Eighty-three counties. One engine.
The platform's first full-state election night: every Michigan county, 268 contests, live boards and a bespoke race tracker in an outside client's war room - on pages shipped the same day.
Statewide toplines from the covered counties.
These are the sums the boards actually serve - counted from the counties on the board (50 of 83 at this snapshot), each one verified to the vote before inclusion. They are not full-state totals; they climb as the certification tail fills the map in. Snapshot: August 6, 2026, 9:29 PM ET.
Unofficial · covered-county sums, not statewide totals · leaders reflect served counties only, not race calls.
Track a whole state at once, for a client watching live.
Georgia in June was one statewide runoff on one vendor feed. Michigan was the real test: a statewide primary across 83 counties that each publish results their own way, with an outside client's war room following two statewide races and two dozen targeted legislative districts in real time. The tracker they watched was assembled and tested the same day - composed from platform services already running, not built from scratch.
There is no format. There is a zoo.
Michigan's counties publish promptly - but each in its own dialect: vendor JSON feeds, per-jurisdiction HTML apps, PDF summary reports in several sub-families, statements of votes cast that only parse by page coordinates, scanned images with no text layer at all. Counties on the same vendor platformencode "which primary is this" five different ways. Reading the state means decoding each dialect once - and once decoded, it is decoded for every future election.
A full analysis of the county-feed differences is being prepared through certification.
Verified to the vote, or not shown at all.
Every county figure the boards served was fetched from the county's own publication, parsed by us, and cross-checked candidate by candidate against an independent reference we treat as a quality oracle and never ingest. Parsers fail closed: a county that cannot prove its own arithmetic does not appear, rather than appearing wrong. A stale-artifact guard refuses any report it cannot date - on this cycle it kept a 2022 general-election file, 267,927 four-year-old votes, off a live board. Where counties publish only scanned images, humans enter the numbers through an attributed, append-only journal that retires itself if a real feed ever appears.
The map keeps filling in.
31 counties were machine-readable on the night itself. Two days later the boards showed 50 of 83 counties - every one verified to the vote - with automated watches sweeping the rest every six hours until the state certifies. Election night is a moment; the coverage is a process, and the engine runs it around the clock.
Sourced from the engine's own retrospective: commit history, service journals and the night-watch record. Every number above was measured, not estimated. Back to the platform