[derek73]

Phase 2f of the vocabulary model (#752): every rule that weighs a roman numeral now asks one predicate, _vocab.reads_as_generation. That predicate reads the vocabulary's own numerals (ii to iv) itself, so no call site copies them. Closes #784, closes #774, closes #819.

What reads differently

Input	Now	1.4.0 through 2.6.0
John Smith VI MD	suffix 'VI MD' (reported)	family 'VI', suffix 'MD'
John Smith VI Jones	suffix 'VI Jones' (S8's run)	family 'Jones'
Smith, John VI	suffix 'VI' (reported)	middle 'VI'
John Smith V, M.J., John Smith VI, M.J.	suffix 'V' / 'VI'	family 'John Smith V' / 'VI'
Georg VI. (GND regnal rows)	suffix 'VI.'	family 'VI.'
Smith, John PhD I, Jr.	middle 'I' (1.4.0 to 2.3.0's reading)	suffix 'PhD I, Jr.' in 2.4.0 to 2.6.0
Nguyen Van Vi	family 'Van Vi' (Title case leans to the name, reported)	suffix 'Vi' in 2.2.0 to 2.6.0

What stays the same: SMITH, JOHN X keeps middle 'X'. A one-letter numeral in the given part after a family comma is the generation only where the vocabulary lists it, which is FEC's constraint. Listed III, Ii and Iii are never weighed, so an initial in front or Title case does not move them.

The three readings Derek agreed (decisions/vocabulary-model, 2026-10-23, item 8)

1. Should Smith, John PhD I, Jr. keep I as a middle initial, as Smith, John V, Jr. does? #819: a lone letter ending the given part ahead of a further comma is the middle initial. The credential run in front stops short of it. Accepted cost: Smith, John PSM I, Jr. splits a certification level from its credential (rules/C9).
2. FEC's constraint: an unlisted one-letter numeral in the given part after a family comma stays the initial.
3. Title case leans to the name: a Title-case numeral in a mixed-case name reads as a name, as a Title-case acronym does.

One reading kept rather than widened — please confirm

John Smith V PhD keeps family 'V'. This is decisions/S10's #721 reading. A one-letter numeral is still read only when it ends the walk; one of two or more letters is read anywhere in the trailing run.

I first read the one-letter numeral with the words-to-spare count in front of a credential too. That moved 6 tse-donors names, and every one was wrong: a lone V or I before JR or JUNIOR that abbreviates a surname (shape Aa A Aa de V JUNIOR gave family 'de', suffix 'V JUNIOR'). Item 8's wording can be read either way.

This leaves John Smith V Ph. D. (today suffix 'V Ph. D.') to the split-credential PR (2e), which has to keep that reading by its own rule.

Measured against master 2af8da8

Measure	Result
Committed snapshot	6 existing names move, all intended: 3 GND regnal rows, Smith VI, AD, and #774's two rows. The other new snapshot names are rule examples.
cldr_agreement.py	GND family 618 → 621 of 801, suffix 779 → 782. Every other record matches.
PSC, 1,039,171 records scored	3 fixed: Vietnamese names ending in the surname Vi, which the Title-case lean keeps a name. 1 broken: a Title-case Viii written as a generation (S10's Accepted bullet). 1 changed, in a title box the scorer sets aside.
ctgov-officials	1 move: a job title ending in a level before a comma, Aa Aa I, A → suffix 'I', reported.
fec-individuals	22 moves, each a numeral ending the given or the family part (A, A VI → suffix 'VI'). About 14 are generations and about 8 are a middle name or a fragment. All are reported.
tse-donors	No moves.

Each local baseline was re-recorded from the master checkout immediately before its check. The cache's snapshot.jsonl is shared, and a sibling session's re-record had once shown its 1,334 joined-credential readings as this change's moves. AGENTS.md now says so.

Spellings swept: V, VI, III, XI, IX, VIII, v, vi, Vi, Ix, Ii, Iii, V., VI., I, X, iii.
Comma shapes: no comma, a full name before the comma, one word before it, nothing before it, and a nickname head.
Under all three orders, with: a title-particle head after a family comma, an initial in front, a maiden clause, trailing titles, and one-case spellings. Two code reviews and a design-docs review ran on top, and a second review of the fix commit.

Remaining split, recorded in decisions/C1: Vega y Lopez V, Mary keeps family 'Vega y Lopez V', where its comma-less twin reads suffix 'V'. The part before the comma joins the connective and the comma-less name does not. Counting a join's words fixed that row but broke Smith and Jones V, Mary, which both spellings join.

Docs

• Rules: S10 (the numeral anywhere in the run, the period, the Title-case lean, the given part and the family part, the kept one-letter reading); S8 (a numeral starts the run); C1 (Should John Smith V, M.J. keep V as the suffix, as John Smith Jr., M.J. keeps Jr.? #774's deviates: rows removed); C9 (Should Smith, John PhD I, Jr. keep I as a middle initial, as Smith, John V, Jr. does? #819); C12 and C5 pointers.
• Decisions: dated entries in S10, S8, C1 and C9, and a vocabulary-model phase 2f entry. Item 8 is annotated where 2f narrows it, and Ask one predicate per question for split credentials, numerals and paired initials (#752) #884's list of remaining sites points at the new entry.
• Other: mechanisms.md, the RFC, and AGENTS.md's local-corpus recipe.
• Release-log bullets: two.
• Case rows: 12 new, 1 moved.
• Rules budget re-recorded: S10 grew most, by the clauses above. Its new Title-case clause reads case, which S10 already read for lower case.

Checks

All pass: uv run pytest, -n0 -m serial, uv run --python 3.15 pytest, mypy, ruff, and the Sphinx doctest build. Frame counts are unchanged on the reference name and Smith, John.

🤖 Generated with Claude Code