Matching methodology
VeroFlag separates candidate retrieval, deterministic scoring, and reviewer decision-making. The matching layer explains why a record appeared, while the reviewer records the final outcome.
Ingestion and query paths use the same normalizer. It applies Unicode NFKC normalization, lowercases names, folds Latin diacritics such as Müller to muller, preserves native-script letters such as Cyrillic, Arabic, and CJK, strips punctuation, and collapses whitespace.
Exact normalized names, token-order keys, surname gates, trigram search, and short edit-distance checks generate candidates. Weak aliases are not normal match drivers unless another signal supports review.
Scores are deterministic and explainable. Country and date-of-birth support can raise confidence, while a date-of-birth year conflict applies a visible penalty and reviewer signal instead of silently hiding a candidate.
The internal evaluation harness runs against the current promoted source graph from CI for matching-related changes. It requires at least 30 seeded cases and fails the workflow if live recall or precision drops below the configured gate.
Minimum cases
30
CI recall gate
>= 0.90
CI precision gate
>= 0.90
Each candidate carries the matched source name, source list, source URLs, match signals, score, confidence band, supporting country or date-of-birth signals, and any date-of-birth conflict flag.
A strong match is still only a candidate until a reviewer confirms it, marks it not relevant, or leaves it pending. The case decision remains a separate human-controlled workflow step.
Native-script names are matched when the source data already includes the same script. Cross-script transliteration is a future enhancement and is not implied by the current normalizer.