Name and address variants that survive normalization

5,410 pairs of recipient renderings filed at the same street and ZIP, one of which the cascade resolved (YES) and one of which it did not (MAYBE). Drawn from the 50,000-case triangulation run over population cb3c78ff0f1a9546, restricted to organization recipients.

Usage

pf_variants

Format

A data frame with 5,410 rows and 57 columns. a_* is the anchor — the best-scoring YES rendering at the address. b_* is the candidate that came back MAYBE.

pair_id, group_id, addr_key: Identifiers. addr_key is the uppercased-alphanumeric street plus ZIP5 that defines the group.
curated: TRUE for 155 exemplars, up to six per name_class x addr_class cell, chosen deterministically by anchor score. Use these for examples and tests; use the full table for regression.
a_name, b_name: The names as filed.
a_name_key, b_name_key, a_dba_key, b_dba_key: The normalized keys the cascade blocks on, from np_normalize().
a_street, b_street, a_street_key, b_street_key: Street as filed, and as normalized.
a_unit, b_unit, a_city, b_city, a_state, b_state, a_zip, b_zip, a_zip5, b_zip5: The rest of the address, filed and derived.
a_ein, a_bmf: The anchor's matched EIN and the BMF name behind it. b_ein is the candidate's own EIN where it has one.
a_score, b_score, a_pass, b_pass: Score and winning pass for each side. a_pass_rank and b_pass_rank grade the pass: 6 for an exact name + ZIP5 hit down to 1 for an unblocked token sweep.
a_grants, b_grants: Grant rows behind each identity.
name_class: What separates the names: identical normalized, spacing or punctuation, leading THE, corporate suffix, and vs ampersand, abbreviation, truncation, token added or dropped, typo, partial token overlap, or different name. and vs ampersand is empty in this build: np_normalize() resolves & to AND upstream, so what survives is an ampersand that was deleted rather than spelled out, which lands in spacing or punctuation.
addr_class: What separates the addresses: identical, zip4 differs, city differs, unit differs, street spelling, or state differs.
name_lev, name_jaccard, shared_distinctive, is_initialism: Edit distance, token Jaccard, count of shared tokens that are not generic org words, and whether one name is an initialism of the other.
street_differs, unit_differs, city_differs, zip4_differs, state_differs: Whether the filed text differs, before normalizing.
name_key_equal, street_key_equal, zip5_equal: Whether the normalized keys agree. Pair street_differs with street_key_equal to see what the street normalizer is buying.
blocking_intact: TRUE when ZIP5 agrees and the state does not disagree, so every blocked pass could still fire. FALSE on 20 rows.
b_outscores_a: TRUE on 726 rows where the candidate's own match beat the anchor's, by score or by a stronger pass. On these the silver label is running against the system's own evidence and should be read before it is trusted.
n_names_at_addr: Distinct renderings at the address.
expected_same_org, label_basis, label_confidence: The label, and how it was arrived at. See "What it does not contain".

Source

The triangulation run in dev/build-triangulation-v2.R, over the deduplicated e-file corpus (efilepf_v3_1, 17,253,199 rows). The normalizer attribute records the np_normalize_signature() the keys were computed under; if it no longer matches, the keys are stale.

Why these pairs are the interesting ones

Every pair here is a difference that survived normalization. np_normalize() already strips a leading THE, rewrites & to AND, expands FDN to FOUNDATION and contracts USPS street suffixes, and pf_dedup() then collapses renderings that share an identity key. So nothing in this table differs only in a way the pipeline already handles: these are the cases that reached the comparator and split. That makes the table a regression corpus. Tighten the normalizer and specific documented pairs should move; the pairs recorded as already closed should stay closed.

What it does not contain

Ground truth. The triangulation verdicts are silver, confidence = "low", and unadjudicated — assigning the anchor's EIN to every other rendering at the address is the assumption under test, not a fact. expected_same_org is therefore TRUE on only 20 rows, where the two names are the same string under a documented cosmetic transformation at an identical address, and NA — never FALSE — on the other 5,390. A pair sharing no tokens may still be one organization (US OLYMPIC AND PARALYMPIC FOUNDATION / USOC); a pair sharing a distinctive token may still be two (MARINE CORPS UNIVERSITY FOUNDATION / MARINE CORP SCHOLARSHIP FUND). label_basis records which rows are waiting for a human.