Guards on observed names: is the written name really this EIN's?

Schedule I and Schedule R give a (name, EIN) pair written by a third party. Both halves can be wrong, and wrong in systematic ways that corroboration does not screen out, because the same habit repeats across filers. networked990's resolver checks every written EIN against its names (R/resolve.R v1.4-v1.6) and in one rebuild dropped 18,326 own or group-central EINs, found 39,016 written EINs whose names share no word with the mention, and 2,879 EINs written for many unrelated names (fiscal sponsors). Loaded unchecked, every one of those becomes an alias here.

Usage

md_guard_observed(
  con,
  sources = NULL,
  common_min_eins = 2000L,
  sponsor_min_names = 5L,
  sponsor_min_filers = 2L
)

Arguments

con: A connection from md_db().
sources: name.source values to check. NULL checks every observed row.
common_min_eins: A word on more than this many EINs' first-party names is too common to show two names are the same organisation.
sponsor_min_names, sponsor_min_filers: Fiscal-sponsor thresholds.

Value

Per-flag row counts by source, invisibly.

Details

md_guard_observed() flags observed rows rather than dropping them, so the raw observation keeps its provenance and the choice stays with the spec.

Columns written on `name`

flag_placeholder_ein: EIN is a placeholder: nine repeated digits, 123456789, 987654321, or a value below 1,000,000 (EIN-00-0xxxxxx; the ingests zero-pad, so a written 0 lands here).
flag_placeholder_name: Name is no name by npmatch's rule, npnormalize::np_name_placeholder(): SEE SCHEDULE O, SEE ATTACHED, VARIOUS, N/A, NONE, SAME AS ABOVE, blank, or no letters - unless it is consistent with the EIN's own names (SEE for Social and Environmental Entrepreneurs).
flag_own_ein: The only filers who wrote this name for this EIN were the EIN itself, and the name is not consistent with that EIN's names (a grantee or related organisation listed under the filer's EIN). A name other filers also wrote for the EIN is not flagged: on Schedule R that is an affiliate listing itself under a name its siblings use too.
flag_group_central: EIN is a group-ruling central (BMF affiliation code 6 or 8) and the name shares no word with the central's names other than denomination words (a congregation written under the PC(USA) EIN).
flag_multi_name: EIN was written on others' Schedule I for sponsor_min_names+ names inconsistent with its own, by sponsor_min_filers+ filers, and this row is one of those names (a project under Tides Center or Fractured Atlas). The EIN's consistent names keep no flag. entity.multi_name_ein marks the EIN itself. Not only fiscal sponsors meet this: universities (departments, centres), denominational centrals and national federations do too, and their rows are flagged alike - a name unlike the EIN's own is not an alias of it either way. Applies to Schedule R rows on the same EIN as well.
name_ein_consistent: The name shares a distinctive word (3+ characters, on no more than common_min_eins EINs' names) or any two words of 3+ characters with one of the EIN's first-party names, equals one (spaces ignored), or is its acronym, long prefix or abbreviation. Apostrophes are dropped before keying (ST MARY'S = ST MARYS). NA when the EIN has no first-party name to compare against. Not a guard by itself: it also fails for real brand names, and for subordinates whose BMF name is the parent's.
guard_any: Any of the five flag_* columns. The spec filter that matches networked990's resolver is NOT guard_any.

First-party names are legal, dba, division and former rows from any source (BMF, return headers, 1023-EZ). Re-run this after every ingest that adds them: flags computed before the return headers load see fewer names.