An example name asserted outside the format it exemplifies pins nothing (`Nguyen Thi Van` is pinned under FAMILY_FIRST)

[derek73]

rules.md#P6 pins Nguyen Thi Van under both family-first orders:

"Nguyen Thi Van"  family-first-given-last  →  given="Van"
"Nguyen Thi Van"  family-first             →  middle="Van"

It is an FFGL-format name. Its FF reading is wrong by construction, so the second line asserts an output nobody wants and nothing should depend on — and freezes a value that a future fix would have to be described as "breaking". The same shape, reversed, applies to Beethoven Ludwig van asserted under FFGL.

An example name carries a format. Asserting it outside that format tests the parser against a premise the name never had.

rules.md has nowhere to record this today: an example line carries an input, an optional policy annotation, a field and a value, but nothing that says which format the name is written in — so there is no way to distinguish "this is the answer for this name" from "this is what falls out if you parse it under a format it was never written in".

Why this bites now, measured

rules.md's ambiguous-particle coverage is one word deep. Of the 37 ambiguous particles in the shipped vocabulary, only 4 appear in an example where the particle stands alone: abu, del, do, van — and van is 12 of those 18 occurrences. von, le, la and freiherr appear only inside chains (Freiherr von Berg, de la Vega).

So a rule about a lone ambiguous particle is effectively tested through one word, and that word drags a second particle (van der) behind it. A format label would make it visible that the coverage is one name deep rather than eight.

Questions

• Should an example line be able to declare its format, and should the doc test refuse to run a name outside it?
• Or is the right unit a Not-an-example-of: note, so the FF line for Nguyen Thi Van can be deleted with a reason rather than silently?

[derek73]

Closing: no notation is needed. The Accepted: clause is already the
right home for this, and both halves of the issue turn out not to hold.

Measured against master at c1a98a9, 2026-09-02.

The example is stale in both halves. #467 deleted the FAMILY_FIRST
line from P6 and left an Accepted: clause saying why. And the output
this issue quoted is no longer what the parser produces:

family-first               family='Van Nguyen'  given='Thi'  middle=''
family-first-given-last    family='Nguyen'      given='Van'   middle='Thi'

So the FF reading is not middle="Van" — it is P6's own particle
attachment firing, folding Van into the family. That is the behavior
P6's earlier Accepted: clause already documents, in as many words: the
ASCII spelling collides with the tussenvoegsel reading and the name
loses its given name. The out-of-format reading is a consequence of a
documented rule, not an unrecorded premise, and "Nguyen, Thi Van" → family="Van Nguyen" is already asserted three clauses above.

No live instance, and no recurrence. Across all 240 example inputs
in rules.md, six are asserted under more than one order — Jong Anke de, Juan de la Vega, Ménil Christophe de, Sheik abdul salam, de Mesnil Jean, de la Cruz Juan Carlos. Every one is deliberate: each
asserts a different field or a different value, showing how the reading
changes with the order. None is the shape this issue describes. A
format: declaration would need explicit opt-ins on all six to catch
nothing, and not-an-example-of: would enforce a convention that has
been violated once and was fixed. mechanisms.md#MAKE-WRONG-STATES-UNREPRESENTABLE
is for a convention that keeps being violated however clearly it is
written down; this is not that.

The coverage half is already answered, and declined. The numbers
have moved and the shape holds: of the 37 ambiguous particles in the
shipped vocabulary, 4 stand alone in an example (van 26, abu 3, do
3, del 1 — 33 occurrences, so van is 26 of 33, not the 12 of 18
filed here), von/le/la/freiherr appear only inside chains, and
29 of the 37 appear in no example at all. But that is per-entry
coverage, which rules.md's Not-in-scope already declines: "what the
tests owe is one exercise per behavioral fork rather than one per entry"
(mechanisms.md#VOCABULARY-EXERCISES-FORKS). The lone-ambiguous-particle
fork has 33 exercises. They are a van monoculture — a real fragility,
since a change to that word's membership moves 26 examples at once — but
the fork is exercised, and a row pins a fork, not a wordlist entry.

Recompute: parse Nguyen Thi Van under both orders through
Parser(policy=...) with the policies in tests/v2/rules_doc.py; for
the other two, walk parse_rules_doc(RULES_DOC.read_text()) and group
example inputs by order-bearing annotation, and intersect the example
words against Lexicon.default().particles_ambiguous.

P6's Accepted: clause keeps its (#470) citation — it points at where
the question was raised, and this comment is the answer.