatever the number of years within them.

Outputs. data/derived/lb_freq_tables.csv, lb_agreement.csv, lb_pressure.csv, lb_placebo.csv.

16 September 2026, party against ideology: which axis divides a public about courts

Where the question came from. The pollster collection surfaced an Environics reading for Canada in 2025 that undercuts a measurement choice running through this whole project. On trust in the Supreme Court, Canadian vote intention separates people by 33 points while ideological self-placement separates them by 7.

On a separate item about whether courts guarantee a fair trial, vote separates by 21 points and ideology by 8.

Why that threatened the European results. Every European figure in this project uses the left-right scale, either as the dispersion measure of polarization or as the division defining two groups. If partisanship carries the division about courts and ideology barely does, then the European nulls could be an artifact of measuring the wrong axis rather than evidence about courts.

The test. The European Social Survey supports both cuts on the same respondents: the government-versus-opposition gap, built from the party a person voted for and the cabinet in office on the interview date, and the right-versus-left gap from self-placement. 267 country-rounds carry both.

The Canadian pattern does not hold in Europe. For courts the two axes divide a public by almost the same amount, and ideology is marginally the wider of the two. The European nulls are therefore not explained away by the choice of axis, which is the answer that leaves them standing as evidence rather than as a measurement problem.

What it suggests instead. Party and ideology come apart in Canada and, on the American evidence in this project, in the United States. In Europe they largely coincide, which is what you would expect where party systems are organised along a left-right dimension and party identification is more a position than an identity. That difference is a candidate explanation for why the United States is the one case here with a court-specific result, and it is a claim about political systems rather than about courts.

Limits. The European partisan cut is government against opposition, which is one particular partisan division and not party identity in general. The Canadian figures are one wave of one survey. Nothing here is a model.

16 September 2026, Canada: the same test, a different answer

What was run. The American battery test repeated on the Democracy Checkup, seven annual waves 2019 to 2025, full samples, confidence in the courts asked beside up to seven other institutions in the same interview, party identification in every wave. src/build_canada_battery.py.

One thing that shapes the reading. The Liberals held the federal government throughout, so the Liberal-minus-Conservative gap is also a government-minus-opposition gap, and the federal government row shows what that alone produces. The American arithmetic ran the other way: the Court's composition changed while control of Congress and the presidency turned over.

A trap worth recording. The item suffix numbers are not stable. In 2020 suffix 5 is public health officials; in 2021 it is Elections Canada and police sits at 7; from 2022 police sits at 9. Mapping by position would have compared public health officials in one year with an election commission in the next. Matching on label text fails too, because Stata truncates a label at 80 characters and the question stem consumes all of it in exactly the waves carrying the federal government item, which is the benchmark row. The script now keys on an explicit per-wave map and uses the label only to check it, reporting any disagreement rather than dropping an item silently.

Result. Liberal minus Conservative, points saying not very much or no confidence at all.

Canada does not look like the United States. Three differences.

The sign is reversed. Conservative identifiers are the more distrustful side about courts throughout, where in the United States it became Democrats.

The size is far smaller. The American Supreme Court gap moved 30 points in six years. The Canadian court gap moved 6.

It is not court-specific. Elections Canada moved 7 points and the provincial election bodies 6, both at least as much as the courts. In the United States the Supreme Court moved 30 and the next institution in the same direction moved 2. Canada shows a cluster of federal appointed and administrative bodies drifting together, not a judiciary singled out.

What does replicate. The federal government gap sits near fifty points throughout against about twenty for the courts. Courts divide a public far less than elected institutions do, which is what the European Social Survey shows across thirty countries.

Outputs. data/derived/canada_battery_gaps.csv. Published on the results page.

16 September 2026, the American case: the clearest result so far

Why this one is different. Everywhere else, asking whether something happened to courts in particular means comparing one survey with another, with different wordings and scales. The General Social Survey asks the same Americans about the Supreme Court and about eleven other institutions on the same three-point scale in the same sitting. The comparison costs nothing.

The measure. Share of Democrats saying they have hardly any confidence in an institution, minus the same share among Republicans. Positive means Democrats are the more distrustful side. Leaners counted with their party, pure independents dropped, survey weights applied. Verified against the raw file rather than taken from the collection agent's summary.

The Supreme Court swings 30 points. The next largest movement in the same direction is organised religion at 2. Everything else moves the other way or barely moves.

Congress and the executive branch are in the table on purpose. Both swing hard whenever the party holding them changes, which is a fact about who is in office rather than about the institution. The Court moves opposite to both, across a period in which its composition changed rather than its control.

The court series.

For half a century the two sides answered almost identically, and where they differed it was Republicans who were marginally the more distrustful. The gap first passes ten points in 2022. Within one two-year step the Democratic share saying hardly any confidence goes from about 17 per cent to about 53.

What this does not establish. A biennial series cannot attribute a change to an event, and the 2021 and 2022 rounds also differ in fielding after the survey's pandemic redesign. The supportable claim is narrower: whatever moved American opinion about the Supreme Court after 2021 did not move opinion about the military, the press, medicine, banks, business, religion, labour, science or schools in the same direction.

Why it matters for the argument. The project's European results are nulls, and the honest reading of them is that we lack the power to see an effect. This result is not a null, it is court-specific, and it survives the placebo test that the European survey measure failed. It makes the United States a genuine case rather than an assertion.

Outputs. data/derived/gss_battery_gaps.csv, data/derived/gss_court_by_party.csv, built by src/build_gss_battery.py. Published on the results page under "The American case".

16 September 2026, the literature: most of this is already published

Why this was checked. Ron asked directly, and said he did not want to present the question as untouched if others have done it. Fifty studies were collected and every one of the 48 with a digital object identifier resolves against Crossref. The four year discrepancies are online-first against issue year.

What is already in print, and close.

What is not there. No published study takes the SPREAD of a court-trust distribution as the outcome across countries, with van der Eijk's agreement statistic or any other ordered dispersion index. The nearest is Bauer and Morisi 2023 in Socius, who run exactly that operation on the European Social Survey 2002 to 2020 across 25 states, using variance rather than the agreement statistic, and for the European and national parliaments. Their paper has no court item. Also absent: expert-coded polarization as the predictor of court-trust spread, and the placebo comparison against schools, hospitals and police from the same interview.

Two studies that put our null in context. Teney and colleagues 2024 in PLOS ONE find that only about a fifth of 817 German attitude items show any distributive polarization, and that at most moderate, measured with the same statistic we use. Gestefeld and colleagues 2022 report that European Social Survey distributions are usually multimodal rather than bimodal. A null on the spread of court trust sits inside an existing pattern rather than contradicting one.

Where this leaves the contribution. The gap analysis is occupied. The dispersion outcome is open, and it is the thing Ron asked for before any of this was checked.

16 September 2026, the frame: courts as sacred and profane

Ron's intended frame is Susan Silbey's legal consciousness, and specifically the sacred against the profane, from Ewick and Silbey's The Common Place of Law (1998) and Silbey's 2019 review in the Annual Review of Law and Social Science.

Why it fits the dispersion measure rather than the level. Their three registers are stances toward law, not degrees of approval. Standing before the law treats law as remote, impartial and above everyday life. Standing with the law treats it as a game to be played. Standing against the law treats it as arbitrary power to be evaded. A public that shares one register agrees about its courts whether or not it rates them highly. A public in which the registers have come apart is divided about its courts in a way an average cannot show. That is the quantity we are already measuring.

Read this way, the spread is a measure of whether a society still holds law as sacred in Durkheim's sense of set apart, rather than as one more arena where interested parties contend. Polarization would then desacralize courts rather than discredit them, which would explain a public whose average rating holds while its agreement falls.

What makes this testable rather than only interpretive. The 2010 justice module carries separate items for all three registers, which the single trust scale does not:

The literature search found no published survey instrument operationalizing the three registers with reported measurement properties. The nearest cross-national instrument is the trust in justice module itself, Jackson and colleagues 2011 in the European Journal of Criminology.

Not yet run. The module is one round, 2010, about 27 countries, so it is cross-sectional.

16 September 2026, Eurobarometer: a court-specific item at last

Why this source is different. Every other court item in this project asks about trust or confidence in general, and the European Social Survey's version does not name a court at all. It asks about "the legal system". The Flash Eurobarometer asks people how they rate the independence of courts and judges in their own country. It runs annually in every member state and feeds the European Commission's Justice Scoreboard.

What has landed. Twenty wave folders, 78 MB, all from the open European Commission portal with no login. Four general-public waves have been parsed into country tables so far, labelled 2021 to 2024, covering 27 to 31 countries. Six earlier general-public waves back to 2015 are downloaded but their question has not yet been matched to the published tables.

How it is processed. The Commission publishes percentages rather than respondent files, which is enough. Van der Eijk's agreement statistic is computed from a frequency distribution over an ordered scale and is invariant to the scale of the frequencies, so published percentages give the same answer as the counts behind them. src/build_eurobarometer.py reconstructs counts from the published base and writes rows in the same shape as the main frequency table, so src/agreement.R computes the measure unchanged. That script now takes input and output file names as arguments so the same verified computation serves both tables.

"Don't know" is dropped rather than placed in the middle of the scale. Treating it as a moderate opinion would invent one, and it would pull the statistic toward consensus in exactly the countries where the item is most contested.

First descriptive result, four waves, general public. How divided each public is about the independence of its own courts, higher meaning more divided.

Serbia at 0.847 and Albania at 0.792 are the most divided publics in the file, both observed only in 2024. Poland is the least expected entry: its public became less divided about judicial independence across these four years, and its average rating rose slightly from 3.56 to 3.60 on 0 to 10, from the lowest base in the Union.

Greece has both the second largest rise in division and the largest fall in rating, from 5.14 to 4.19. Spain's rise in division coincides with the deadlock over its judicial council.

Why this is not yet analysed against polarization. Four time points per country cannot carry a year trend, which costs two degrees of freedom. The six earlier waves would take the series to about ten annual points per country from 2015, which is the whole reason for pursuing this source. The collection agent has been asked to prioritise matching them.

Caveat. The published tables give one distribution per country per wave, so this source cannot be broken down by party, vote or region. It measures how divided a whole public is, not which groups differ. The year attached to each wave is its issue date, which trails fieldwork by a few months.

Outputs. data/derived/eb_freq_tables.csv, data/derived/eb_agreement.csv.

16 September 2026, fourth run: the direct test, after Ron's objection to the subtraction

Ron's objection, which was correct. Subtracting the placebo answers whether the court effect is distinctive. It does not answer whether there is an effect on courts. A public that divides about courts and about universities has still divided about courts.

Why the subtraction was there. The survey-based polarization measure shares respondents with the court item, so a year in which a sample spread out on every ordered scale produces a correlation with no content. V-Dem's expert measure is coded without reference to any survey and cannot produce that artifact, so the court correlation stands on its own against it. The subtraction should never have been applied to the expert measure.

The direct test. Correlation of expert polarization with the spread of trust in courts, within countries, year trend removed, one value per country.

The interval includes zero and includes a modest positive effect. With 32 countries and six to eleven observations each, the test cannot separate them.

The same quantity for every institution.

Courts sit third. The institutions whose publics divide most as polarization rises are hospitals and schools, not courts and not parliament. Nothing in this ordering is separable from noise.

Position. The honest claim is that we cannot yet detect an effect of polarization on either the level or the spread of trust in courts outside the United States, and that the absence is as consistent with a small effect we lack power to see as with no effect.

16 September 2026, third run: academics added to every survey that has them

Question. Ron asked whether academics appear in surveys beyond the World Values Survey.

What was found. A label sweep across every raw file we hold:

The European items are satisfaction with a public service rather than confidence in an institution, so their levels are not comparable with the trust items. Their spread is, which is all a placebo needs, because both are 0 to 10 scales put to the same respondents in the same interview.

What changed when they went in. The placebo set for the European cases went from the police alone to the police, education and health.

Israel leaves the outcome set. Its court-spread correlation with polarization is +0.38 and its placebo correlation is +0.55, so Israeli attitudes to schools and hospitals spread out more than attitudes to courts did. Israel still has the largest fall in the level of trust in courts of the 29 European countries, and still the second largest rise in spread. What it does not have is a rise in spread that is specific to courts.

The United States survives at +0.43 against a placebo of +0.22.

The conditions still predict nothing, and one runs backwards.

Countries that grew more polarized across their own survey window are half as likely to show courts coming apart from other institutions as countries that did not.

Outputs. data/derived/cases.csv, data/derived/truth_table.csv.

16 September 2026, later: which countries are the cases, and what predicts them

Question. Ron asked which countries show courts coming apart from other institutions under polarization, and what distinguishes them, suggesting qualitative comparative analysis if the case count is small. He also asked that academics appear as a comparison outcome rather than only as the people conducting the study.

What was built. src/build_cases.py, which produces a case table and a truth table rather than fitting a model. The outcome is the detrended within-country correlation between V-Dem polarization and the spread of trust in courts, minus the same correlation for a placebo institution.

Academics now sit in the comparison. E069_54 confidence in universities and E069_03 confidence in the education system were added to the values-survey extract, and CONEDUC to the General Social Survey alongside CONARMY and CONSCI. The European Social Survey asks about neither universities nor schools, so its cases still rest on the police.

Result: none of the four conditions predicts the outcome.

The configuration in which all four conditions are present holds six cases, Argentina, Bulgaria, Spain, Hungary, Poland and Slovenia, and has the lowest consistency of any configuration at 0.17. Where governments attacked courts in public and high-court independence actually fell, confidence in courts moved together with confidence in the police rather than apart from it.

The countries with the largest differentials are Iceland (+0.68), Israel (+0.61), South Korea (+0.54), Japan (+0.53) and the Czech Republic (+0.42). Two of those five, Iceland and Japan, sit below the median on every condition, which is the clearest sign that the ordering is not picking up judicial conflict.

Reading. Polarization does not divide publics about courts in the countries where courts are under the most visible political pressure. One reading is that a government campaign against courts moves the whole institutional field at once rather than singling out the judiciary. Another is that the differential is too noisy at six to eleven observations per country to rank countries at all. The two are not separable with what is here.

Caveat that limits this hard. The outcome is a sign test on a difference of two noisy correlations, and it splits 16 of 32 cases. A measure that divides the cases in half by construction, then fails to correlate with anything, is weak evidence of a null rather than strong evidence. Before this goes anywhere it needs the threshold varied and the differential bootstrapped.

Outputs. data/derived/cases.csv, data/derived/truth_table.csv.

16 September 2026: does the American pattern hold in Israel and across Europe?

Question. In the United States, confidence in the Supreme Court both falls and splits as polarization rises. Ron asked whether that holds in Israel and in European countries, and suggested bringing in an external measure of polarization.

What was built. src/build_by_country.py, which runs the two quantities separately for every country rather than pooling, and uses two measures of polarization side by side:

src/build_pressure.py gained a GSS-US entry in to_iso3; without it the American series was silently dropped, because pivot_table discards any row with a missing index level.

Israel

Israel has the largest fall in the level of trust in courts of the 29 ESS countries, and the second largest rise in its spread.

Both directions match the United States. Israel's left-right spread did not rise, so whatever divided its public about courts is not captured by left-right dispersion.

Europe

The falling half of the American pattern is absent. Average trust in courts is lower at the last round in only 6 of 29 countries; the large falls are Israel (−2.04), Cyprus (−1.72) and Greece (−0.84), and the large rises are Lithuania (+2.25), the Czech Republic (+1.72), Bulgaria (+1.47) and Slovakia (+1.40), which is post-accession recovery rather than anything about polarization.

The splitting half is wider: the spread of court attitudes is greater at the last round in 12 of 29 countries, and both spreads rose together in six (Germany, Spain, Hungary, Lithuania, Poland, Slovakia).

The placebo test, which overturns an earlier claim

Detrended within-country correlation of each polarization measure with the spread of three institutions, median across countries:

The survey-measured relationship is not specific to courts: the spread of trust in the police moves with left-right spread exactly as much (+0.37 for both). A year in which a sample is more dispersed on one ordered scale is a year in which it is more dispersed on the others, which is a property of responding rather than of attitudes to courts.

This undercuts the earlier "+0.23 Europe" figure, which was computed pooled, with the survey measure, and without a placebo institution to compare against. The European relationship between polarization and the splitting of court attitudes should now be reported as near zero on the external measure, and as not distinguishable from the police on the internal one.

The United States is the exception and survives the test: on the external measure the spread of Supreme Court confidence moves with polarization at +0.43 while confidence in Congress moves at −0.35, so the two institutions part company.

Ranking European countries by how much more the court spread tracks polarization than the police spread does, on the external measure, puts Iceland (+0.68) and Israel (+0.61) at the top and Lithuania (−1.05) and Bulgaria (−0.70) at the bottom.

What this means for the argument

Polarization as a general pressure on courts is not supported across Europe. What is supported is a small set of cases, the United States and Israel most clearly, in which court attitudes come apart from attitudes to other institutions. That is a claim about particular countries and particular judicial conflicts, not about polarization as such.

Caveats. Between four and eleven observations per country. Country-years are not independent. Detrending costs two degrees of freedom and is only computed from six observations up, which is why the values surveys drop out of the detrended columns almost entirely. Nothing here identifies an effect.

Outputs. data/derived/by_country.csv, data/derived/by_country_series.csv.

16 September 2026 — Inside a country: European regions and Canadian provinces

Ron asked whether Alberta can be compared with Ontario, or New York with California. Every figure in this project treats a country as one public, and that is an assumption rather than a finding.

Europe, one wave. The World Justice Project's EUROVOICES survey, 64,089 respondents across 110 NUTS regions in 27 countries, carries trust in judges and a left-right self-placement. Of the variation in the regional right-minus-left gap, 86 per cent sits between countries and 14 per cent inside them. Germany is the exception: its internal range is 2.18 points, twice the standard deviation across all European regions, and the spread comes from the eastern right rather than from the east generally. Unweighted, because EUROVOICES carries no survey weight.

Canada, seven waves. The Democracy Checkup, 2019 to 2025, gives what one European wave cannot, a province watched over time. Confidence in the courts is highest in Quebec (6.14 of 10) and lowest in British Columbia (5.31), and the Conservative-minus-Liberal gap runs from −1.00 in Quebec to −1.60 in Alberta, with New Brunswick at −2.08 on only three waves. Conservative identifiers trust the courts less in every province and in every year.

Here the decomposition points the other way from Europe: 38 per cent of the variation across province-years sits between provinces and 62 per cent within a province over time. The national gap was −1.30 in 2019, narrowed to about −1.02 through 2020 to 2022, then widened to −1.65 in 2024 and −1.43 in 2025. Nova Scotia widens fastest (−0.27 a year), Saskatchewan not at all (+0.01).

Is it about courts or about Ottawa? The Liberals held the federal government across the whole period, so a Liberal-Conservative gap is also a government-opposition gap, and Alberta and Saskatchewan are where the federal government was least supported. Carrying the federal government item alongside settles it: the court gap is 0.37 to 0.44 of the federal gap in eight of nine provinces, including Alberta at 0.44. Alberta's wider court gap is its federal politics in the same proportion as everywhere else. New Brunswick alone stands out at 0.60, on three waves.

Building it, twice. The province codes in the Democracy Checkup are alphabetical and they shift between waves: Alberta is 1 in 2022 and 14 in 2019. A hard-coded numeric map put Yukon and the Northwest Territories among the largest samples and dropped Ontario entirely. The map is now read from each file's own value labels.

Outputs. data/derived/regional_gaps.csv, data/derived/canada_provinces.csv. Built by src/regional_gaps.py and src/canada_provinces.py.

16 September 2026 — The Rule of Law Index: polarization reaches the courts through the executive

Ron asked for this on the ground that judges will want it. The World Justice Project's index is the measure a judicial audience already knows, and it is the only source in this project whose outcome is not what a public believes but what practitioners and households report about how courts actually work. 135 countries match V-Dem, 2013 to 2025, twelve editions for the median country, 53 factors and sub-factors.

Why the index can answer a question the surveys cannot. The review data says polarization changes what governments will agree to about their courts; the survey data says it does not change what publics think. If that is right, polarization should reach the sub-factors about GOVERNMENT INFLUENCE over the courts and leave alone the sub-factors about access, delay and enforcement, which describe a court's ordinary functioning rather than its relationship with the executive. The index separates them.

Every one of the 53 factors goes through the same model, so courts sit among the other things the index measures rather than being examined alone. Score and polarization are both demeaned by country and by year, so the estimate is within a country and within a moment, and standard errors are clustered by country. Coefficients are index points out of 100 per unit of V-Dem's division-into-camps scale.

The direction holds and the contrast is not significant on its own

The two influence sub-factors rank 7th and 8th of 53 on how far they fall as polarization rises. The functioning sub-factors sit at essentially zero. Put in one model, the influence sub-factors fall an extra 0.67 points relative to the functioning ones, at t −1.3, so the contrast is a direction rather than a demonstrated difference.

Where it is significant is the middle of the distribution and the top of the range

Quantile regression on the score itself, with t from a bootstrap that resamples countries rather than rows, because twelve editions of one country are not twelve independent observations:

The first quantile pass used the fit's own standard errors, which assume independent observations and produced t values up to −5.4. Those are not comparable with the clustered figures elsewhere in this project, and the bootstrap replaces them.

The threshold, which is where the review data found its result too

Letting the effect differ above the 80th percentile of polarization, 5 of the 53 factors break downward and none breaks upward:

Two of the five are criminal-justice sub-factors and the criminal justice factor as a whole is close behind at −1.05, t −1.9. Courts are prominent here and they are not alone: fundamental rights and labor rights break too. The claim this supports is that the most polarized fifth of countries is where the index moves at all, not that it moves only for courts.

What this adds to the project

The three sources now say the same thing from different directions. Governments in polarized states stop accepting outside advice about their own courts (review data). Their publics do not divide any further about courts (survey data). And what deteriorates in the index is the courts' insulation from the executive rather than their speed, their accessibility or their enforcement. Polarization reaches courts through the executive rather than through opinion.

What could make it spurious. V-Dem's polarization coding and the index's expert questionnaire are both expert judgment, and an expert forming a general impression of a country's political health could move both without either causing the other. The index's own source table splits the sub-factors: 8.6 is built from experts only, while 7.4 and 1.2 combine experts with a household poll, so the effect is not confined to the pure expert measure. That speaks to the problem without settling it. Nothing here identifies a cause.

Outputs. data/derived/roli_polarization.csv, roli_quantile.csv, built by src/roli_polarization.py on data/wjp_rule_of_law/wjp_index.csv.

16 September 2026 — The Eurobarometer battery lands: the durability and sequencing reruns

The harvest that was running all day finished: all eight trust items, 35 waves from October 1997 to April 2026, 51 countries, 39,743 rows, with a log giving a reason for every workbook-item not read. Two claims on the site rested on eleven European Social Survey rounds and could now be rerun on roughly three times the European resolution. src/eb_battery_gaps.py turns it into 7,122 country-wave-institution gaps.

It also fixed a defect in the justice table: GE is Germany in the old waves and Georgia from EB101, and the legacy code table folded Georgia into Germany, so Georgia's band rows for four waves were lost. The battery table carries them.

How far apart the two sides are, by institution

Courts divide European publics less than the government, parliament or the press, and more than parties, police or doctors. That is the same ordering the European Social Survey gives, now on 35 waves.

Durability: real, and not distinctive

Each institution judged at its own median gap, no smoothing, persistence measured against 500 shuffles of each country's own sequence.

Every p is under 0.002. Courts rank fourth of eight, third among the six items asked across the whole period. So a public divided about its courts is more likely than chance to still be divided at the next wave, and the same is true of every institution in the battery, with parliament and government ahead of courts. This is the third source to reach that conclusion, after the European Social Survey and Latinobarómetro, and the corrected version of the durability claim now rests on three sources rather than one.

Sequencing: the claim holds, and the denominator nearly sank it

The alphabet is the one already in use (N nothing divided, E elected divided, EC elected and coercive, C coercive only, J the public split about its courts), on annual averages of the waves.

Both cuts are shown because neither is innocent. A common cut is cleared by the government most of the time and by the courts rarely, given gaps of 21.9 against 11.6, so it tilts toward finding courts late. Judging each institution at its own median removes that tilt and adds another, since every family then sits above its line half the time.

The error worth recording. The first run of this reported 11 of 27 at the own-median cut and concluded the ordering claim was an artifact of the threshold. It was not: 10 of those 27 countries were already divided about their courts in the first year they were observed, which says nothing about what came first. Counted and set aside, both specifications agree.

What does depend on the cut is how many countries have anything to explain. At the common cut, 6 of 29 never divide about their courts at all; at the medians, 2 of 29.

Outputs. data/derived/eb_battery_gaps.csv, eb_durability_placebo.csv, eb_sequencing.csv, built by src/eb_battery_gaps.py, src/eb_durability_placebo.py and src/eb_sequencing.py.

17 September 2026 — Published polls: the American gap changes sign, the Canadian one does not

Ron's reason for collecting these: a judicial audience is far more likely to have seen an Ipsos or a Leger number than any microdata, and every partisan gap in this project so far comes from raw files the audience cannot inspect. 3,756 published figures from 22 pollsters across 79 countries, 1972 to 2026, each carrying its question wording, field dates, base and source URL.

Of those, 424 are broken out by party, vote or self-placement, and they yield 54 polls where a right-minus-left gap can be computed inside a single poll. Only two countries have them: the United States and Canada.

The measure matters, and separating it is the finding. Almost all the American figures are Gallup's question about whether one approves of the way the Supreme Court is handling its job, which is not the same quantity as confidence in the judiciary as an institution.

The American gap tracks the Court's majority rather than a view about courts. Republicans led by 17 to 29 points through 2006 to 2008, Democrats led by 19 to 58 points from 2009 to 2016, and Republicans have led since 2017, reaching +61 in 2022, +57 in 2024 and +65 in 2025. The sign reverses twice, at the two changes in the Court's composition.

That qualifies the American result already on this site. The 30-point divergence between the American Supreme Court and eleven other institutions is real, and the published series says it is not one side turning against courts as such. Both sides move, and each approves of the Court when it is theirs.

Canada does not do this. Six polls between 2019 and 2024, all Angus Reid asking about confidence rather than approval, put Conservative identifiers 20 to 35 points below Liberal, NDP and Green identifiers, with no reversal. That is the same direction and roughly the same size as the Democracy Checkup microdata gives, which is the point of the exercise: the Canadian finding does not depend on our processing of anyone's raw file.

Levels, for 57 countries. Denmark 88 per cent, Finland 85, Austria 84 and Germany 82 rate their justice system good in the 2026 Eurobarometer; Bulgaria is at 24. Ipsos finds 21 per cent in Colombia, 18 in Chile, 16 in Argentina and 13 in Peru calling the judiciary trustworthy. These rank within a question and never across them.

What this cannot do. Different pollsters, different questions, different scales. Levels are never averaged across pollsters here, and only the gap inside one poll is comparable, because both sides answered the same question on the same day. Two American rows use confidence rather than approval and both favour the right, which is too few to say whether the reversal would appear in a confidence series as well.

Outputs. data/derived/pollster_gaps.csv, pollster_levels.csv, pollster_coverage.csv, built by src/pollster_gaps.py on data/pollsters/pollster_statistics.csv.

17 September 2026 — The Silbey registers: the division is about equal justice, not about reverence

On the list since the first day and finally run. Everywhere else in this project a court is a single trust score, which cannot separate reverence from satisfaction. The 2010 ESS justice module asked ten questions of the same 52,458 respondents in 27 countries that fall into Ewick and Silbey's three stances toward legality.

before the law dbctvrd duty to back the court's verdict, ctfrdc courts decide fairly with the law wevdct willingness to give evidence, bplcdc and doplcsy duty to the police against the law ctprpwr courts protect the rich, ctinplt political pressure, jdgcbrb bribes, wraccrp and wraccrc the poor and the ethnically other are likelier to be convicted

Each item is put on a common 0 to 1 scale with a stated direction before anything is averaged, because the module uses four different response scales.

Which register divides a public, and it is not the one the frame expects

Paired across the same countries, the against-the-law gap is wider than the before-the-law gap by 0.021, t +4.76, in 23 of 27 countries. Against the police placebo it is +0.008 at t +1.16, which does not separate them.

So the left-right division about courts in 2010 Europe was a division about equal justice rather than about deference. The right defers more, to courts and more strongly to the police. The left is markedly more likely to say courts protect the rich and powerful, bend to political pressure, take bribes, and convict the poor and the ethnically other. The register that divides publics most is the one about what courts DO, not the one about what courts ARE.

That the police placebo divides as much says part of this is a disposition toward authority in general. What is court-specific is the asymmetry between the two court registers: 0.044 against 0.023.

Desacralization: the right signs, no significance

A more divided public stands a little less before the law and a little more against it, which is what desacralization predicts, and on 27 countries in one year not one interval excludes zero. The gaps behave the same way: the against-the-law gap correlates +0.34 with polarization, the before-the-law gap −0.25, and the police gap −0.08, which is the placebo staying flat.

Where the countries sit

Denmark (0.75), Sweden (0.71) and Cyprus (0.69) stand furthest before the law; Ukraine (0.53), Slovenia (0.56) and Portugal (0.58) least. Greece (0.67), Ukraine (0.66) and Bulgaria (0.66) stand furthest against it, Denmark (0.32) least. The right stands further before the law in Greece, Ireland and Croatia; the left stands further against it in Ireland, France and Croatia.

What it does not settle

One year, so this compares countries and cannot watch one change, and every correlation rests on 27 points. The registers are built from the items the module happens to carry, which is not the same as the registers Ewick and Silbey read out of interviews. And a cross-check on the trust series shows the direction is not new: the right trusted courts more on average in 2010 (+0.41) and in the 2020s (+0.41), with large movement inside individual countries, Germany from +0.16 to −0.88 and Hungary from +0.79 to +2.20.

Outputs. data/derived/silbey_registers.csv, silbey_gaps.csv, built by src/silbey_registers.py.

17 September 2026 — Optimal matching on the Eurobarometer sequences: one trend, and no clean types

TraMineR on the 31 European sequences built from the trust battery, 1997 to 2026, with substitution costs taken from the observed transition rates rather than chosen by hand, and partitioning around medoids so each group has a real country at its centre.

The share of European publics split about their courts went from 16 to 42 per cent

Across the whole period the average country spends 8.3 years with nothing divided, 7.3 with only its elected institutions divided, 3.7 with elected and coercive, 6.4 split about its courts, and 0.8 in the off-path state where coercive institutions divide a public and elected ones do not. The 1997 column is all "nothing divided" because the alphabet uses a three-wave rolling mean, which cannot flag a division in the first waves of the series; that column is an artifact of the start and not a claim about 1997.

The clustering does not find a typology, and that is the result

Two, three and four groups fit equally badly, and 0.25 is weak structure by any reading. European trajectories differ by degree rather than falling into kinds. The Latin American run produced a cleaner separation, which is worth keeping in view: the same method on the same alphabet types one region and not the other.

Taking the two-group solution anyway, it separates by how much time a public has spent split about its courts rather than by the shape of the route:

14 countries, 3.5 years split on average, 14.0 with nothing divided: Netherlands, Germany, Sweden, Austria, Belgium, Ireland, Italy, Portugal, western Germany, Romania, Denmark, Finland, Luxembourg, United Kingdom. 17 countries, 8.8 years split, 3.6 with nothing divided: Malta, Türkiye, Czechia, Greece, Hungary, France, Cyprus, Slovenia, Lithuania, Poland, Spain, Estonia, Slovakia, Croatia, North Macedonia, Bulgaria, Latvia.

The split is close to the older member states against the 2004 and later entrants, with France in the second group and Romania in the first. Malta spends 22 of its 30 years split about its courts, Türkiye 20, Czechia and Greece 17; Denmark, Finland, Luxembourg, the United Kingdom, Bulgaria and Latvia never enter that state at all.

Outputs. data/derived/sequence_clusters_eb.csv, sequence_distances_eb.csv, sequence_plot_eb.png, built by src/build_sequences.py and src/sequences.R.

17 September 2026 — Direction: governments move first, publics divide afterwards

Every other result in this project is a correlation, a persistence figure or a threshold, and none of them says which end moves first. For a judicial audience that is the question underneath the advice about acting early: if a public divides and a government then moves against the courts, the division is a warning worth watching; if the government moves and the public then divides, the division is a consequence and watching it is watching the wrong dial.

What makes the test possible. V-Dem codes, for 182 countries from 1975, how often a government publicly attacked the integrity of its judiciary and how independent the high court is. That coding comes from different experts and different material than any survey here. The attack variable is reversed in the original: Denmark sits at +1.8 and Poland falls from +2.3 in 2014 to −4.3 in 2018, so a high score means FEW attacks. It is negated so that the variable counts attacks, and the check is in the script rather than taken on trust.

The court gap is pooled from Eurobarometer, the European Social Survey and Latinobarómetro, each standardised by its own standard deviation because the three use different scales.

The cross-lagged panel, and it points one way only

Country and year effects absorbed, clustered by country, every variable standardised over the same sample, so the two directions are directly comparable.

812 country-years, 38 countries, 2003 to 2025. The same pattern holds within Eurobarometer alone (attacks +0.084 at t +1.5, independence −0.232 at t −3.0) and within the European Social Survey alone (attacks +0.337, t +2.1). In every sample the reverse direction is indistinguishable from zero, with coefficients between +0.002 and +0.015.

So the ordering is one-way. What governments do to courts, and how divided a country's politics already is, both forecast how divided its public will be about courts a year later. How divided the public is about courts forecasts nothing at all about what the government will do next, about whether polarization will rise, or about whether judicial independence will fall.

The event study agrees, with a lag of about three years

Episodes are years when a government's attacks jumped sharply after at least three calm years. Episode fixed effects, so an unbalanced window cannot manufacture a profile out of which episodes happen to be observed when; clustered by country.

At the larger threshold, 29 episodes inside a survey series, the gap is flat before the attack and through the year of it, and then widens by about eight tenths of a standard deviation three years later. The flat run-up is the part that matters: a public was not already dividing before the government moved.

The episodes are recognisable: Poland 2015 and 2018, Hungary 2010, Türkiye 2013, Ukraine 2014, Romania 2017, Slovakia 2024, Slovenia 2012 and 2020, Italy in four separate years.

What this does and does not support

It rules out one of the two orderings. A public's division about its courts is not an early warning of an attack on courts, because it does not predict one. It is what follows.

It does not identify a cause. A cross-lagged coefficient is a prediction, and an event study without a control group describes what happened around a date. The post-event widening appears only at the larger threshold and only from the second year, and several horizons were examined, so the timing is softer than the ordering. And nothing here rules out a third thing moving both.

Outputs. data/derived/direction_crosslag.csv, direction_events.csv, direction_event_profile.csv, built by src/direction.py.

17 September 2026 — New York against California: the American states, and a correction

Ron asked in the first place whether polarization about courts could be compared inside a country. The European answer used 110 regions and the Canadian one ten provinces. I reported the American version as impossible because the General Social Survey codes region rather than state and restricts its state identifiers. That was true of the GSS and wrong as a conclusion. The Cooperative Election Study carries state, seven-point party identification and approval of the U.S. Supreme Court on samples of about 60,000, and its files are public on Harvard Dataverse with no login.

Which variable is the court one, established rather than assumed. The 2020 battery runs CC20_320a to d with no wording in the file. The Supreme Court item was identified by matching the guide's published marginals against the data: 5,295 strongly approve, 22,206 somewhat, 16,076 somewhat disapprove, 8,522 strongly, 8,839 not sure, N 60,938, which CC20_320c matches exactly. 2018 is CC18_308c, matched the same way. The script refuses to run if either check fails.

What the states show. 121,000 respondents, 2018 and 2020, 39 and 40 states passing a rule of at least 100 respondents on each side.

Across state-years, 34 per cent of the variation in the partisan gap sits between states and 66 per cent within a state from 2018 to 2020. The gap narrowed in 37 of the 38 states with both years, by 15.6 points on average, and a state's 2018 gap predicts its 2020 gap at only +0.35.

So the American answer is the Canadian answer. What moves is the national level, not the ranking of states. This is the third country to give that answer, against Europe, where 86 per cent of the variation in the regional court gap sits between countries rather than inside them.

A detail worth keeping. The partisan gap about the Supreme Court is widest in Democratic states, not Republican ones: Connecticut +55, Wisconsin +53, Massachusetts +52, Washington +47, Maine +46, against Oklahoma +16, New Hampshire +18, Nevada +28, Arkansas +29. Connecticut Republicans approve at 88 per cent while Connecticut Democrats sit at 33.

Two cautions. This item asks how the Court is handling its job, and the published-poll section shows that quantity reversing sign with the Court's composition, so it is not the trust measure used elsewhere here. And the two years bracket a change in the Court: 2018 runs during the Kavanaugh hearings, 2020 immediately after Barrett's confirmation, which is most of why the gap moved.

Outputs. data/derived/us_states.csv, built by src/us_states.py on the CES common content files in data/usa/ces/, which are git-ignored at 420 MB and refetchable from Dataverse file ids 14075808 and 4949558.

17 September 2026 — The placebo on the direction result, which corrects it twice

The direction finding was published this morning without the test every other claim in this project has been given: the same cross-lag on the other institutions measured the same way. The Eurobarometer battery carries eight, so the test is available. It corrects the finding in two places, and the second correction is worth more than the original claim.

The division about courts is not distinctive. It is the weakest case of a general pattern.

Attacks on the judiciary predicting a widening left-right gap next year, by institution, annual, country and year effects, clustered by country:

Courts rank sixth of six, and are the only institution whose estimate does not reach a t of 2. At wave frequency, where the lag is the previous wave rather than the previous year, the ordering is the same: courts sixth of seven at +0.061, t +1.3.

So a government attacking its judiciary is followed by its public dividing about everything, and about courts least of all. The claim that attacks forecast a division about courts is true only in the weak sense that they forecast a division about institutions in general.

But on the LEVEL of trust the asymmetry reverses, and courts are the clearest case

The gap is a difference between two sides. The battery also carries the national level of trust, and running the same test on the level gives a different answer:

Courts are the strongest of the eight, and the only institution where the public-to-government coefficient is three times the government-to-public one. A public losing trust in its courts precedes a government attacking them, by about a year, and courts are the clearest case among every institution the battery covers.

What has to change in what was said this morning

The claim published this morning was that a public's stance toward its courts forecasts nothing about what a government will do next. That is right about the left-right GAP and wrong about the LEVEL. Falling public trust in courts does forecast attacks on courts, more clearly than falling trust in anything else does.

The statement that survives is narrower and more precise. How far apart a country's two political sides sit about courts tells you nothing about what its government will do; how much the public as a whole trusts its courts does. Those are different quantities and the project had been treating the gap as the interesting one throughout.

What this does not settle. V-Dem's attack coding is produced by experts who live in these countries and may read the public mood, so a trust decline and an expert's judgment of attacks could share a source. A year is a coarse lag. And nothing here identifies a cause.

Outputs. data/derived/direction_placebo.csv, built by src/direction_placebo.py.

17 September 2026 — Israel: the one case where the court gap closed

Israel is the case a judicial audience already has in mind, and the only country here that can be watched across a constitutional fight over its own courts. The European Social Survey has fielded there eight times: 2002, 2008, 2011, 2012, 2015, 2016, 2020, and again from December 2023 to October 2024, which is after the overhaul fight and after 7 October. About 17,000 respondents.

Trust in the legal system, 0 to 10, weighted

The level fell two points across two decades. The gap ran near one point for fifteen years, deepened to 1.74 in 2020, and then all but vanished in 2024. It closed from both ends: the right rose from 3.38 to 4.49 and the left fell from 5.12 to 4.59. Israelis of the left and the right now report almost identically low trust in their legal system.

And it closed while every other gap in the country exploded

Change in the right-minus-left gap, 2020 to 2024:

Every gap moved the same way, toward the right being more positive, which is what happens when a public's own side holds the government during a war. On courts that movement closed a gap; on everything else it opened one. In 2024 the Israeli right sits 2.60 points above the left on the police, 2.30 on parliament and 3.59 on satisfaction with government, and 0.10 below on courts.

The expert coding says the same thing about polarization

V-Dem's division-into-camps score for Israel runs 1.02 from 2018 through 2023 and then falls to 0.02 in 2024. Attacks on the judiciary peak in 2023, at the highest level in the series.

So Israel is a within-country demonstration running the other way from everything else in this project: polarization fell sharply, and the division about courts closed with it. Most of the evidence here watches polarization rise. A case where it falls, and the court gap falls too, is harder to find and harder to explain away.

What could break this reading

Round 11 rests on 875 answers to the courts item against about 2,500 in each earlier round, and its fieldwork ran through a war. That is the noisiest point in the series and the one the reading depends on.

The composition of the two sides is not the explanation, which was worth checking. The right is 50.4 per cent of those placing themselves in 2024, against 52.1 in 2002 and 50.4 in 2010, with a mean placement of 6.0 in both 2002 and 2024. The unusual round is 2020, where the right is only 43.5 per cent, and that is the round with the deepest gap.

Nothing here separates the overhaul fight from the war. Both happened between the 2020 and 2024 readings, and one survey cannot tell them apart.

The data that is still out of reach

The Israel Democracy Institute's data portal at dataisrael.idi.org.il returns 403 to every request, with a transaction identifier characteristic of a web application firewall rather than a login page. That is bot detection, so nothing was attempted around it; the institute's main site answers normally. The Israel National Election Studies site at ines.tau.ac.il does not resolve at all from here. Both are things Ron can open in his own browser in a minute, and the 33 IDI figures already in the pollster table came from published reports rather than from either.

Outputs. data/derived/israel_series.csv, built by src/israel.py.

17 September 2026 — The Israeli Democracy Index 2025: the division is about the Supreme Court, not the legal system

The Israeli finding published this morning rested on one ESS round of 875 answers fielded through a war, and the stated weakness was that it needed a second source. Ron had the second source in his downloads: the Israel Democracy Institute's 2025 survey, 1,588 respondents, with trust in the Supreme Court by name, the same battery for four other institutions, and the judicial reform asked directly.

It does not confirm the convergence. It explains it.

Trust, share answering high or very high, 2025

By vote in November 2022, which is the split the reform fight actually ran on, the chasm is wider still: 12 per cent of coalition voters and 73 per cent of opposition voters report high trust in the Supreme Court, a gap of 61 points, while the same coalition voters sit 39 points above on the Government and 23 on the Knesset.

The two questions the ESS cannot ask

What this does to the ESS reading

The ESS asks about "the legal system" and found the left-right gap closing to −0.10 points in 2024. The Democracy Index asks about "the Supreme Court" and finds a 57-point chasm in 2025. Those are not in conflict, and read together they are sharper than either: Israelis of the left and right have converged on a middling view of their legal system in general, and are as far apart as any two groups in this project about the Supreme Court in particular.

The within-survey placebo is what makes that reading hold rather than being an artifact of two different scales. In the same survey, on the same 4-point scale, in the same month, the Israeli right is 34 points above the left on the Government, 20 on the Knesset, 18 on the Police and 10 on the IDF, and 57 points below on the Supreme Court. The sign flips for one institution and one only.

Two details

Arab respondents report higher trust in the Supreme Court than Jewish respondents, 54 per cent against 39, weighted within each sample. And 56 per cent of all Israelis say the Court should be able to strike down a Knesset law while 33 per cent support the reform, so the reform is a minority position and judicial review is a majority one, even with the right at 29 per cent on the first question.

Two coding traps in this file

The left-right scale runs 1 = RIGHT to 7 = LEFT, which is the reverse of every other source in this project. Trust runs 1 = no trust to 4 = very high trust and is not reversed, with 98 meaning don't know. Reading either backwards would have produced a confident, publishable, wrong result.

A third trap was mine: the reform item runs 1 to 5 while the judicial-review item runs 1 to 4, and clipping both at 4 dropped every respondent who strongly supports the reform, putting support at 40 per cent among the right instead of 59. The valid range is now read from each item's own value labels.

Outputs. data/derived/israel_idi_2025.csv, built by src/israel_idi.py. The microdata sits in data/israel_idi/2025/ and is git-ignored pending confirmation of the institute's terms.

17 September 2026 — Israel over twenty-six years: the country did not change its mind, it sorted

Ron downloaded the whole Israel National Election Study archive, twenty years of files from 1969 to 2025, run by the Cohen Institute for Public Opinion Research at Tel Aviv University. Six of them carry a trust item for the Supreme Court on the same four-point scale and the same left-right placement, and three carry the judicial-review item. That turns the Israeli case from a snapshot into an arc.

A correction first. The 2025 file was described in this log yesterday as the Israel Democracy Institute's. It is not. All of these, 2025 included, are the Cohen Institute's surveys in the election-study series. The Democracy Institute figures in the pollster table are a separate source that happens to agree, so the agreement between them is corroboration rather than a check that the file was read correctly.

Trust in the Supreme Court, share reporting high or very high

In 1999, 85 per cent of Israelis reported high trust in their Supreme Court and the right and the left were two points apart. By 2019, four years before the reform was announced, the gap was 37 points. By 2025 it was 57, and right-wing trust had fallen to 19 per cent while left-wing trust sat at 76, almost exactly where it was in 1999.

So the division opened long before the reform fight and widened through it. The right's collapse does the work: 85 per cent in 1999, 41 in 2019, 19 in 2025. The left has barely moved in twenty-six years.

Should the Supreme Court be able to overturn a law of the Knesset

This is the sharpest result in the Israeli case and possibly in the project. The country as a whole has not changed its mind about judicial review. 61 per cent supported it in 2009 and 56 per cent support it in 2025. What changed is that a six-point difference between right and left became a sixty-four-point chasm. Israel did not turn against judicial review. It sorted into two camps about it, with the aggregate holding almost still while the two halves flew apart.

That is polarization in the strict sense, and it is visible here in a way it is nowhere else in this project, because the level and the division can be watched separately over sixteen years on the same question.

The placebo, in the years that carry one

The government gap flips sign with who holds the government, from +47 under Netanyahu in 2019 to −12 under the Bennett-Lapid coalition in 2022 and back to +34 in 2025. The Supreme Court gap never flips. It is negative in every year and deepens.

What bounds this

1999 and 2009 are unweighted, and 2009 splits its questionnaire across subsamples: the people who answered the judicial-review item answered neither of that year's two left-right scales, with an overlap of exactly zero, so 2009 is split by the political-stream question instead. That is a different instrument, which makes the 2009 row comparable in direction and not in level. Each survey has its own mode and house effects. And 1969 to 1996 carry no court item at all, so the arc cannot be taken back further.

Two traps in these files. Variable numbers are not stable across years: v128 is Politicians from 2019 to 2022 and the Police in 2025, v129 is family status in 2019 and the IDF in 2025, so every institution here is found by reading each file's own labels. And the 2022 value labels carry invisible left-to-right marks, which made that entire year vanish from the series until the marks were stripped.

Outputs. data/derived/israel_ines_series.csv, israel_judicial_review.csv, israel_ines_2025.csv, built by src/israel_ines_series.py and src/israel_ines_2025.py. The microdata sits in data/ines/ and is git-ignored pending the terms of use.

17 September 2026 — Eighteen countries: trust in courts follows who governs, support for judicial power does not

The Israeli case rests on two questions that turned out to behave differently: how much people trust the Supreme Court, and whether they think it should be able to strike down a law. The Comparative Study of Electoral Systems Module 6 asks both of those, in eighteen national election studies fielded between 2022 and 2024, alongside six other institutions in the same battery. The file was downloaded for this project weeks ago and never opened.

F3004_2 the courts should be able to stop the government acting beyond its authority F3007_3 trust in the judiciary placebo parliament, government, scientists, parties, traditional media, social media

Judicial review is popular everywhere, and the right likes it less almost everywhere

Support runs from 58 per cent in Brazil to 90 in Slovenia, with a median near 69. The right is less supportive in 13 of the 17 studies that fielded the left-right item, and the gaps are modest: −19 in Brazil and Türkiye at the widest, +8 in Denmark at the other end. Nothing approaches Israel's −64.

Trust in courts, by contrast, flips sign with who holds them

Across the studies the partisan gap in trust in the judiciary correlates +0.68 with the partisan gap in trust in the government. Whether it is the right or the left that trusts the courts is largely a question of whose government the courts are sitting under.

The two questions are unrelated to each other

The gap in judicial-review support correlates −0.18 with the gap in trust in the judiciary. So a country can have a right that trusts its courts and still wants them constrained, which is what Türkiye, Poland and the United States all look like: the right trusts the judiciary more than the left does, and is still the side less willing to let courts stop a government.

Trust in courts follows control. Support for judicial power follows ideology. That is the sharpest statement this project has about what "polarization about courts" actually means, and it explains the Israeli case rather than sitting beside it: the Israeli right's trust collapsed while the courts blocked its government, and its support for judicial review fell at the same time and much further.

The placebo

Courts rank fourth of seven, behind the government, the press and parliament, which is the same ordering every other source in this project gives.

And it tracks polarization, in opposite directions

Each election year carries V-Dem's coding of division into hostile camps, produced independently of the survey.

The two questions move opposite ways as polarization rises. The more divided the country, the MORE the right trusts the courts relative to the left, and the LESS it wants those courts able to stop a government. Both intervals exclude zero, barely.

And the level of support for judicial review does not move at all, at −0.03. That is the Israeli pattern generalised across seventeen countries: a polarized public does not change its mind about whether courts should be able to stop a government, it separates about it.

Seventeen countries is seventeen countries, the intervals are wide, and the polarized cases here are also the ones where the right holds or recently held the courts, which this design cannot disentangle from polarization itself.

Limits. Module 6 is a snapshot, 2022 to 2024, so this compares countries and cannot watch one change. Taiwan did not field the left-right item and drops out; France did not field the trust battery. Two of the three scales are reversed in the original. Nothing here identifies a cause.

Outputs. data/derived/cses_courts.csv, built by src/cses_courts.py.

17 September 2026 — Losing faith or sorting: two diagnoses one average hides

A country whose trust in courts falls twenty points on both sides has a legitimacy problem. A country whose average holds still while the two sides pull thirty points apart has a sorting problem. Reported as an average they look identical, and this project has been reporting one or the other without separating them. Israel showed the second in its pure form and the CSES snapshot showed the same shape between countries; this runs the split on the three sources with real time depth: 765 country-institution series across 84 countries.

Each series compares its first five years with its last five, on the level of trust and on the SIZE of the left-right gap, both standardised within source because a percentage point of Eurobarometer and a point of the European Social Survey are not the same quantity. A country is called sorting or losing faith when it passes half a standard deviation.

Courts lose faith less often than other institutions and sort about as often

The cases that matter are the ones where trust ROSE and the gap widened anyway

In these six countries trust in the justice system rose by 6 to 18 points while the distance between right and left grew by 10 to 27. Germany is the clearest: ten points more trust and twenty points more division, over the same twenty-nine years and the same question.

The countries where both went the wrong way are the ones already in the literature: Türkiye (level 64 to 47, gap 17 to 44), Cyprus, Greece and Venezuela.

So the usual story, that publics are losing faith in courts, is the wrong description of most of Europe. Trust rose in most of these countries. What happened is that it stopped being shared.

And the regions differ

Not one Latin American public sorted about its courts without also losing faith in them; the mean change in level there is −0.54 points against +7.9 in Europe. Europe divided about courts it trusts more than it used to. Latin America lost confidence without dividing.

Polarization predicts sorting, and does so for courts more than for anything else

Change in V-Dem's division into camps over the same window, against the change in the gap:

Where polarization rose, the court gap widened about three times as strongly as the gaps for the other institutions did. That is the clearest court-specific statement the project has, and it is about the DIVISION rather than the level, which is the whole point of the decomposition.

Limits. 34 court series is 34 countries; the correlation is computed across country-institution series that share countries; the classification cut is a choice and moving it moves the counts, though not the correlations. Eurobarometer publishes no overall level before 2000 and only patchily to 2004, so the level window for those countries starts at their first year with a published figure, which is recorded per series. Nothing here identifies a cause.

Outputs. data/derived/level_vs_sorting.csv, level_vs_sorting_camps.csv, built by src/level_vs_sorting.py.

17 September 2026 — The principle was always contested. The institution became contested.

Mining the whole Israel National Election Study archive, rather than searching English variable labels for one phrase, turned up a question nobody in this project had noticed. It was asked in six surveys across twenty-seven years:

There are situations in which there is a contradiction between preserving the rule of law and protecting national security interests. In the case of such a contradiction, should national security interests prevail or should the rule of law prevail?

That is the rule of law asked about directly, as a thing that can lose, rather than as trust in an institution.

The answer has not moved in twenty-seven years

Between a fifth and a quarter of Israelis have said the rule of law should prevail in every situation, in every survey since 1988, and the right has been 23 to 30 points below the left throughout. There is no trend in either column. The 1988 to 1996 surveys use a seven-point scale and the 2009 to 2015 ones a three-option version, so the levels are not comparable across that break, but the gap is computed inside each survey and is stable across both.

Which is the opposite of what happened to the Supreme Court

In 2009 Israelis disagreed far more about the principle than about the Court's power, by 27 points against 6. By 2025 that had inverted, at 64 points on the Court's power against a principle gap that was still about 25 the last time anyone asked.

So the abstract question of whether law should yield to security has divided Israelis along the same line, by the same amount, for as long as the question has been asked. What changed, and changed enormously, is disagreement about the institution: whether the Supreme Court can be trusted and whether it may strike down a law.

That distinction is the one worth putting to judges. A public can be stably divided about a principle without that division touching the court, and the two can come apart by fifty points in fifteen years.

Limits. The question was not asked after 2015, so the most divided years of the court series have no reading on the principle, and the inversion rests on 2009 as the overlap year. 2009 splits its questionnaire across subsamples, so its gap rests on the 388 people who answered both the item and the placement. Every left-right variable is a different variable in each survey and each was checked one by one; all of them run 1 right to 7 left.

Outputs. data/derived/israel_rule_of_law.csv, ines_inventory.csv, built by src/israel_rule_of_law.py and src/ines_inventory.py.

17 September 2026 — The Eurobarometer reader audit: two of the three failures were not failures

The open items listed three reader defects: EB57 failing entirely, thirty-one workbooks failing in EB92, and part of EB93. Reading the harvest log against the source volumes shows that description was wrong in two of the three cases, and that the one real defect was somewhere else.

EB57 is a limit of the source, not of the reader

Its Volume C is nine sheets named by question range, plus demo and proto. The justice item appears on the question sheets, whose banner columns are sex, age, education and occupation. The left-right scale appears only on the demo sheet, and it appears there as a ROW variable crossed by those same demographics. EB57 never published the trust battery by left-right, so there is nothing to read. The 152 logged failures are the wave, not the code.

EB92 was harvested and its failures are a module volume

The wave is in the table: 844 rows, 36 countries, six institutions, read from the two country volumes. 240 of its 317 logged failures are attempts on eb_92_volume_C_AI.zip, which holds 52 sheets covering the QF digital module and the demographics and carries no trust battery at all. The reader was trying a volume that could not contain the item.

EB93 was harvested too

868 rows, 33 countries, eight institutions. Its 47 failures are the T182 stacked sheets and three workbooks rejected by the add-up guard.

The 43 missing court readings are mostly questionnaires that drop the item

Where another institution was read in a wave but justice was not, the cases are dominated by the Turkish Cypriot Community and the candidate countries. Albania's EB90 workbook has no justice item in any of its 115 sheets: its battery runs the written press, the internet, social networks, political parties, regional authorities and the government, and stops.

The one real defect, and it was in the guard itself

Three EB93 national total rows were published with trust and not-trust summing to 100 and a don't-know of 100 on top, because a total column was read as an answer row. The add-up guard that exists to catch exactly this explicitly skipped the all band, so it never inspected the one band that was broken.

repair_band now inspects every band, including the national total, and drops a don't-know or no-answer that makes a band sum past 115 while the two trust answers already sum to 100. The three rows in the harvested table were repaired in place and marked in source_note rather than re-harvesting for thirty minutes. No published figure changes: the three rows are national totals for the army, political parties and health staff in one wave, and every gap on the site is computed from the three-band rows.

What this changes about the open items

The Eurobarometer reader has one fixed defect and no outstanding ones. What remains is an absence in the sources: EB57 has no left-right crosstab for the battery, and the non-EU questionnaires drop items that the EU ones carry. Neither is repairable by code, and both are now documented instead of sitting on a list as if they were bugs.

Outputs. src/eurobarometer_volc.py gains repair_band; data/eurobarometer/derived/eb_battery_by_leftright.csv has three repaired rows.

17 September 2026 — Three merges: the index, the review, and the regions the decomposition could not reach

1. A divided public does not mean courts that work worse

The surveys say how far apart a country's right and left sit about its courts. The Rule of Law Index says what practitioners and households report about how those courts actually work. They come from different people and different instruments and had never been put in one model. 321 country-years, 36 countries, 2013 to 2025.

Within a country, with country and year effects absorbed and both sides standardised:

The judicial factors average −0.028 and everything else averages −0.038. One of nine judicial factors reaches a t of 2; three of six non-judicial ones do. Between countries the pattern is the same: judicial factors correlate −0.36 with the gap on average and the rest −0.40, with Fundamental Rights the strongest at −0.56.

So a country whose public is divided about its courts scores lower on the whole index, and not especially on the parts about courts. The division is not a symptom of courts performing badly. The one judicial factor that does reach significance within a country is whether government powers are effectively limited by the judiciary, which is the executive relationship again rather than court performance.

2. The polarized refusers are mostly outside the survey world

The review finding is that governments in the top fifth of polarization stop accepting recommendations about their own courts. The sorting decomposition splits countries into those losing faith and those dividing. Merging them keeps 35,189 of 111,656 recommendations and 57 of 172 states.

In one model with the state and the cycle absorbed, the interaction between a court recommendation and a sorting public is −0.043 at a t of −0.7. Nothing here is distinguishable from noise.

The reason is worth more than the null. The merge keeps 15 of the 51 most polarized states and 306 of their 825 court recommendations. The states that drove the original finding are largely outside Europe and Latin America, which is where the surveys are. This test cannot be run properly until the survey side reaches the countries the review already covers.

3. The regions the first decomposition could not reach

The sorting decomposition measures division as the gap between a left and a right, which needs a left-right scale, which Africa and the Arab world do not have in any source here. The SPREAD does not need one: it is how far a public is from agreeing with itself, from van der Eijk's agreement measure, and it exists everywhere. 134 country series, 84 countries, six sources, 1973 to 2024.

Mean change in the level: Latin America −0.81, Middle East and North Africa −0.46, Europe −0.00, the Anglosphere +0.19. Mean change in the spread: the Anglosphere +0.024, Sub-Saharan Africa +0.023, the Middle East +0.016, Europe −0.003, Latin America −0.068.

Three regional patterns, and they are different problems. Europe holds its level and a quarter of its publics spread out. Latin America lost confidence and became MORE agreed, not less. The Arab world did both at once, and it is the only region where "both" is the largest category.

That matches the gap decomposition, which found not one Latin American public sorting without also losing faith, and it extends it: what Latin America has is a legitimacy problem, what Europe has is a division problem, and the Middle East has both.

The countries that spread out most are Pakistan, Israel, Jordan, Bulgaria, Iraq, Hungary, Slovakia, Sudan and Taiwan. Israel goes from a level of 6.59 to 4.54 and a spread of 0.66 to 0.88 between 2002 and 2024, which is the same arc the Israeli case shows from its own election studies.

Africa is still barely here. Afrobarometer appears in the pooled table for 2022 only, so the two African series come from the values surveys. Saying anything about the continent needs the Afrobarometer rounds harvested wave by wave, which has not been done.

Outputs. data/derived/roli_vs_gap.csv, upr_vs_sorting.csv, sorting_global.csv, built by src/roli_vs_gap.py, src/upr_vs_sorting.py and src/sorting_global.py.

17 September 2026 — What governments are asked to do about their courts, and which asks they refuse

Every use of the review data here has treated a court recommendation as one thing. It is not. "Guarantee full independence of the judiciary" and "provide more resources to improve response time to cases" are both court recommendations, and they ask for opposite things: one asks a government to let go of its courts, the other asks it to fund them.

That is the distinction the Rule of Law Index makes between a court's insulation from the executive and its ordinary functioning, where polarization moved the first and not the second. Here the outcome is not an expert assessment but a government's own recorded decision at the United Nations. 3,502 recommendations about a state's own courts, 172 states, 2008 to 2024, classified by keyword against the text with the rules and a sample of each family published.

Independence is the hardest thing to be asked for

And in the most polarized fifth, capacity does not move at all

Put in one model with the state and the cycle absorbed, independence recommendations are accepted 10.1 points less than capacity ones everywhere, at a t of −3.6. The extra refusal in the most polarized fifth is −7.9 points at a t of −1.5, so the interaction is a direction rather than a demonstrated difference, and the flat capacity line is what carries the reading: the most divided states go on accepting advice to train and fund their judges at exactly the rate everyone else does, while refusing advice to leave them alone.

Two sources with nothing in common but the subject now say the same thing. The index says what deteriorates under polarization is a court's insulation from the executive and not its speed, cost or enforcement. The review says what governments stop agreeing to is insulation, and not training, funding or delay.

Who hands out this advice, which extends the corridors argument to courts

Each recommendation carries the state that made it. Attaching what that state is itself:

States that recommend independence sit 0.52 higher on judicial independence and 0.11 higher on democracy than states recommending capacity, and are less polarized themselves.

And independence advice travels downhill more steeply than any other kind.

Nearly nine in ten recommendations about judicial independence go from a state with more independent courts to one with less, against seven in ten for advice about training and funding. Independence is the most hierarchical thing said about courts in the review: it is handed down, where capacity advice is passed around.

That is the corridors argument in Sendroiu and Levi applied to courts. Partnership patterns there are about who reaches whom; here the pattern is about what can be said by whom, and the subject with the steepest gradient is the one this project has been tracking all day.

What bounds it

Keyword classification, not a model and not a person; the rules and a random sample of each family are in the script and on the site so a reader can judge them rather than trust them. The residual "other" is 43 per cent of court recommendations and falls furthest under polarization, which means some of what is being refused at the top is not classified. Nothing here identifies a cause.

Outputs. data/derived/upr_what_is_asked.csv, upr_family_threshold.csv, built by src/upr_what_is_asked.py.

17 September 2026 — Africa, at last: twenty-one years and forty-seven countries

Africa was the hole in this project. The global decomposition put Sub-Saharan Africa at two country series because Afrobarometer sat in the pooled table for one round, which meant the site described Europe and Latin America and quietly implied that was the world.

The merged round files for Rounds 2 to 9 are public and were fetched: 330,000 respondents, 2002 to 2023, up to 39 countries a round. "Trust courts of law" is asked in identical wording in every round under eight different variable names, and nine other institutions are asked alongside it. Every institution is matched by its label rather than its number, because a fixed map would silently compare the police in one round with the army in another.

The spread is van der Eijk's agreement measure computed by the same R package on the same definition as every other source here, so the African numbers sit on the European scale.

Trust in courts across the continent has fallen

Among the 29 countries with at least three rounds spanning eight years, the level fell in 18 and the spread rose in 18, with a mean change of −0.31 points and +0.068.

But courts fell less than the elected and executive institutions

The army is the only institution on the continent that gained trust over twenty-one years, and it is also the one African publics agree about most. That is the Arab Barometer finding again from a different survey and a different region: where confidence in the political institutions falls, the coercive one holds or rises.

The four boxes

Courts: 14 per cent spreading, 17 per cent losing faith, 21 per cent both, 48 per cent neither. Every other institution: 13, 15, 15 and 57. Africa has more "both" than Europe, where it is 5 per cent, and is closer to the Middle East, where it is 36.

Lesotho and Senegal fall hardest and land in "both": Lesotho from 6.51 to 4.80 with the spread going 0.598 to 1.045, Senegal from 7.23 to 5.72 with 0.537 to 0.827. Tanzania and Zambia rise, from 6.74 to 7.98 and from 5.16 to 6.44.

What Africa cannot answer

Afrobarometer has no left-right self-placement, so the gap between a political right and left cannot be computed and the sorting result proper is not tested here. It does carry a party identification item, which with a ruling-party crosswalk for 47 countries across eight rounds would give a government-opposition gap on the model of the European one. That crosswalk does not exist and is the obvious next step.

A repair. Rounds 8 and 9 renamed the within-country weight from withinwt to withinwt_ea, and the first version of the harvest found neither and ran them unweighted, which would have put an unweighted 2023 against a weighted 2005 inside the same trend.

Outputs. data/derived/afro_freq_tables.csv, afro_agreement.csv, afro_panel.csv, africa_courts.csv, built by src/build_afrobarometer_panel.py, src/agreement.R and src/africa_courts.py.

17 September 2026 — The values surveys join the decomposition, and the court-specific edge narrows

The European Values Study and the World Values Survey were built on 16 September and never brought into the sorting decomposition, which is why it covered Europe and Latin America. They carry the same left-right split on the same 0 to 10 confidence scale, for 107 countries back to 1981.

Adding them takes the decomposition from 765 series in 84 countries to 1,122 series in 120 countries, and adds 61 court series in 61 countries, among them Nigeria, South Africa, India, Japan, Korea, Brazil, Mexico, the United States, Morocco, Thailand, Taiwan and the Philippines. This is the first time the gap version of the analysis has left Europe and Latin America.

What changes, and it is not nothing

The headline holds: courts lose faith less often than other institutions and sort about as often.

But the court-specific claim about polarization weakens. Change in polarization against change in the gap was +0.37 for courts and +0.12 for everything else on the European and Latin American sample. On the global one it is +0.32 for courts and +0.22 for everything else, on 95 court series against 480. Courts are still the institution where polarization shows up most as division, and the margin over everything else is now about half what it looked like.

That is the honest cost of leaving Europe. The claim that survives is that polarization goes with publics dividing about institutions in general, courts somewhat more than most, rather than with courts being singled out.

Where the new countries land

Of the 61 court series the values surveys add, 35 are in "neither", 16 are losing faith, 6 are sorting and 4 are both. The world outside Europe and Latin America is doing less of everything, on three waves a country rather than eleven rounds or thirty-five, which is the other thing to say about it: the values surveys run about every nine years, so a series of three is a twenty-year span with three points in it.

A note on the source. The European Values Study file is under GESIS terms that prohibit AI processing of their data. Ron's reading, on 17 September, is that running a regression locally is not what that prohibits. Every script here runs on this machine and no row of the file is sent anywhere.

Outputs. data/derived/level_vs_sorting.csv and level_vs_sorting_camps.csv rebuilt by src/level_vs_sorting.py, which gained a values_surveys() reader.

17 September 2026 — Is confidence in courts falling worldwide? Not everywhere, and duration is what bites

Ron asked the simplest question in the project and the one nobody had asked: pooling every source, is trust in courts going up or down, and does polarization explain it. Nine sources, 260 country series, 139 countries, 1973 to 2026. The sources use a 0 to 10 scale, a four-point scale and a percentage, so each country's trend is expressed as per cent of its own source's scale per decade.

There is no worldwide decline. There is a split.

One series per country, averaging the sources that cover it: 71 of 139 countries falling, 51 per cent, with the middle half running −4.4 to +4.6. The European sources and Gallup rise; the Arab world, Africa, Latin America and the values surveys fall.

Of the four ways polarization could matter, duration is the strongest

And at the threshold: countries in the most polarized fifth lose 1.0 per cent of scale per decade while everyone else gains 2.9, a difference of about four points of scale per decade.

Duration beats intensity, which is the answer to the question Ron posed on 16 September about whether courts should hope to nip polarization in the bud. How long a country has been divided predicts the loss better than how divided it is now.

The second row is the awkward one. Countries whose polarization ROSE show trust rising, against the other three measures. The post-communist states are the likely reason, rising on both at once as they left the 1990s. It is reported rather than dropped.

The figures

Four, in light and dark, as PNG and SVG in site/figures/: the trend by source with every country series shown, duration against the trend, the top fifth against the rest, and the four tests with their intervals. The palette is two colours checked through the colour-vision validator rather than chosen by eye; a three-colour version in the lab's gold failed the normal-vision separation floor against the red.

What a slope cannot say. It is a straight line through three to thirty-five points, so a country that fell and recovered looks flat. Sources differ in question, scale and coverage, which is why the per-source rows matter more than the pooled one. Nothing identifies a cause.

Outputs. data/derived/global_trend.csv, global_trend_tests.csv, site/figures/*, built by src/global_trend.py and src/figures.py.

Still to do

• The Eurobarometer reader is done. EB57 carries no left-right crosstab for the trust battery and the non-EU questionnaires drop items the EU ones carry; both are limits of the sources rather than open defects.
• Tone in the Canadian Hansard. The published Hansard result is volume and vocabulary because nothing codes whether a mention of courts is supportive or hostile. src/hansard_corpus.py writes the passages needed to code it; the classification itself is not done and needs Ron's go-ahead on the API spend.
• Artificial intelligence as a second possible threat to courts, which Ron raised on 16 September 2026 without a specific measure in mind. His existing ai-in-canadian-courts project is the nearest available material.

Published to the site, 16 September 2026

Every finding in this log now has a section on the results page. The sweep added five that had been recorded here and never published: the 29-year Eurobarometer series and its durability rerun (#eurobarometer), the Latinobarómetro and Arab Barometer placebos (#placebo), the threshold test on survey data (#threshold), and the European regions and Canadian provinces (#within). The per-country results and the police-spread reversal sit in #spread and #context.