Verify outcomes
Reconcile county totals to official statewide control totals and retain provenance for every edition.
02EVIDENCE STANDARD
Useful because
it is bounded.
Strategic intelligence should make uncertainty legible. This release separates verified evidence, modeled context, and unresolved gaps.
Outcome Audit v31 · Geography Pilot v2
Last analytical release: September 2026
A clean edge around what the evidence supports.
The 2020 cross-section contains 3,102 released counties across 50 states, representing 153,650,737 two-party votes. Ten source geography rows remain outside the release rather than being repaired with invented allocations.
MeasureReleasedHeldCoverage
2020 counties3,1021099.68%
Stable historical counties1,448—2012–16
Historical state-years848Strict gate
Contextual model, not election forecast.
The 2020 ridge model estimates county Democratic two-party share from public economic, demographic, and state context. Its residuals help identify counties that diverge from the national pattern. Because training and evaluation use the same election year, these metrics describe interpolation—not future predictive power.
ModelCounty MAEVote-wtd. MAEWinner accuracy
National mean13.09 pp21.11 pp82.85%
State mean10.14 pp15.99 pp84.72%
Ridge + ACS + state4.78 pp5.20 pp93.68%
The honest reading
A small residual says a county resembles the national relationship encoded by the model. A large residual says the local result needs more explanation. Neither establishes why voters behaved as they did.
Audit first. Join second. Model last.
Every analytical row passes a layered release process. The sequence matters: a sophisticated model cannot repair unsupported geography or conflicting official editions.
Stabilize geography
Separate fixed counties from boundary changes, nonadditive districts, and combined reporting units.
Validate joins
Match economic context only after identifiers, years, missingness, and coverage pass explicit tests.
What remains unresolved is part of the record.
These eight historical state-years are withheld from strict county analysis. Each requires new official evidence—not a statistical patch—to cross the release boundary.
Official Edition Conflict
44 D / 26 R statewide votes lack county attribution
Evidence needed: Dated official county erratum or certified county canvass allocating the residual
Official Edition Conflict
Official 82-county register edition is D350/R386 above later FEC edition
Evidence needed: Official county-level erratum or certified recanvass allocating the residual
Nonadditive Geography
Forty official district reports reuse shared special-vote batches; D69,340/R101,378 overcount
Evidence needed: Official non-overlapping precinct or ballot-allocation file with stable geographic keys
Three Edition Conflict
Election-night official, county candidate, and certified statewide editions disagree
Evidence needed: Official county-level certified final canvass or revision ledger
Official Edition Conflict
373 D / 75 R statewide votes lack county attribution
Evidence needed: Dated official county erratum or certified county canvass allocating the residual
Correction Provenance Gap
Saved official county and precinct PDFs agree but a later Lake correction is not officially pinned
Evidence needed: Dated official correction or certified Lake County canvass
Official Edition Conflict
670 D / 137 R statewide votes lack county attribution
Evidence needed: Dated official county erratum or certified county canvass allocating the residual
Official Edition Conflict
392 D / 78 R statewide votes lack county attribution
Evidence needed: Dated official county erratum or certified county canvass allocating the residual