Name and address variants

How recipient names and addresses actually vary in the 990-PF filings, which of those variations the pipeline already absorbs, and which ones decide whether a grant resolves to an EIN.

Everything here is computed from pf_variants: 5,410 pairs of recipient renderings filed at the same street and ZIP, across 4,291 addresses. In each pair the cascade resolved one rendering (YES) and returned MAYBE on the other. They come from the 50,000-case triangulation run over population cb3c78ff0f1a9546.

Read this first: these pairs survived normalization

The pairs are not a random sample of variation. They are the residue.

Before any pass runs, npmatch::np_normalize() has already uppercased, dropped a leading THE, rewritten & as AND, expanded FDN to FOUNDATION and contracted USPS street suffixes; pf_dedup() has then collapsed renderings sharing an identity key into a single recipient. So a pair that differs only in a way the pipeline handles never reaches the comparator as two things and never appears here.

What that buys: the table is a regression corpus. Every row is a live case that reached the comparator and split. Tighten the normalizer and specific rows should move; the rows recorded as already closed should stay closed.

What it costs: pf_variants tells you nothing about variation the normalizer already fixes. For that, compare the *_differs columns (the filed text) against the *_key_equal columns (what the normalizer made of it).

	pairs	share
ZIP5 agrees inside the pair	5,410	100.0%
filed street text differs	88	1.6%
…but the normalized street key agrees	9	10.2%
normalized name keys agree anyway	46	0.9%
blocking intact (ZIP5 and state both usable)	5,390	99.6%

ZIP5 agrees in every pair because the group is defined by street-plus-ZIP5, which is also why this table cannot be used to study ZIP variation.

It is not ground truth

The triangulation verdicts are silver, confidence = "low", and unadjudicated. Writing the anchor’s EIN onto every other rendering at the address is the assumption under test, not a fact.

So expected_same_org is TRUE on 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. Two examples of why a string test cannot decide it:

• US OLYMPIC AND PARALYMPIC FOUNDATION and USOC share no tokens and are the same family of organization.
• MARINE CORPS UNIVERSITY FOUNDATION and MARINE CORP SCHOLARSHIP FUND share a distinctive token and are two different charities.

label_basis records which rows are waiting for a human.

Name variants

class	pairs	share	median edit distance	median token Jaccard	name keys agree
different name	4,080	75.4%	23.0	0.00	1
token added or dropped	595	11.0%	15.0	0.50	1
truncation	428	7.9%	19.5	0.50	23
partial token overlap	274	5.1%	12.0	0.50	4
leading THE	14	0.3%	4.0	0.75	14
typo	8	0.1%	1.0	0.60	0
abbreviation	4	0.1%	7.0	0.60	3
spacing or punctuation	4	0.1%	1.0	0.49	0
corporate suffix	2	0.0%	4.0	0.50	0
identical normalized	1	0.0%	0.0	1.00	0

The cosmetic classes everyone expects to dominate are almost absent, and that is the normalizer working: it has already closed them upstream. What is left is dominated by names that are genuinely different strings.

name keys agree is the column to watch. Where it is non-zero the normalizer did close the name difference and the two renderings still reached the comparator separately — covered in “A seam” below.

Exemplars

Six per class, from the curated subset.

different name

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
ALLIED ARTS COUNCIL OF ST JOSEPH	ARTS FUND	FALSE	25	0.143	0.98	0.720
DANIELLES ANIMAL SAFE HAVEN INC	ANIMAL RESCUE	FALSE	22	0.167	0.98	0.720
JEWISH COMMUNITY FEDERATION - EAST BAY	BERKLEY HIGH HOLIDAYS SERVICES	FALSE	31	0.000	0.98	0.780
ALLIANCE FRANCAISE DE MILWAUKEE	BIG BROTHERS BIG SISTERS	FALSE	25	0.000	0.98	0.720

token added or dropped

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
BOY SCOUTS OF AMERICA	GNFC - Boy Scouts of America	FALSE	5	0.800	0.98	0.760
COMO FRIENDS	COMO ZOO AND CONSERVATORY SOCIETY - COMO FRIENDS	FALSE	34	0.333	0.98	0.780
COVER HOME REPAIR	COVER	FALSE	12	0.333	0.98	0.720
The Humane Society of Broward County	BROWARD HUMANE SOCIETY	FALSE	25	0.500	0.98	0.751

truncation

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
CORNERSTONE CHRISTIAN SCHOOL AND EDUCATIONAL SERVICES	CORNERSTONE CHRISTIAN SCHOOL	FALSE	25	0.500	0.98	0.720
FRIENDS OF YOUTH SERVICES AND PALM BEACH COUNTY INC	FRIENDS OF YOUTH SERVICES	FALSE	26	0.444	0.98	0.720
FAITH CHURCH ST LOUISCOM INC	FAITH CHURCH	FALSE	16	0.400	0.98	0.720
COLLEGE BOUND ST LOUIS	College Bound	FALSE	9	0.500	0.98	0.720

partial token overlap

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
MOON TOWNSHIP LIBRARY FRIENDS	MOON TOWNSHIP PUBLIC LIBRARY	FALSE	13	0.600	0.98	0.980
DELTA PHI EPSILON EDUCATIONAL FOUNDATION	DELTA UPSILON EDUCATIONAL FOUNDATION	FALSE	5	0.500	0.98	0.720
A BETTER WISCONSIN INSTITUTE	A BETTER WISCONSIN TOGETHER	FALSE	8	0.600	0.98	0.980
ALLIED ARTS COUNCIL OF ST JOSEPH	ALLIES ARTS OF ST JOSEPH	FALSE	9	0.571	0.98	0.760

leading THE

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
The Pasadena Educational Foundation	Pasadena Educational Foundation	TRUE	4	0.750	0.98	0.720
THE JACKSONVILLE HUMANE SOCIETY	JACKSONVILLE HUMANE SOCIETY	TRUE	4	0.750	0.98	0.739
THE WESTERN PENNSYLVANIA CONSERVANCY	WESTERN PENNSYLVANIA CONSERVANCY	TRUE	4	0.750	0.98	0.759
THE ONESTAR FOUNDATION	Onestar Foundation	TRUE	4	0.667	0.98	0.720

typo

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
MUSEUM OF THE AMERICAN REVOLUTION	MUESEUM OF THE AMERICAN REVOLUTION	FALSE	1	0.667	0.98	0.760
YOUNG WOMENS CHRISTIAN ASSO OF MISSOULA	YOUNG WOMENS CHRISTIAN ASSOC OF MISSOULA	FALSE	1	0.714	0.98	0.780
LONG BEACH COMMUNITY FOUNDATION	LONG BEACH COMMUNITY FOUNDATON	FALSE	1	0.600	0.98	0.765
YESHIVATH BETH YEHUDAH	Yeshiva Beth Yehudah	FALSE	2	0.500	0.96	0.960

abbreviation

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
COMMUNITY FOUNDATION FOR LOUDOUN	COMMUNITY FDN FOR LOUDOUN	TRUE	7	0.600	0.980	0.780
MEDICAL FOUNDATION OF NASHVILLE	MEDICAL FDN OF NASHVILLE	TRUE	7	0.600	0.960	0.780
WHITE PLAINS HOSPITAL FDN	WHITE PLAINS HOSPITAL FOUNDATION	TRUE	7	0.600	0.960	0.733
LOS GATOS HIGH NEW MILLENNIUM FOUNDATION	LOS GATOS HIGH NEW MILLENNIUM FOUND	FALSE	5	0.714	0.888	0.780

spacing or punctuation

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
BOYS & GIRLS CLUB OF WASHINGTON COUNTY	BoysGirls Club of Washington County	FALSE	1	0.571	0.980	0.720
US NAVAL ACADEMY	U S NAVAL ACADEMY	FALSE	1	0.400	0.960	0.902
US Marshalls Museum	U S Marshalls Museum	FALSE	1	0.400	0.949	0.760
HEALTH CARE FOUNDATION OF GREATER KANSAS CITY	HEALTHCARE FOUNDATION OF GREATER KANSAS CITY	FALSE	1	0.625	0.900	0.780

corporate suffix

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
BRAC USA	BRAC	FALSE	4	0.5	0.96	0.720
CARE	CARE USA	FALSE	4	0.5	0.82	0.720

identical normalized

resolved (YES)	unresolved (MAYBE)	name key agrees	edit dist	jaccard	YES score	MAYBE score
BOYS & GIRLS CLUBS OF MANATEE COUNTY INC	BOYS GIRLS CLUBS OF MANATEE COUNTY INC	FALSE	0	1	0.98	0.780

Address variants

class	pairs	share	blocking still intact
identical	4,278	79.1%	100.0%
zip4 differs	534	9.9%	99.6%
city differs	285	5.3%	99.3%
unit differs	229	4.2%	98.3%
street spelling	72	1.3%	100.0%
state differs	12	0.2%	0.0%

The address is scored, but inside a group it is saturated

Blocking and scoring are separate layers and the address behaves differently in each. pf_staged_passes() blocks on name_key or dba_key crossed with zip5 or state, and nothing else:

pass	criteria
A1-name-zip5	name + ZIP5 (exact)
A2-name-state	name + state (exact)
A3-name-dba	name == dba, same state
A4-dba-name	dba == name, same state
A5-dba-dba	dba == dba, same state
B1-token-state	shared token, same state
B2-token-concat	shared token + compound, same state
C1-name-any	name == name, ANY state
C2-dba-any	name == dba, ANY state

So no pass blocks on street, unit or city. But scoring is a different layer, and there the address carries real weight. pfmatch runs method = "hier", whose score is

score = 0.6 * name_key_similarity + 0.4 * geo_score

and geo_score is the maximum over a tier ladder:

geo term	weight	fires when
zip9	1.00	ZIP+4 agrees
unit	0.97	the suite agrees, where it discriminates (a multi-tenant building, not a crowded mail drop)
street	0.95	street number agrees AND normalized street key similarity >= 0.9, not a PO box
zip5	0.90	ZIP5 agrees
pobox	0.90	PO box number agrees
zip3	0.55	ZIP3 agrees (zeroed on a state conflict)
city	0.50	city Jaro-Winkler, floored at jw_threshold (zeroed on a state conflict)
state	0.20	state agrees

street_key — npmatch’s normalized street — is compared directly and is worth 0.95, behind only ZIP+4 and the suite. The address normalizer is very much in use.

The catch is the max(). Inside a triangulation group ZIP5 agrees by construction, so geo_score is already at least 0.90 before the street is considered, and a street hit can lift it only to 0.95. That is 0.05 of geo, 0.02 of the final score. City at 0.50 and state at 0.20 are dominated outright and cannot bind at all while ZIP5 agrees.

So the address is not inert — it is saturated, and only in this corpus. Outside it, wherever ZIP5 does not agree, street_hit is the best geo term available and street normalization is carrying the match. That is the opposite of what these pairs alone would suggest, and it is why pf_variants is the wrong instrument for judging the address comparator’s value.

Consequences that do hold here:

• A city spelling difference — FT LAUDERDALE against Fort Lauderdale — cannot change the outcome while ZIP5 agrees, because the 0.50 city term is dominated by the 0.90 ZIP5 term. There are 290 such pairs, and blocking stays intact in all of them.
• A suite line could not separate two organizations in one building when these pairs were cut: .np_hier_score() had no unit term, and the street_unit_full weight profile that does carry one needs 50% unit coverage against an actual 3.84%. npmatch has since added the unit rung above — so the mechanism now exists, gated on the suite discriminating at that address rather than merely being present. The pairs in this table predate it, which is why pf_unit_evidence() exists here at all. Its own parser has since been retired in favour of npmatch’s; see dev/UNIT-UPSTREAM-SYNC.md.
• The state costs twice when wrong: it blocks stages A2–A5 and all of stage B, and a state conflict zeroes the city and ZIP3 geo terms. It disagrees in 20 pairs, all adjacent-letter filing errors:

resolved (YES)	state	unresolved (MAYBE)	state
AICE	MO	JDC	MD
UK PHILANTHROPY	NY	UNIVERSITY OF KENTUCKY	KY
FEATHER RIVER COLLEGE FOUNDATION	WA	FEATHER RIVER COLLEGE - ELLEN HOLLIFIELD SCHOLARSHIP	CA
ASSOCIATION OF BALTIMORE AREA GRANTMAKERS	MD	BALTIMORE COMMUNITY FOUNDATION INC	MA
FREE ARTS FOR ABUSED CHILDREN OF NEW YORK CITY INC	NY	MOTHERS2MOTHERS INTERNATIONAL INC	PA
NEW VENTURE FUND	DC	HOPE & HEAL FUND	CA

Since ZIP5 is present and agrees in every one of these pairs, a ZIP5-to-state repair would fix them outright.

Where street normalization does pay

Not inside a group — by construction every pair already keyed to the same street. It pays in forming groups at all. The filed street text differs in 88 pairs here, and the normalizer reconciles 9 of them:

filed (YES side)	filed (MAYBE side)	normalized key
4550 WEST 77TH STREETSUITE 200	4550 West 77th Street Suite 200	4550 W 77TH ST
1360 MISSION STSUITE 300	1360 MISSION ST SUITE 300	1360 MISSION ST
506 2ND AVESUITE 1400	506 2ND AVE SUITE 1400	506 2ND AVE
620 MOUNTAIN VILLAGE BLVD STE 2B	620 MOUNTAIN VILLAGE BLVDSTE 2B	620 MOUNTAIN VILLAGE BLVD
6321 Porter RoadSuite 8	6321 PORTER ROAD SUITE 8	6321 PORTER RD
8 S MICHIGAN AVENUE SUITE 500	8 S MICHIGAN AVENUESUITE 500	8 S MICHIGAN AVE

A missing space before SUITE, Street against ST, ROAD against RD. Measured over the whole corpus rather than these pairs, normalizing the street collapses 528,862 of 2,326,459 distinct address keys — 22.7% — and brings 58,815 addresses above the two-rendering threshold that are currently split across spellings and so form no group at all.

The caution: a normalizer that drops the secondary-unit designator to achieve that will also merge STE 100 with STE 200, which are different tenants. Distinct units are rare at these addresses, so the trade is probably worth making, but it is a trade.

The suite line as an adjudication tie-breaker

The one address fact that bears directly on a triangulated label is the door. Two renderings at one street giving different suites are very likely different organizations, so the anchor’s EIN should not be written onto the candidate. When these pairs were cut the matcher could not use this at all; npmatch has since given the hierarchical scorer a unit rung (0.97) and annotates every candidate pair with geo_unit, so the mechanism now exists there too. A reviewer still needs the four-way reading below, which geo_unit does not give: it cannot separate “one side filed nothing” from “neither did”.

pf_unit_evidence() compares the two lines. The parse is npmatch’s — pf_street_unit() hands line 2 to np_normalize() and reads street_unit back, so pfmatch reads a suite the same way the matcher does. What is still pfmatch’s is the two rejections the field needs, because it is not a unit field: it holds street continuations, city names, PO boxes, and in one case a sequential counter written once per grantee. The verdicts tabulated below were recorded under the normalizer the corpus was built with; retiring the local parser moved 22 of the 5,410 pairs, every one traced in dev/UNIT-UPSTREAM-SYNC.md.

evidence	pairs	share	how to read it
absent	5266	97.3%	neither side filed a parseable unit
one_sided	108	2.0%	one side did; silence is not disagreement
agree	27	0.5%	same door — NOT evidence of one organization, see below
conflict	9	0.2%	different doors — the label is probably wrong

anchor (YES)	unit	candidate (MAYBE)	unit
CITY GROWERS	1105	PRATT CENTER FOR COMMUNITY DEVELOP	701
Lankenau Medical Center Foundation	340	MAIN LINE HOSPICE	100
COURAGE WORLDWIDE	433	RRUFF - ROCKLIN RESIDENTS UNITE FO	125
AUTISM ALLIANCE OF MICHIGAN	520	WOMEN OF TOMORROW	155
BALTIMORE ARCHITECTURE FOUNDATION	STREET SUITE P-101	JOHNS HOPKINS UNIVERSITYJOHNS HOPK	STREET SUITE 234
LEUKEMIA & LYMPHOMA SOC	SUITE 115	KORE HAITI INC	SUITE 100
WEIMARANER RESCUE OF NORTH TEXAS	184	THOMAS JEFFERSON ALUMNI ASSOCIATIO	150
MCCOOK ECONOMIC DEVELOPMENT CORP	STE 301	MCCOOK COMMUNITY FOUNDATION	RM 315
CITY GROWERS	1105	DIEU DONN	602

Two cautions, both measured rather than assumed.

agree is not support. Generic suites are shared by unrelated tenants — MIDDLE EAST CHILDRENS ALLIANCE and IRISH REPERTORY THEATER are both at Suite 100, GLOBAL FUND FOR WOMEN and THAKKER CHARITABLE TRUST both at Suite 400 — a fiscal sponsor’s suite is shared by everything it sponsors, and some values are not units at all (15219 is a Pittsburgh ZIP code on line 2). It is recorded for completeness and carries nothing.

conflict concentrates. Over the whole triangulation run it fires on 69 of 14,790 anchor/candidate pairs (0.47%), but one address produced 31 of them: a fellowship program writing 291 BROADWAY STE 1901, STE 1902, … one per individual grantee. Always read the raw columns, never the flag alone. In the adjudication queue the flag lands on 121 of 58,680 rows, 26 of which sit in the zero-flag band — cases that looked clean until the address was read.

What decides YES against MAYBE

side	median score	exact name+ZIP5	token-blocked
resolved (YES)	0.96	56.0%	19.8%
unresolved (MAYBE)	0.78	9.1%	53.2%

A seam: identity is keyed on one thing, blocking on another

pf_dedup() keys identity on name_full — deliberately, so that an organization is not merged with its affiliate by a name_key that has already thrown the affiliate suffix away. The cascade then blocks stage A1 on name_key crossed with zip5. The two disagree, and 46 pairs land in the gap: same normalized name key, same street key, same ZIP5, two identities, two different tiers.

resolved (YES)	unresolved (MAYBE)	shared name key	YES score	MAYBE score
COMMUNITY FOUNDATION FOR LOUDOUN	COMMUNITY FDN FOR LOUDOUN	COMMUNITY FOUNDATION FOR LOUDOUN	0.98	0.780
CHILDREN’S HOME SOCIETY	CHILDRENS HOME	CHILDRENS HOME	0.98	0.680
The Pasadena Educational Foundation	Pasadena Educational Foundation	PASADENA EDUCATIONAL	0.98	0.720
THE JACKSONVILLE HUMANE SOCIETY	JACKSONVILLE HUMANE SOCIETY	JACKSONVILLE HUMANE	0.98	0.739
THE WESTERN PENNSYLVANIA CONSERVANCY	WESTERN PENNSYLVANIA CONSERVANCY	WESTERN PENNSYLVANIA CONSERVANCY	0.98	0.759
THE ONESTAR FOUNDATION	Onestar Foundation	ONESTAR	0.98	0.720

When the unresolved side has the better match

The most useful flag in the table. Triangulation writes the anchor’s EIN onto the candidate as correct_ein. On 726 pairs (13.4%) the candidate’s own match was better than the anchor’s — a higher score, or an exact name-plus-ZIP5 hit against a token-blocked one. On those the silver label runs against the system’s own evidence.

anchor (YES)	score	pass	candidate (MAYBE)	score	pass
FRIENDS OF SPL	0.900	D1-acronym	Friends of the Scarsdale Library Inc	0.98	A1-name-zip5
CHILDREN’S HOME SOCIETY	0.900	D1-acronym	BOYS AND GIRLS CLUB OF MIAMI-DADE INC	0.90	A2-name-state
REGENTS OF THE UNIVERSITY OF MINNES	0.900	D1-acronym	UNIVERSITY OF MN FOUNDATION	0.90	A2-name-state
BROWARD ALLIANCE INC (THE)	0.913	B1-token-state	GREATER FORT LAUDERDALE ALLIANCE	0.98	A1-name-zip5
CENTRAL GWINNETT JROTC	0.905	B1-token-state	PARENT BOOSTER USA INC	0.98	A1-name-zip5
PRESTON HOLLOW ELEMENTARY SCHOOL FUND	0.958	B1-token-state	PTA TEXAS CONGRESS	0.98	A1-name-zip5
FRIENDS OF HAYWOOD PUBLIC LIB	0.893	B1-token-state	HAYWOOD COUNTY PUBLIC LIBRARY FOUNDATION INC	0.98	A1-name-zip5
JENSEN BEACH COMMUNITY PLAYGROUND	0.876	B1-token-state	PTA FLORIDA CONGRESS	0.98	A1-name-zip5

BETHEL COLLEGE against BETHEL UNIVERSITY is the shape to recognise: a renamed organization, where the rendering being labelled wrong found the better reference row. Filter on b_outscores_a before trusting a triangulated label.

Using this as a regression corpus

The pairs carry both the raw filed strings and the keys np_normalize() made of them, so the keys can be recomputed and compared. tests/testthat/test-variants.R does exactly that, in two layers:

• Invariants that must hold under any sane normalizer: a leading THE never decides a name key; streets differing only by suffix spelling or a missing space normalize together; streets recorded as genuinely different stay different; a city difference never breaks blocking.
• A pinned check that every recorded key still reproduces, scoped to the np_normalize_signature() the fixture was built under (v1-6dcb298522cc5118). It skips when the normalizer moves, because an intentional improvement should move keys — and the skip message tells you to rebuild with data-raw/make-pf-variants.R.

To test a change, re-run the fixture build against the new normalizer and diff the class table: cosmetic classes should shrink, name_key_equal should grow, and no pair should move from a different-street class into an agreeing one unless you meant it to.