The night the turnout numbers stopped making sense.

Setting the scene

The campaign team for Senator Jane Mwangi - running on the Civic Path coalition ticket in a senatorial race in a Kenyan county we will call Nakuru for this composite - had been on Election Edge for eleven weeks. They had mapped 281 polling stations, recruited 312 agents, and run two full dry-runs against the staging environment. By 18:00 on election day, 246 stations were reporting on schedule. Thirty-five were quiet. That gap was the first thing the manager dashboard surfaced.

The candidate they were watching most carefully was David Kiprono, the Future Forward party challenger, who had built his ground game around three sub-counties where Civic Path historically underperformed. The campaign chief of staff, Aisha Hassan, had decided weeks earlier that the senate race was tight enough that station-by-station tally discipline mattered more than county totals. She wanted every Form 34A confirmed, every variance flagged, every agent reachable.

Election Edge was running in its normal mode: agents submitting via the offline-first mobile app, OCR pulling figures off Form 34A photos, supervisors confirming before any number rolled into the campaign tally. The election anomaly engine was watching for the patterns the team had pre-configured - turnout spikes above the historical band, conflicting submissions for the same station, and stations going dark.

21:47 - the anomaly fires

At 21:47, the anomaly desk lit up. Twelve stations across one sub-county were reporting turnout numbers between 94 and 99 percent. The historical band for that sub-county sat around 68 percent. Two of those twelve stations had two different Form 34A submissions in the supervisor queue - one from the official campaign agent, one from a roving observer the team had added that morning as a backup. The two sets of numbers did not match.

On the old workflow - WhatsApp groups, a shared spreadsheet, a phone tree - this is the moment the campaign loses control. Two numbers, no provenance, no timestamps, no way to ask the agent which photo was theirs without scrolling through 600 messages. By the time anyone has the truth, the figure has been quoted on radio and the party legal team is reacting to numbers nobody on the campaign can defend.

On Election Edge, the dashboard already had what Aisha needed. Each submission was tied to a station ID, an agent ID, a server-side timestamp, and the original Form 34A photo. The supervisor queue showed the variance in red. The map showed the cluster - twelve stations in a tight geographic band, all hitting the same window.

22:10 - the response

Aisha pulled three people into the anomaly room: the field operations lead, the legal desk lead, and one of the senior supervisors. They worked the queue in parallel. The supervisor opened both Form 34A photos for the two contested stations and compared them against the IEBC paper trail the agents had also uploaded. One set of figures - the roving observer's - had a transcription error on the rejected ballots line. The OCR had actually flagged the discrepancy at intake but the variance crossed threshold only when the second submission arrived.

Field ops called the agents in the affected sub-county. Six were reachable. Four were at the tallying centre, two had moved on. The agents who answered confirmed what the corrected numbers showed: real turnout in that sub-county was high but not 94-99 percent high. Three of the twelve stations had a transcription error in the same direction. The remaining nine were genuinely high turnout, but inside a wider plausible band once the model was re-fit against the corrected three.

Legal got a snapshot. The export bundle - station-scoped, with agent identities preserved, with photo hashes, with the audit log of who confirmed what and when - went to the party general counsel at 22:34. It was forty-seven minutes from anomaly fire to a defensible record on the legal team's desk.

01:50 - the outcome

By two in the morning, the campaign tally for Senator Mwangi was stable and reconciled. The senate race in this composite went to a margin tight enough that an unresolved twelve-station variance would have changed the post-election conversation. Because the corrections were made on the live record - with an audit trail showing the original submission, the variance, the supervisor confirmation, and the corrected figure - there was no version-of- events problem the next morning. The numbers the campaign quoted were the numbers in their own archive.

More importantly: when the opposing campaign filed a station- level complaint at the IEBC two weeks later citing the same sub-county, Civic Path's legal team responded with the same export the campaign had already sent on election night. The evidence bundle - RFC 3161 timestamps, photo hashes, agent chain of custody - meant the response was a citation, not a defence.

Takeaways

The point of this scenario is not that the technology was clever. It is that the operational record existed before anyone needed it. Three observations the team made afterwards:

• Variance flags only work if a second source exists. The roving observer programme that surfaced the anomaly cost the campaign almost nothing to run, because Election Edge could accept parallel submissions per station without breaking the tally.
• Supervisors do not scale by adding more supervisors. They scale because the queue is filtered to the things that actually require human judgement. The OCR did the boring work first.
• The legal export is the product. Every operational decision on election night was shaped by the fact that whatever the campaign acted on would be in the bundle that went to court if the race got contested. That discipline was the difference.