STRATAHELM · CAERIHomeCountiesArticlesMethodologyContact

StrataHelm Community AI Exposure & Resilience Index (CAERI)

Methodology Specification v1.0.0

Classification of this document. This specification is published in full

except for two deliberately withheld blocks, each fenced below and marked in the

source. What is published: the pillar structure, every data source, the pillar

weights and their rationale, the exposure-ensemble weights, the confidence-rating

formula, the validation design and results, the known limitations, and the

relationship of this index to prior published research (Sections 1-4, 5.1, 6.1,

6.2, 6.3, 7, 8, 9). What is withheld: the county occupation estimation procedure

(Section 5.2) and the indicator construction, normalization and time-phasing

mathematics (Section 6.9). Why: the estimation procedure is the engineering that

makes a county-level index possible at all and is the asset a competitor would

otherwise copy directly; the withheld scoring internals are the indicator-level

formulas, not the weights, and the weights are what a reader needs to judge or

reproduce the index's structure. Everything withheld is described in enough

detail to state what it does, what it is built from, and how it was validated.

A published index whose weights are secret is not citable, so the weights are

published.

1. What the Index Measures (and What It Doesn't)

CAERI is a 0–100 composite score, computed per U.S. county (and rollable to municipalities, workforce development areas, and states), answering:

"How exposed is this community's economy to AI-driven labor disruption, and how well-positioned is it to adapt?"

It is an exposure and resilience index, not a displacement forecast. We explicitly do NOT claim to predict how many jobs will be lost by a given date. We claim, defensibly:

1. Which communities have employment concentrated in occupations that published, peer-reviewed research identifies as highly exposed to current AI capabilities.

2. Which communities have economic structures (concentration, low diversification, weak adaptive infrastructure) that historically amplify labor shocks.

3. How these two interact, relative to all other communities in the state and nation.

Scores are presented as percentile ranks with relative-standing bands and confidence indicators — never as false-precision point predictions. On the national basis the bands are Top 10% / 75th–90th / 50th–75th / 25th–50th / Bottom 25% — cuts on a county's national percentile, i.e. a statement of relative standing among all U.S. counties, not a condition assessment and not a forecast. The in-state basis is reported as rank and percentile with no band, because percentile bands are not meaningful within a single state (some states have only a handful of counties).

2. Index Architecture: Five Pillars

*Baseline weights; see Section 6.2 for the weighting rationale and Section 7 for the sensitivity-analysis requirement and its result. P5 is also sold standalone as the premium Fiscal Resilience module with a deeper model.

3. Data Sources (All Public, All Federal or Peer-Reviewed)

3.1 Employment & Occupation Structure

3.2 AI Exposure Research Base (occupation-level scores)

We use an ensemble: each occupation's exposure score is a weighted blend of the above, with disagreement across sources feeding the confidence rating (Section 6.3); the ensemble weights are published in Section 6.1. We separately track automation share (task substitution) vs. augmentation share (task assistance) — a county heavy in augmentation-dominant occupations scores materially lower risk than one heavy in automation-dominant occupations at the same raw exposure.

3.3 Concentration, Capacity, Buffer, Fiscal

4. Pipeline Overview (Public Description)

[Federal APIs/files] → Ingest → Clean/Conform → County Occupation
Estimation → Exposure Scoring → Pillar Indicators → Normalization →
Weighted Aggregation → Tiering + Confidence → Scorecards/Dashboard

Refresh cadence: quarterly (QCEW/LAUS), annual (OEWS, CBP, ACS), event-driven (new exposure research, WARN notices).

5. County Occupation Estimation

5.1 What is published, and how it was validated

The core engineering problem: OEWS publishes occupation detail only at MSA and

nonmetro-area level, but the product is county-level. Bridging that gap is the

single largest modelling step in the index, and it is the step withheld

(Section 5.2). What is published here is what the step consumes, what it

produces, and — at length — how accurate it turned out to be.

Inputs and universe. County employment by industry from County Business

Patterns with a QCEW cross-check; national industry-to-occupation staffing

patterns from the BLS National Employment Matrix; county occupation-group and

industry marginals from the ACS; detailed occupation and wage marginals from

OEWS at area level. The estimation universe is wage-and-salary employment, all

ownerships, including federal, state and local government. Unincorporated

self-employed workers, unpaid family workers, and farm production employment

are outside it, because no county-level occupation source for them exists.

Output. An estimated employment count and wage bill for each detailed

occupation in each county. These are model-based ESTIMATES, not surveyed

counts, and every surface that publishes them labels them as such and carries

the county's confidence rating.

Validation: the honest account.

The design originally specified in our build plan was a direct accuracy test:

for counties that are coterminous with their own OEWS area, compare the

estimates against the directly observed OEWS values, and report the error by

occupation major group against a target of under 12%.

*That test could not be run on the Minnesota pilot, because no Minnesota county

is coterminous with its OEWS area.* A substitute design was adopted

(2026-07-04): leave-one-anchor-out. Each area's estimates are rebuilt with

that area's OEWS anchor held out of the fit, rolled back up to area level, and

compared in shape against what OEWS actually observed. The error is

employment-weighted, reported at occupation major-group level, and excludes

farm production.

Both designs are now computed nationally, and they do not agree:

Figures are mean absolute percentage error, employment-weighted, across

521 area tests in 51 states (10,916 area-by-occupation-group cells), farm

production excluded for the reason given below. Minnesota, the pilot state, is

8.0% as production estimates and 12.0% strictly held out.

⚠ The per-state figure on each county page still uses the earlier comparison

described in the second correction below (Minnesota 8.5% rather than 8.0%).

Those figures are regenerated as a batch with the next county-page rebuild

rather than in isolation; the difference is a fraction of a percentage point and

in the favourable direction.

Why two columns. The estimator applies a calibration that reconciles how

people describe their own occupation to the Census against how employers code it

on payroll, and that calibration is built from the whole state's observed data.

When a single area dominates its state, holding out that area's anchor while

still calibrating on an aggregate the area largely is makes the test easier

than it should be. The second column removes the held-out area from the

calibration as well.

Neither column is simply right. As production estimates reflects what a real

county estimate genuinely has access to — the area's observed data does inform

the statewide calibration, legitimately. Strictly held out withholds

information a real estimate would have, and is therefore conservative. The truth

is between them, and we publish both rather than choose.

The gap is concentrated, not general. It is negligible where no single area

dominates its state, and severe where one does: Nevada moves from 4.4% to 18.1%,

Hawaii from 6.7% to 16.6%. **Washington DC and Rhode Island have exactly one

OEWS area each**, so for those two the statewide aggregate is the held-out

area and no strictly-held-out figure can be computed at all. Under the strict

column, 17 states exceed the 12% target rather than 5.

**Stated plainly: the test originally specified comes in above the 12% target,

and the figure reported on this site and on county pages uses a different

test.** The leave-one-anchor-out design measures how well the method recovers

an area's occupation shape when that area's anchor is withheld; the coterminous

design measures how close a single county's estimates land to observed values.

The second is the harder and more directly relevant test, and it misses the

target by 1.0 point. We publish both rather than the flattering one.

The 12% target is self-set. It came from our own build plan. There is no

external standard, no regulatory threshold and no peer-reviewed convention

behind that number, and it should not be read as one. It is the bar we wrote

for ourselves before we had results, kept unchanged afterwards.

**Correction, 2026-09-19 (second) — the comparison itself was discarding data

on both sides.** It matched our estimate to OEWS on detailed occupation codes

before rolling up to groups. That dropped 4.2% of our estimated employment

(codes OEWS does not publish at detail in that area — assemblers, laboratory

technicians, home-health aides) and compared what remained against a sum of OEWS

detail rows that is itself 3.1% short of OEWS's own published group totals,

because OEWS suppresses cells it cannot disclose. Both sides are now compared at

group level against OEWS's published group row, which discards nothing. The

effect is small and favourable — 9.3% becomes 9.1%, coterminous 13.3% becomes

13.0% — and it does not change any conclusion.

**Correction, 2026-09-19 (first) — the national figure was previously reported

as 8.7% and was understated.** 43 OEWS areas straddle a state line, and because each

state validates the areas it touches, those areas were entering the national

average twice. They are large metropolitan areas, and large areas are estimated

more accurately than small ones, so double-counting them pulled the national

figure down. Counting each area once raises it from 8.7% to 9.3%. We found

this ourselves, in the course of preparing this methodology for external review,

and we are publishing the worse number rather than the one that had already been

in print. The coterminous figure is unaffected — a single-county area cannot

straddle a state line — and no score, rank or band moved, because this is a

diagnostic of the estimator, not an input to it.

The error is not evenly distributed, and it is not all noise. A mean

absolute error cannot distinguish error that averages out from error with a

sign, so we measured both. In aggregate the estimator is very nearly unbiased —

its mean signed error is +0.1% against a mean absolute error of 9.1%. But

that aggregate nets off large, opposing biases at occupation-group level:

12 of 21 occupation major groups carry a systematic bias of 2 percentage

points or more (judged by a bootstrap confidence interval on the weighted mean

signed error, because per-area errors are neither independent nor normally

distributed).

*These are measured on the corrected comparison described above. On the earlier

comparison two further groups appeared biased and several looked worse than they

are — personal care and service most of all, at −6.7% against −0.6% once nothing

is discarded. 12 of 21 groups carry a material bias, not 13. The pattern and

every headline case survive the correction; the rank correlation between the two

sets of group estimates is 0.97.*

This matters most for the occupation figures themselves. Where a county page

reports estimated local employment in a named occupation, that figure carries its

group's bias: a county's estimated legal employment is, on average, high by about

a tenth, and its estimated production employment low by about an eighth. Anyone

using those counts directly should read them with that in mind, alongside the

county's confidence rating.

It matters much less for the index. The bias leans in the direction that

would concern us most — the overstated groups are disproportionately the

high-exposure ones, so the concern is that P1 is inflated. We tested it: deflating

every biased group by its measured bias, holding each county's observed total

employment fixed, rebuilding P1 and rescoring the whole country changes the

national ranking by a rank correlation of 0.9993. The median county moves

15 places out of 3,142, 50 counties move more than 100, and **6 counties leave

the top decile** (6 others enter it). The reason is structural: P1 measures

shares of a fixed county total, and the exposed set contains both overstated

groups (legal, computer, business) and understated ones (education, life

sciences), so the errors substantially cancel before they reach a score.

The likely source is the documented difference between how the Census's

self-reported occupation data and the employer-reported OEWS data code the same

job. Our procedure already calibrates for that difference at state level; what

remains is presumably the part that varies within a state, which a statewide

calibration cannot remove. That is an open item, not a solved one.

Further disclosures on the same measurements:

• At detailed occupation level, rather than major-group level, the

Minnesota diagnostic error is 30.5%. Detailed-occupation estimates for a

single county are substantially noisier than the major-group figures, and

should not be relied on individually.

• 6 of 22 Minnesota occupation major groups are individually above 12%,

even though the weighted overall figure is 8.5%. The weakest are farm

(SOC 45, 40.0%), construction and extraction (SOC 47, 18.1%) and life,

physical and social science (SOC 19, 16.2%).

• **Five states are at or above the 12% gate on the leave-one-anchor-out test

itself**: Montana 13.8%, Louisiana 13.1%, North Dakota 13.1%, Wyoming 12.9%

and Alabama 12.8%. Each of those states' county pages reports its own figure.

• Every county page reports its state's own leave-one-anchor-out figure as a

number, not as a pass or fail against the target.

5.2 Withheld — the estimation procedure

*This subsection is withheld. It specifies the procedure that turns

area-level occupation data into county-level estimates: the sequence of

steps, the constraints each step is fitted to, and the treatment of

suppressed source cells. What it consumes, what it produces and how

accurately it does so are published above in Section 5.1; the validation

results reported there are results of this procedure.*

6. Scoring: Published Weights and Formulas

6.1 Exposure ensemble weights

Each occupation's exposure score is a weighted blend of three independent

published research bases, each rescaled to a common 0-1 range before blending:

Rationale for the split. Observed usage carries the largest weight because

it is the only measure of what is actually happening rather than what is

technically possible; it is not given a majority because it reflects today's

adoption, which is uneven and partly a function of who has access to the tools.

AIOE carries the second weight for its peer-reviewed standing and its complete

coverage of the SOC structure. Eloundou et al. carries the third because it is

task-granular and the most capability-forward of the three. These are a

judgment, not an estimate: no procedure fixed them, and we make no claim that

they are optimal. They are published so that a reader can disagree with them

specifically, and so that the index can be reproduced.

Automation adjustment. Exposure is then scaled by the share of an

occupation's observed AI usage that is task substitution rather than task

assistance. An occupation whose AI usage is entirely assistance is scaled to

half its raw exposure; one whose usage is entirely substitution keeps its full

exposure; the scaling is linear between those ends. Occupations with too little

classified usage to support a ratio fall back, in order, to a model fitted on

work-activity content and then to occupation-group means; every published score

carries which of the three tiers produced it.

Cross-source disagreement — the spread across the three rescaled sources —

is carried forward into the confidence rating (Section 6.3) rather than being

averaged away.

6.2 Pillar weights and rationale

The composite is the weighted sum of the five pillars' national percentile

ranks, with P3 and P4 inverted before aggregation, divided by 100. Weights are

held in configuration, not in code, so that the sensitivity analysis in

Section 7 can perturb them.

These weights are a judgment and are published as such. They were not

derived from an optimization, because there is no outcome variable to optimize

against — the index measures position, not a realized result. The defence of

them is not that they are correct but that they are stated, versioned, stable

under perturbation (Section 7), and changed only under a published governance

policy.

6.3 Confidence rating

Every county carries a High / Medium / Low confidence rating, computed from

three components:

Each component is percentile-ranked nationally so that higher means less

confidence, and the three are combined with equal weight (1.0 / 1.0 / 1.0).

The resulting 0-100 composite is cut as follows:

The cuts are even thirds. No county is exempt from a rating, and the rating is

published alongside every score rather than only on request.

6.9 Withheld — indicator construction, normalization and time-phasing

*This subsection is withheld. It specifies how each pillar's underlying

indicators are computed from the source data, how they are normalized before

aggregation, and the time-phasing model. The weights applied to those

indicators' pillars, the exposure-ensemble weights and the confidence formula

are published above in Sections 6.1 to 6.3 — what is withheld is the

indicator-level arithmetic, not the weights.*

7. Validation & Defensibility Program

This is the full program as specified. **Two of the four items have not yet

been run**, and are marked as such rather than implied to be complete.

Items 1 and 2 are the two that would test whether this framework detects

real-world labor disruption, as opposed to being internally consistent. Until

they are run, the index's claim rests on the defensibility of its inputs and

its structure, not on demonstrated predictive performance, and it should be

read that way.

Every figure in this specification is generated, not transcribed. The

statistics quoted throughout are produced from the pipeline's own artifacts

into a single machine-readable file, and the published page is tested against

that file rather than against literals — so a figure that moves when the data

is refreshed fails the build until this document is brought with it.

8. Known Limitations

Published deliberately; candor is a moat. Several of these carry numbers from

our own internal analysis of the index, and are stated whether or not they

flatter it.

8.1 Scope limitations

• Exposure research measures technical capability overlap, not adoption decisions, regulation, or new job creation.
• County occupation estimates are modeled, not surveyed, below MSA level (mitigated by the Section 5.1 validation).
• The index is comparative (relative risk), not absolute (jobs lost).
• Self-employed (unincorporated) and unpaid family workers — roughly a tenth of the workforce — are outside the county occupation estimates; no county-level occupation source exists for them. Farm production employment likewise (v0.2.0).

8.2 P1 and P3 are substantially correlated, and partly cancel

This is the most material limitation in the index. A county's Direct

Occupational Exposure pillar and its Adaptive Capacity pillar are not

independent measurements:

The mechanism is educational attainment. The occupations that current exposure

research scores as most exposed are disproportionately degree-requiring, and

educational attainment is the single heaviest indicator in P3. High-exposure

counties are therefore systematically high-capacity counties, and because P3

enters the composite inverted, it works against P1 rather than adding

independent information.

**Quantified: a county one percentile higher on P1 is on average 0.43

percentiles higher on P3, so roughly 30% of P1's nominal 35-point weight is

systematically cancelled by P3. P1's effective weight against a county's

exposure standing is closer to 24 points than 35.**

This is not an artifact of our ensemble. The same relationship appears, more

strongly, when exposure is measured using the AIOE scores alone: rank

correlation 0.715 against educational attainment. It is a property of current

AI exposure measures generally.

The open question we cannot yet answer. Educational attainment is used here

as a proxy for a workforce's ability to retrain and an economy's ability to

regenerate. That proxy was established against earlier labor shocks, in which

the destination for displaced workers was disproportionately degree-requiring

work. It is not established that it remains valid when the exposure measure it

nets against is itself degree-weighted, and when the retraining destination

market may be the exposed market. If it does not hold, P3 is discounting

exactly the communities whose exposure is most concentrated. We flag this as

unresolved rather than assert either answer.

8.3 Pillar contribution is uneven

Recomputing the composite five times, each time dropping one pillar and

renormalizing the remaining weights, then rank-correlating each variant against

the full composite. A lower correlation means the pillar is doing more work:

Two findings follow, neither of them flattering:

• P2 does more work than P3 despite carrying less weight (20% against 25%).

P2 is essentially uncorrelated with exposure, so all of its weight buys

movement, while a substantial part of P3's is spent offsetting P1 (Section

8.2).

• P4 is close to inert. At a rank correlation of 0.979 and a median move of

113 places out of 3,142, the Regional Buffer pillar is the weakest

contributor in the index — and it is also the pillar with the known data

compromises: no travel-time data, out-of-state commuting destinations dropped

from the shed mix, and border counties' buffer consequently understated.

**These figures are a candidate basis for a v1.1 recalibration of the P2, P3

and P4 weights, under the governance & recalibration policy. No weight has been

changed. v1.0.0 stays frozen, and this analysis is the documented reason a

change waited rather than being made quietly alongside the disclosure.**

8.4 Among highly exposed counties, the ranking is driven by resilience

Restrict the country to the 315 counties in the top decile of P1 — our own

exposure pillar — and the rank correlation between the CAERI composite and P1

within that group falls to 0.268 (Kendall's tau-b 0.183). Once a county is

in the most-exposed tenth, most of what separates it from its peers in the

CAERI ranking is resilience, not exposure.

This is the index working as designed, not a defect — an exposure-and-

resilience index is supposed to separate communities that share an exposure

profile by their capacity to absorb it. It is stated here because it is

counter-intuitive, because a reader who discovers it unaided will reasonably

suspect the exposure measure is not doing anything, and because it means a

top-decile CAERI list should never be described as a list of the most exposed

places.

*Correction, 2026-09-18: this subsection previously reported 0.197, computed

against a strict employment-share-weighted AIOE ranking rather than against P1.

That was inconsistent with Section 9, which declines to treat that construction

as a benchmark, so the figure is now computed against our own exposure pillar.

No score, rank or band was affected.*

9. Relationship to Prior Work

CAERI is not the first measure of geographic AI exposure, and P1 is not an

original construct.

Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and

geographic exposure to artificial intelligence: A novel dataset and its

potential uses. Strategic Management Journal, 42(12), 2195-2217.

Felten, Raj and Seamans construct an AI Occupational Exposure (AIOE) score for

each occupation, and aggregate it to places by weighting occupations by their

local employment shares — their AI Geographic Exposure (AIGE) measure. **P1 is

conceptually close to AIGE**: both ask what share of a local economy's work

sits in occupations that published research scores as exposed, and AIOE is one

of the three inputs to our exposure ensemble.

What CAERI adds:

1. An ensemble rather than a single source. Exposure is blended across

AIOE, Eloundou et al. and observed usage from the Anthropic Economic Index

(Section 6.1), with the disagreement between them carried into a published

confidence rating rather than discarded. AIOE alone is an expert-assessment

measure of technical capability overlap; the ensemble adds a

revealed-usage measure and a task-level one.

2. Four non-exposure pillars. Concentration, adaptive capacity, regional

buffer and fiscal sensitivity (Section 2). This is the substantive addition:

AIGE measures exposure, CAERI measures exposure against the capacity to

absorb it.

3. A maintained refresh cadence. Quarterly, annual and event-driven source

refreshes under a published governance policy, versioned and reproducible,

rather than a one-time research dataset.

4. A fiscal layer. P5 scores the local government's own revenue exposure —

a question the labor-exposure literature does not address, and the one local

officials ask first.

How far the pillars actually move the ranking. Against a ranking built on

our own P1 pillar alone, the full composite has a rank correlation of

0.7269 (Kendall's tau-b 0.5197) across all 3,142 scored counties. That is

the honest measure of what the four non-exposure pillars add: substantial

reordering, not a relabelling of an exposure ranking — and not independence

from it either.

What we deliberately do not claim. We have separately built a strict

employment-share-weighted AIOE ranking for comparison, and it correlates less

well with CAERI than the figure above. We do not publish that number as our

distance from AIGE, because our own analysis shows it is not a clean benchmark:

no county-level OEWS exists, so any county-level AIOE ranking must ride on the

same modelled county occupation estimates CAERI uses; 57 occupation codes

covering 7.18% of employment fall outside the AIOE-to-current-SOC crosswalk and

are dropped; and the AIOE vintage predates the large-language-model period that

two of our three ensemble inputs are built to capture. Quoting a larger

separation from a benchmark we have ourselves shown to be unsound would

overstate the difference, which is the failure this section exists to avoid.

About the county estimates

County-level occupation figures are model-based ESTIMATES assembled from the federal sources above; every county carries a Confidence rating reflecting source agreement, disclosure imputation, and county size (Section 6.3). How those estimates were validated — including the test that comes in above our target — is set out in full in Section 5.1, and the index's known limitations, with figures, are in Section 8. The estimation procedure itself (Section 5.2) and the indicator-level scoring formulas (Section 6.9) are the two withheld blocks; everything else in this specification is published, including the pillar weights and the exposure-ensemble weights. This is CAERI v1.0.0. The index is comparative and not a forecast.

How this methodology is governed

How versions work, how often weights may change, how corrections are handled, and the standards any recalibration must meet are set out in the Governance & recalibration policy.