RuralSim

A reproducible rural-county digital twin: five pilot counties as OpenUSD 3D worlds, a national statistical table for 3,144 counties, and an experiment harness that scores interventions on a rural quality-of-life index (RQLI), always as a distribution with hard equity constraints.

Results are screening-grade. Engines are reduced-order with assumed device parameters; no experiment result is a finding until confirmed with full-physics models. See docs/LIMITS.md.

Quickstart

pip install -r requirements.txt

# rebuild a pilot county's 3D stage from the committed raw layers, then validate it
python3 tools/build_stage.py 16081
python3 tools/validate_county.py 16081

# run experiment cards (20 seeds = promotion-eligible; baselines cached to experiments/baseline_cache/)
python3 -m tools.harness.run_experiment EXP-0002 --seeds 20
python3 -m tools.harness.run_experiment --county 01131 --seeds 20

# rebuild the national table
python3 tools/build_national.py

# full rebuild including live data fetches (needs network)
python3 tools/build_county.py 31173

# build any other current U.S. county the same way (boundary cached to counties/<fips>.geojson)
python3 tools/build_county.py 48301      # Loving County, TX

# rebuild the website into public/ (Vercel runs this on every push to main)
python3 tools/build_site.py

Pilot FIPS: 31173 Thurston NE · 16081 Teton ID · 54047 McDowell WV · 20019 Chautauqua KS · 01131 Wilcox AL. Live site: https://www.ruralsim.io. The data tables in data/ and pilots/ were built by the scripts in build_scripts/ (see its README for sources).

The live fetchers depend on public services that are sometimes down or rate-limited (OpenStreetMap Overpass, Open-Meteo, CropScape). Each fetch is resumable: rerun the same command and it picks up where it stopped.

License

Code: Apache 2.0. Data: see DATA_LICENSE.md. Map data derived from OpenStreetMap is © OpenStreetMap contributors under the ODbL; other public data keeps its publisher's terms; RuralSim's own specs, results and writing are CC BY 4.0. Reproducibility check (2026-11-12, clean machine): all five stages, all ten cards at 20 seeds, and the national table reproduced identically to the original build.

---

Original handoff bundle v0.1

Built 2026-11-10. Everything Kimi K3.1 Pro needs to start building the 3D rural world model, except the items under "You must supply".

Build status 2026-11-12 — all five pilot 3D twins built, validated and render-verified (worlds/, validation/); Tier A national statistical twin built (national/, 3,163 counties, 8/8 checks pass); experiment harness operational (tools/harness/, ten seed cards in experiments/ledger/); improve-quality-of-life skill packaged in skills/. See docs/ARCHITECTURE.md §8, docs/DECISIONS.md D-001…D-026, docs/LIMITS.md.

Start here

1. Unzip this bundle into the working directory of an agentic Kimi session (Kimi Code or equivalent, with shell and network).
2. Fill the compute budget in PROMPT.md, then paste PROMPT.md as the first message.
3. Kimi stops at Gate 1 after profiling the data. Review before it proceeds.

Contents

Path	What it is	Rows
PROMPT.md	The build prompt, pre-filled with this bundle's paths and pilot counties
data/county_master.csv	One row per county: rural codes 1974–2023, Census rural share, typology, area, centroid, population 1900/2000/2010/2020/2024, poverty, income, hospitals, last three presidential margins	3,163 (3,144 current + 19 retired codes)
data/county_year.csv	Population and components of change, 2010–2024	47,150
data/county_profile.csv	254 demographic, jobs, income and veteran indicators (USDA ERS Atlas)	3,148
data/county_decade.csv	Decennial population 1900–2020	40,159
data/county_election.csv	Presidential returns by county, 1868–2024	116,878
data/state_decade_power.csv	State population, House seats, electoral votes, Senate / House / Electoral College weight indices, 1910–2020	588
data/national_power_summary.csv	Malapportionment indices by decade; nonmetro- vs metro-weighted Senate and Electoral College weight (2020)	12
data/facility_hospital.csv	Every Medicare-registered hospital with address, type and county	5,419
data/codebook.csv	Every column: description, source, vintage, provenance tag	375
pilots/pilot_counties.csv	Five pilot counties with selection rule, profile and bounding box	5
pilots/pilot_county_boundaries.geojson	Real county polygons (Census 2023 cartographic boundaries)	5
specs/geodata_manifest.yaml	Every 3D layer: source, access point, auth, provenance	22 layers
specs/rqli_spec.yaml	Quality-of-life index: 9 domains, 21 indicators, 6 hard constraints
experiments/experiment_card_schema.yaml	The format every test must follow
experiments/seed_experiments.yaml

Checks run on the data

Check	Result	Reference
2020 population, current counties	331,449,281	Matches the 2020 Census count
Nonmetro / metro counties (RUCC 2023)	1,958 / 1,186	Matches USDA ERS
Nonmetro share of population	13.8%	Matches USDA ERS
Census rural population	66,300,254 (20.0%)	Matches Census Bureau
Key uniqueness, fips5	Unique in master
Hospital-to-county match	97.4% of facilities	2.6% unmatched on county name
Nonmetro presidential margin	D+1.5 in 1992, R+15.0 in 2000, R+32.1 2016, R+32.0 2020, R+35.6 2024	2024 is consistent with published ~R+36
Nonmetro share of votes cast	16.7% (1992) → 14.2% (2024)	Counties classified by their 2023 code in every year

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗