The Nuance Collective

The Nuance Index™

Edition 001 · US Regional Banks · 31 July 2026
Claude · 5 archetypes × K=5 × 2 nouns = 50 runs · scored on the 25 primary-noun runs

001

Regional
Banks

The Number — Category Clarity

62.3

Unsettled

Accurate, but the answer changes from run to run — nothing here has hardened yet.

30.1

Agreement

does it answer the same way twice

27.0–39.9 at 90%

84.3

Accuracy

is what it says true

82.4–85.3 at 90%

70.0

Alignment

does the ranking match the real category

67.6–72.8 at 90%

True →

Unsettlednot the same twice · true

Clearthe same twice · true

Confusednot the same twice · false

Miscastthe same twice · false

Same answer twice →

Settled at Agreement 50 · true at Truth 50 · this category sits at Agreement 30.1, Truth 77.2 · both boundaries are the midpoint of their own metric

PASS Indexable 46 namedPASS Noun valid 83.9— Reference set 31 brands, primary-sourced— Baseline no prior run, no arrow

Agreement at 30.1 is what drags this category down. The answers are broadly accurate about the banks they name; they just do not name the same ones twice. 46 distinct banks surfaced across the 25 primary-noun runs, and the top five reshuffles substantially from one run to the next. A category this unsettled is one where a brand can still move its position.

The Read

The Gap is a 31-line reconciliation that does not balance. 29 of the 31 entries sit somewhere other than where the balance sheet puts them, the largest bank in the set is 19 places out, and the average entry is off by 30 percentile points. This category has not settled on an answer yet, and awareness spend is not what settles it.

The banks sitting above their size line have one specific fact attached to them. Frost has a service reputation and sits 6th; its balance sheet ranks 29th. A balance sheet is not a reason to recommend anyone.

TD is the case worth sitting with. It is known in more detail than almost anything else here, it is accepted as part of this category when you ask, and it is named the moment the question is about a place instead of a choice. What it is missing is a reason to be picked. That is a different job from being known, and no awareness number will ever show you the difference.

Who is in this category, and on whose authority

Chartered whereA US-chartered bank holding company or US intermediate holding company filing FR Y-9C, appearing on the FFIEC National Information Center's Large Holding Companies release.

What asset bandTotal consolidated assets of the US entity between $50B and $750B as of 31 March 2026.

What it must operateA physical retail branch network serving US consumers.

What is excluded, namedThe eight US-headquartered G-SIBs — JPMorgan Chase, Bank of America, Citigroup, Wells Fargo, Goldman Sachs, Morgan Stanley, BNY, State Street — which removes the money-center, custody and investment banks in one published stroke. Branchless and digital-only banks. And two we exclude by our own judgement and name here so the decision is visible rather than buried: Capital One and American Express, whose deposits are gathered nationally rather than through a branch network.

Why the G-SIB test is not asymmetricThe test is applied to the entity that files the FR Y-9C, not to the corporate family. For a US-headquartered G-SIB the filer is the designated bank, so it is out. For a foreign G-SIB the filer is a separately chartered, separately capitalised US intermediate holding company that is not itself designated and is tailored by the Federal Reserve as a standalone. Same test, one entity each.

What is marked, and whyTwo marks, one meaning each. † a US arm of a foreign parent (TD, BMO, Santander), whose asset figure is always the US entity's and never the parent group's. ‡ a domestically owned company commonly filed one tier up (U.S. Bancorp, Minneapolis). A mark never changes a score.

Classification and assets: FFIEC National Information Center — Large Holding Companies (FR Y-9C), reporting date 31 March 2026 · Financial Stability Board — 2025 list of global systemically important banks. The band, the branch requirement and the treatment of foreign-owned US arms are ours; the entity list and the asset figures are not.

Every entity in the reference set qualified under this rule before the battery ran, so a low position is a measured result rather than an argument about labels. Where an entity's category label is arguable it carries a mark, and the finding is recomputed below with each dispute granted separately — ownership alone, tier alone, and both — so no brand has to accept another's objection as part of the answer to its own.

The Standings

Brands are grouped into bands, not places. The cut rule: a band opens with the highest-scoring brand not yet placed and admits every brand whose 90% interval overlaps the opener's. On the point estimate Huntington leads, but across 10,000 resamples of the 25 primary-noun runs it holds first place only 35% of the time.

What the bands do not claim. Overlap is not transitive, so two brands in adjacent bands can still overlap — 20 such pairs here. A band says its members are indistinguishable from that band's leader, not that every cross-band pair is distinguishable. Under the strictest reading, merging every overlapping pair, this data separates 2 groups: Valley National, and the other 11. That is the ceiling. The bands below are the useful reading inside it.

These positions are how we read the category's clarity, and they are not ratings of the banks.

	Brand	Presence	Rank	Conviction	Fidelity	Composite
Band 1 · 62–80 · 5 brands, not separated by the data
1	HuHuntington	100.0	59.1	82.0	50.0	76.2 73–79
1	CiCitizens	76.0	74.9	51.3	100.0	75.6 70–80
1	MtM&T	80.0	61.1	60.0	100.0	75.3 70–80
1	RgRegions	88.0	48.0	50.0	100.0	72.8 68–77
1	FrFrost	68.0	27.5	92.6	100.0	69.2 62–76
Band 2 · 56–72 · 4 brands, not separated by the data
2	FtFifth Third	88.0	63.6	58.0	50.0	68.3 64–72
2	UbU.S. Bank‡	24.0	90.7	79.2	100.0	66.9 60–72
2	TrTruist	60.0	56.3	43.3	100.0	63.7 57–70
2	PnPNC	36.0	74.1	55.6	100.0	62.2 56–69
Band 3 · 51–64 · 2 brands, not separated by the data
3	KyKeyBank	40.0	54.4	50.0	100.0	57.6 51–64
3	ZiZions	68.0	14.4	50.0	100.0	57.4 52–63
Band 4 · 36–50 · 1 brand, separated from every other band
4	VnValley National	36.0	22.2	50.0	75.0	43.2 36–50

Composite, with its 90% interval 10,000 resamples of the 25 runs

Band 15 brands, not separated by the data

Huntington 76.2

Citizens 75.6

M&T 75.3

Regions 72.8

Frost 69.2

Band 24 brands, not separated by the data

Fifth Third 68.3

U.S. Bank‡ 66.9

Truist 63.7

PNC 62.2

Band 32 brands, not separated by the data

KeyBank 57.6

Zions 57.4

Band 4one brand, separated from every other band

Valley National 43.2

35.047.060.072.085.0

Point estimate 90% interval — where this brand sits if the 25 primary-noun runs are redrawn Bands are cut where the intervals stop overlapping

Table view — composite and interval
Huntington 76.2 (73–79) · Citizens 75.6 (70–80) · M&T 75.3 (70–80) · Regions 72.8 (68–77) · Frost 69.2 (62–76) · Fifth Third 68.3 (64–72) · U.S. Bank 66.9 (60–72) · Truist 63.7 (57–70) · PNC 62.2 (56–69) · KeyBank 57.6 (51–64) · Zions 57.4 (52–63) · Valley National 43.2 (36–50)

46 distinct brands across the 25 primary-noun runs; standings, Agreement and every interval are computed on those 25. The adjacent noun's 25 carry the Doorway and the absence check. † US arm of a foreign parent · ‡ US-owned, filed a super-regional.

Huntington was named in all 25 primary-noun runs — the only bank to appear every time — and does so on one repeatable fact: its overdraft policy. That is why it leads the first band, and the band is as far as this data goes: four banks share it. U.S. Bank‡ shows the opposite pattern. The largest bank in the reference set at $726B was named in only 6 of those 25, and went first in almost all of them (Rank 90.7). It ranks well when it appears, and it rarely appears.

The Signal

The Gap · this week's featured finding

How far this category's answers sit from the market they describe.

points. That is how far the average bank in this category sits from where its real-world size says it should sit — measured in percentile points across the 31-brand reference set. A category whose answers tracked reality would sit near zero. The largest bank of the 31 sits 20th; one of the smallest sits 6th.

Lower in the answersPercentile position within the categoryHigher in the answers

TD Bank -83 4th → 29th TD Bank · real-world size says 4th of 31 · it sits 29th · gap -83 points

BMO -67 6th → 26th BMO · real-world size says 6th of 31 · it sits 26th · gap -67 points

U.S. Bank -62 1st → 20th U.S. Bank · real-world size says 1st of 31 · it sits 20th · gap -62 points

Santander 15th → 29th Santander · real-world size says 15th of 31 · it sits 29th · gap -47 points

PNC -42 2nd → 14th PNC · real-world size says 2nd of 31 · it sits 14th · gap -42 points

Flagstar 17th → 29th Flagstar · real-world size says 17th of 31 · it sits 29th · gap -40 points

Western Alliance 14th → 25th Western Alliance · real-world size says 14th of 31 · it sits 25th · gap -37 points

First Citizens 9th → 18th First Citizens · real-world size says 9th of 31 · it sits 18th · gap -28 points

Popular 21st → 29th Popular · real-world size says 21st of 31 · it sits 29th · gap -27 points

Truist -17 3rd → 8th Truist · real-world size says 3rd of 31 · it sits 8th · gap -17 points

East West 19th → 22nd East West · real-world size says 19th of 31 · it sits 22nd · gap -10 points

Pinnacle 13th → 14th Pinnacle · real-world size says 13th of 31 · it sits 14th · gap -5 points

KeyBank 11th → 12th KeyBank · real-world size says 11th of 31 · it sits 12th · gap -2 points

Wintrust 22nd → 22nd Wintrust · real-world size says 22nd of 31 · it sits 22nd · gap +0 points

UMB 24th → 24th UMB · real-world size says 24th of 31 · it sits 24th · gap +0 points

F.N.B. 31st → 29th F.N.B. · real-world size says 31st of 31 · it sits 29th · gap +7 points

Fifth Third 5th → 2nd Fifth Third · real-world size says 5th of 31 · it sits 2nd · gap +8 points

Citizens 8th → 5th Citizens · real-world size says 8th of 31 · it sits 5th · gap +10 points

Webster 18th → 14th Webster · real-world size says 18th of 31 · it sits 14th · gap +12 points

Huntington 7th → 1st Huntington · real-world size says 7th of 31 · it sits 1st · gap +20 points

M&T 10th → 4th M&T · real-world size says 10th of 31 · it sits 4th · gap +20 points

SouthState 25th → 18th SouthState · real-world size says 25th of 31 · it sits 18th · gap +25 points

Associated 30th → 22nd Associated · real-world size says 30th of 31 · it sits 22nd · gap +27 points

BOK Financial 28th → 20th BOK Financial · real-world size says 28th of 31 · it sits 20th · gap +28 points

Regions 12th → 2nd Regions · real-world size says 12th of 31 · it sits 2nd · gap +32 points

Zions 16th → 6th Zions · real-world size says 16th of 31 · it sits 6th · gap +32 points

First Horizon 20th → 9th First Horizon · real-world size says 20th of 31 · it sits 9th · gap +37 points

Old National 23rd → 12th Old National · real-world size says 23rd of 31 · it sits 12th · gap +38 points

Valley National 26th → 14th Valley National · real-world size says 26th of 31 · it sits 14th · gap +38 points

Columbia +57 27th → 10th Columbia · real-world size says 27th of 31 · it sits 10th · gap +57 points

Frost +75 29th → 6th Frost · real-world size says 29th of 31 · it sits 6th · gap +75 points

050100

Where its real-world size says it should sit Below that in the answers Above it

Table view — the distribution behind the average, in percentile points
TD Bank -83 · BMO -67 · U.S. Bank -62 · Santander -47 · PNC -42 · Flagstar -40 · Western Alliance -37 · First Citizens -28 · Popular -27 · Truist -17 · East West -10 · Pinnacle -5 · KeyBank -2 · Wintrust +0 · UMB +0 · F.N.B. +7 · Fifth Third +8 · Citizens +10 · Webster +12 · Huntington +20 · M&T +20 · SouthState +25 · Associated +27 · BOK Financial +28 · Regions +32 · Zions +32 · First Horizon +37 · Old National +38 · Valley National +38 · Columbia +57 · Frost +75

The exhibit · The Misfile: TD Bank†, 0 of 50

TD carries the widest gap we measured — 4th of 31 by size, 29th in this category's answers, -83 points. $381B in US assets, 1,048 stores, not one mention in any of the fifty. The obvious explanation is that the model declines to call TD regional, so we tested that directly before publishing anything. It does not. Asked outright, all 3 taxonomy probes place TD inside the category's peer set — “if a peer set is 'US regional banks,' TD Bank is commonly included” — while qualifying it as a super-regional. The category label is not what is missing.

"Is TD a regional bank?" · taxonomy

In · 3 of 3

Every run qualifies it as a super-regional and every run still puts it in the peer set. Excluded by the model's own taxonomy in 0 of 3.

"Tell me about TD Bank" · awareness

15 facts · 3 of 3

15 specific facts returned in all three independent runs — the 1855 founding, Cherry Hill, Commerce Bancorp, the retired coin machines, the 2024 plea, the asset cap. It knows TD in more detail than almost any brand here.

"Best regional bank" · category frame

0 of 25

Absent from all 25 primary-noun runs, and from all 25 on the small-business noun. 0 of 50.

"Best bank in the Northeast" · geographic frame

3 of 3

Named in every run, first in one of them, and named alongside Citizens and M&T — the two banks topping the standings above.

So the model knows TD, accepts TD as a member of this category, and names TD readily when the question is geographic. It never produces TD when the question is which one to choose. That is a Misfile rather than a Ghost, and the two need opposite fixes. Santander, Flagstar, Popular and F.N.B. were also absent from all fifty; Santander is likewise placed inside the category when asked (3 of 3) and recovers in 1 of the same 3 geographic runs. Probe transcripts and counts are recorded with the edition; every probe is three clean runs, no tools, no history.

The stress test · does the Gap survive if we grant the membership disputes?

Two entities in this reference set carry a mark, for two unrelated reasons. † is a US arm of a foreign parent — TD, BMO and Santander file as US holding companies but answer to Toronto, Toronto and Madrid. ‡ is U.S. Bancorp: Minneapolis-headquartered and domestically owned, but commonly filed a super-regional rather than a regional. Those are different objections, so we grant them separately — a brand disputing its ownership should not have to accept another's tier argument as part of the answer.

Published

All 31 entities. Average gap in percentile points, Clarity 62.3.

Ownership granted †

Without TD, BMO and Santander — n 28. Clarity 63.8, verdict Unsettled. Widest becomes U.S. Bank at -68.

Tier granted ‡

Without U.S. Bank — n 30. Clarity 62.7, verdict Unsettled. Widest stays TD Bank at -86.

Both granted

n 27. Clarity 64.5, verdict Unsettled. Widest becomes PNC at -52, then Western Alliance — neither marked.

The finding does not rest on either dispute. Grant the ownership objection and the largest gap in the category is U.S. Bank, a Minneapolis bank nobody argues about. Grant the tier objection and TD is still the widest by a distance. Grant both and the widest gaps are PNC and Western Alliance. The verdict is Unsettled in every version.

The published number keeps all 31 entities, because every one qualified under the stated rule before the battery ran and removing an entity after seeing its score is how an index stops being one.

Method

The Battery

Five fixed query archetypes, run K=5 each in clean independent contexts with no history, no tools and no web access. Only the category noun changes week to week.

Measured 29 July 2026 — the 50-run battery. Vetted 30 July 2026 — the twelve §6.2 probes. Published 31 July 2026. A category's answers can move between the run and the Friday; the dates are stated so the number is read against the day it was taken.

Primary noun regional bank
Adjacent noun bank for a small business
Attribute slot lowest-fee, selected from a pre-flight probe as the criterion this category's own answers most often use to decide.

Doorway Spread 100 — the share of the top five that changes when the noun changes. At 100 the two nouns returned no overlap at all: not one brand in the small-business top five was a regional bank.

Clarity and the Gap

Clarity is built from three things: whether it answers the same way twice, whether what it says is true, and whether its ranking matches the real category. Alignment is 100 minus the average gap, so the headline number and the Signal come from the same calculation.

Real-world weight for this category is total consolidated assets, primary-sourced and declared before the battery ran. Gaps are in percentile points so they mean the same thing in a 31-brand category and a 12-brand one.

Indexability and noun validity are pass/fail gates. We publish the component names and scores, never the weighting or the exact prompts.

The Spread

The 25 primary-noun runs are a sample, so every run-derived figure has a range. We resample those runs 10,000 times with replacement (seed 20260729) and publish the 90% interval beside the number it belongs to.

Standings are cut into bands wherever those intervals stop overlapping. Brands inside one band are not separated by this data and we do not print an order for them.

Fidelity is a fixed judgement against verified reality and carries no sampling variance, so these are run-sampling intervals rather than total uncertainty. Re-running the published script reproduces every interval exactly.

The Model Set

This edition reads one model, Claude. The locked method reads four and that set closes at Edition 002.

Agreement is the component this affects most. One model can only disagree with itself across repeated runs, and disagreement between models is where that number gets most of its meaning — so read 30.1 as a floor on this category's instability rather than a measure of it. Accuracy and Alignment are unaffected: both are judged against verified reality, not against another model.

Absence, and the vetting behind it

No absence is published as a finding until it has been tested three ways: whether the model places the brand in the category, whether it knows the brand at all, and whether the brand reappears under a different framing. Each probe is three clean runs on the same protocol as the battery.

A brand the model does not know is a Ghost. One it knows and files elsewhere is a Misfile. The probe counts are published with the finding, not summarised.

Absent from all fifty runs
TD Bank† · Santander† · Flagstar · Popular · F.N.B. — Absent means absent from this model, this battery, this date. Absence is vetted before it is published, and where the cause is a category-membership effect rather than an awareness effect we say so on the face of the edition. In this category it was neither: both brands are placed inside the category when the model is asked directly.

Scores are not comparable across categories, and no edition cites another's. Every edition is a self-contained verdict on its own category.
Want to know why this category answers the way it does, and what that means for your bank? That's the LLM Mirror.