43. English and IPA made commensurable, and the shopfront's third shape

Context

ADR 0042 concluded that the blend needed no grapheme-to-phoneme model, and it was right about that. It also recorded two things left unreachable, and named one of them wrongly as a curiosity rather than as the main event: the cross-lingual blend.

Going to the attested collections (docs/research/2026-09-10-…) settled how large that gap is. The commonest real French salon name in existence is one this project could not check. tif.hair counts ~150 salons called Imagin'hair, and the whole -hair family — Atmosph'air, Caract'hair, Bulles d'hair, L'Hair du Temps, Changement d'Hair — is English spliced into French. Roughly half the attested German names are the same shape in the other direction: Hairforce One, Open Hair, Bel Hair, Kamm Back, Wellkamm.

The obstacle was not phonology. It was notation. CMUdict answers in ARPABET and the German and French packs answer in IPA, so hair is HH EH R and French air is ɛ ʁ, and an edit distance over those two symbol sets is meaningless rather than merely inaccurate.

Separately, a third shape of shop name had been excluded by both rows and never catalogued: A Cut Above, where nothing is displaced and nothing spliced, and the trade alone supplies the second reading.

Decision

D1. A 39-entry ARPABET-to-IPA table, and nothing more clever. phonetics.ARPABET_TO_IPA maps CMUdict's phoneme set to IPA; German and French pass through unchanged because they are IPA already. A table a reader can check line by line against the CMUdict inventory, with nothing that can be silently wrong — the property that ruled out a learned g2p in ADR 0042 rules in a lookup table here.

D2. Rhotics fold across languages and only across languages. CROSS_LINGUAL holds one class: {ɹ, ʁ, r, ɐ̯, ɐ, ɚ}. English ɹ, French and German ʁ, German's vocalised ɐ̯ are one category realised differently, and a pun turns on the category. Kept separate from the per-language EQUIVALENT, because folding ɹ and ɐ̯ together inside English would assert something about English that this table is not asserting. Measured: hair against French air is 0.333 with the fold and two edits apart without it.

D3. The splice may come from another pack, named by a parameter. portmanteau.splice_lang resolves the spliced word in its own language's pack and compares through phonetics.across_languages. No new capability is declared: what the row needs is phonemes on two packs, which is the capability it already requires, asked of a second language. Naming a phonemes.bilingual capability here would claim a pack provides something no pack provides — the bridge is in core.

This does not unblock homophonic_translation. That row is still marked blocked on phonemes.bilingual and stays so. The notation half of its problem is now solved; judging that a whole target text follows the sound of a whole source across a language barrier is the other half, and much the larger.

D4. A host may be more than one word. Mona Lisa behind Monhaarlisa, Hart am Limit behind Haart am Limit. Pronunciation is now assembled word-by-word and run together, all-or-nothing: one unknown word in a host would otherwise silently shorten it and move every distance computed from it.

D5. portmanteau's default band opens to 0.7, from 0.5. Measured on the attested corpus, real names sit above the old ceiling — Haarwaii at 0.667, Coiff'Hair at 0.667 — because a splice genuinely may cover a stretch loosely. A blend may also name its own max_distance, which one shipped name uses: Monhaarlisa is a blend you see rather than hear, Haar covering the a of Mona and sounding nothing like it, and it asks for 1.0 rather than being dropped for failing a test of the wrong sense.

D6. amphibologia is a row of its own. (Withdrawn 2026-09-10, the day it was decided. The row was built, evaluated and cut in one session, and the reason is worth more than the row was: its checker is a proxy.* Polysemy is not a phrase reading two ways — A Cut Above and A Room with a View both carry a word of the trade with more than one sense, and only one of them is a pun. The sense counts it leaned on are an editor's decision besides: Open English WordNet gives cut seventy senses and another dictionary would give another number. A row whose definition says "can be taken two ways" while its checker asks "is this word polysemous" is the ADR 0015 error in a subtler dress than the two this session already caught, and the honest response to finding it a third time is to delete rather than to reword.

Puttenham's Amphibologia, or the Ambiguous is real and was verified — he files it among the vices of style, beside Pleonasmus and Bomphiologia — and the figure may well deserve a row one day. It does not deserve one whose checker cannot see it. The paragraph below is left as written, because an ADR records what was decided when.) Puttenham, The Arte of English Poesie* (1589): speaking "doubtfully, and the sense may be taken two ways". A phrase carrying a word of the trade, where that word is polysemous, so a second reading exists for the trade to activate. requires: [tokens, lexicon.glosses], checkability: self, kind: restrictive.

Puttenham files it among the vices of style, beside Pleonasmus and Bomphiologia, and it is left there rather than quietly promoted — a shop sign wants the doubtful reading caught, so the vice is the point.

It is a row and not a mode because it states a relation of a different kind: paronomasia and portmanteau relate two texts, and this relates a text to a trade. Folding it into either would make that row's definition describe two things, which is the drift ADR 0015 exists to prevent.

D7. Both new rows generate, and neither invents from nothing.

portmanteau.apply proposes every way of splicing each trade word into the host — replace any stretch, including none — and hands each candidate to this row's own check. What survives is a blend because the checker said so; the generator only decides what to ask about. That is the project's thesis used as a search, and it is what makes the missing grapheme-to-phoneme alignment survivable: not knowing where to splice stops mattering if you can afford to try everywhere. The search is small — w * (n+1)(n+2)/2 candidates, at most 859 for English paraphernalia, about 0.2s.

amphibologia.apply selects rather than invents, which is what that figure is: it is noticed in something people already say, not built from parts. The reader supplies the phrases one per line, per ADR 0020, and the row returns those that read two ways for the trade. The round trip is closed by construction.

D8. The blend ranking is (sound, host-recoverability, seam), chosen by measurement and shipped with its failures named.

Measured against the attested corpus with the shipped hair vocabulary: the real salon name comes back first for 10 of 12 hosts and is in the first ten for 11. The misses:

• paraphernalia gives phairaphernalia ahead of hairaphernalia, by one place.
• Chamäleon gives ChKammäleon ahead of Kammäleon, at rank 23. It is a seam stutter: keeping Ch and adding Kamm retains more of the host than replacing Cham does, so it wins on recoverability while being a word no German speaker would write.

Three fixes were tried and each made the whole result worse, and they are recorded so nobody spends the afternoon again:

1. Filter on seam similarity. A bound tight enough to remove the stutters (0.75) keeps 37 of 38 attested names but loses Coiff'Hair, which replaces ure with hair and resembles it not at all.
2. Rank by absolute length difference. It began preferring coinages that delete host letters — Haaronie, hairtage, SHaara — which mangle the host worse than the stutters do.
3. Rank by the length of the replaced stretch. Fixed Chamäleon and broke heritage, which then preferred hairtage over hairitage.

Twelve attested names is not enough to tune a ranking against, and each of those three was a rule that fit the examples in front of it. test_portmanteau.py pins the 10-of-12 floor and pins the Chamäleon case as a known failure, with a note saying that if it ever starts passing the test should be deleted rather than kept as a monument.

D9. The blend's search window is seeded from a real window, and two shipped names were resting on the fact that it was not.

_closest_window began at (1.0, 0, 0) and returned that untouched when no window scored below 1.0 — a zero-length window, reported as evidence: h aː ɐ̯ for at 1.000, a comparison against nothing. It read as a loose match and was no match at all.

Two blends shipped on it, both with max_distance: 1.0 and a comment calling them blends you see rather than hear. With the bug fixed the measurement is plain: Monhaarlisa is 1.000 from every stretch of Mona Lisa, and Pawsome is 1.000 from every stretch of awesome. They are not blends this row can see, so both are removed from the shipped data, and the per-blend max_distance that existed only to carry them is removed from the model with them.

This is the sharper form of what ADR 0042 already admitted. There is a real class of name — Monhaarlisa, Pawsome, Yes We Kämm — whose joke is visual and cultural, where the letters carry it and the sound does not. No phonetic rule reaches them, because there is no phonetic fact to reach. What this row checks is the heard relation, and it should say so rather than open its band until anything passes.

Consequences

Ambiguity is read off a sense count, which is an editor's decision. Open English WordNet gives cut seventy senses and German Wiktionary gives Haar three; neither number would survive a change of dictionary. What survives is the ordering, so the row asks for more than one rather than scoring them, and xylophone — one sense — is the case that shows the test is not vacuous.

amphibologia ships no apply and is kind: restrictive. Generating one means choosing a phrase from a corpus of things people say, and ADR 0020 gives the corpus to the reader. There is nothing here for a generator to search that the reader has not already supplied.

Fifteen of the sixteen names the maintainer asked for now check, across all three rows. The one that does not is Pony & Clyde: Bonnie is in no dictionary here, and the only honest routes to it are a caller-supplied pronunciation or the g2p ADR 0042 declined. It is left failing and said so.

Coverage. 124 implemented rows become 125 and the catalogue 158 becomes 159; six constructed cases take the corpus from 583 to 589, and the externally sourced share falls to 11.2% — arithmetic, for ADR 0041 D7's reason, which the research trip did not change: the attested corpus is entirely post-1929.

The two limits that remain are unchanged in kind. A proper-noun host, which is most celebrity blends; and the found pun's cousin where the phrase is not one anybody says, which no test here can see.