# DATA_NOTES — Colorado AI Exposure Atlas, 2026 edition

Auto-generated by the pipeline (`pipeline/build_data.py`). Employment data
year: 2025. Every number in this bundle is produced by the open
pipeline in this project's repository; nothing is hand-edited.

## Source vintages

| Source | Vintage | Role |
|---|---|---|
| BLS QCEW county x industry employment | 2025 annual averages (NAICS 2022) | County industry employment |
| BLS OEWS national staffing patterns | May 2025 (NAICS 2022 x SOC 2018) | Industry -> occupation shares |
| BLS OEWS national occupation file | May 2025 | National employment weights, mean wages |
| BLS OEWS state file | May 2025 | Validation target (reconciliation below) |
| Eloundou et al., "GPTs are GPTs" | O*NET 27.2 (2023); human β ratings | Primary exposure score (β x 100) |
| Felten/Raj/Seamans AIOE | 2021 (SOC 2010, crosswalked to SOC 2018) | Robustness check — derived statistics only (no license; raw values are never published) |
| Anthropic Economic Index | Sept 2025 release + labor-market companion files | Observed-usage triangulation (methodology page; not in county scores) |
| Census cartographic boundaries | 2024 vintage, 1:500k, simplified | Map geometry, county names |
| Census ACS 5-year estimates (public domain) | 2020-2024 (tables B01003, B20017) | County context only — population, median worker earnings; never an input to any exposure computation |

## What the employment numbers cover

QCEW counts jobs covered by unemployment insurance — it excludes most
self-employment, many farm workers, and some others. Every employment figure
here is **UI-covered employment**, not "total jobs."

## Suppression, imputation, and fallback routing

QCEW suppresses small county x industry cells. Suppressed cells are imputed
top-down from disclosed parent totals, distributed proportional to the
state-level industry mix (floors from disclosed children honored). Separately,
employment in industries with no published OEWS staffing pattern (public
administration, parts of agriculture, private households) is routed to
documented blend patterns ("fallback-routed" below; FOLLOWUPS #15-#17).

`data_quality` thresholds (decided Aug 6, 2026; FOLLOWUPS #3): below
10% of employment imputed = `full`;
10%-50% = `partial_imputed`;
above 50% = `heavily_imputed`.

Statewide: 8.6% of employment imputed,
6.5% fallback-routed. County distribution:
9 `full`, 38 `partial_imputed`,
17 `heavily_imputed`.

| County | FIPS | Employment imputed | Fallback-routed | Imputed cells | Data quality |
|---|---|---|---|---|---|
| Adams | 08001 | 6.9% | 4.5% | 81 of 273 | full |
| Alamosa | 08003 | 39.4% | 12.6% | 103 of 150 | partial_imputed |
| Arapahoe | 08005 | 3.6% | 4.8% | 77 of 270 | full |
| Archuleta | 08007 | 33.3% | 9.0% | 100 of 146 | partial_imputed |
| Baca | 08009 | 38.9% | 13.2% | 56 of 67 | partial_imputed |
| Bent | 08011 | 55.4% | 15.9% | 42 of 55 | heavily_imputed |
| Boulder | 08013 | 6.0% | 5.0% | 78 of 268 | full |
| Broomfield | 08014 | 17.5% | 4.0% | 106 of 211 | partial_imputed |
| Chaffee | 08015 | 21.0% | 11.3% | 104 of 170 | partial_imputed |
| Cheyenne | 08017 | 57.6% | 19.7% | 41 of 53 | heavily_imputed |
| Clear Creek | 08019 | 58.2% | 12.2% | 89 of 116 | heavily_imputed |
| Conejos | 08021 | 53.8% | 18.8% | 64 of 81 | heavily_imputed |
| Costilla | 08023 | 53.6% | 29.6% | 48 of 60 | heavily_imputed |
| Crowley | 08025 | 54.8% | 37.0% | 33 of 38 | heavily_imputed |
| Custer | 08027 | 52.8% | 13.0% | 72 of 91 | heavily_imputed |
| Delta | 08029 | 25.4% | 13.3% | 119 of 183 | partial_imputed |
| Denver | 08031 | 1.8% | 6.1% | 65 of 283 | full |
| Dolores | 08033 | 54.2% | 21.9% | 50 of 60 | heavily_imputed |
| Douglas | 08035 | 8.1% | 3.3% | 87 of 245 | full |
| Eagle | 08037 | 21.7% | 5.2% | 96 of 197 | partial_imputed |
| Elbert | 08039 | 26.7% | 9.1% | 96 of 148 | partial_imputed |
| El Paso | 08041 | 5.1% | 5.3% | 79 of 275 | full |
| Fremont | 08043 | 24.2% | 27.9% | 115 of 179 | partial_imputed |
| Garfield | 08045 | 16.0% | 7.5% | 95 of 208 | partial_imputed |
| Gilpin | 08047 | 86.4% | 8.3% | 65 of 82 | heavily_imputed |
| Grand | 08049 | 17.7% | 8.0% | 97 of 146 | partial_imputed |
| Gunnison | 08051 | 20.8% | 8.8% | 97 of 165 | partial_imputed |
| Hinsdale | 08053 | 47.7% | 22.4% | 37 of 45 | partial_imputed |
| Huerfano | 08055 | 57.9% | 13.1% | 74 of 89 | heavily_imputed |
| Jackson | 08057 | 65.2% | 29.8% | 45 of 54 | heavily_imputed |
| Jefferson | 08059 | 6.0% | 7.4% | 77 of 273 | full |
| Kiowa | 08061 | 47.5% | 12.6% | 26 of 35 | partial_imputed |
| Kit Carson | 08063 | 45.5% | 19.4% | 70 of 95 | partial_imputed |
| Lake | 08065 | 54.4% | 15.4% | 77 of 98 | heavily_imputed |
| La Plata | 08067 | 19.3% | 11.3% | 106 of 219 | partial_imputed |
| Larimer | 08069 | 6.7% | 6.2% | 75 of 269 | full |
| Las Animas | 08071 | 39.4% | 15.6% | 98 of 131 | partial_imputed |
| Lincoln | 08073 | 41.9% | 30.6% | 67 of 79 | partial_imputed |
| Logan | 08075 | 33.7% | 20.8% | 91 of 141 | partial_imputed |
| Mesa | 08077 | 13.6% | 6.6% | 93 of 248 | partial_imputed |
| Mineral | 08079 | 63.1% | 10.2% | 41 of 50 | heavily_imputed |
| Moffat | 08081 | 47.2% | 11.5% | 110 of 140 | partial_imputed |
| Montezuma | 08083 | 31.7% | 14.0% | 108 of 173 | partial_imputed |
| Montrose | 08085 | 16.6% | 7.9% | 110 of 205 | partial_imputed |
| Morgan | 08087 | 48.0% | 13.5% | 117 of 171 | partial_imputed |
| Otero | 08089 | 45.0% | 12.5% | 94 of 134 | partial_imputed |
| Ouray | 08091 | 60.1% | 14.8% | 85 of 113 | heavily_imputed |
| Park | 08093 | 29.8% | 14.9% | 95 of 137 | partial_imputed |
| Phillips | 08095 | 56.8% | 31.0% | 64 of 80 | heavily_imputed |
| Pitkin | 08097 | 12.5% | 10.9% | 93 of 163 | partial_imputed |
| Prowers | 08099 | 37.2% | 20.4% | 95 of 127 | partial_imputed |
| Pueblo | 08101 | 18.0% | 8.2% | 100 of 228 | partial_imputed |
| Rio Blanco | 08103 | 50.8% | 15.1% | 79 of 97 | heavily_imputed |
| Rio Grande | 08105 | 47.8% | 18.5% | 97 of 133 | partial_imputed |
| Routt | 08107 | 18.2% | 7.4% | 116 of 201 | partial_imputed |
| Saguache | 08109 | 45.6% | 29.8% | 66 of 86 | partial_imputed |
| San Juan | 08111 | 75.4% | 12.0% | 48 of 56 | heavily_imputed |
| San Miguel | 08113 | 46.6% | 8.9% | 101 of 136 | partial_imputed |
| Sedgwick | 08115 | 45.4% | 25.5% | 54 of 64 | partial_imputed |
| Summit | 08117 | 27.6% | 8.5% | 99 of 177 | partial_imputed |
| Teller | 08119 | 42.1% | 9.6% | 108 of 159 | partial_imputed |
| Washington | 08121 | 38.3% | 23.6% | 64 of 77 | partial_imputed |
| Weld | 08123 | 8.4% | 9.0% | 69 of 265 | full |
| Yuma | 08125 | 42.5% | 24.7% | 77 of 112 | partial_imputed |

## Validation results

Gates enforced by `pipeline/validate.py` (build fails on any):

1. **County-to-state QCEW total.** Sum of the county QCEW totals =
   2,860,276 vs the published state total (area
   08000) = 2,891,093;
   difference 1.0659% (tolerance
   1.5%). The published state total includes employment
   at locations QCEW cannot assign to any county (the "Unknown Or Undefined"
   area, 08999); the county files definitionally exclude
   it, which accounts for the
   30,817-job
   difference.
2. **Completeness and ranges.** All 64 counties
   present in every emitted file; every published number non-null; scores in
   [0, 100]; ranks a permutation of 1..64; slugs
   unique.
3. **OEWS reconciliation.** The imputed statewide occupation distribution vs
   the published OEWS state file, compared as shares by SOC major group
   (levels are not comparable across the two programs; shares are — OEWS is
   QCEW-benchmarked). Max |delta| = 2.35 pp,
   mean |delta| = 0.57 pp. Tolerances:
   2.5 pp max, 0.75 pp mean —
   set after observing this baseline, frozen for future editions, where they
   become binding (FOLLOWUPS #4, #26).

| SOC group | Title | Imputed share % | OEWS share % | Delta pp |
|---|---|---|---|---|
| 11-0000 | Management Occupations | 7.50 | 5.44 | +2.06 |
| 13-0000 | Business and Financial Operations Occupations | 6.90 | 9.25 | -2.35 |
| 15-0000 | Computer and Mathematical Occupations | 4.27 | 4.76 | -0.50 |
| 17-0000 | Architecture and Engineering Occupations | 1.95 | 2.52 | -0.57 |
| 19-0000 | Life, Physical, and Social Science Occupations | 0.89 | 1.13 | -0.24 |
| 21-0000 | Community and Social Service Occupations | 1.63 | 1.94 | -0.31 |
| 23-0000 | Legal Occupations | 0.78 | 0.94 | -0.15 |
| 25-0000 | Educational Instruction and Library Occupations | 5.57 | 5.35 | +0.22 |
| 27-0000 | Arts, Design, Entertainment, Sports, and Media Occupations | 1.33 | 1.42 | -0.09 |
| 29-0000 | Healthcare Practitioners and Technical Occupations | 5.68 | 5.35 | +0.34 |
| 31-0000 | Healthcare Support Occupations | 4.36 | 3.73 | +0.63 |
| 33-0000 | Protective Service Occupations | 2.46 | 2.30 | +0.16 |
| 35-0000 | Food Preparation and Serving Related Occupations | 9.25 | 9.45 | -0.20 |
| 37-0000 | Building and Grounds Cleaning and Maintenance Occupations | 3.19 | 3.10 | +0.09 |
| 39-0000 | Personal Care and Service Occupations | 2.45 | 2.72 | -0.27 |
| 41-0000 | Sales and Related Occupations | 8.75 | 10.38 | -1.62 |
| 43-0000 | Office and Administrative Support Occupations | 11.63 | 10.55 | +1.08 |
| 45-0000 | Farming, Fishing, and Forestry Occupations | 0.47 | 0.18 | +0.30 |
| 47-0000 | Construction and Extraction Occupations | 4.95 | 4.73 | +0.22 |
| 49-0000 | Installation, Maintenance, and Repair Occupations | 4.06 | 3.97 | +0.10 |
| 51-0000 | Production Occupations | 3.77 | 3.26 | +0.51 |
| 53-0000 | Transportation and Material Moving Occupations | 8.13 | 7.53 | +0.60 |

4. **Eloundou-vs-AIOE Spearman.** Rank correlation between the primary
   (Eloundou) and robustness (Felten AIOE) county scores:
   0.934 (floor 0.7, FOLLOWUPS #8).
5. **Determinism.** The full emit, run twice, produces byte-identical files
   (checked by `validate.py` on every build).
6. **Committed outputs current.** The committed `site/src/data/` and download
   bundle must match a fresh build byte-for-byte; a stale committed dataset
   fails the build.

## Caveats

- **Exposure is not job loss.** The score measures overlap between an
  occupation's tasks and what current AI systems can do — where change
  arrives first, augmentation and automation both. It is not a layoff
  forecast.
- **County occupation mixes are estimated, not observed.** Establishment-based
  county occupation data does not exist; each county's mix is imputed from its
  industry mix and national staffing patterns (the established
  Brookings-lineage method). Detailed-occupation county estimates are deliberately never
  published — occupation detail stops at the SOC major group.
- **Jobs, not people.** QCEW counts jobs; a person holding two covered jobs
  is counted twice, so employment shares are shares of jobs, not of workers.
- **ACS context columns are residence-based.** The `population` and
  `median_earnings` columns (ACS 2020-2024 5-year) describe where people
  live, while every employment figure describes where jobs are located — so
  they contextualize a county without entering any exposure computation.
  They are the bundle's only nullable columns: blank when the ACS raw file
  was unavailable at build time, and `population_moe` is blank wherever the
  ACS total-population estimate is controlled (no sampling error published —
  most counties; only the smallest carry a real MOE).
- The high-exposure threshold (score >= 50.0,
  the top quintile of the national employment-weighted score distribution)
  was decided Aug 6, 2026 (FOLLOWUPS #2, #21); the 2026 cutoff sits on a
  16-occupation tie mass.
- Public-administration employment is routed to OEWS state/local government
  staffing blends; uncovered agriculture to crop/animal support-activity
  blends (FOLLOWUPS #15, #16). Both are counted in "fallback-routed" above.
- Felten AIOE raw values are unlicensed and are never published; only derived
  statistics (the Spearman correlation above) appear in this bundle.

## License

Data: CC BY 4.0 (see LICENSE in this directory). Please cite:
"Colorado AI Exposure Atlas, 2026 edition. Christopher Martin.
https://coloradoaiexposureatlas.com/"