Beneficiary de-duplication

Names were transliterated inconsistently, dates of birth were often approximate, and no shared identifier existed across partners. Exact matching found almost nothing; loose matching produced false positives that would have wrongly excluded real households.

A probabilistic matching service scores candidate pairs across name, date of birth, location and household composition, then routes anything in the uncertain band to a human reviewer rather than deciding automatically.

Exclusion is never automatic. The service flags; a caseworker decides.

Reach figures became defensible, and the review queue gave partners a shared, auditable record of why a household was or was not treated as a duplicate.