[derek73]

Based on master since #1371 merged; master is merged in, and the diff is this PR's own.

What

The group suggester's common given names (_COMMON_GIVEN in nameparser/config/_order_names.py) are now ranked from ORCID's public data file (CC0), not from nam_dict's country columns (GNU FDL). After this PR, _order_names.py holds no nam_dict data.

tools/vocab/order_names.py builds the table like this:

• For each country an ORCID holder states (250 in the local corpus), it ranks the words typed first in the given-name box.
• Only the first word is ranked. A ranking of every word let in East Slavic patronymics (викторовна).
• A word must be typed that way at least 3 times in the country to count, and each country contributes its top 500.
• Each word is counted under the spelling the suggester looks it up by: _suggest._EDGES stripped, then full stops, then folded. A test holds the two spellings together.

The TABLES filter is unchanged: at most a 5% family-box share over at least 20 uses, two letters or more, and no listed surname. ORCID needs two more refusals:

• a word holding a full stop (dr.ahmed, م.م);
• a word the default lexicon reads as a title outside titles_ambiguous (prof, dra, lic, engr, assoc, asst, lady). Holders type these before the name, and in the first build Prof. Li Jing cast a given-first vote.

The table holds 6,409 names, where nam_dict's held 4,262: 4,235 shared, 2,174 new and 27 left. The new names are romanized Chinese and Korean, Indian, African and Central Asian given names, plus Cyrillic spellings. The table is the project's own measurement over a CC0 file, so it carries no licence of its own (decisions/shipped-data-licences, FACTS WE MEASURE).

Licences

#1372 (merged) dropped nam_dict's familiar pairs, so with this PR no nam_dict data is left in the package. The merge commit b6816da therefore removes the licence material:

• LICENSE-nam_dict;
• README's bullet;
• GFDL-1.2-no-invariants-or-later from license, and the file from license-files;
• its _THIRD_PARTY row in tests/v2/test_licensing.py.

The expression is now LGPL-2.1-or-later AND Apache-2.0 AND CC-BY-4.0 AND OGDL-Taiwan-1.0, and no release ever named the GFDL. The release-log bullet from #1370 says so.

Moves

uv run python tools/differential/suggest_measure.py --base <checkout of 5a713720>, run against #1371's tree, whose tables differ only in _COMMON_GIVEN:

Order table. 159 cells moved: 147 up and 12 down. The largest fall is 0.6 points.

• Same-order rows. Four same-order cells fell, each by one group in 300. The worst same-order row outside the three anti-goal rows still reads 98%.
• PatentsView's Indian inventors are first read right in 95% of groups at 50 names, where it took 200. PSC's Indian filers still take 200, and no other population's first size moved.
• Largest gains. These are where nam_dict's columns were thin. PSC's Indian filers written given-first under a family-first base went from 30.0 / 59.7 / 93.3% to 45.7 / 84.0 / 99.3% at 10 / 20 / 50 names. PSC's Nigerian filers written family-first went from 35.3% to 62.7% at 10 names. PatentsView's Indian inventors went from 52.3% to 77.7% at 20.

Per name. The table cast 48,202 right and 1,087 wrong given-name votes, against 45,079 and 973 before (2.21% wrong, from 2.11%).

• Within a population, the wrong share held or fell in most rows.
• The names read wrong rose most for PSC's Nigerian filers (16→25 and 19→26 of 3,000) and for PatentsView's Indian inventors written family-first (42→52).

Parliaments. The order counts of 12 lists moved, and one suggestion changed. Mongolia's Khural is now decided family-first by two names that end in a listed given name, which is how locales.MN reads it (#1237); two is the order rule's minimum. The Duma casts 383 given-first votes, and its suggestion, decided by the patronymic rule, does not move.

Anti-goal rows. The rows of ORCID holders who typed their given name in capitals rose: from 26.7% to 30.0% at 2 names. Parser.suggest's docstring now quotes 30%.

Caveat. The orcid-* populations are drawn from the same ORCID sample, so their rows are not an independent test of this table. The anti-goal populations (PatentsView, PSC, the Czech candidates, the Vietnamese names) are.

Docs

• decisions/group-suggestion:
  • a 2026-11-12 GIVEN NAMES FROM ORCID entry, marked proposed, for review;
  • two Standing lines;
  • a head note on the TABLES entry.
• decisions/shipped-data-licences: an entry, and a head note on Name the licences of the shipped data in the package metadata #1370's table.
• Parser.suggest's docstring and docs/suggest.rst: the Indian figures, re-taken.

Checks

• uv run pytest (after merging Replace Unihan with Wikidata for locales.ZH's simplified forms and the suggester's pinyin #1371 and master): 46,817 passed, 38 skipped, 3 xfailed.
• uv run python tools/vocab/build.py --check: current.
• uv run pytest -n0 -m serial: 38 passed.
• uv run mypy and uv run ruff check: clean.

Reviews

• design-docs-reviewer. Its findings were all fixed:
  • the 27% stale in Parser.suggest;
  • 13 lists, not 12;
  • the Standing line written as decided;
  • Name the licences of the shipped data in the package metadata #1370's head note.
• pr-review-toolkit code-reviewer. It found that the tool's spelling differed from the suggester's (ala' was unreachable), the docstrings missed the title refusal, listed_titles' docstring did not match its code, and the floor was stale. All fixed.
• A code-reviewer pass on the fix commit. It found that mst and ace entered the table, and since a listed name is never read past as a post-nominal, Mary Smith MST casts no vote. This is recorded as an accepted cost beside asa and cem. It also found that the spelling test checked a copy of the tool's formula; it now checks every row against _suggest._name_words.
• A review of the merge that removes the nam_dict licence. It found the release-log sentence calling the familiar forms rebuilt where they were dropped; fixed.

🤖 Generated with Claude Code