Institutional quality is several separable capacities, and the score that ranks them is mostly a choice of weights

Institutional quality is the phrase that arrives when a committee, a ministry or a lender has run out of things it can count. It is used as though it named one property, graded from good to bad and held by a jurisdiction the way a rating is held by an issuer. Operationally it is not one property. It is a small number of separable capacities that fail independently of one another, and a jurisdiction can be strong in any of them while being weak in the rest. A single score is what happens when those capacities are forced onto one axis. The operation that forces them there is arithmetic that anyone can perform, and almost nobody shows it.

Where this is written from

I write as an analyst inside The Cefa Brief, working only from published statistics and public documents: the annual federal series cited below, retrieved for this site on 19 September 2026, the state files held in this repository, and this site's own State Development Index together with the weight audit generated alongside it. That position lets me read a method. It shows the unit a series is stated in, the vintage of an estimate, the weighting a composite declares, and the band a rank moves across when that weighting changes. It cannot see a permitting office, a procurement panel, a court or a treasury desk, and no series in front of me records whether a named task was done at the stated cost in the stated time. So I cannot tell a reader which jurisdiction is well governed. I can show what the number claiming to tell them is made of.

Two things follow, and both narrow the piece. Nothing here is a finding about any jurisdiction's administration, because no series in front of me observes one. And where the argument turns on how institutions behave rather than on how a number is built, it is an argued position, and the text says so. What the position is good for is construction. An index is a published object, its inputs carry their own documentation, and the arithmetic joining them can be re-run by anyone who dislikes the result.

Four things, and only one of them is a measurement

Move the question off the adjective and onto the task. A jurisdiction is asked to do something specific: authorise a connection, issue a permit, award a contract, settle an invoice, hear an appeal. Four conditions make that task measurable. The task has to be named, so that everyone agrees what was asked. A cost has to be stated in advance. A time has to be stated in advance. And the outcome has to be published afterwards, saying whether the thing was done, at what cost and within what period. The first three are specifications and cost nothing to publish. Only the fourth is a measurement, and the fourth is the one almost nobody publishes. Without it an observer is left holding activity.

Construction spending on manufacturing plant, series TLMFGCONS, is published by the Census Bureau in billions of dollars at a seasonally adjusted annual rate, which means the source already states each period at a yearly rate, so each annual point is a yearly rate that approximates the year. It reads 75 in 2020, 82 in 2021, 125 in 2022, 202 in 2023, 244 in 2024 and 215 in 2025. That is a record of money going into structures. It records no completion, no schedule and no cost set against an estimate. Housing units authorised by permit, series PERMIT, is the same shape in a different unit, thousands of units at a seasonally adjusted annual rate: 1,479 in 2020, 1,734 in 2021, 1,685 in 2022, 1,516 in 2023, 1,474 in 2024 and 1,431 in 2025. An authorisation is a permission, not a building. Both series are honest about what they count. Neither closes the loop the definition needs.

What a composite removes first

Because the fourth condition is missing from every source here, the common move is to gather several of the third-best measures and add them together. Adding requires the units to go, and the units are not decoration. Research and development inside GDP, series Y694RC1Q027SBEA from the Bureau of Economic Analysis, is billions of dollars at a seasonally adjusted annual rate, so each annual point is a yearly rate: 754 in 2020, rising to 1,081 in 2025. New business applications, series BUSAPPWNSAUS from the Census Bureau, is thousands of applications as a weekly average, which means the source publishes a value for each week and the year's point is the average of those weeks rather than the year's total: 84 in 2020 and 109 in 2025. U.S. semiconductor production, series IPG3344S from the Federal Reserve Board, is an index on a 2017 base of 100 with five annual points, 121 in 2021 and 165 in 2025, and it contains no dollars and no count of anything. Three figures, three kinds of unit. The unit is the part of each figure that says what it is.

Here is the operation performed on three of those series, small enough that every step can be checked. Take research and development, permits and factory construction across 2020 to 2025 and scale each onto a range of nought to a hundred against its own six-year high and low: subtract the series minimum, divide by the series range, multiply by a hundred. Research and development runs from 754 to 1,081, a range of 327. Permits run from 1,431 to 1,734, a range of 303. Factory construction runs from 75 to 244, a range of 169. For 2022 that gives research and development at (942 minus 754) divided by 327, times 100, which is 57.5; permits at (1,685 minus 1,431) divided by 303, times 100, which is 83.8; and factory construction at (125 minus 75) divided by 169, times 100, which is 29.6. For 2025 the same three steps give 100, nought and 82.8. Averaged with equal weights, 2022 scores (57.5 plus 83.8 plus 29.6) divided by 3, which is 57.0, and 2025 scores (100 plus nought plus 82.8) divided by 3, which is 60.9. That last figure is not billions of dollars, not thousands of units and not an index on any base. It is not a number of anything. It also hides the fact that one of its three inputs sits at its own six-year low in that very year: permits read 1,431 in 2025, the lowest of the six observations, which is exactly why they score nought.

Change the weights, change the order

Equal weights are a choice rather than a neutral position, and the choice decides the ranking. Give permits half the weight and the other two a quarter each, which is a defensible view for anyone who thinks the ability to get a building approved is the capacity at issue. Then 2022 scores 41.9 plus 14.4 plus 7.4, which is 63.7, and 2025 scores nought plus 25 plus 20.7, which is 45.7. Under equal weights 2025 stands 3.9 above 2022. Under the second weighting 2022 stands 18.0 above 2025. No observation changed between those two sentences. Nothing was revised, rebased or corrected. The order of the two years is a property of the weights, and the weights are the part of the method that no data supports.

The same test, on a published index

This site publishes a composite of its own, which makes it a fair thing to test, because it also publishes its construction. The State Development Index scores all fifty states and the District of Columbia across eight pillars, named on the page as starting up, population health, productive base, labour market, prosperity, things get built, fiscal standing and openness. It uses percentile ranks rather than min-max scaling, so that one very large state cannot set the scale for everyone. It expresses anything that scales with population per head, so that size does not masquerade as quality. It applies equal weights across pillars as a declared choice, and it scores a state on the pillars it has rather than imputing an indicator it lacks. Each of those decisions is written into the file that computes the score, with its reason attached.

Alongside the score the site publishes an audit of the one decision the data cannot settle. Twenty thousand weight vectors are drawn uniformly over the eight pillars, the composite is recomputed for each draw, and all fifty-one jurisdictions are re-ranked; the draw is seeded and deterministic, so it can be reproduced. The file then reports, for each jurisdiction, the band of ranks it occupies under weightings where every pillar sits between half and twice its equal share, which is the set of weightings a person could argue for in public. Colorado ranks first under the published equal weighting with a score of 71, and across that set it ranks between first and third. New Mexico ranks fifty-first with a score of 33, and across the same set it ranks fifty-first and nowhere else. Massachusetts ranks nineteenth with a score of 56, and across the same set it ranks anywhere from ninth to thirty-ninth.

Where the rank is a finding, and what date it carries

Across the fifty-one jurisdictions the narrowest of those bands is nought places wide and the widest is thirty, with a median of eighteen, and forty-two of the fifty-one move ten places or more on the weighting alone. The ends of the table are robust and the middle is not. That is the general shape of a composite, and it is the opposite of what a reader needs. Where a jurisdiction is plainly strong or plainly weak, the score repeats what the underlying figures had already said without it. Where help is actually wanted, in the middle of the table, where most jurisdictions sit and most decisions are taken, the rank is largely a record of somebody's weighting. Nineteenth is not a fact about Massachusetts.

A second property of composites is arithmetic rather than judgement, and it is rarely stated on the page. A score is as current as its oldest input and never as its newest. The files behind this index carry their own dates, each proper where it sits: per-state metrics fetched 7 September 2026, with monthly series running to July 2026, the house price index to April 2026 and the annual output and per-capita income series to 2025; state tax collections from the Census Bureau's fiscal 2025 survey; population from the Census Vintage 2024 estimate; adult obesity prevalence from CDC chronic disease indicators for 2024; and, behind the fiscal standing pillar, a table of credit ratings the repository itself marks as indicative, as of 2025, with an instruction to verify it against current agency reports before use. Added together they yield one number carrying no date at all, and it will be read as current.

What to publish instead

The alternative is not a better index. It is a list, in which each item carries the four things the score strips out: the quantity, the unit it is stated in, the agency that publishes it and the date it was retrieved, and against each item the gap, meaning what that series does not cover. This site's index already does that in one case, and the case is the instructive one. It records that it cannot measure regulatory burden, because the works that do measure it are copyrighted and their scores, weights and component structures cannot be reproduced, and that the honest answer is the gap, stated on the page. That is the general condition rather than the exception. The capacity a reader most wants graded is usually the one with no citable series, and a composite's chief service is to conceal that by producing a number regardless. So the claim here stays narrow. Institutional quality resolves into separable capacities; the record holds inputs to those capacities rather than outcomes of them; a composite makes the inputs addable by deleting what each one measured; and the order it produces is set in large part by weights nothing in the data supports. From where this is written I cannot say which jurisdiction is well governed. A score whose construction is not published should be read as an opinion with an arithmetic finish.