Not official results. Only INEC declares election results in Nigeria. Binchika is a citizen transparency layer. No party ads · equal space for all 22 parties · our rules
--:--:--Abuja · WAT (UTC+1)

AI data governance

AI suggests. Rules and people decide.

Nothing an AI produces reaches the public ledger without passing the same published rules as everything else, and every correction needs the human BinchikAdmin.

Where AI is used

LayerAI taskOutput is…
IntakePhoto quality check (blur, glare, cropping)a hint to retake the photo
ReadingReading party totals from Form EC8Aa proposed tally with a confidence score per field
Re-investigationRe-reading every photo for a flagged unita summary for the OpsAdmin, with no authority
ManifestosSummaries with page and paragraph citationsa draft, reviewed before publishing
TranslationHausa, Yoruba, Igbo and Pidgin guidancea draft, reviewed by native speakers

Never used for: consensus, comparison with IReV, baseline checks, ledger writes, verifier selection, party order or colours, news selection, or saying who is "winning".

Limits we design around

  • Handwriting, low light, folds and stamps cause misreads, so every low-confidence field goes to a person.
  • AI cannot spot a doctored sheet on its own. Independent reporters agreeing, comparison with IReV and the pre-election baselines catch fraud.
  • Language models can invent text. Summaries must quote and cite, and any unsupported sentence is removed.
  • Error rates are monitored by state; if one region misreads more, it goes to people.

The AI constitution

  1. Suggest, never decide.
  2. Label everything. Every AI output is marked and logged with the model, version, input hash and time.
  3. No personal data. Hashed handles, stripped photo metadata and blurred faces before any model sees an image. Nothing identifiable goes to outside AI services.
  4. Rules beat models. If they disagree, the case goes to a person.
  5. Say how sure it is. Below 0.9 confidence per field, a person confirms.
  6. No politics. No predictions, no persuasion, and the same pipeline for every party.
  7. Auditable. Anyone can ask to re-run an AI step on the public evidence.
  8. Least power. AI writes only to a review queue, never to the ledger or the results.
  9. Kill switch. Any AI step can be turned off and the system keeps working with people.
  10. Corrections are human. Only the BinchikAdmin approves an overwrite, and it becomes a new public ledger entry.

Human checkpoints

StepAI?Human checkpoint
Photo quality hintYesNone (advisory only)
Reading the tallyProposesRequired below 0.9; a random 5% sample above it
Consensus & ledger entryNoPublic rules, in code
Discrepancy flagNoSZV raises it, NRV reviews it
Re-investigationSummarisesOpsAdmin reviews
CorrectionNoBinchikAdmin, always