Live results: the expected-vote methodology
Our denominator and our call rule are published in full. Open methodology is a feature, not an apology — it is how a self-sourced, math-only desk earns trust against the incumbents.
What a race looks like on the night
The share of expected vote, a running tally, which counties have reported, and an honest call label. Every live number is provisional and can be revised — usually downward — as counties finish counting.
Counted votes read solid; the central estimate is the bar’s length; the faint hatched tail is the conservative upper bound — the remaining vote, made visible, so “why isn’t this called yet” is a picture, not a footnote.
The denominator — a dual expected-vote estimate
A useful early read needs an estimate of how many votes are still out. We publish two numbers and use each for a different job.
- Central estimate — our best estimate of the final total, blending each county’s precinct-reporting completeness with a prior-midterm turnout baseline. This drives the “% of expected vote in” you see.
- Conservative upper bound — a deliberately high ceiling on the votes still to come. This is the only number our call rule is allowed to use. Using a central estimate as the call denominator is exactly how a race gets called early; the upper bound makes the call rule provably safe.
The call rule — math-decisive only
For 2026 Akashic makes no probabilistic projections and shows no needle.
A race is marked called only when it is arithmetically impossible for the trailing field to catch up, given the conservative upper bound on the remaining vote:
(leader’s votes − strongest other candidate’s votes) > conservative upper bound on votes remaining
Until that holds, the label is honest: too early, leading, or too close to call. Ranked-choice contests are never auto-called — a first-choice plurality can be overturned by reallocation. A call can be retracted if a correction undoes the decisive margin; retractions are versioned and visible.
The turnout baseline
Grounded in public MIT Election Lab precinct returns, published per state.
The expected-vote baseline is the most recent comparable midterm (2022, then 2018) county turnout for the office — never a presidential cycle, which would inflate the denominator by a third.
Provisional → certified
The live layer can churn freely without corrupting history.
Live numbers are provisional and never overwrite Akashic’s canonical historical record. After the canvass, certified results fold into the static atlas through a separate reconciliation step.
What we deliberately do not cover live
Coverage follows feed quality, never ambition. Saying what we will not do is part of the methodology.
- U.S. House races — no live House coverage on night one. A self-sourced desk cannot responsibly cover 435 districts across heterogeneous feeds in its first cycle; every district page links to its state’s official results, and certified results backfill into the atlas after the canvass.
- States outside the night-one set — where we cannot self-source a feed we can verify, we link to the official Secretary of State results rather than relay numbers we cannot stand behind.
- Ranked-choice contests (Maine, Alaska) — first-choice tallies only, labeled uncallable; reallocation can overturn a first-choice plurality, so no auto-call is possible by construction.
- Probabilistic projections — none in 2026, anywhere. Calls are arithmetic or they are not made.
Other states report on a coverage ladder by feed quality: where we can self-source a clean feed we show live numbers; otherwise we link to the official source and backfill certified results.
What we will not report live, and why
Akashic reports live results for a deliberately narrow set of contests. Everything outside it is declined on the record, with a reason, rather than left to look like an outage on the night. These are the standing exclusions for 2026.
- Akashic does not report U.S. House results live in 2026. House races are counted at a district level that most state feeds do not publish on election night, and a district that crosses county lines cannot be assembled from county totals without guessing. Rather than show a number we would have to estimate, we show nothing and link to the state's own returns. House coverage follows feed quality, not ambition.
- Akashic reports live results for a deliberately small set of states in 2026 — the ones whose official feeds we have fetched, parsed and reconciled against a real prior election. For every other Senate and Governor race we link to the state's own official returns instead. A state is added when its feed is proven, not when the race becomes interesting.
A declined contest still gets its certified result once the state publishes one. Declining to report a number live is not declining to report it.
Win probability — what it is, and what it is not
On a contest we are reporting live, Akashic may show a win probability beside the vote count. It is a model estimate of how the remaining vote is likely to fall, given what has been counted, where it came from, and how comparable places have voted before. It moves as votes arrive and it can reverse.
It is not a call and it is not a projection. Akashic’s call — the label that says a race is decided — remains math-decisive only: it appears when the votes still outstanding cannot change the outcome, never because a model is confident. A formal projection is a separate, stronger claim that requires a named human to approve it on the record. A number in the nineties is still a probability, not a result.
A probability is published only while every publication gate passes: validated results have arrived, the official feed is healthy and fresh, a grounded candidate-level prior exists, evidence capture and ledger persistence are complete, minimum counted-vote and reporting-unit diversity thresholds are met, and no correction cooldown, unresolved high-leverage precinct, ballot-method skew, source suppression, or contest freeze is outstanding. If any gate closes, the number is withdrawn rather than frozen — a missing probability means a gate is closed, not that the race stopped moving.
Each probability is shown with a 95% lower bound and the model and policy versions that produced it, so any number on the page can be traced to the code that made it.