Verse Dialects

Glossia can render a payload as metrical verse — lines of a fixed syllable count, optionally scanning as iambic pentameter. The lines are real: a reader can hear when one is wrong.

That is the point. Glossia is a readable encoding, and the properties it sells are readable, speakable, transcribable, verifiable. Meter is the only one of those a human can check without a machine. A line that does not scan is a line that lost or gained a word.

Using one

A verse dialect is an ordinary dialect. Feed it a mnemonic and a poem comes out:

$ glossia --dialect haiku legal winner thank yellow zoo zebra youth wrist wrap world work word

Our ally is
legal. Winner thank our
yellow zoo. Zebra
may get youth. Wrist wrap
a world work to word.

Three ship for English:

glossia --dialect syllabic ...  # ten-syllable lines
glossia --dialect haiku    ...  # 5 / 7 / 5, repeating for longer payloads
glossia --dialect iambic   ...  # iambic pentameter (what literature calls blank verse)

Two things differ from a prose dialect, both automatic. Output breaks on the meter rather than at --width columns — a verse dialect ignores --width. And --best-of defaults to 4 instead of 1, because whether a rendering's lines come out whole is the tie-break among equally dense candidates, so a single draw leaves it to chance. An explicit --best-of always wins. Raising it to 8–12 buys a better poem; past that the CLI's memory use climbs steeply, which is a pre-existing property of --best-of rather than anything about verse.

Decoding is unchanged, because nothing about the payload changed: filter the words against the payload wordlist and read them in order. A verse rendering decodes with exactly the code that decodes prose:

$ glossia --from-ascii "meet me at dawn" --dialect haiku
Ability hope
clinic to fringe. Dolphin stone
tea. Mother set our
magic lottery.
Good may arm to swear.

$ ... | glossia --decode
meet me at dawn

What it costs

Measured over 50 payloads at best_of=4, against the body dialect scored the same way:

Syllable counting is nearly free. Stress meter costs between an eighth and a fifth of the density — the price of a filler that must satisfy the beat as well as the line.

The ordering among the stress meters is not the obvious one. Triple time ought to be harder than duple, since only one position in three carries a beat, and anapests bear that out at 44%. But dactyls scan more often than iambs — because English content words fall. A third of the payload wordlist scans 10 against a ninth scanning 01, and a dactylic foot (DUM-da-da) takes those words at the head of every foot, where an iambic line can only use them off the beat. The falling foot runs with the vocabulary; the rising foot argues with it.

Scan rate and density trade against each other, so read both columns together: a form that spends more cover words buys more freedom to fix the beat. dactyl scans more often than iambic and costs more density, which is the same mechanism seen from either end.

How it works

Three rules, in order of precedence:

1. The payload is untouchable. Payload words and their order are fixed before any of this runs. A payload word that will not scan where it lands is placed anyway and the line breaks. Meter never removes, reorders, or substitutes a payload word, which is why decoding is unaffected.
2. Cover words carry the meter. The generator already spends about a third of its words on cover. A verse dialect spends the same words differently: at each cover slot it asks the wordlist for a word of the slot's POS and a syllable count that keeps the line completable.
3. Best-of-N picks the winner. Density stays primary; among equally dense drafts, the one whose lines came out whole wins.

The mechanism that makes this work at all is small: a monosyllable is metrically flexible, so placing one always scans and always flips the parity of everything after it. Parity is therefore repairable at any slot offering both a one- and a two-syllable word — and every content-bearing POS in the cover list offers both.

Before committing to a word, the filler runs a backward pass over the sentence's slots to find which metrical positions can still finish. Without it, a cover word that fits locally can leave the line unfinishable; with it, the forward walk never backtracks.

Declaring one

A dialect becomes a verse dialect by declaring a meter: in grammar.yaml. With no rules: of its own it inherits the base grammar, so verse output is ordinary Glossia prose that has been filled to scan — not a different grammar:

  dialects:
    iambic:
      meter:
        lines: [10]        # syllables per line, cycling
        stress: lenient    # free | lenient | strict
        foot: iamb         # iamb | trochee | anapest | dactyl | amphibrach

stress: free counts syllables only. lenient additionally forbids a primary stress on a weak beat — the standard reading of "does it scan", and the one English verse observes; monosyllables float, as they do in practice. strict also forbids an unstressed syllable of a polysyllable on a strong beat, which is stricter than real verse and fails far more often.

foot is a repeating beat pattern read modulo its own length, so duple and triple feet are the same machinery. Line lengths must be a whole number of feet whenever a stress rule is in force — ten syllables is five iambs but three anapests and a limp — and a meter: block that breaks that rule is treated as no meter at all. (rise: true|false still parses, as the older shorthand for iamb and trochee.)

Because line lengths cycle, forms built from unequal lines need no extra machinery. Common metre is lines: [8, 6, 8, 6] with foot: iamb; iambic tetrameter is lines: [8]; an alexandrine is lines: [12].

A dialect with no meter: never loads the prosody data and generates exactly as it did before verse dialects existed.

The data

languages/<lang>/prosody.yaml annotates both wordlists — payload and cover — the way semantics.yaml does. English's is built from CMUdict by experiments/prosody/build_prosody.py, which covers 2047 of 2048 BIP39 payload words without hand-annotation:

stress: {"donkey": "10", "record": "010|100", ...}
rhyme:  {"donkey": "AONGKIY", ...}

The stress string does double duty: one digit per syllable, so its length is the syllable count. A | separates pronunciation variants, which are free slack for the fitter — record is both RE-cord and re-CORD. The filler's index is (POS, refinement, syllables, parity) → words, with parity derived from the stress string rather than stored.

Like semantics.yaml, the file annotates an existing wordlist rather than defining one, so it is regenerable rather than append-only — except that a canonical version naming the language freezes it, since verification re-renders.

Why not rejection sampling

The obvious approach — generate normally, keep the drafts that happen to scan — does not work for stress meter, and the measurements are in experiments/prosody/. Over 10,240 candidate renderings, strict iambic pentameter appeared zero times at 32 and 64 bytes.

The reason is not the vocabulary. Only 3 of 1440 polysyllabic words cannot scan at some position: a trochee is a perfect iamb when it starts on the beat. The reason is that a word's metrical position is fixed by the cumulative syllable count of everything before it, so every polysyllable is one more constraint on a running parity — about three times as many constraints as syllable counting imposes, each roughly a coin flip.

Those constraints are not unsatisfiable, though. They are unsatisfiable by filtering, because filtering arrives after the cover words have already been spent on grammar. Building for the meter instead of filtering for it is what turns 0% into 78%.