Election night: a live results show built in a day, and the 40 seconds it got wrong
A local news outlet streamed election night with live results from the official county count, a rotating cast of hosts, and candidates joining from home. Nine AI sessions ran the back of house. Here is what worked, the one wrong number that aired, and what changed the next morning.
The problem
Election night is the most-watched night of the year for local news, and the hardest one to fake. Viewers want the numbers the moment they move. Candidates want to be treated the same. And one wrong number on screen is not a typo, it is the story people remember.
The outlet had a primary-night dashboard from the spring. It was a screen share of a rotating chart. For the general election the goal was a real results show: a scene for every contested race, hosts on camera, candidates dropping in live, and audience comments on screen. The build window was one evening and the next morning.
What the system does
One source of truth. The state's results site started refusing scripted requests the day before. We did not work around the block. The county clerk publishes its own unofficial results page, openly, so that became the feed. A small script read it every 60 seconds and wrote one results file. Every number on air came from the latest clean read of that page, stamped with the clerk's own run time.
Stale beats wrong. If a read failed or the page changed shape, the screen said "results paused" with the time of the last good update. It never showed zeros and never filled a gap with an estimate. When the state's site went down for about an hour and forty minutes mid-show, the race cards that used it fell back on their own to a "county votes only" label. Nobody had to touch anything.
Scenes from data, not by hand. The clerk page held 49 races. The build produced a scene for each of the 24 contested ones plus summary boards, more than 50 scenes in all. Every candidate card was identical: ballot name, party, votes, percent, bar. "Leading" meant most votes in the current count, nothing more. No race was called on air.
One command per change. When a co-host sat down, one command switched the host layout inside every race scene at once. Nobody rebuilt 24 scenes mid-show.
People in the loop where it counts. Going live was a human click. Featured comments went on screen only when a host clicked one. Candidate invites were drafted and sent by a person, with the same wording and time cap for everyone in a race.
Nine AI sessions, one inbox. Separate sessions handled the live rig, the multi-county totals, a forecast view, the guest desk, comments, TikTok prep, a configuration audit, and the website. One master session relayed every question to the producer. More on why that last part mattered below.
What happened
- On air 6:50 PM to 8:31 PM. All 21 polling places were reporting by the 8:05 PM clerk update.
- About 0.3% of frames dropped over the whole stream, with one burst at 8:27 PM that recovered on its own.
- Four candidates joined live.
- The audience posted 165 comments from 84 people. Hosts featured the best ones on screen.
- After the show, the results page on the outlet's website was rebuilt from the clerk's final run and checked race by race: 51 of 51 matched.
The 40 seconds we got wrong
A ballot question on the clerk page was formatted differently from the races. Its heading did not carry the "(Vote for N)" line that every race had. The parser did not see it as a new race, so the question's YES and NO totals were added to the city council race above it.
The parser reported no errors all night. From its point of view, every read succeeded. A configuration-audit session flagged the problem at 4:47 PM. The fix did not land until 8:54 PM, after the show. The bad card was on screen for 40 seconds, from 8:28:10 to 8:28:50 PM.
The flag got lost because it reached the producer as one more message in a busy afternoon, from a session nobody was watching at the time. The fix itself took about three minutes once the producer said "fix it."
What changed the next morning
The next day, a separate session read every log from the night: the full OBS log, the feed log, clerk page snapshots, eight session transcripts, and the comment history. It came back with 12 ranked proposals. The producer approved all but one that afternoon, and most were built and tested the same day.
- The results feed now checks itself on every read. It runs against saved copies of the real clerk pages from that night as test fixtures, and a race that suddenly gains extra rows or candidates fails loudly instead of quietly. 22 tests pass on the production machine.
- Every question from any session goes through the master, with a clock on it. An unanswered ask now escalates instead of sitting in a tab.
- Pre-show checks are getting stricter. Stream outputs are now audited before air. Two more fixes are built and tested offline and still need a live rehearsal: a confirmation step before the stream can be stopped, after an accidental stop ended the main broadcast early, and guest links that let co-hosts hear guests, which they could not on the night.
What this means for your business
Any live, time-boxed event has the same shape: a product launch, an earnings call, a storm-response update, a big sale day. Three rules carry over.
- Pick one source of truth and show its timestamp. If you cannot say where a number came from and when, it does not go on screen.
- A "successful" run can still be wrong. Check that the output makes sense, not only that the job finished. The parser never errored. It was confidently wrong.
- Route every alert through one place a human is actually watching. We had the right warning four hours early. It was in the wrong window.
The outlet and the candidates are not named. The workflow and the numbers are real. If you are planning a live event and want automation you can trust on the night, get in touch.