How we verify election information
Strata-AF™ Elections keeps one verified record for each contest. The calendar, dossiers, updates and results all read from that record, and every fact in it carries its own source and the time it was last checked.
What each label means
| Official | Published by the legally responsible body. |
| Verified | Corroborated by two independent sources. |
| Expected | Inferred from the constitutional timetable. |
| Disputed | Credible sources conflict. |
| Claimed | Asserted by an interested party, not verified. |
| Provisional | Announced, but not yet legally final. |
| Final | Certified by the legally competent body. |
Verified means two independent sources. Two outlets repeating the same wire story count as one.
Source hierarchy
| Tier | Source | Examples | Used for |
|---|---|---|---|
| 1 | National authority | Electoral commission, constitutional court, official gazette, parliament, presidency | Authoritative for dates, eligibility, candidate lists and certified results |
| 2 | Continental and regional bodies | African Union, ECOWAS, SADC, EAC, IGAD observation missions | Process assessments and observation findings |
| 3 | Specialist datasets | EISA, IFES ElectionGuide, International IDEA, ACE Electoral Knowledge Network | Discovery, history and electoral-system data |
| 4 | Independent reporting | Credible national media, civil-society monitors, international reporting | Context and dispute flags; never silently overrides tier 1 |
| 5 | Interested parties | Candidate posts, party statements, eyewitness material | Treated as claims until independently verified |
Aggregators help us discover what is coming. Nothing is published from them automatically; an editor confirms each fact against a primary source.
How often we re-check
| Stage | Re-verified |
|---|---|
| Expected | Every 30 days |
| Announced, campaign | Every 7 days; every 3 days in the final month |
| Voting, counting | Every 4–6 hours |
| Contested | Every 24 hours |
| Certified, transition | Every 14 days |
When a dossier passes its check window, it shows a Check overdue label until an editor confirms the record again. We show this on purpose.
Two editors for anything final
Any change that makes a value official or final, names a winner, or changes a result figure needs a second editor’s approval. The database enforces this rule; no single editor can bypass it.
What we do not do
- We do not give elections a “free and fair” score. We document indicators, such as candidate exclusions, internet disruptions and observer access, and we attribute every assessment.
- We do not publish unofficial tallies. Party and candidate totals appear only as attributed claims. Projections are kept apart from results and appear only on the Projection tab.
- We do not run a minute-by-minute live blog. During voting we publish a slower, verified update stream.
The Projection tab
For selected contests (Nigeria 2027, Cabo Verde 2026, The Gambia 2026, Kenya 2027 and Angola 2027), a Projection tab sits between Campaign and Vote & count. Each is built from that country's last two presidential elections, region by region, and uses that country's own winning rule. It is kept apart from results and is always shown with its uncertainty.
- The starting record. Every region's results in the last two presidential elections (states in Nigeria, counties in Kenya, provinces in Angola, administrative areas in The Gambia, electoral circles in Cabo Verde), with registration and turnout, each figure linked to the electoral authority's declaration or, where that was not available, the best published compilation, which the tab names.
- Weekly evidence. Editors review tagged evidence, such as defections, rallies, endorsements, observer reports and economic shocks. Each piece is recorded as one of six forces (strength, conversion, exasperation, reassurance, resilience, apathy) for a named candidate in a named state, with a source.
- The books must balance. Every point a candidate loses in a state goes to a named rival or to voters who will not vote. If a state does not balance, that week is held.
- Monday Movement Report. What moved this week, where, and why. It makes no win claim.
- Friday Forecast. The election simulated 10,000 times under the country's real rule. Nigeria: the most votes and at least 25% in 25 of the 37 jurisdictions (we also test the reading 24 of 36 states and the FCT). Kenya: more than half the vote and at least 25% in 24 of the 47 counties. Cabo Verde: more than half the valid votes. The Gambia: the most votes, one round. Angola: the party list with the most votes. We show each chance with an 80% range, the chance of a run-off where the rule has one, and each condition separately.
- When we hold back. If uncertainty is too high, a state is shown as too uncertain to call, and the national headline is withheld. We name a likely winner only when one candidate wins at least 60% of simulations.
- Our own readers do not count. Strata-AF™ readers' activity is shown on Mondays as readers, not voters, and never moves the forecast.
- Before launch. No forecast is published until the model has been tested against the country's two previous elections (run from the earlier one, scored against the later one) and an admin has signed off that test. The result of each test is shown on the tab. After that, each weekly release is published either by an editor or by sAI when every guardrail passes (see Run by sAI below).
Run by sAI
sAI is Strata-AF™'s election engine. It covers every presidential contest we follow across Africa: it reads the news, files campaign events, checks each one against its source, and, where a contest has a projection model, turns verified evidence into the weekly movement and forecast. Strata-AF™ editors set the rules it works to and audit its work.
- How sAI checks a claim. It re-reads the source page and asks whether the page reports the claim. It must copy the supporting sentence, and that sentence must really be in the page. It checks that the candidate, the place and the date match, looks for other publishers reporting the same thing, and looks for duplicates already counted. Large claims, such as a governor defecting, need independent reporting.
- What sAI may do on its own. Each contest has its own level, from checking only (editors decide everything) to publishing routine campaign events and the weekly reports. Defections, endorsements, controversies and anything sAI cannot verify always go to an editor.
- Guardrails. A weekly report is held for an editor if it fails any check: the numbers must balance, enough evidence must be behind it, no single story may drive most of the week, and no state may swing implausibly. A forecast also needs a signed-off backtest and a published Monday report.
- Earned trust. sAI publishes nothing on its own until editors have reviewed enough of its calls, over enough weeks, and agreed with almost all of them.
- Who published it. Every weekly report says whether sAI or an editor published it. Corrections and retractions are listed on the Projection tab.
Corrections
No fact is edited in place. Every change is a new, sourced entry, and the previous value stays visible in the corrections log. To report an error, email legal@strata-af.com with the page and the source.