Phase 4: order and shapes, the rules a pack can choose, and the output model #754

Activity

1. derek73

   self-assigned this

   on Oct 19, 2026
2. derek73

   added

   design

   vocabularyPer-word vocabulary model and language packsPer-word vocabulary model and language packs

   on Oct 19, 2026
3. derek73

   added a parent issue on Oct 19, 2026

   • Vocabulary model and language packs (tracking) #747
4. derek73

   added sub-issues on Oct 19, 2026

   • Should suffix split into generation and credentials? (Smith, Jr., PhD) #326
   • Should a two-word given name like María José or Min Ho fill given, instead of given + middle? #704
   • Should Ramona Navarro vda. de Villegas keep vda. de Villegas in the family name? #565
   • Under a family-first order, should a lowercase von/van den opening the name start the family? van den Berg Cornelia gives family van #670
   • Tariq Abdul Aziz: should a two-word surname like Abdul Aziz stay whole after the given name? #694
   • Should an ALL-CAPS word in a mixed-case name mark the family name? DUPONT Jean gives given DUPONT #659
5. derek73

   added a commit that references this issue on Oct 21, 2026

   Standing block for decisions/vocabulary-model (#762) (#800)

   ...

   b6c8ffc
6. derek73 commented on Oct 23, 2026

   

   derek73

   on Oct 23, 2026

   OwnerAuthor

   More actions

   A ranking for the surname-span rule, from an unselected population

   The RFC's surname-span measurement (315 of 390 given-first double surnames disagreeing, GND and Wikidata) "was taken over corpora built to hold second surnames, so it shows that the gap exists, not how it ranks against the others". The local corpus psc-individuals (#849) can rank it. It holds Companies House's people with significant control, 1 in 8 of the snapshot dated 2026-10-22: 1,040,708 filed names, nobody chose them, and each filer typed the forename, middle name and surname into separate boxes.

   Size. 29,023 names (2.8%) have a two-word filed surname. 22,168 of them read with the surname's first word as a middle name (Ana García López → middle García, family López). It is the largest disagreement shape in the score.

   Who. Only 28% of the 22,168 (6,225) are Spanish, Portuguese or Latin American nationals. The rest are not:

   nationality	records	two-word surname	read with a middle	share of records
   British	680,873	10,236	7,688	1.1%
   Spanish	6,101	3,940	3,820	62.6%
   Indian	21,132	1,206	1,151	5.4%
   Portuguese	5,343	1,204	932	17.4%
   Italian	11,790	1,046	571	4.8%
   Moroccan	3,577	550	541	15.1%
   Pakistani	16,532	523	481	2.9%
   French	9,249	609	433	4.7%
   Brazilian	1,927	560	389	20.2%
   Sri Lankan	1,974	392	360	18.2%
   Mexican	405	271	262	64.7%
   Colombian	379	264	258	68.1%

   What it means for this phase.

   • Where an es pack would help: in Spanish, Mexican and Colombian names, nearly two in three names are misread. That is the strongest case in the data for an es pack's surname-span rule.
   • Where it would not: a rule chosen by a language pack reaches only callers who know the language, and most of these names come from nationalities no es/pt pack would be chosen for. Nationality is not a naming convention, so the British, Indian, Moroccan and Sri Lankan counts mix several conventions (South Asian two-word surnames are Tariq Abdul Aziz: should a two-word surname like Abdul Aziz stay whole after the given name? #694's question). Even so, they say the shape is not only Iberian.
   • What that leaves the default: the only general signal is the comma, and the RFC's own figure shows that index form disagrees on 0 of 60.

   Recompute. uv run python tools/differential/local_agreement.py psc-individuals reports the shape as middle 0->1, family 2->1 and middle 1->2, family 2->1. The per-nationality rows count the records in agreement.jsonl whose filed surname has two words and whose parsed family is its second word. The names stay in the cache; this comment prints counts only.