40. Canonical verse is read strictly, and the strictness is a parameter

Context

The corpus provenance split — added 2026-09-07 in feat: the corpus says how much of it is evidence about the reading, and carried by no ADR of its own — reported 38 of 532 golden cases (7.1%) sourced outside this repository. The rest were written for the suite by its author, which is the failure the 0.1.0 review named: an implementation and its own examples agree by construction, and a misreading of a form is present in both or neither.

The attempt to raise that number found why it is low. The commit test: canonical verse, and what the checkers make of it (2026-09-08) put seven canonical pentameter lines through their rows, found three not read as ten syllables, and named synaeresis as the sharp cause — disobedience is five syllables to CMUdict, exactly and with no estimate, and four to Milton. It shipped as characterisation and deferred the editorial question here, because widening the metre model touches all 41 heuristic rows.

Two research batches, both 2026-09-08, then supplied 23 verified public-domain texts across 17 rows in three languages, with per-item confidence notes and two deliberate gaps (haibun, for want of any pre-1929 English translation rendering a prose-plus-haiku unit). That is enough text to answer the deferred question on a measurement rather than on a sample of seven. This record rests on 111 canonical lines, and they do not say what the seven said.

What 111 lines measure

	exactly on metre	one over	other
English, 69 lines	78.3%	8.7%	13.0% (all one under)
German, 42 lines	45.2%	52.4%	2.4%

Nothing in English is off by more than one syllable. Of the 15 English outliers, 12 contain a word the pronouncing dictionary had to guess — louely, heauen, siluer, shielde, bloudy and the rest of the 1590/1609 orthography, Chaucer's Middle English, and the proper nouns Sion, Aonian, Camöens, Thesiphone. Only 3 of 69 lines are disagreements between the poet and the model with no guessing involved. The seven-line sample pointed at prosody; 69 lines point at the lexicon.

German is a different story with a single cause: 52.4% of German canonical lines are exactly one syllable over, because the klingende Kadenz is the norm in German blank verse and the alexandriner, not an exception.

Decision

D1. Feminine endings become expressible, and the bar was met — on line length. The bar, chosen before the work in the manner of ADR 0034's D5: at least 95% of German canonical lines must fall within their declared syllable count. Measured 97.6%, against 45.2% today. The bar is stated on length because length is what was measured; the stricter question is reported below rather than folded into it.

A feminine ending is an additional acceptable pattern, not a different one. line_metre already scans a line against a set of readings — that is how a dactyl may be a spondee — so permitting a klingende Kadenz means offering {P, P + "0"} where the row offers P. A fixed longer pattern would not do: Gryphius' sonnet rhymes feminine Erden/Herden (13 syllables) against masculine ein/seyn (12) in one poem, and no single string spans both.

shakespearean_sonnet, rhyme_royal, spenserian_stanza, blank_verse, curtal_sonnet, alexandrine and sonnet hard-code their metre; they gain a parameter admitting the feminine variant, defaulting off.

Under the full metre scan — stress pattern as well as length — the same texts go from 35.7% to 83.3% (15 of 42 to 35 of 42). That is the honest figure for "does this line scan", it is below the 95% bar, and the residue is a stress question that D4 addresses rather than a length one D1 can reach. English is unchanged line for line under the same change (Sonnet 18 stays 10 of 14, Milton 12 of 16), which is D6 holding by construction.

D2. English synaeresis is declined, on the measurement. The bar: it must fix at least 10% of English canonical lines. The ceiling is 3 of 69, 4.3%, because elision reduces a count and 9 of the 15 English outliers are already one syllable under. Implementing it would move three lines and touch all 41 heuristic rows. The lever English actually wants is dictionary coverage, and Report.evidence already names the guessed word, which is what makes that diagnosable.

D3. The rhyme model is not widened. Measured over the canonical texts with their correct schemes: 11 of 366 rhyme constraints violated, 3.0%, with zero false positives. Spenser and Gryphius are perfect. The 11 are Sibbald's slant rhymes (4), Wordsworth's genuinely disputed scheme (5), temperate/date (1) and Chaucer's ye/joie (1). Accepting slant rhyme would trade a model that currently never fires spuriously for eleven cases, five of which sit in a text its own research note recommends replacing. Three of the four affected texts — Sibbald, Sonnet 18 and Chaucer — ship satisfied: false with the reason in source, which is a true statement about how this package reads them. The fourth, Wordsworth's disputed sonnet, is in no golden fixture; it exists only in tests/fixtures/canonical_lines.yaml.

D4. Stress gets no general change; two rows declare a limitation instead. 59 canonical lines raise 13 wrong_stress violations, but they concentrate: outside curtal_sonnet and elegiac_couplet, 44 lines raise 3. curtal_sonnet does not read sprung rhythm. elegiac_couplet already models dactyl and spondee as satisfiability per ADR 0014, and fails on German accentual hexameter, where word stress and metrical ictus diverge — Schiller's own Hexameter is stressed 0100 and takes the ictus anyway. Both rows say so in the catalogue rather than the model moving under them.

D5. A corpus case is transcribed from its named edition and never edited toward the checker. Measured, because the temptation is real and looks helpful: hand-modernising Sonnet 18's orthography moved it from 10 of 14 lines on metre to 8 of 14, since expanding the Quarto's ow'st to owest both adds a syllable and leaves the dictionary. Tuning the text to the checker destroys the only property that makes an external corpus worth having. A rejection is the finding.

D6. No default moves. Every parameter this record adds defaults to today's reading, so the 532 existing cases are unchanged by construction and the gate cannot go green on a widening nobody asked for.

Consequences

Seven rows gain surface, and a silent wrong answer with it. A caller who omits feminine_ending gets today's verdict, which is right for English and wrong for a German alexandriner — and gets it without a warning, because omitting a parameter is not an error. D6 buys backward compatibility at exactly this price. (alexandrine carries no metre parameter at all; feminine_ending is the whole of what it gained.)

One default did move, and it is a message rather than a verdict. line_metre reports a wrong_line_length against the candidate it judged closest, and that candidate's expected named only its own length — true of the candidate, false of the call, which had offered several. It now names every length the options allow. For a single candidate the text is byte-identical, so D6 holds where it matters; but two rows that were already scanning substitutable feet before this record existed — dactylic_hexameter and elegiac_couplet — now say 13, 14, 15, 16 or 17 syllables where they used to name one. That is a change to published output, made without a bar and justified only by the old message being false, which this project counts as a defect rather than a preference.

The external corpus rises by less than 23 — correction, measured at 13. This paragraph predicted an upper bound before the corpus was built; the implementation raised it from 38 to 51 (7.1% to 9.4%), which is a smaller rise than 23 and confirms the bound rather than contradicting it, so the number below is a correction in the sense of filled in, not overwritten. Three texts ship satisfied: false under D3 (Wordsworth's disputed sonnet, the fourth, is in no golden fixture — only in tests/fixtures/canonical_lines.yaml), and the four French sonnet and ballade cases cannot ship at all — sonnet and ballade want stress, which French does not have, so they sit inside the 18-row ceiling ADR 0034's D3 recorded. A satisfied: false case is evidence about the reading, but it does not demonstrate a row accepting real verse, which is the thing 7.1% was low on: of the 13, only five demonstrate exactly that — Crapsey, Bentley, Verlaine, du Bellay and Kuhlmann pass outright. The other eight, three of them German (Gryphius twice, Goethe once), ship satisfied: false with the measured reading recorded in source, which is the same shape this paragraph predicted, at the scale this record can now state rather than bound.

English canonical verse still fails, and this record declines to fix it. 21.7% of English canonical lines are not read on metre after D1. Anyone reopening D2 should re-measure first and over more than 69 lines: 4.3% is a ceiling on this sample, and the sample is six poems.

The package still rejects Sonnet 18. D3 makes that a position rather than an oversight, and it will go on looking like a bug to anyone who meets it without this record. That is the cost of a rhyme model with no false positives.

rondeau is untouched and still cannot express its own history. Its fifteen lines are module constants, so Charles d'Orléans' thirteen-line rondeau — the older form — raises wrong_line_count and broken_rentrement and has nowhere to go. That is a catalogue question about form variants, not a strictness question, and it belongs in its own record.

Two of this record's inputs were wrong before they were measured. The Gryphius sonnet looked like a rhyme-model defect firing on eight correct rhymes; it was the pentameter default, and with the right scheme and metre the rhyme is perfect. The Kuhlmann abecedarian ships satisfied: false in its research note on the ground that the source is stanza-initial, but the note excerpts only each stanza's first line, so the text it actually supplies is line-initial and the checker's acceptance is correct. Both were caught by running the texts rather than reading about them, which is the only reason this record's numbers are worth anything.

A defect the corpus found: lowercase du reads as two syllables. word_stress returns ['10', '?'] for du and ['?'] for Du and every other German pronoun, though syllable_count("du") is correctly 1 — case-dependent, and the same family of defect as the find_word casing mismatch fixed in device.segment. It lives in the denckring-de-data distribution, not in this package, so this record cannot fix it. It makes one of the 42 German canonical lines — Gryphius, "Es ist alles eitel", l.1 — report as 14 syllables in a wrong_line_length violation's found figure, where line_syllables reads the same line correctly at 13. Parked rather than fixed: sonnet/gryphius-eitel-sonett's golden case carries a min_score floor rather than an exact score for exactly this reason, so a later fix to the pronunciation data raises the score instead of breaking the fixture.

A scoring artefact, parked. metre_violations returns total=1 from its wrong-length branch and total=len(words) from the branch where a line matches a candidate's length, so turning feminine_ending on changes the score's denominator as well as its numerator — a six-line Iphigenie excerpt goes from 20/24 to 38/39 for the same line-for-line reading. Pre-existing: the single-pattern path has always returned 0/1 for a wrong-length line, and D6 holds exactly, because with feminine_ending=False there is one candidate and the accumulation is byte-identical to before this record. No satisfied verdict is affected, since that turns on whether violations exist rather than on the ratio; only score moves. Documented in line_metre's own docstring. Fixing it would change what score means for every metre-bearing row, and ADR 0005 makes the continuous score load-bearing, so it is its own decision record rather than a side effect of this one.

Alternatives considered

Widening the model to fit the poets. Implement English and German elision, accept feminine endings unconditionally, tolerate slant rhyme. Canonical verse passes, and every one of the 41 heuristic rows accepts strictly more than it did — which blunts the eval as a discriminator in exactly the dimension the provenance split exists to measure. D1 takes the one piece of this that a bar justified and leaves the rest.

Characterisation only: measure the gap, never move it. The shape the 2026-09-08 canonical-verse commit already took, extended to all 23 texts. Honest and cheap, and it leaves German failing 54.8% of its own canonical lines for want of a parameter the sonnet row has had all along. D1 is too well evidenced to defer again.

Modernising the texts so the dictionary can answer. Refuted by measurement under D5, and it would have been the wrong move even had it worked.