Skip to content

Read Dutch Oude by default and ten under a new locales.NL pack (#1455) - #1460

Merged
derek73 merged 5 commits into
masterfrom
claude/issue-1455-dutch-ten-oude-a4bac3
Nov 19, 2026
Merged

derek73 merged 5 commits into
masterfrom
claude/issue-1455-dutch-ten-oude-a4bac3

Conversation

@derek73

@derek73 derek73 commented Nov 19, 2026

Copy link
Copy Markdown
Owner

Closes #1455

Builds the three decisions recorded on the issue on 2026-11-18:

  • oude is a particle in the default vocabulary (nl.toml), with the case mask Oude. parse("Piet Oude Luttikhuis") gives family Oude Luttikhuis, where 1.4.0 through 3.1.0 gave middle Oude.
  • ten is an nl.toml preposition written default = false, with a given reading beside it. The given reading is declared in zh.toml (see below). The default parse of a name holding ten is unchanged.
  • locales.NL ships, a pack choosing nl alone. parser_for(locales.NL).parse("Jan ten Brink") gives family ten Brink.

The kind chosen for oude, and why

oude uses the generic particle kind (particle = {}).

  • It is an adjective ("old": the Twente and Achterhoek farm names, Oude Luttikhuis, Oude Elferink).
  • The README's particle kinds each name what a word does in a name (preposition, article, patronymic, teknonymic, clan, rank, saint, alias). None of them describes an adjective.
  • Every particle kind compiles into the same set, PARTICLES, apart from rank and saint. Those two are written with a capital by kind, and oude is neither. So the choice is descriptive only and moves no parse.
  • Alternative weighed and not built: a kind of its own (qualifier, say) for the Low Saxon farm-name qualifiers (Groot, Klein, Lütke). That adds a public Lexicon.words(kind=...) value for one vetted word, so I left it for when those words are vetted. The vocab README now names oude as the one generic particle reading.

What the decisions did not settle (my calls, open to review)

  • oude carries a given reading. The vetting found nam_dict lists Oude as a given name (band 1, the Netherlands). The README's C-i rule treats one attested bearer as settling the may-be-given side.
    • oude is therefore a particle that may be given. Opening a name it stays the given name and reports the fork (Oude Luttikhuis → given Oude, particle-or-given), as van does.
    • The vetting scored never-given and may-be-given alike (16 fixed / 1 broken either way). Dropping the reading would read a surname-only Oude Luttikhuis as family, unreported. That is one line in nl.toml.
  • ten's given reading is in zh.toml, not nl.toml.
    • Where it is borne as a given name: Singapore, Taiwan, Malaysia and Bhutan. ORCID and PatentsView given boxes open with it 8 times (scratch probe over the cached extracts, 2026-11-18). No Dutch filer means a given name by it.
    • The decision says "with a given reading beside it". I read that as on the word, placed in the file whose names carry it. This follows how Latin ben (given reading in en.toml) and sushri (or.toml) were placed, and the RFC's in/to plan. No file holds Dzongkha, so the Bhutanese bearers are noted in zh.toml's entry.
    • Consequence: the plain nl choice keeps a leading Ten given, reported. A strict nl choice reads no zh.toml, so a leading Ten is the family name (Ten Kao Liu → family Ten Kao Liu). ten therefore joins STRICT_UNGATED["nl"] in tests/v2/test_reading_index.py.
    • Putting the given reading in nl.toml instead would keep strict NL from ungating it.
    • The build accepted the shape as first written; nothing was refused.
  • The pack drops the credential VD. Choosing nl turns off other files' vocabulary readings of words nl.toml holds, so vd (van der) loses en.toml's VD (Volunteer Decoration): John Smith VD gives family VD under the pack. After a comma it costs more: John Smith, VD reads the whole name as the family (as John Smith, de does by default). This is the same mechanism as locales.AR and ABD. The docstring, docs/languages.rst and the release log say so, with the en remedy as a doctest, and case rows pin both forms.
  • Docs example moved. docs/customize.rst used ten as its example of an unlisted particle. It now uses Dutch in (in 't Veld), which no file holds.

Measurements (all 2026-11-18, py3.11, against master 688991e7)

The comparator for master is expr:Parser(lexicon=Lexicon.default().remove('oude')). Its 36 compiled sets equal master's, dumped in both trees and compared. The only difference is oude's case mask, which no compared field reads.

Default (oude)

Check Result
tests/v2/test_snapshot.py No committed corpus name moves. The snapshot gains only the four new rules examples (re-recorded with snapshot.py --record).
uv run python tools/differential/cldr_agreement.py --record The default records are unchanged ("accepting: no change").
config_diff.py diff "expr:…remove('oude')" default plain:NL strict:NL --committed --local all --match '(?i)(^|[\s,.(])(ten|oude|vd)([\s,.)]|$)', 370 names at head The default moves 26, all oude middle→family: 25 local (the vetting's 22 plus 3 in ctgov-contacts, a corpus fetched since, #1454) and 1 committed, this PR's own rules/R4 example. No name holding ten moves by default.
config_diff.py score "expr:…remove('oude')" default plain:NL strict:NL --cap 20000, 434 groups ≥100 The default fixes 16 / broken 1: ORCID NL 10, PatentsView NL 5 and DE 1. The break is a PatentsView filer who typed Oude into the given box. This confirms the vetting's 16/1.

The pack (locales.NL)

Check Result
Same diff plain:NL moves 250 names (holding ten 206, oude 26, vd 18); strict:NL moves 254 (208, 26, 20). The two read 34 names differently, 24 holding ten: under strict a leading Ten, or one opening the given part after a family comma, reads as the family.
Same score plain:NL fixes 102 / 10 against master. Subtracting oude leaves ten (and vd) at 86 / 9, matching the vetting: Netherlands and Dutch filers 66 / 4 (ORCID NL 31/0, PatentsView NL 28/4, PSC Dutch 7/0), Belgium 2/0, everyone else 18/5. Read one by one, two of the nine breaks are the mid-name syllable (PatentsView MY Kheng Ten Choo, SG Terence Ten Teck Hong); the other seven are filers who typed ten at the end of the given box.
config_diff.py gaps nl Strict against plain over the Dutch-language groups: 456 broken / 54 fixed. The breaks are other languages' Latin particles dropped (el, di, al, le, da); ten is not among the top 20 words.
CLDR locale view The nl file reads 22/24 with and without the pack; recorded in cldr_agreement_locales.json.

Suggester and name match (oude joins the particle set both read)

  • suggest_measure.py --base <master checkout>: 0 cells moved.
  • match_diff.py --base <master checkout>: 0 of 690,382 rows moved before any test pair held oude. With the two pairs this PR adds to tests/v2/test_match.py, written as tuples so the tool draws them: 1,254 of 690,426 rows move (57 under each of the 22 settings). Every one is a pair holding Piet Oude Luttikhuis, whose keys gain oudeluttikhuis;p and whose middle Oude empties. Piet Oudeluttikhuis is now consistent (family joined) where master read a conflict, and Piet Luttikhuis sets the particle aside. The key guarantee holds: 0 consistent pairs share no key.

Strict

  • parser_for(locales.NL, strict=True) reads ten as a never-given particle. A leading Ten becomes the family name (Ten Brink, Ten Kao Liu, , Ten Brink, Brink, Ten Jan → family Brink Ten Jan).
  • After a family comma, a given part opening with Ten folds into the family under strict (Brink, Ten Jan, , Ten Brink, SWACKHAMER, TEN EYCK). docs/languages.rst and the decisions entry record this as a cost.
  • oude keeps its given reading, since that reading is in nl.toml (pinned by a test).
  • STRICT_UNGATED["nl"] moves from {den, tho} to {den, ten, tho}.

Sweeps (nameparser/AGENTS.md)

Both words, each comma shape × each name_order × {default, locales.NL, strict NL}, one-case and mixed-case, plus case repair plain and forced. The full grid is reproducible with a short script (in the session scratchpad). Summary:

  • ten, default: unchanged from master in every shape.

  • ten, under the pack:

    • Mid-name: Jan ten Brink → family ten Brink under every order. Under FAMILY_FIRST it becomes given ten Brink, the particle chain.
    • At the head it stays given and is reported, under every order (Ten Brink, Ten Wei Ming, Ten K. Smith, Ten John).
    • After a family comma: Brink, Jan ten → family ten Brink, reported (P6). Brink, Ten Jan keeps given Ten.
    • Initials stay safe: J. Ten Smith keeps given J.
    • The mid-name cost the decision accepted: Park Ten Hwan and Kheng Ten Choo → family Ten … (now in rules/P2's Accepted).
  • ten's given reading moves no default parse: a parametrized test compares Parser() with the reading removed, over 12 shapes × 3 orders, with a negative control under the pack.

  • ten, the vetting's section-5 shapes that read wrongly without a given reading: Ten Wei Ming, Ten K. Smith, Ten John, Brink, Ten Jan and , Ten Brink all read given Ten under plain NL. Under strict NL they read as family, which is the strict cost above.

  • oude, all configs:

    • Mid-name it reads family Oude Luttikhuis under every order; after a family comma it reads family, reported.
    • At the head (Oude Luttikhuis, Oude Wei Ming) it stays given and is reported.
    • Jan de Oude Brink → family de Oude Brink. Van Oude Luttikhuis → given Van (reported), family Oude Luttikhuis.
  • Case repair (capitalized() and capitalized(force=True)):

    Input Default NL / strict
    PIET OUDE LUTTIKHUIS, piet oude luttikhuis Piet Oude Luttikhuis Piet Oude Luttikhuis
    JAN DE OUDE BRINK Jan de Oude Brink Jan de Oude Brink
    VAN OUDE LUTTIKHUIS Van Oude Luttikhuis Van Oude Luttikhuis
    JAN TEN BRINK, BRINK, JAN TEN Jan Ten Brink Jan ten Brink
    TEN BRINK Ten Brink NL: Ten Brink (given); strict: ten Brink (the whole family)
    Jan Ten Brink left as written unless forced forced: Jan ten Brink
  • Found, not fixed (an existing shape): a possibly-given particle opening a name before a suffix comma folds into the family unreported. Under the pack, Ten Smith, Jr. → family Ten Smith; by default on master, Van Smith, Jr. reads the same. Recorded under the decisions entry's Left out.

Docs and rules amended

  • decisions/vocabulary-collisions: two Standing lines and a 2026-11-18 Should ten and oude be particles? (Jan ten Brink gives middle ten, family Brink) #1455 Log entry carrying every recipe above.
  • rules/P2 (particles join forward): the Chinese-syllable cost under the pack, with examples.
  • rules/R4 (case repair reads the unjoined mark): an Oude example.
  • User docs:
    • docs/languages.rst: a "Dutch particles" section.
    • docs/locales.rst: the pack's two rows.
    • docs/migrate.rst: a "Changed in 3.2" entry.
    • docs/release_log.rst: two 3.2 bullets.
    • docs/customize.rst: the example change above.
    • README.rst: available().
  • Tests: case rows for both words (fix/feat(#1455), parity where the fields match the 1.4.0 wheel; parity_check.py reports 0 of 628 rows reading otherwise), plus the recorded sweep counts in tests/v2/test_cases.py, moved with a dated note.
  • The name match moved too: see the measurements above. The same pair is now a conflict for a caller who drops particles-aside, as Jan van Berg / Jan Berg already is.

Checks

  • uv run pytest (50,068 passed after the review fixes) and uv run pytest -n0 -m serial, on 3.11. Neither environment has the ja extra installed.
  • The same suite on 3.15 in its own environment.
  • With the ja extra, in its own environment: test_locales, test_cldr_agreement and test_rules_doc.
  • uv run mypy, uv run ruff check, uv run python tools/vocab/build.py --check.
  • uv run sphinx-build -b doctest docs <tmp>, and the html build: clean.
  • Each new test run alone with -n0.

Review aspects run

All run on the two implementation commits; every finding was fixed in 10cb9859, which was then reviewed again.

  • pr-review-toolkit:code-reviewer: no bug and no missed roster. It found the undocumented John Smith, VD cost, a lower-case claim for ten resting on one source of three, and a P2 count its cited entry did not hold.
  • pr-review-toolkit:pr-test-analyzer: every vocabulary property fails a test when reverted (5 mutations, each run). Gaps filled: the two-run cost, VD, the name match, the zh reading's no-default-move guard and strict Oude.
  • pr-review-toolkit:comment-analyzer: Oude is not always followed by another word (De Oude); migrate.rst's 3.1 scope; languages.rst overstating which Dutch particles the default reads; wording.
  • design-docs-reviewer: the diff recipe no longer reproduced its figures after the rules examples landed, and "13 more" was wrong (re-taken, and counted by the word held); builder's choices were recorded as decided (now separated); strict comma costs were missing.
  • Second review of the fix commit: the fix holds: every re-run figure reproduces and the new tests are not vacuous. It found three prose slips, fixed in the last commit: "20 outside the Netherlands and Belgium" (it is 20 outside the Netherlands), P2's 9-against-66 pairing (now 9 against 86), and a -particles-aside verdict that moves True → False for Piet Oude Luttikhuis / Piet Luttikhuis, as van already does.
  • Not run: silent-failure-hunter and type-design-analyzer, since the change adds no error handling and no type; code-simplifier, since the only code is a 40-line pack module that mirrors ar.py.

🤖 Generated with Claude Code

derek73 and others added 4 commits November 18, 2026 16:07
)

oude is a particle on by default (nl.toml), with a given reading for
nam_dict's rare Dutch given name and the case mask Oude, so Piet Oude
Luttikhuis gives family 'Oude Luttikhuis'. ten is nl.toml's preposition
written default = false, with zh.toml declaring the Chinese given-name
syllable it opens, so locales.NL, a new pack choosing nl alone, reads
Jan ten Brink with family 'ten Brink' and a leading Ten stays given,
reported, save under strict=True. The pack's choice also drops en's
credential VD for vd's particle. Registered, with rotators, contents,
pickling, case-repair and strict tests, the CLDR locale view's nl file
(22/24 with and without the pack), and ten in the reviewed table of
words a strict choice ungates.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
decisions/vocabulary-collisions gains two Standing lines and a dated
entry with every figure's recipe: oude 16 fixed to 1 broken by default,
ten 86 to 9 under the pack (the Netherlands 66 to 4), strict nl's gap.
rules/P2 records the Chinese-syllable cost under the pack, rules/R4
oude's mask; the rules corpus and snapshot gain the four examples, no
existing name moving. The vocab README names oude as the one generic
particle. docs/languages.rst gains a Dutch section, docs/locales.rst
the pack's rows, migrate.rst a Changed in 3.2 entry, the release log two
bullets; customize.rst's unlisted-particle example moves from ten to in.
Case rows for both words, checked against the 1.4.0 and 3.1.0 wheels.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The review findings. Records: the decisions entry re-takes its diff at
head (the default 26, plain:NL 250, strict:NL 254, the two reading 34
names apart) and counts moved names by the word they hold; says which
calls were the builder's (oude's kind and given reading, ten's given
reading in zh.toml) apart from the two decision comments; states ten's
case on all three sources; drops 'always before another surname word'
for oude (two boxes end in De Oude); names the two scored mid-name
breaks; records the strict comma costs and the pack's VD comma cost,
and the name match's moves. rules/P2's bullet carries that count, dated,
with only its cost and a boundary in its fence.

Docs: languages.rst's scope, counts and costs; migrate.rst's 3.2 scope
and table row, and Oude's new report; the release log's VD comma form;
nl.py's docstring; the vocab README's wording; zh.toml's Bhutanese
names.

Tests: case rows for the two-run cost and VD (bare and after a comma);
strict keeps a leading Oude given; zh's given reading of ten moves no
default parse (with its negative control under the pack); the match
pairs, written as tuples so match_diff draws them (1,254 of 690,426
rows move, each a pair holding Piet Oude Luttikhuis).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 20 fixes are those outside the Netherlands, not outside it and
Belgium; rules/P2 pairs the nine breaks with the 86 fixes over the same
records; the match account adds the one verdict that moved the other
way (Piet Oude Luttikhuis against Piet Luttikhuis under
-particles-aside, as van already reads) and the two plain/strict
differences outside the shapes it named.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@derek73 derek73 added this to the 3.2 milestone Nov 19, 2026
@derek73 derek73 added vocabulary Fixed by a word-list entry: per-word vocabulary and language packs particles Particles and surname prefixes (rules/P) locale-packs Locale packs and language choices (nameparser.locales) labels Nov 19, 2026
@derek73 derek73 self-assigned this Nov 19, 2026
@derek73

derek73 commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

The four choices listed under "What the decisions did not settle" are accepted as made (2026-11-18): the given reading on oude, the given reading of ten placed in zh.toml (so a strict Dutch choice ungates it), the credential VD dropped under the pack with its documented remedy, and the generic particle kind for oude.

@derek73
derek73 merged commit 3175314 into master Nov 19, 2026
12 checks passed
@derek73
derek73 deleted the claude/issue-1455-dutch-ten-oude-a4bac3 branch November 19, 2026 02:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

locale-packs Locale packs and language choices (nameparser.locales) particles Particles and surname prefixes (rules/P) vocabulary Fixed by a word-list entry: per-word vocabulary and language packs

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Should ten and oude be particles? (Jan ten Brink gives middle ten, family Brink)

1 participant