Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

0.3.1 - 2026-09-22

Added

• chimera, the third pos row and the first that generates it (issue #22). A text's nouns, verbs and adjectives stripped out and refilled from three different donors. The parameter shape it was said to be blocked on is three role-keyed fields — nouns_from, verbs_from, adjectives_from — beside the inherited source, and no change to core/base.py was needed: the premise that SourceParams "carries one source and chimera needs three" was wrong, since the source is the frame, the text this one was made from, and the donors are lexical stock rather than sources. An ordered sources: list[str] was rejected for hiding the role in a position params_schema() cannot describe.

Auditing it found the row's one real defect before it shipped: the generator produced text its own checker rejects, because a drawn word's class is a fact about where it lands — salt is a NOUN in the donor and PROPN at the head of a sentence. _produce now re-tags its own output and redraws the positions whose class did not survive. Four questions the definition leaves open — an empty donor, repetition, capitalisation, and how the frame's surviving words are compared — are ruled on in the module docstring, each with a test that has to change if the ruling does. ADR 0046 records the parameter-shape decision ADR 0045 declined to make.

Two limits are measured and stated rather than papered over. apply's redraw loop is greedy and incomplete: a refusal is that search failing, not a proof that the frame cannot be filled, and a sweep of 400 random cases found 21 refusals of which at least 8 were false. The message says so. And a donor pool is the tagger's reading of the donor, which can be far smaller than the donor text looks — so each candidate carries forced_positions beside target_positions, counting the positions where exactly one donor word survives and seed therefore changes nothing. - Thirty French-sourced rows now declare French (ADR 0047). meta.languages is authored editorial scope and runs_in is computed from capabilities; measured on 2026-09-21, 174 (row, language) pairs ran without being editorially claimed, of which 92 were French. This closes the thirty whose own source: field is the evidence — lipogram is Perec's La Disparition and declared [en, de] — and leaves the other 144 open rather than sweeping them. Eleven French names and twelve French definitions were written for the rows that had none, and 24 French golden cases, because a declared language must be one the row has actually been run in. French editorial scope goes from 14 rows to 44; runs_in does not move. The same test over German-sourced rows returns zero: every row sourced to a German original already declared de. Measuring for this found the README's denckring[fr] count stale at 103 against a measured 106, and its illustration of the distinction resting on belle_absente — a row whose languages was wrong, which is how it could illustrate anything. clerihew replaces it. - tests/test_declared_language_strings.py, asserting through describe that no declared language is ever answered in a substitute. tests/test_catalogue_quality.py already held the rule against the YAML; this holds it at the surface a caller sees, where the per-field fallback to English is what does the damage. - s_plus_7 and n_plus_7's two descriptions of s_plus_7 now agree (#23, documentation half). catalogue.yaml called the row "N+7 generalised to any part of speech"; the module docstring called it "N+7 generalised to any displacement", the step "left to the writer" — only one can be the row's point. Settled on the catalogue's side, since a definition here arrives with a source and the docstring's phrasing described an implementation instead. Neither description was ever built as a difference: n_plus_7 already accepts offset and dictionary, SPlus7Params adds no field, and _check/_produce both delegate — both rows' notes now say so, in both directions, because denckring show and describe_procedure over MCP surface notes to every caller. Still open on #23, not decided here: whether the row's definition itself, sourced to Oulipo's La Littérature potentielle (1973), should change — the honest reading of the code is a description of an implementation, not of a form.

Fixed

• Four guards that could not fail, and the state they leaked (#17). A catalogue test named for mutation never mutated anything and was true whether or not Meta was frozen; an MCP assertion was a disjunction almost any dict satisfied; a preflight test was named for the changelog branch it never reached. The two concurrency tests asserted that a thread had not finished yet — a claim about the scheduler, not the lock — and restored module globals with trailing statements, so a mid-test failure left _REGISTRY/_PACKS empty for everything after. They now wait on an instrumented lock probe and restore in a fixture finalizer. test_version_fallback.py restores in finally, so a raising reload no longer leaves __version__ at "0+unknown" for the rest of the session.
• Dead code, an unexplained re-export, a test-only default and an unchecked duplicate (#18). differ_by_one moved out of paragram.py into the test that is its only caller; import importlib as importlib says why it is written that way; _install's sources is required, removing a branch only a test reached and with it a write into the real _SOURCES global. _download_integrity.py exists as five byte-identical copies of which one was checked — a new test derives the set from the filesystem and asserts they hash alike.
• The release gate depended on the runner's ambient Python (#16). preflight ran a bare python3 with no toolchain step while every adjacent job set one up explicitly; the script needs 3.11+ for tomllib, so the job that exists to gate a release was the one that would break, during a release. It now sets up uv and runs uv run python, which provisions a Python satisfying requires-python regardless of the image.

• The audit guard checked one direction only (#19). It caught a registered procedure with no audit file, but not a stale file for a removed one, and not that docs/audit/README.md lists what is on disk. Strengthening it found real drift: homosyntaxism and verbless_prose were named in the README's intro and filed under no classification section.

• test_the_sdist_stays_small measures the commit rather than the checkout (#14). uv build reads the working tree, so a gitignored site/ of mkdocs output took the archive to 9.5 MB against a 950,000 cap and the guard failed on developer machines for reasons unrelated to the bound. Every implementer who hit it flagged it as pre-existing and worked around it — a guard training its readers to ignore it. The fixture now builds from git archive HEAD, so a local run and CI measure the same bytes. It skips where there is no repository, because the sdist ships tests/ and a downstream packager runs this suite from an unpacked tarball. The exclude entries in pyproject.toml are left in place: they no longer matter to the test but still protect a hand-run uv build in a dirty checkout.

• The sdist bound is raised from 950,000 to 1,050,000 bytes, by maintainer decision, after merging this release's three branches crossed the old one for real: the 35,963 bytes of headroom recorded at 0.3.0 were spent by legitimately shipped prose and fixtures, not by a data file creeping in. Re-measured from git archive HEAD, the safe window is (953,687, 1,203,604) — the clean build up to that build plus the smallest data file the guard exists to catch (graded_words.txt.gz, 249,917 bytes) — and 1,050,000 leaves 96,313 bytes of headroom while still catching that file or the roughly 300,000-byte mutants/ class of accident.

0.3.0 - 2026-09-21

Added

• pos, a part-of-speech capability, and the two rows it unblocks (ADR 0045, issue #22). LanguagePack.pos_tags() reads a whole sentence and returns one PosTag(upos, verb_form, known) per token. verbless_prose bars finite verbs; homosyntaxism checks that a text keeps its source's word classes position by position while replacing its open-class words.
• denckring-en-pos, a seventh distribution, supplying pos for English: an averaged perceptron trained on UD English-EWT v2.18, shipped as a weights table, pure Python at inference time and with no model download. Its own distribution because the treebank is CC BY-SA where denckring-en-data's data is not, and ADR 0013 quarantines a data licence. pip install denckring[pos].

It is a model, so its verdicts are estimates, and they are not hidden. Measured on the treebank's held-out test split: joint accuracy 0.9269, UPOS 0.9342, and for finite verbs precision 0.9482, recall 0.9498, F1 0.9490. So about one finite verb in twenty escapes verbless_prose. PosTag.known carries this per token, every report carries an undecided_words metric, and describe() reports reading.determinacy as heuristic for both rows. verbless_prose's golden fixtures include three satisfied: false cases that are false positives on genuinely verbless Dickens, recorded rather than avoided. - core.text.sentence_spans(), splitting on a new _SENTENCE_END constant. Neither existing splitter was right for this: _TERMINAL_PUNCTUATION's comment says "sentence punctuation" while the constant carries ,;:, and clause-level splitting measurably breaks the appositive prose verbless_prose exists for. - check_text and apply_procedure on the MCP server now refuse a text over DENCKRING_MCP_MAX_CHARS characters (default 50000) with a new text_too_long error code, rather than running an algorithm sized for an interactive call against an arbitrarily large remote input (P2-06).

Changed

• Meta.aliases, Meta.languages, Meta.requires and Meta.apply_requires are now tuple[str, ...] rather than list[str], and Meta itself is frozen (P2-01): catalogue.get() returns the same cached object to every caller, and a mutable Meta let one caller's introspection silently change what a registered procedure required for every future check() call.
• Report now raises ValidationError if constructed — or, as of this release, mutated — with satisfied disagreeing with score == 1.0 (P2-02). This was previously only a documented convention that individual procedures were trusted to honor.
• describe.runnable() and apply_runnable() now propagate exceptions other than UnknownLanguage instead of swallowing all of Exception, so a bug in a procedure's own code is no longer indistinguishable from that procedure legitimately not running in a given language.
• denckring-en-data's en entry point is now the factory denckring_en_data:pack rather than the class EnglishDataPack, so the tagger distribution can be composed in without claiming the language itself — the move German made under ADR 0030. The class is still exported and still works; what changed is which object the entry point names.
• describe._SOFT now contains pos, so the two new rows report reading.determinacy as heuristic. A tagger never looks anything up, so a pos row can no more be exact than a syllabic one; without this, describe() told callers a statistical verdict was certain.

0.2.0 - 2026-09-11

Added

• paronomasia — the pun, as the relation between a text and the phrase it displaces. The row is checkability: source: the caller declares the phrase, and the check decides that at least one word was displaced, that no more than max_displacements were, and that each displacement lands inside a requested band of phonetic distance. Whether the result is funny is the writer's claim, the position kangaroo_word already takes on synonymy.

The band has two edges on purpose. 0.0–0.0 asks for a homophone and refuses anything else; 0.4–0.7 asks for a pun that has to work for it and refuses a homophone as too easy. "Make it worse" is a writing constraint the same way "avoid the letter e" is, and a ceiling alone could not express it.

What ships is narrower than the figure, and the row says so rather than implying it. A displacement is checkable only when the word that lands is itself in the pronouncing dictionary — bread for Brad, French Diminu'tif for diminutif. The blends that dominate real punning shop names put a coined word there, and nothing here can pronounce a word it has never seen: hairways, hairitage, barberella, Haarmonie, Chaarisma and atmosphair all raise MissingCapability, measured rather than assumed. That family needs a phonemes.g2p capability, named on the row with no pack behind it — the honest state ADR 0015 describes, and the one homophonic_translation is already in for phonemes.bilingual.

New: core.phonetics, an unweighted Levenshtein over phoneme symbols normalised by the longer sequence. Deliberately not PanPhon's feature-weighted distance, which is more perceptually faithful and would put numpy and pandas behind a package whose dependencies are pydantic, pyyaml and typer. The cost is admitted in ADR 0041: this measure charges the same for /t/→/d/ as for /t/→/m/. ARPABET's stress digits are stripped first, or the reads as a pun on the; the rule strips a trailing digit, so it is a no-op for German and French IPA.

Measured while building it, and recorded because none of it was guessed: the generator needs no search budget, because prefiltering on phoneme length and on a shared first or last phoneme takes English's 45,884 pronounceable graded words down to between 143 and 2,424 per position. Homophone density is an orthography difference rather than a data gap — at distance exactly 0.0, French air offers 13 displacements and temps 7, English air offers 3, and German Komm and Haar offer none.

• Ten trades, up from three: bar, butcher, coffee, fishmonger, florist, nails and pets join bakery, hair and optician — 366 vocabulary words and 48 blends across three languages. A trade now ships a vocabulary and need not ship phrases at all: TradeWords.phrases is optional, because requiring it is what pushed the first three trades into carrying phrases invented for reachability. Seven words were dropped as unpronounceable, including perm, which had been in the hair vocabulary since it shipped and was therefore never usable.

• A bug in the blend's window search, and the two names that were resting on it. _closest_window started at (1.0, 0, 0) and returned it untouched when nothing scored below 1.0 — reporting a zero-length window as its evidence, h aː ɐ̯ for at 1.000, which is a comparison against nothing. It read as a loose match and was no match.

Monhaarlisa and Pawsome shipped on it, each with max_distance: 1.0 and a comment calling them blends you see rather than hear. Measured honestly they are 1.000 from every stretch of their hosts, so both are removed, and the per-blend max_distance that existed only to carry them is removed from the model. All 48 remaining blends check at the row's own default band.

The class is real and no rule reaches it: Monhaarlisa, Pawsome, Yes We Kämm are carried by the letters and by what the reader already knows, not by sound. This row checks the heard relation and now says so instead of opening its band until something passes.

• Both new rows generate now. portmanteau.apply proposes every way of splicing each trade word into the host and hands each candidate to its own check — the project's thesis used as a search. Not knowing where to splice stops mattering if you can afford to try everywhere, and you can: at most 859 candidates, about 0.2s. Kamera gives Kammera, airline gives hairline, curtain gives curltain.

Measured against the attested corpus, the real salon name comes back first for 10 of 12 hosts and is in the first ten for 11. It ships at that number rather than tuned: three ranking fixes were tried and each made the whole result worse — a seam filter loses Coiff'Hair, length-delta starts preferring coinages that delete host letters, replaced-length fixes Chamäleon and breaks heritage. Twelve names is not enough to tune against. ADR 0043 records all three, and a test pins the Chamäleon failure with a note to delete it if it ever starts passing.

amphibologia.apply selects rather than invents, which is what that figure is: it is noticed in something people say, not built from parts. The reader supplies the phrases, one per line (ADR 0020), and the row returns the ones that read two ways for the trade. Given fourteen ordinary phrases and a salon: A Cut Above, Head over heels, Curl up with a book — and not Sheer Delight, because sheer is not the trade's word. Both rows join PARAMETER_GATED, since neither is reachable by a fuzz harness that supplies no trade, and both carry their own round-trip test.

• The attested shop names are quotable, and eighteen of them are now evidence. The maintainer's ruling of 2026-09-10 settles the question ADR 0041 D7 deferred: a business name is a name and not a work, the sources are journalism and public registers rather than anthologies, and quoting one is nominative use.

Externally sourced cases rise from 66 to 84 (11.8% to 13.8%) — the largest single movement in that figure since the corpus began. portmanteau carries 11 across all three languages, paronomasia 5, amphibologia 2. Haarmonie, Kammäleon, SaHaara, Imagin'hair, Diminu'tif, Atmosph'air, Hairport, Hairaphernalia, Curl Up and Dye, Grateful Head, Jack of all Fades, A Cut Above — each naming where it was seen.

Five are satisfied: false and they are the useful ones. Monhaarlisa at 1.000 is the class of name no phonetic rule reaches, kept as the evidence for that rather than admitted by opening the band until it passes. I Kneaded This fails because CMUdict carries knead and needed but not kneaded — a gap in the lexicon, not in the pun. Caldius Biberius Mero fails on Latin proper nouns. Carroll's slithy and chortle fail because a fusion is not a splice.

One correction found while measuring: Atmosph'air splices the French air, not the English hair — the coinage carries air and the joke is what is heard behind it. It checks at 0.000; an earlier attempt named the wrong splice and blamed the row.

• Three defects the suite could not see, found by evaluating the rows against the project's own standards rather than by running them. All three passed 589 golden cases and every gate.

paronomasia cited Peacham's Garden of Eloquence (1577) for a figure that could not be verified to be in it — the house rule is to record what is checked. It now cites Puttenham (1589), "Prosonomasia, or the Nicknamer", whose own example is the row's operation exactly: Claudius Tiberius Nero becomes Caldius Biberius Mero.

portmanteau was primary/author-stated for Carroll and cannot check a single one of Carroll's portmanteaux: slithy, mimsy and chortle all fail splice_not_present. His are fusions, where neither word survives and no host is left standing; this row checks a splice into a host that stays readable. That is the ADR 0015 error — a definition promising more than its checker verifies — so the row is now reference/codified, says so in its notes, and carries two external golden cases recording the rejection.

Two definitions promised what no checker decides: paronomasia said "a known phrase" and amphibologia said "can be taken two ways", where the checker verifies only that a trade word is present and polysemous. Both rewritten in antigram's house style, which states the figure and then names what is left to the writer.

Externally sourced cases rise from 66 to 69 (11.7%), and all three came from checking the citations rather than from adding fixtures. All three are satisfied: false: Puttenham's Latin names are in no pronouncing dictionary, and Carroll's fusions are not splices — each a limit met on the figure's own founding example.

• amphibologia was built, evaluated and cut in the same session, and the reason is worth more than the row was. A phrase that reads two ways because of what the shop sells — A Cut Above over a salon — sourced to Puttenham's Amphibologia, or the Ambiguous (1589), which he files among the vices of style.

It is gone because its checker was a proxy. Polysemy is not a phrase reading two ways: A Cut Above and A Room with a View both carry a trade word with more than one sense, and only one 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 already caught this session, and the honest response to finding it a third time is to delete rather than to reword.

The catalogue holds at 158 rows, 124 implemented; generators go back to 31. ADR 0043's D6 is marked withdrawn and left as written.

• English and IPA are now commensurable, and it was notation all along. CMUdict answers in ARPABET and the German and French packs in IPA, so hair was HH EH R against French ɛ ʁ and no distance between them meant anything. A 39-entry ARPABET_TO_IPA table fixes it — a lookup a reader can check line by line, which is what makes it admissible where ADR 0042 refused a learned model. Rhotics fold across languages and only across languages ({ɹ, ʁ, r, ɐ̯, ɐ, ɚ}), so hair sits 0.333 from French air instead of being incomparable.

This matters more than it sounds. The commonest real French salon name was one this project could not check: tif.hair counts ~150 called Imagin'hair, and the whole -hair family is English inside French. Roughly half the attested German names are the same shape the other way. portmanteau takes splice_lang to reach them; no new capability is declared, because what it needs is phonemes on two packs and that is what it already required. homophonic_translation stays blocked — the notation half of its problem is solved and the larger half is not.

Also in portmanteau: a host may be more than one word (Mona Lisa, Hart am Limit), pronounced word-by-word and all-or-nothing; and the default band opens to 0.7, because attested names sit above the old ceiling. A blend may name its own max_distance, which Monhaarlisa uses — a blend you see rather than hear, Haar covering the a of Mona and sounding nothing like it.

• portmanteau — the coinage that carries one word inside another, and the half of the punning shopfront paronomasia could not reach: Haarmonie, Hairitage, Föhnix. ADR 0041 named phonemes.g2p as what this family was blocked on. It was wrong, and measuring is what showed it. A blend is never an arbitrary string — the host word and the word spliced into it are both ordinary dictionary entries — so the coinage is verified as a derived form and nothing unknown is ever sounded out. No new dependency, no espeak-ng, no learned model, no GPL.

What actually blocked German was the distance metric, not the missing pronunciation. German vocalises a syllable-final /r/, so Haar is h aː ɐ̯ where Harmonie opens h a ʁ — the same segment spelled two ways, which unweighted Levenshtein charges a full substitution for. core.phonetics.EQUIVALENT now records, per language, the symbols that are one segment: German gets {ʁ, ɐ̯, ɐ, r} and Haarmonie moves from 0.667 to 0.333. Vowel length is deliberately not folded — aː and a are contrastive, Staat against Stadt, so collapsing them would buy a distance of 0.000 by discarding a distinction German makes. English and French get no entries at all, because ARPABET has no vocalised-r symbol and Lexique's SAMPA maps to a single ʁ; a table for them would be a guess dressed as data.

Verified additive before it shipped: all nine German pairs the existing fixtures and domain phrases turn on score identically with and without the fold, and a test pins that. The prefilter had to learn the same fold and briefly did not — a prefilter stricter than the measure behind it hides candidates the band would have accepted, silently.

A blend whose coinage is its host is refused as identical_to_host. Crustacean really does contain crust and Breadwinner contains bread, and both passed an earlier draft at a spelling distance of 0.000 — because nothing had been spliced. That is the found pun, which states no relation between two texts because it is one text, and paronomasia's notes already exclude it for the same reason. Found while authoring the shipped blends, not by reasoning about the code.

Still out of reach, and neither is g2p's fault: a host the lexicon does not carry (Barbarella is absent from CMUdict, which is most celebrity blends) and the cross-lingual blend (Atmosph'air), which needs phonemes.bilingual. apply is not shipped though kind says both — generating a blend is choosing where to splice, which is Deri and Knight's multitape-FST problem; the docstring says so where a caller will meet it.

• A pun needs to know what the shop sells, so paronomasia's generator takes a domain. Three trades ship in core — bakery, hair, optician — each carrying, per language, a vocabulary and a set of phrases worth punning on. Every word and phrase was written for this project, so there is no third-party licence and nothing to quarantine under ADR 0013; core/domain.py is the loader, and domain_words takes a vocabulary of your own for a shop nobody catalogued.

The ranking was wrong and is fixed. Candidates were ordered by SCOWL band first, so bread (band 20) came 74th behind brand, bad, it and put (all band 10) whatever their distance — the row that catalogues the figure could not generate its own canonical example. The order is now trade, then distance, then commonness, and Bread Pitt is first. Measured across the three trades, all 37 shipped phrases now displace onto their own trade, each pinned by its own test.

The phrase lists were audited against the attested collections on 2026-09-10 and expanded from 37 phrases to 56, with 33 blends. The audit found six entries that passed every mechanical check and should not have shipped: Klar gewinnt, Weit gesehen and brosse à part were invented and are not phrases anyone says, and Life of Pie and Site for Sore Eyes were hosts that were already the pun, so the displacement restored the original instead of departing from it. Whether a phrase is one a reader knows is the single property here that no test can reach — each is real words in real grammar landing real trade vocabulary — so it is audited by hand and the standard is written into each file.

What replaced them is attested: Curl up and die (Atlas Obscura counts thirty American salons called Curl Up & Dye), Grateful Dead, Jack of all trades, Kopf oder Zahl, and the German blends SaHaara, Kammäleon, Kammbodscha, Kammpus, Philhaarmonie and Scheitelkeiten from a list of 300+ that gives each pun's host. Curl up and die had been dropped by an earlier note arguing that a phrase containing a trade word is "the joke already made" — wrong, because curl being present does not stop die becoming dye.

The commonest real French salon name is one this project cannot check. The -hair family — Imagin'hair at ~150 salons, Atmosph'air, Caract'hair — is English inside French, refused as splice_not_a_word because hair is not a French word. That needs phonemes.bilingual, not phonemes.g2p. Roughly half of the attested German names fall to the same boundary or to a proper-noun host. Recorded in docs/research/ rather than inferred from the ones that worked.

Two things decide output quality that no rule in the checker can express, both found by measuring rather than reasoning: displacing a function word produces nonsense — French la is 0.333 from laque, which offers laque vie en rose for half the language — and a phrase that already contains its own trade word has nothing left to displace (Curl up and die, Das Brot). A third applies to German alone: every German frame is article-free, because a German article agrees with its noun's gender and nothing here knows a noun's gender, so displacing the noun out of Der Wille or Zum Wohl left the article stranded and produced Der Welle and Zum Welle, which are not German. Komm rein, Kopf hoch and Locker bleiben cannot break that way whatever lands in them — a constraint on the phrase instead of a gender table nobody has. The shipped phrases are chosen against all three. German nouns get their capital from the trade file, because German's graded words are an entirely lower-case frequency corpus and a displacement spelled as that list spells it reads Alles haar.

The catalogue goes to 157 rows and 123 implemented, the golden corpus from 560 cases to 572. The externally sourced share falls from 11.8% to 11.5%, which is arithmetic and not a regression: no case here claims external provenance, because no example has been read off a page in this pass and an example nobody has read is the invented citation the house style refuses. The attested modern corpus of the figure is entirely post-1929, so it waits on the same decision about quoting in-copyright material that research batch 3 already has pending.

0.1.1 - 2026-09-08

Fixed

• A refrain is the same line returning, whatever punctuation its position wants. form_report, rondeau and pantoum compared repeated lines after nothing more than strip().casefold(), so a refrain that came back with a different stop was reported as broken_refrain. That is not an edge case, it is how the fixed forms are printed: Passerat's villanelle (1606) closes its refrain Tourterelle:, then Tourterelle., then Tourterelle, as the syntax around it changes, and Ranchin's triolet returns du mois de mai once bare and once as de mai !. The three sites now compare core.text.line_identity, which strips terminal sentence punctuation and nothing else — apostrophes stay, because in French the elision is part of the word (l'âme, i'oy) and folding it would compare something other than the line.

Four canonical texts went from rejected to accepted: Ranchin's triolet, Leconte de Lisle's pantoum and McCrae's rondeau all reach 1.000, and Passerat's villanelle rises from 0.922 to 0.980 — its one remaining broken_refrain is a real textual question, aprés ell against aprés elle, which the fix correctly does not paper over. No existing verdict moved: all 34 constructed cases in the refrain-bearing rows repeat their refrains byte-for-byte, because the same author typed both copies, which is exactly why none of them could ever have found this.

• mutants/, which mutmut fills with 5.6 MB and leaves behind, is excluded from the source distribution and from git. Untracked, and uv build reads the working tree rather than the index, so 312 of its files landed in the archive and pushed the sdist from 0.9 MB to 1.2 MB; test_the_sdist_stays_small is what noticed. The fourth instance in this repository of untracked is not absent, after superpowers/, stage_mockups/ and feedback/. The exclusion was verified with a 6.1 MB mutants/ present rather than after deleting it.
• A text's Unicode normalisation form no longer changes a verdict. letter_spans walked the string character by character, so a decomposed ä — a followed by U+0308 — presented its base letter, which is alphabetic, and dropped the combining mark, which is not. With fold_diacritics off, the setting whose documented purpose is that ä stays ä, the diacritic was folded anyway and only for text that arrived decomposed: check("lipogram", "Bär", forbidden="a", fold_diacritics=False) was satisfied in NFC and unsatisfied in NFD. prisoners_constraint carried the same defect independently, where folding is never right at all — Bär scored 0.333 in NFC and 0.667 in NFD, and the decomposed reading listed one violation where there are two. Both now read base characters together with the combining marks that belong to them, at the base character's offset, so a Violation still points into the text the caller passed.

Added

• Fifteen texts from research batch 3 enter the golden corpus across thirteen rows, taking externally sourced cases from 51 of 545 to 66 of 560 (9.4% to 11.8%) and the rows carrying any external evidence from 40 to 54 of 122. Eight pass — Wagner's alliterative verse, Gaunt's anaphora, Lincoln's epistrophe, Platen's ghazal, Leconte de Lisle's pantoum, McCrae's rondeau, Ranchin's triolet, and courtship for the charade. Seven ship satisfied: false with the reason read off the run rather than written from memory: the chronogram's numerals sum to 4204 and not the 1603 it claims, homoteleuton compares final letters where Peacham's figure turns on a syllable, and spoonerism declines because neither Kinquering nor Congs is in the pronouncing dictionary.

Passerat's villanelle is the honest one. It scores 0.980 on a single broken_refrain, line 24 reading aprés ell where line 3 reads aprés elle, and its source says in as many words that whether this is the 1606 spelling or a transcription slip is not verified against the print. The case ships as supplied, with the question open.

Three of the batch did not land, and none for want of a checker: wechselsatz was given the string Kuhlmann 1671 where the parameter wants the template of interchangeable slots, denckring was given Harsdörffer's instructions to the bookbinder rather than words built from the rings, and the sestina was six lines of a thirty-nine-line form.

EXTERNAL_FLOOR rises from 51 to 66, and the new satisfied: false cases carry min_score floors — except where the floor would have been 0.0, which guards nothing.

• Mutation testing, via mutmut, configured and deliberately outside the four-command gate (ADR 0039). Coverage says a line ran; this says whether a test would notice the line being wrong, which for a library of checkers is the question that matters. No score is published yet: the first run was cut short by its own timeout, and its most promising survivor turned out to be an equivalent mutant — verified over 9,757 generated inputs — so survivors are a worklist requiring judgement, not a defect count.
• Report.provenance and Production.provenance: the package version, a schema version that moves independently of it, which pack answered and which data distributions it was reading, the fold_diacritics policy in force, and the seed a drawing generator used. Stamped once at check and produce rather than in each of the places a report is built, so pangram, which builds its own, carries it too.
• LanguagePack implementations declare data_distributions. A tuple rather than one name because German is three of them — GermanWiktionaryFrequencyPack layers the lexicon, the pronunciations and the frequency bands — and a single version would be a lie about two. Read with getattr rather than added to the LanguagePack protocol, which is runtime_checkable and a published contract, so a third-party pack that predates this keeps satisfying it.
• Report.evidence: which words a syllabic verdict counted, what it made them, and whether the number came from a dictionary or a spelling heuristic. metrics["estimated_words"] said how many words were guessed and never which, which is enough to know a verdict is soft and not enough to repair it or audit it. basis is a closed set rather than a confidence score — the pack knows which of two ways it got an answer and does not know a probability, and attaching one would invent precision no measurement here supports. Empty on the exactly decidable rows, which have nothing to explain. Metre reports the stress each word was read as, rhyme the keys each line ending was judged on — including both when a word has two pronunciations, since reporting one would make the verdict look arbitrary to anyone checking it against a dictionary. 36 of the 41 heuristic rows carry it; assonance_constraint, proteus_verse and spoonerism reach their verdicts by paths of their own that this did not thread, and are named here so the gap is a record rather than a surprise.
• describe().reading: the three things a specification should record that describe did not — normalization (the fold policy in words), tokenization (the pack's word pattern), and determinacy, exact or heuristic, which is 81 rows against 41. Derived rather than authored, for the reason runs_in is: a hand-written answer drifts from the code it describes and nothing notices.
• LanguagePack.syllable_evidence, the per-word breakdown behind line_syllables' two numbers. FrenchDataPack overrides it to return one line-scoped entry, because a final mute e elides or counts depending on what follows, so a word's contribution is not a property of the word (ADR 0034) and there is no per-word breakdown to report. Read with getattr, so a third-party pack written before it reports nothing rather than failing.
• tests/test_canonical_verse.py, a conformance corpus in miniature: seven pentameter lines from out-of-copyright poems, none written for this suite. Three of the seven are not read as ten syllables, for three separate reasons — Milton's synaeresis (disobedience is four syllables to him and five to the dictionary, exactly and with no estimate), a feminine ending in Keats, and a dictionary miss on Pope's Pierian. The second is a stated limitation of the catalogue row; the third is now visible through Report.evidence; the first is undocumented and has no treatment at all — French got elision in ADR 0034 and English and German never got the equivalent. The tests pin what the package reads rather than judging the poetry, so a future change shows as a diff here.
• denckring.core.text.clusters, the base-plus-marks reading the fix above is built on.
• Golden cases record a provenance — external, constructed or self-generated — and denckring eval reports the split on a third line: 38 externally sourced (7.1%), 491 constructed, 3 self-generated of 532. The pass count says how much agrees with the implementation; only the first of these says how much of that agreement is evidence the constraint was read correctly. tests/conftest.py loads source instead of dropping it, and a guard now refuses a case that names none.
• The value is explicit rather than inferred from the source prose: deriving it by looking for an English "constructed" prefix mis-filed a German case reading konstruiertes Beispiel, and three cases that say they were produced by the row's own apply and accepted back by its own check.
• tests/test_normalisation.py, pinning that the two forms agree, and pinning what was already true of zero-width spaces, byte-order marks and non-breaking spaces — that they are not letters, do not join the cluster before them, and do not shift offsets.
• ADR 0040, deciding how strictly canonical verse is read, and MetreParams.feminine_ending, the parameter that answers its first question: a klingende Kadenz — one unstressed syllable past the counted length — becomes an admissible reading rather than a rejected one, on the seven rows that hard-code a metre. Defaulting to false, it reaches blank_verse, shakespearean_sonnet, rhyme_royal, curtal_sonnet, sonnet and spenserian_stanza through with_feminine, which offers line_metre the pair {pattern, pattern + "0"} rather than a fixed longer string — a single string cannot span both, since Gryphius rhymes a feminine Erden/Herden (13 syllables) against a masculine ein/seyn (12) in the same sonnet. alexandrine, which counts syllables rather than scanning stress, takes the same flag as pattern_result's extra. Measured against 42 German canonical lines, against the bar ADR 0040 D1 set before the work — at least 95% within their declared length or one over — the reading came in at 97.6%; the full stress-and-length scan, a stricter question D1 leaves open, rises from 35.7% to 83.3% over the same lines. Every default is today's reading, so the 532 pre-existing golden cases are unchanged by construction (D6).
• curtal_sonnet and elegiac_couplet gain a catalogue notes field declaring the verse each cannot scan (ADR 0040 D4), rather than the stress model moving to fit it. Hopkins wrote "Pied Beauty" in sprung rhythm, which curtal_sonnet reads against plain iambic pentameter and so rejects — the poem that named the form does not satisfy the row named after it. elegiac_couplet stays declared for English only, because German accentual hexameter puts the ictus where word stress does not fall — Schiller's own Hexameter scans 0100 and carries the beat regardless — and no German fixture lands for the row.
• tests/fixtures/canonical_lines.yaml, 111 canonical lines — 69 English, 42 German, none written for this suite — and tests/test_canonical_verse.py asserts both bars ADR 0040 measured against them: the German length floor (95%; measured 97.6%) and the English ceiling on true disagreement with no dictionary guess involved (3 of 69, 4.3%).
• 13 canonical texts enter the golden corpus — Crapsey, Bentley, Hopkins, Verlaine, Chaucer, du Bellay, Shakespeare, Spenser, Dante via Sibbald's translation, Kuhlmann, and Gryphius and Goethe in German — taking denckring eval --all from 532 to 545 passed and the externally sourced share, ADR 0040's opening measurement, from 38 (7.1%) to 51 (9.4%). Five of the thirteen pass outright; the other eight ship satisfied: false with the measured reading recorded in source — Hopkins' sprung rhythm, a Middle or Elizabethan English word the dictionary cannot place, a rhyme (Chaucer's, Sibbald's, Shakespeare's) this package's strict model does not accept, and three German cases whose feminine ending closes the length gap but leave a stress or rhyme violation standing. A rejection earns its place in the corpus the same way a pass does — tests/test_corpus_provenance.py now checks that a satisfied: false case's source names only a violation rule the checker actually raised for it, after three drafts of these cases were caught citing a rule that never fired. EXTERNAL_FLOOR moves from 38 to 51 with them.

Changed

• line_metre's wrong_line_length violation names every syllable count a candidate metre would accept, not only the winning candidate's. alexandrine's syllable path already reported "12 or 13 syllables" with feminine_ending on; the stress path reported only "12 syllables", so a caller who had turned the parameter on was told to write a line the code would in fact have accepted one longer than. Generalised to N candidates rather than the two feminine_ending introduces, so a row with more than one substitutable foot now reports the full list, e.g. "13, 14, 15, 16 or 17 syllables".

0.1.0 - 2026-09-07

Added

• Core protocol: Violation, Report, Meta and the LanguagePack interface, all serialising to stable JSON for non-Python callers.
• BaseProcedure, a template method that validates language, capability and parameters before delegating, so a procedure module cannot forget those checks.
• Autodiscovering registry over denckring.procedures, enforcing one module per procedure with the module name equal to the procedure id.
• Catalogue of 31 sourced procedures in src/denckring/data/catalogue.yaml, licensed CC BY 4.0, as the single source of truth for procedure metadata.
• English language pack, shipping in core with no data files, declaring tokens, alphabet, fold_diacritics and letter_shapes.
• Batch 1 — twelve lexicon-free restriction procedures: lipogram, univocalic, tautogram, pangram, heterogram, palindrome, snowball, reverse_snowball, prisoners_constraint, beau_present, acrostic, telestich.
• Three-level eval harness: golden fixtures, Hypothesis strategies, and a scoreboard that exits non-zero on regression.
• CLI: check, apply, describe, list, search, show, status, eval, catalogue export.
• scripts/new_procedure.py, which scaffolds module, test, strategy, fixture and catalogue row from templates. It is a contributor tool and stays outside the package: it writes into src/denckring/procedures/ and appends to this repository's catalogue.yaml, so as the shipped denckring new it put a command in every user's CLI that could do nothing for them.
• Eight architecture decision records under docs/adr/.
• German language pack, shipping in core with no data files and discovered through the denckring.lang entry-point group — the same path a third-party pack takes.
• fold_diacritics, a per-call parameter on the nine letter-comparing procedures, deciding whether ä counts as a and ß as ss. Defaults to true.
• LanguagePack.exceeds_x_height, and golden cases in German for all twelve procedures.
• Two further ADRs, 0009 and 0010, on the fold parameter and entry-point discovery.
• Catalogue grown from 31 to 145 sourced entries across eight families.
• family, aliases and attribution on every catalogue entry. attribution distinguishes a traced origin (primary) from a form documented in a standard reference (reference) from one with no single origin (traditional).
• denckring search matching ids, names and aliases; denckring list --family.
• denckring catalogue export --format json|csv emitting the dataset as a standalone file, and LICENSE-DATA carrying the CC BY 4.0 notice.

• A catalogue quality suite: every row must have a source, an English name, definition and prompt hint, a family and an attribution; aliases must be unique catalogue-wide and must not collide with any procedure id.

• checkability on every catalogue entry: self, source, or none for forms with no computable acceptance criterion. Coverage is now measured against the implementable subset, so denckring status no longer promises a gap that cannot close.

• 31 further procedures implemented, taking the total from 12 to 43 — letter and alphabet constraints, position and repetition constraints, the end-word forms (sestina, quenina, pantoum), and the source-relative procedures.
• SourceParams and denckring check --source FILE, for procedures decidable only against the text they were made from.
• apply() on cut_up and every_nth_word, and a round-trip suite asserting that check(apply(text)) is satisfied — the seed's central property, deferred since the Batch 1 spec for want of any constructive procedure.
• Constructive, a runtime-checkable protocol for procedures that can generate.
• ADR 0011 on checkability.
• Syllable counting: syllables.heuristic in core, with syllable_count returning a count and whether it was looked up or estimated. Every syllabic report carries metrics["estimated_words"].
• denckring-en-data, installed by pip install denckring[en], adding syllables.dictionary from the CMU Pronouncing Dictionary (BSD-2-Clause, vendored). On Bashō's frog-pond haiku, core alone estimates 13 words; with the data package, none.
• Eight syllabic procedures: haiku, tanka, senryu, cinquain, syllable_count, monosyllabic_prose, hendecasyllable, alexandrine.
• ADRs 0012 and 0013 on the two syllable capabilities and on data as separate distributions.
• Rhyme and metre: rhyme_key, rhyme_keys, stress_pattern and stress_patterns on the data pack, plus the stress capability. Twelve procedures — rhyme_scheme, iambic_pentameter, trochaic_tetrameter, blank_verse, heroic_couplet, terza_rima, villanelle, triolet, limerick, shakespearean_sonnet, petrarchan_sonnet, rhyme_royal.
• Metre is checked as a satisfiability question: polysyllabic stress is fixed, monosyllabic stress is free, and every listed pronunciation is searched. These are the first procedures a core-only install genuinely cannot run — without denckring[en] they raise MissingCapability naming the extra.
• ADR 0014 on metre as satisfiability, and on what strict scansion does not accept.
• Release machinery: GitHub Actions CI (lint, typecheck, a 3×3 test matrix, a separate denckring eval --all gate, a core-only install job, a docs build and a distribution build), tag-triggered Trusted Publishing with attestations, and Dependabot.
• A generated documentation gallery: one page per catalogued procedure, built from the catalogue and the golden fixtures, so every worked example is one the test suite enforces.
• CITATION.cff, SECURITY.md, CODE_OF_CONDUCT.md, issue and pull request templates, and a .pre-commit-config.yaml mirroring the CI lint and typecheck jobs.
• lexicon.words and lexicon.nouns capabilities, backed by WordNet's 55,239 single-word noun lemmas (Princeton licence) added to denckring-en-data.
• Five procedures: n_plus_7, s_plus_7, charade, semordnilap, word_square. N+7 reports ambiguous_words, because a word list cannot tell you that run is a verb in this sentence.
• ADRs 0015 and 0016 on lexicon capabilities and on one data pack per language.
• apps/explorer, a local FastAPI browser for the catalogue and bench for the procedures. Outside the distribution: nothing published depends on it.
• Combinatorial devices: ordered slots of alternatives, asked either to segment a word or to select one alternative per position. denckring, cent_mille_milliards and wechselsatz implemented on it, with Harsdörffer's 1651 rings shipped in core. denckring also generates, so spinning the rings round-trips through its own checker.
• Catalogue rows for ars_combinatoria and llull_figure, the tradition the Denckring descends from.
• ADR 0017 on a device's data shipping with its procedure.
• attested on every catalogue entry: author-stated, codified or reconstruction. It records whether anyone ever set the procedure down as a rule, which attribution — about how the source was established — does not answer.
• serial_lipogram, the whole-work form attributed to Tryphiodorus and excerpted by Jean Paul in 1783.
• ADR 0018 on attestation.
• notes on catalogue entries, for contested figures, reception history and caveats — so a definition can go back to being a definition.
• Combinatory figures: an alphabet read at several levels at once. Llull's ternary Ars ships with all six of its tables, and Kircher's Ars Magna Sciendi as a preset of the same mechanism. llull_figure checks and generates chambers; counts are computed from the letters (84 ternary chambers, 36 pairs) rather than quoted.
• A corpus layer: Corpus, Entry and a parser reading JSON or plain text. No corpus ships with the package — the loader is here, the reading is the reader's.
• ideenwuerfeln, reconstructed from Jean Paul's notebook of February 1795, and witz_metaphor catalogued as the step it stops short of. The first entries to carry attested: reconstruction.
• ADR 0020 on corpora not being shipped.
• pasigraphy, Kircher's universal writing: a sentence sent across a numbered vocabulary, with the losses counted rather than smoothed — words with no number, numbers with no entry at the far end, and words reachable from several numbers, where two distinct things arrive as one. The vocabulary is supplied by the caller.
• arca_musarithmica: a phrase is measured, the tablet for that length is found, and the chosen pattern must be one that tablet offers. The patterns are never read, so Kircher's pitch numbers and a set of stress patterns work equally well in the same table. No tablet ships.
• ADR 0021 on implementing the transferable part of a procedure rather than the artefact.
• belle_absente: Perec's form, one line per letter of a dedicatee's name, each line omitting exactly that letter and using every other letter of the alphabet, so the missing letters read down the poem as the name itself.
• serial_lipogram: the whole-work form attributed to Tryphiodorus, as many parts as the alphabet has letters, the omitted letter walking one step per part. paragraph_spans in core/text.py gives a part more than one line to work with, the way Tryphiodorus's twenty-four books do.
• pangrammatic_window: the shortest naturally occurring stretch of text holding every letter of the alphabet, scored against a caller-stated max_length rather than a hidden constant — without a bar this is pangram under a second name.
• paragram, with a lexicon-gated generator: both halves of a one-letter word swap are left in the text for check to find, and apply ranks every candidate swap by pronounceability, noun membership and length instead of taking the first match. Checking needs no lexicon; generating does, so apply carries its own capability requirement rather than the catalogue row claiming one check never needed. The new NoCandidateWord error is raised when the search finds nothing to swap, rather than returning text with no swap in it for check to then reject.
• dactylic_hexameter: six feet, the first four either a dactyl or a spondee, the fifth a dactyl, and the sixth a spondee or trochee. This yields thirty-two readings spanning 13–17 syllables.
• elegiac_couplet: a dactylic hexameter answered by a pentameter. The pentameter is two hemiepes; only the first admits the spondee. That asymmetry is the form.
• sapphic_stanza: three hendecasyllables and an adonic, 11/11/11/5.
• alcaic_stanza: two hendecasyllables, an enneasyllable and a decasyllable, 11/11/9/10.
• double_dactyl: two quatrains, with one line of the second being a single six-syllable word. The nonsense opening, the named person and the rhyme are not checked, and the docstring says so.
• renga: alternating 5-7-5 and 7-7 stanzas. Its links parameter is a minimum, not an exact count.
• haibun: prose and haiku alternating.
• The four new metrical procedures—dactylic_hexameter, elegiac_couplet, sapphic_stanza, alcaic_stanza—are checked as English accentual verse. Classical quantity is vowel length, which English does not have, so stress stands in for it. This is the batch's central editorial decision.
• ? now means "either" on the pattern side as well as the word side—the classical anceps. Before, it was honoured only for a free monosyllable, so a pattern written with anceps rejected every fixed-stress word at that position. Sapphics and alcaics both need it.
• A word not in the pronouncing dictionary no longer aborts the scan. pack.stress_patterns raised MissingCapability for such a word—naming a capability the pack actually provides—so one ordinary English noun killed the whole check. Longfellow's own hexameter could not be checked. Unknown words now scan as free and are counted in a new estimated_words metric, mirroring the contract the syllable path has always had. This changed previously deliberate behaviour: a test named test_an_unknown_word_raises_rather_than_guessing was replaced. Callers who want strict scanning can reject on estimated_words > 0.
• word_stress, feet and stanza_violations added to core/prosody.py. word_stress handles the out-of-dictionary repair; feet enumerates substitutable foot readings, making a metre a set of readings rather than one; stanza_violations applies one pattern per line, since a sapphic changes shape from line to line.
• estimated_words now reported by all twelve form-report callers—blank_verse, heroic_couplet, iambic_pentameter, limerick, petrarchan_sonnet, rhyme_royal, rhyme_scheme, shakespearean_sonnet, terza_rima, triolet, trochaic_tetrameter, villanelle—where before it was computed and discarded.
• Catalogue rows for the five procedures that scan stress — the four metres and double_dactyl — gained stress in requires, since scanning stress reaches pack.stress_patterns. renga and haibun did not — they count syllables only.
• describe(id) and summaries() in core, returning the catalogue as a machine reads it: a definition, the prompt hints, whether this install can actually run the row and what it lacks if not, and the parameter schema as JSON Schema. Four callers needed this assembly — the CLI's show, the explorer, the MCP server and the skill — so it lives in one place rather than being computed four times and drifting four ways. list_procedures() is untouched: it is public and still returns list[str].
• A stable code on all fourteen error classes and a to_dict() on the base, so a failure crosses a process boundary as {"code": "invalid_params", "detail": {...}} rather than as a traceback. A model can act on the first and cannot act on the second. UnknownProcedure additionally carries near-match suggestions, so a mistyped lipogam answers with lipogram instead of a dead end.
• denckring describe <id> --json and denckring list --json, so the skill and any other non-Python caller reach the same two functions the MCP server does. --json refuses to combine with --kind or --status, which filter the catalogue rather than the implemented set: the flag would otherwise be accepted and silently ignored.
• An MCP server behind pip install denckring[mcp], run as denckring-mcp. It exposes four tools — list_procedures, describe_procedure, check_text, apply_procedure — taking the procedure as a parameter, rather than one tool per procedure. Eighty-six tool definitions would sit in every client's context, and model performance degrades sharply with tool count. No tool raises: each catches DenckringError and returns its dictionary form.
• A Claude Code skill at skills/denckring/SKILL.md, which shells out to the CLI and so needs no extra beyond the package. Both transports are deliberately thin — neither computes anything, so the two cannot disagree about what a procedure is.
• Twenty-eight procedures implemented, taking the total from 86 to 114. Eleven fixed forms: sonnet, ottava_rima, spenserian_stanza, ballade, rondeau, ghazal, curtal_sonnet, englyn, clerihew, alliterative_verse, assonance_constraint. Three clusters of source-comparison rows: letter-class (homoconsonantism, homovocalism), selection (diastic, mesostic, column_reading, haikuization) and rearrangement (boustrophedon, text_folding, mathews_algorithm). And a remainder: fold_in, univocalic_translation, lipogrammatic_translation, larding, multiple_constraint, tmesis, spoonerism, word_ladder.
• core/source_compare.py: letter_class_report, selection_report and rearrangement_report, the three shapes eight of this batch's rows share. Each was extracted from a row already written, not designed ahead of one — homoconsonantism was written by hand first, and writing homovocalism against that code showed the entire body was identical but for one predicate, which is why letter_class_report exists at all. Precedent: form_report, extracted from real sonnets the same way.
• lines= on form_report, so a fixed form with a prescribed length reports wrong_line_count and stops rather than scanning a scheme or metre against lines that are already the wrong ones — matching stanza_violations's existing behaviour.
• multiple_constraint, Oulipo's general composite-constraint form. It replaces univocalic_lipogram_pair, whose definition described any composite constraint rather than the one pair its id named. denckring search finds the row under the old id, under the old id written as prose, and under the row's previous published English name, Compound constraint; all three are carried as aliases. denckring show univocalic_lipogram_pair still raises UnknownProcedure, because registry.get resolves ids and never consults aliases — but it names multiple_constraint as the suggestion.
• German declared on seventeen rows, each decided on whether the procedure means something in German rather than whether it merely runs, and each shipping a German golden fixture — a declaration with no fixture behind it is an assertion, not a claim. Every one also carries names.de and definitions.de, so the German a row claims appears in the row a German reader reads, and a test in the catalogue quality suite requires that of every declared language. semordnilap, which declared de with no German name since long before this batch, is corrected with them. cent_mille_milliards joins them from outside the batch: check already accepted a German poem against German strips while the catalogue still declared languages: [en], an assertion wrong in the opposite direction from the others'. It now declares languages: [en, de] with names.de and definitions.de, backed by a German case pair in its golden fixture — a full fourteen-position poem the checker accepts, and the same poem with one line swapped for an alternative from a different position, which it does not — built from the real strips written for the stage scene. Rows declaring German: 38.
• tests/test_requires_honesty.py, guarding requires in both directions. One test watches through the real check() path whether a row's golden fixtures ever reach the capabilities it declares; it caught limerick on its first run against every implemented procedure, not just this batch's. The other runs each row's own fixtures against a pack carrying exactly what the row declares and nothing more, so a row that reaches further than it promised raises rather than passing quietly on the full en pack. Their shared blind spot, documented in the module: a row that calls a capability's method and discards the answer is invisible to both — the gap haikuization fell into below.
• lexicon.glosses on LanguagePack: every definition Open English WordNet records for a word, or an empty sequence when the lexicon cannot resolve it at all. Every sense, not the first — which sense a writer meant is not recoverable from the text, so a reader that accepts any of them is the honest one, the same reading rhyme_keys and stress_patterns already take. It costs 1.03 MB over shipping first senses only. Resolution tries the word as given, then its lemma, then plain s/es stripping; irregulars such as went stay out of reach by design and are disclosed rather than guessed at.
• packages/denckring-en-data/scripts/build_lexicon.py, the build the data package never had. nouns.txt was an opaque blob extracted by hand at some point in the past with no record of how; it is now generated, alongside glosses.txt.gz and a metadata.json carrying source, date and row counts, by a script in the repository. A test pins the shipped nouns.txt against what the script produces, so a hand-edit is caught rather than inherited.
• kangaroo_word: a word carrying a synonym of itself inside its own letters, in order (encourage carries urge). It needed no thesaurus — the synonym is a synonym parameter, the writer's claim, and what the row decides is the decidable half: that the claimed synonym is a real word, hidden in order, and not the word itself.
• definitional_expansion: each substantive word of the source replaced once by one of its dictionary definitions. The comparison was chosen by running it against real gloss text rather than at the desk — glosses carry parentheses, semicolons and colons no writer retypes character for character, so a gloss counts as present when its own lower-cased tokens appear as a contiguous run, punctuation and case folded away, word choice and word order kept. Occurrences are counted rather than merely found: a word appearing twice needs two expansions, and each match consumes the span it used.
• definitional_literature: the same procedure fed back into itself, which is the form Bénabou and Perec described. check is given the source and the finished text and nothing in between, so it discovers the depth — one round, then two, then three — and reports the first that accounts for every resolvable source word in an iterations metric. Below the top level a round means "some gloss of this word has all of its own substantive words present one round shallower"; word order is checked where the gloss must appear as a run and not below that, because pinning the nesting positions would mean parsing a text where every definition is stitched to the next with nothing between them. The cap of three rounds is measured, not guessed: against a genuine three-round expansion of the cat sat (3,679 tokens), depth 3 takes 0.15 s, depth 4 takes 18 s and depth 5 takes 219 s, because the cost falls on the searches that fail and cannot stop early.
• lexicon.synonyms and lexicon.antonyms were specified, measured, and deliberately not built, with the measurements recorded in the catalogue rows that wait on them rather than merely omitted. WordNet synonymy is synset co-membership: of twelve substitutions a writer would naturally make, five are co-members, and only four of eleven content words in plain prose have any synonym at all. Antonymy is thinner still — 6,633 lemmas have any antonym, and 2 of 11 content words in plain prose do. "Every substantive word replaced by a synonym" is not satisfiable by ordinary text, and a larger thesaurus does not change that, so synonymic_substitution, antonymic_substitution and antonymic_translation say so in their notes.
• The code is licensed Apache-2.0 rather than MIT, decided before the first release because relicensing afterwards needs every contributor's consent. What MIT does not do is the reason: it asks only that a notice be retained, so a fork that keeps a line in a file has met it while presenting the work as its own, and it is silent on trade marks, leaving the project's name — the one thing a fork cannot take without renaming — unprotected. Apache-2.0 adds a NOTICE every derivative distribution must carry, a section reserving the name, and a patent grant. The catalogue stays CC BY 4.0, and ADR 0006's data rule is untouched; ADR 0024 records the change. Per-file licence headers are deliberately not used.
• Each distribution now declares what it actually contains rather than only what its author wrote: denckring is Apache-2.0 AND CC-BY-4.0 and ships LICENSE-DATA in the wheel, so the catalogue's attribution terms reach people who install the package and never see the repository; denckring-en-data is Apache-2.0 AND CC-BY-4.0 AND BSD-2-Clause for Open English WordNet and the CMU Pronouncing Dictionary; denckring-de-data is Apache-2.0 AND CC0-1.0 for Wikidata Lexemes. Both data packages gained a NOTICE naming which file carries which terms and saying plainly that the derived lexicons are modifications of their sources.
• The core source distribution no longer carries the workspace members. It was 7.1 MB because it bundled packages/ and apps/ wholesale — the CMU dictionary, the WordNet glosses and the German lexicon a second time, under a licence expression that did not describe them and inside a package that never reads them. Excluding both leaves 0.6 MB. tests/ and docs/ stay, since a downstream packager building from source runs the suite. tests/test_packaging.py builds the sdist and reads it, because a test asserting that the pyproject contains an exclude pattern would pass while the build ignored it.
• A stability statement in the README, naming the four surfaces treated as contracts through 0.x — procedure ids, Report as JSON, the catalogue export schema and the denckring.lang entry-point group — and what is expected to move: violation rule strings, metrics keys, message wording and denckring.core. The README already promised "stable JSON for non-Python callers" and left the scope of that promise to be guessed at.
• A naming policy in the README. Apache-2.0 section 6 reserves the name and says nothing about permitted use, so the reservation is stated as something that can be complied with: redistribute unmodified under the name, describe compatibility freely, rename before publishing a modified version.
• denckring --version. __version__ was exported from the library and unreachable from the command line.
• poesie_automat, Enzensberger's Landsberger Poesieautomat: a flap-board of six lines by six modules by ten alternatives, so 10^36 six-line poems. check cuts each line into its six modules and asks whether every piece is one of that module's ten alternatives; apply presses the button. The count is recomputed from the shipped modules rather than quoted, and Device.combinations returns it as an integer — metrics["combinations"] is a float and cannot hold a 37-digit number exactly. Enzensberger died in 2022 and his word lists are in copyright until 2092, so all 360 fillers in data/devices/poesieautomat_2000.yaml were written for this project; the mechanism is what the row catalogues, on the same footing as the substitute strips the cent_mille_milliards fixture carries for Queneau. Three design rules make every one of the 10^36 readings grammatical German rather than most of them: the article is folded into the noun, everything after the subject is an adjunct, and every verb is intransitive third person singular present. A fourth constraint is about the display rather than about German: the board renders letter by letter into character cells, so no filler exceeds eleven characters and no line the modules can assemble is wider than 71 columns — both pinned by test_no_flap_is_wider_than_the_board. Two further constraints keep the strangeness on the right side of broken: no word reaches a line from two modules, compared with German inflection folded, and each family of mutually exclusive time anchors — when on the clock, since when, from when, how long ago, which month — sits in one module per line, so Der Frost gedeiht is possible and um fünf ... um drei is not, while roles that differ still stack.
• An optional line on Slot, with Device.lines and Device.for_line, so one device can hold several lines. Additive: a device file that never mentions line — which is every existing one — reads as a single line and behaves exactly as before.
• An optional separator on device.segment, defaulting to the empty string. The Denckring's rings concatenate with nothing between them; a board whose flaps carry whole words sets it to a space, and the same backtracking walk then splits a line into modules rather than a second segmentation being written.
• DENCKRING_DEVICE_PATH, a colon-separated list of directories device.load searches before the packaged one — a device's word-lists no longer have to be ours. Directories earlier on the path are tried first, so one can both add a new device id and shadow a packaged one under an existing id, which is documented as deliberate rather than left to be discovered. Unset, behaviour is unchanged: the packaged directory is still all that is searched. A directory on the path that does not exist or cannot be read is skipped quietly, so a stale entry does not stop a packaged device from loading. What it does not add: no schema versioning beyond ordinary validation, and no record of where a resolved device actually came from — a caller who needs to know whether a packaged device was shadowed must control what it puts on the path, because load cannot say so after the fact. The environment is read on every call rather than cached at import time, and the on-disk read load used to memoise by device_id alone is now memoised by the resolved path instead — a cache keyed on the id would have kept answering a changed environment with whatever device it saw first. A device or figure id is validated to a bare filename stem before it is ever joined to a directory (see Fixed, below), so a cartridge lives in a directory on the path and cannot be addressed by a path of its own; a relative directory on the path resolves against the process's working directory at call time, so changing directory after setting the variable can turn into silent misses under the same skip-quietly contract.
• ConstructiveProcedure, the template method for apply that check has always had: it resolves the pack, enforces both capability lists, validates parameters and refuses output that misrepresents what ran — identical to the input, or empty from input that was not — so a generator cannot forget any of it. parse_apply_params also refuses a caller-supplied source: the text being transformed is the source, so a second one names two texts for one argument, and passing both now raises InvalidParams instead of one silently winning. ConstructiveProcedure.ignores_input opts the three devices that never read their text — denckring, poesie_automat, llull_figure — out of the identity half alone, where the comparison is against an argument nothing read. ADR 0025.
• Meta.apply_requires, so a generator can declare a capability its checker does not need — anagram generates only with a word lexicon and checks with core alone, and one list could not say both.
• Description.constructive and Summary.constructive, reporting whether this install has a generator, distinct from kind, which says whether the form admits one. Nine rows are honestly both with no generator here.
• Description.apply_params, the generator's own JSON Schema beside the checker's params, so denckring describe and MCP's describe_procedure finally show seed and allow_identity — which live on apply_params_model() and reached no non-Python surface at all. Empty for a row with no generator. The two schemas stay separate because source is a checker parameter apply supplies for itself and refuses from a caller.
• DegenerateOutput and InputTooShort.
• Production, Report's counterpart for the generating half: procedure, texts (best first, never empty), truncated, metrics. ADR 0026.
• produce() on every generator, and denckring.produce() at the package level, returning every result a generator found rather than only the first.
• denckring.apply() at the package level. check has been exported since the beginning and apply never was, so every Python caller reached through the registry to generate anything.
• denckring apply --json, emitting the Production, matching describe --json.
• max_results on ApplyParams, defaulting to 10.
• NotConstructive, replacing a dict literal the MCP server built by hand — one failure mode in this package was not a DenckringError and could not be caught with the others.
• Candidate and Production.candidates: a result carries the metrics it was ranked by, so an order a caller is asked to trust arrives with its reason attached. anagram is the generator that needed it — its covers carry words and max_band; a generator with nothing to add wraps its texts through plain() and carries empty metrics, which is every other one so far. texts is unchanged in meaning and value, now a computed field over candidates, and is still what apply returns and what every existing consumer reads. Both are named in the README's stability contract. ADR 0027.
• Produced, what _produce hands back in place of list[str]: its candidates, and whether it abandoned its own search budget. The spine can see that max_results capped a result set and cannot see that a search gave up, because giving up makes the set smaller rather than larger — Production.truncated is now either event. ADR 0026's design spec named this case in advance; anagram's node budget is it.
• lexicon.graded_words on LanguagePack: every word with the SCOWL size band it first appears at. Separate from lexicon.words rather than replacing it, because ADR 0015's rule is one capability per question the lexicon is asked — semordnilap and charade ask membership and are right to keep asking the broad oracle, and a broad oracle is exactly what cannot rank. ADR 0028.

• graded_words.txt.gz in denckring-en-data: 77,078 words from SCOWL 2020.12.07 at sizes 10 through 60, one word<TAB>size line each, 244 KB in the wheel. 60 is SCOWL's own line — the largest size its documentation states it is confident holds no misspellings — and nothing above it ships, so anagram's max_size is capped at 60 rather than promising bands the data does not hold. *-proper-names and *-abbreviations are excluded, which is the point: they are why known_words() cannot tell room from romito. Of the 26 single letters, only a is kept, on SCOWL's bands rather than on an intuition. scripts/build_graded_words.py is the derivation, the whole of SCOWL's Copyright file ships as LICENSE-SCOWL, and the package's licence expression gains HPND-sell-variant.

• dictionary on NPlus7Params, the ordered list N+7 and S+7 displace within, defaulting to the pack's nouns. It sits on the check model rather than the apply model so both halves walk the same list; a generator displacing through a list its own checker could not see is the defect class the round-trip property exists to catch. The row's definition has said "the seventh noun following it in a chosen dictionary" since it was written, and until now there was no choice. Entries are validated as non-empty and single alphabetic words, which is ADR 0015's rule for nouns() for the same reason. A supplied list is matched casefolded and nothing more: pack.noun_index lemmatises and a bare list has no lemmatiser behind it. SPlus7Params now inherits NPlus7Params whole instead of restating offset, so it gains the field with it. A dictionary does not unlock either row for a language whose pack has no noun list — capabilities are enforced before parameters are parsed — and ADR 0029 names conditional capabilities as what would fix that.

• ambiguous_nouns on the same two rows: free (the default, and what shipped before), undecidable or strict, deciding what an unchanged word that the dictionary lists does to the score. A word list cannot tell you that run is a verb in this sentence, so accepting the position was a policy, reported as ambiguous_words and chosen by nobody. The three names are RhymeParams.unknown_rhyme's, which answers the same shape for a different undecidable. ambiguous_words counts the position under every reading, so the metric does not depend on which one was asked for. Under undecidable, a text whose every position is undecided returns an ambiguous_nouns_undecidable violation rather than scoring a vacuous 1.0 over nothing weighed.
• allow_subset on AnagramParams, accepting the transposal convention — a candidate built from some of the source's letters. Also on the check model, and for the same reason. It relaxes one half only: a shortfall stops being a violation, a surplus letter never does, since relaxing both leaves a check no text could fail. A candidate with no letters against a source that has some fails as empty_transposal rather than scoring 1.0 on a zero denominator. Each cover now carries letters_used in its Candidate metrics beside words and max_band. ADR 0029.
• FrenchPack, the third built-in default beside English and German (ADR 0022's shape), carrying no data files and declaring tokens, alphabet, fold_diacritics and letter_shapes. œ and æ fold to oe and ae explicitly, because neither has a casefold mapping or an NFKD decomposition the way ß does. 70 of the 119 implemented rows run in French on core alone; the other 49 want syllables.heuristic (29 rows), phonemes (21), stress (17), lexicon.words (6), lexicon.glosses (2) or lexicon.nouns (2). There is no denckring[fr] distribution, and this is not one.
• Description.runs_in, the languages this install can actually check a row in, computed from the packs' capabilities rather than authored. Distinct from languages, which is the row's editorial scope — wechselsatz is German by nature, not merely by capability — and the two visibly diverge on anagram, which is languages: [en] and runs in all three. Two fields answering two questions is a surface a reader can confuse; one field could not answer both.
• ADR 0029 on all of the above, and an amendment banner on ADR 0028, whose ranking paragraph describes three keys where allow_subset made four.
• German counts syllables. GermanPack gains syllables.heuristic, which it never had — it inherited the raising stub, so every syllabic form in the catalogue was unreachable in German. Vowel-group counting, always reporting itself as estimated. Deliberately simpler than the English heuristic rather than a port of it: a final "e" is pronounced in German, so the silent-e rule and the -le rule that exists to compensate for it are both dropped, and the diphthongs need no special case because each is already a run of adjacent vowels. It undercounts a vowel sequence spanning a morpheme boundary — Museum gives 2 where German says 3 — which a test pins rather than leaves to be found later. German goes from 78 of 119 runnable rows to 89.
• German runs every row. denckring[de-wiktionary], a fourth distribution, vendors 837,689 pronunciations and 179,831 glosses extracted from the German Wiktionary dump, giving the German pack phonemes, stress, syllables.dictionary and lexicon.glosses. German goes from 89 of 119 runnable rows to 119, and denckring eval --all from 365 passing cases to 401. Syllable counts and stress are read off the transcription rather than sourced separately, so stress can never be present without phonemes. Measured 94.0% token coverage over 340,958 tokens of Wieland, Goethe, Kafka and Mann; the residue is pre-1901 orthography and proper names. ADR 0030.
• denckring-de-data gains a pack() factory, and the de entry point resolves to it rather than to a class. Core refuses two packs claiming one language and ADR 0013 forbids merging CC BY-SA data into a CC0 distribution; a factory satisfies both, and every pack's capabilities stays a fixed ClassVar rather than becoming computed.
• 36 German golden cases across 30 rows, including Goethe's opening hexameter from Hermann und Dorothea, Voß's from the Odüssee and Heine's trochaic tetrameter, which scan under the checkers unmodified.
• proteus_verse, the 120th implemented procedure and the 28th generator. A line whose words permute into many metrically valid variants: check asks that the line scans as written and that its words admit at least minimum orderings that also scan, scored separately because a writer can fix them separately. produce returns the orderings. Voß's opening hexameter admits 1,728 of its 40,320. No new prosody abstraction: the scansion is core.prosody's, and the handover that proposed a Meter/Syllabifier pair was answered by what already existed — a pack's stress_patterns already returns every reading a word has, which is the ambiguity that proposal existed to model. The factorial is guarded by counting rather than enumerating: a walk over (words placed, syllables filled) is bounded by 2^n where scanning each ordering is n! * 32, and lines above max_words are refused rather than searched.
• InputTooLong takes the noun it counted. It said "letters" unconditionally, which is right for anagram and wrong for a row that rearranges words.
• A device is an address space. Device.at and Device.address are the two directions of its index — Piṅgala's naṣṭa and uddiṣṭa, which makes the odometer reading an attested procedure rather than a modern gloss — over the mixed base Device.radix. The first slot is the most significant digit, stated as a choice because sources differ. Abbildung is reading 52,699 of Harsdörffer's 103,680,000.
• Device.mask records which readings are attested, after al-Khalīl's muhmal marking: enumerate the space, then flag which members are real. hold is the default because it is what the source does. denckring apply reads it under attestation: mask, and refuses in both directions — a device with no mask, or a pack that cannot answer. Measured: 0 of 20,000 spins of the rings is a word the German lexicon knows, while 4.40% of German nouns are spellable on them.
• denckring spins max_results times rather than once, because a generator that flags which of its own outputs are real needs more than one to flag.
• Device.disputed_totals carries the counts the literature asserts that the inventory does not support, with who asserts each. combinations stays computed. ADR 0031.
• hemeling, the 155th catalogued row and the 121st implemented. Johann Hemeling's Arithmetische Letter- oder BuchstabWechslung (1653) asks for an anagram of a name und durch Reime zu erklähren — explicated beneath it in rhymed verse — and both halves are mechanically checkable, which is what makes it the tightest rule in the surveyed European material. The first line is judged by anagram's multiset comparison and the rest by rhyme_scheme's scheme walk, imported rather than reimplemented, so a text this row accepts is one those two accept. The halves are scored separately, so a broken rhyme over a sound anagram reads 0.5 and says which. The primary text has not been consulted; Kilcher, who read it, is named in the source.
• docs/expansion_ideas/ is excluded from the source distribution, on the argument the existing exclusions already make: those are working documents — proposals, handovers and research notes addressed to whoever picks the work up — not documentation of what this package does. They stay in git, because the ADRs cite them.
• The sdist bound is raised from 1,000,000 to 1,200,000 bytes, by maintainer decision after chapter 4 crossed the old one for real. A side effect worth naming: the whole suite now passes on a working tree, where test_the_sdist_stays_small had failed locally and passed in CI for as long as the bound existed — the untracked 68 KB blob the local build swallows no longer pushes the archive over. The new number is a stated trade: 236,778 bytes of prose headroom against roughly 13,000 bytes of margin on the other side, which is what still lets the bound catch the smallest vendored data file being re-included. There is no per-file bound that separates the two, because uv.lock is larger than graded_words.txt.gz.
• docs/expansion_ideas/ is now kept out of the sdist by a name-based test as well as by the size bound, since a raised bound would let a silent re-inclusion fit underneath it.
• Device.from_text builds rings from a text the reader supplies, with scripts/rings_from_text.py writing one out for the DENCKRING_DEVICE_PATH cartridge path — so denckring apply poesie_automat -p device=notes runs on rings drawn from notes.txt. Deterministic, because check and apply derive it separately: falling frequency, ties broken by first appearance. Words are dealt round-robin so every ring spans the frequency range rather than the first ring holding the commonest words. No catalogue row: the technique has no historical precedent, and neither attribution nor attested has a value meaning "contemporary, with no source". Amendment to ADR 0031.
• release.yml gains a TestPyPI rehearsal on workflow_dispatch. PyPI refuses a re-upload of a version that already exists, so without one the first genuine v* tag would also be the first time this workflow had ever run end to end.
• release.yml's docs job runs before publish rather than after it. As written, a docs failure left the release on PyPI and undocumented, and unrecoverable — and mkdocs gh-deploy fails outright on a private repository, so the old order guaranteed the first release would publish and then go red.
• Three docstrings that counted the generators no longer do. The worst was the apply_procedure MCP tool description, read by callers at call time, which said twenty-seven while the twenty-eighth was landing; list_procedures answers it exactly and cannot go stale.
• denckring-fr-data, installed by pip install denckring[fr] — a fifth distribution giving French lexicon.words, lexicon.nouns, lexicon.glosses and lexicon.graded_words: 125,343 words, 44,746 nouns, six frequency bands and 510,973 gloss headwords. French goes from 70 to 80 of the 121 implemented rows, unblocking charade, definitional_expansion, definitional_literature, kangaroo_word, n_plus_7, s_plus_7, semordnilap, tmesis, word_ladder and word_square, with 22 golden cases across eleven rows. denckring eval --all goes from 417 to 439. Two sources, both CC BY-SA 4.0, so one distribution rather than German's two: Lexique 3.82 for membership, nouns and frequency, and the frwiktionary dump for definitions. French registers its own entry point resolving to a class, not to a pack() factory — there is no second French distribution to choose between. ADR 0032, which records why Wikidata Lexemes could not carry it: 12,768 French noun lexemes against German's 188,948, 6.8%, measured before anything was built.
• French is the second pack with lexicon.graded_words, so apply anagram now runs in it. Lexique's frequencies are inverted at build time into SCOWL's direction, where larger means less common, because that is what the capability's contract says and what anagram sorts by; a test pins the direction. The bucketing into six bands is arbitrary — Lexique gives a continuous frequency — so roughly 20,900 words share each band with no ordering inside it, which is coarser than English's editorially-graded bands.
• A french-without-data CI job, mirroring german-without-pronunciations: the data-free French install is a real shape the suite never runs, and a lexical row in it must refuse by naming lexicon.nouns rather than guessing.
• French declares no prosody, and 41 rows stay blocked — 18 of them permanently. Lexique's orthosyll, phon and syll columns are present and deliberately unused: the 23 syllabic rows need a line-level count, because a French mute e counts before a consonant, elides before a vowel and never counts at line end, so summing citation forms undercounts systematically. The other 18 are accentual metres and French has no lexical stress, so no future data source closes them. ADR 0032 D5.
• MissingCapability now names pip install denckring[fr] as the remedy for a French lexicon row, and denckring.lang's docstring no longer says no such extra exists. Both were true when written and were falsified by the distribution above.
• LanguagePack.line_syllables, a pack answering how many syllables are in a whole line and how many of its words were estimated. BasePack implements it as the per-word sum syllable_count.line_syllables ran inline before, so English and German are unchanged by construction and a test asserts it. French overrides it, because a French final mute e is a syllable before a consonant, elides before a vowel or a mute h and never counts at the end of a line — Lexique gives femme, une, belle and porte all one syllable in citation form where verse frequently wants two. ADR 0034 D1; the seam is on the pack because "how long is this line" is a question about the language.
• denckring-fr-data gains phonemes, syllables, syllables.dictionary and syllables.heuristic from a 125,653-row table of Lexique's nbsyll, SAMPA phon and orthosyll, plus a 3,370-entry aspirated-h list Lexique cannot supply — extracted from {{h aspiré}} in frwiktionary and expanded through Lexique's lemme column, which adds only 85 forms but among them hais, the high-frequency verb forms verse actually uses. French goes from 80 to 103 of the 121 implemented rows, unblocking all 23 syllabic ones — alexandrine, arca_musarithmica, assonance_constraint, cinquain, clerihew, englyn, ghazal, haibun, haiku, hemeling, hendecasyllable, limerick, monosyllabic_prose, renga, rhyme_scheme, rondeau, senryu, spoonerism, syllable_count, tanka, terza_rima, triolet and villanelle — with 48 golden cases across them. denckring eval --all goes from 439 to 487. French is the first pack that can honestly declare syllables: the capability means a segmentation of the written word, orthosyll is one (car-ros-se), and ADR 0030 removed the same capability from denckring-en-data, which had declared it and never implemented it. ADR 0034.
• The French line count is an estimate and always will be (exact=False). Diérèse is etymological rather than spelling-derived, and the rule that ships is narrow: a glide after any consonant, gated on the -ion(s) suffix or the dia- prefix. Its measured effect is Racine's Mithridate 89.3% → 90.3% and Hugo's Hernani 79.3% → 80.2% of interior alexandrines scored at exactly twelve, against ceilings of 96.6% and 84.8% for a perfect per-site oracle — 1.0 of the 7.3 points available on Racine, short of the 94% accept bar, and recorded as a STOP rather than a success. The chapter plan's own proposed rule (a glide after a consonant cluster ending in a liquid) measured 84.5%/74.7%, five points below doing nothing. ADR 0034 D5 records the failed variants so the plausible ones are not proposed again.
• French coverage stops at 103, and the remaining 18 rows are blocked on stress alone — now measured mechanically against meta.requires rather than projected. They are accentual metres and French has no lexical stress, so no data source closes them. ADR 0034 D3, confirming ADR 0032 D5.
• The french-without-data CI job now also asserts that a syllabic row (alexandrine) refuses by naming syllables and a rhyme row (rhyme_scheme) by naming phonemes. The second has no other home: French gaining phonemes means no installed pack lacks that capability any more, so tests/test_prosody_robustness.py moved its refusal test to sonnet/stress and a core-only install is the last place a phonemes refusal runs at all.
• layer on every catalogue row — verfahren or instrument — and a second counter beside the five-part coverage line. ADR 0033 took the split with no instrument entry in hand; this implements its D1 and D3 and still adds no row, so denckring status prints the unchanged 155 catalogued · 130 implementable · 121 implemented · 121 validated · 25 not mechanically checkable and then 0 instruments catalogued. The zero is printed rather than hidden, because a counter at zero is the expected state after that ADR and a line that appeared from nowhere with the first instrument would read as a regression. Every field of Coverage counts verfahren alone — including implemented, which is now the registry intersected with the verfahren ids, so an instrument shipping a checker of its own could not move the denominator that every coverage figure in ADRs 0025 through 0032 was recorded against. catalogue.ids() unnarrowed still returns both layers: list, search and export describe the whole dataset, and only the counters read one layer at a time.
• LanguagePack.vowel_inventory(), the base vowel letters a language names, split out of vowels() — which answers a different question and therefore carries the accented forms. All three packs inherit aeiou and none overrides. vowels() keeps its name and its meaning because word_ladder._alphabet widens an alphabet with it and asks nothing about vowelhood; the five other call sites do ask, and keep the written set, which is the orthographic reading ADR 0035 D2 entitles them to. The third reading the one method was conflating — a contextual, phonetic classification, where y in yoyo is a consonant — is designed and deliberately not built: ADR 0035 D2 licenses it for spoonerism alone, the one row declaring phonemes, and D5 records that _letter_onset still splits on flat vowels() membership until then.
• calculator_word — a word a seven-segment display can write, entered as digits and read by turning the machine over: 7353 is ESEL. The 156th catalogued row and the first whose constraint is a property of a machine rather than of an orthography or a lexicon. Checks a text against a digits parameter, generates by partitioning a digit string into words segments, and runs in all three languages with no new data. ADR 0037 records the three decisions a reader would otherwise take for defects: the folk 2 → Z is refused because 2 is rotationally symmetric on a seven-segment display, which costs roughly half the corpus in every language; fold_diacritics defaults to false on this row alone, because a calculator cannot write Geheiß or blessé; and 6 and 9 both give G, so apply is not the inverse of check for about a third of every corpus.
• denckring-de-frequency, a sixth distribution: German frequency bands from the Leipzig Corpora Collection, under CC BY. 101,296 graded words, 434 KB, between English's 77,078 and French's 125,343. It gives German lexicon.graded_words, so apply anagram runs in German — Lebensmittel → mitbestellen. ADR 0038 records the case rule it turns on: German capitalises common nouns, so a noun is counted only from a corpus's capitalised rows, and a source that lowercases everything cannot tell German Tag from English tag at all.
• denckring.core.calculator — the display table and its two pure functions, shared by the procedure and the explorer's stage scene so neither imports the other's internals.
• A ninth stage scene for it, drawing seven real segments per digit and rotating the whole display, so the digits and the letters a viewer reads are the same marks. It carries a keypad of its own — 7 8 9 / 4 5 6 / 1 2 3 / 0, a real machine's layout — and arrives showing the number rather than the word, so turning it over is the viewer's gesture and not the page's opening state. Pressing a key redraws the display in the browser and clears the previous verdict; every judgement still comes from the library. 2 is on the keypad and worth pressing: it lights like any other digit and yields no letter, which is ADR 0037 D2 made pressable rather than argued.
• Coverage is enforced rather than reported: CI fails below 97%, the figure the README's badge states, so the badge cannot drift from the code without the build going red. Measured 97.1% — 6,015 of 6,194 statements.
• Description.untranslated, naming the fields served in a substitute language. Localisation has always fallen back to English silently and per field — measured over 155 rows, names de 95 / fr 98, definitions de 95 / fr 95, prompt_hints de 4 / fr 0 — so a French caller received English prose in a field typed as French with nothing marking it.
• Over MCP — an explorer page where a local model uses denckring's own MCP tools. The first consumer this repository has for that server, and a real stdio client rather than an in-process shortcut. Only gemma4 of the four installed Ollama models can call tools at all; the three German and European ones are listed with Ollama's own reason for refusing them. explorer/models.py is a second model provider beside the Anthropic one.
• An overview recording — docs/showcase/ — of the catalogue, all nine stage scenes and two live MCP tool calls, produced by apps/explorer/tests/browser/overview-gif.mjs and scripts/build_overview_gif.py. On the documentation site, out of the sdist.
• apply_procedure over MCP now returns candidates and metrics alongside texts. Without them ADR 0031's attestation: "mask" was a no-op over that surface, since the mask's entire output is a per-word mark on Candidate.metrics.

Changed

• slenderizing.apply(..., fold_diacritics=false) now refuses where it used to return text. The old _produce ignored the parameter and always folded, so apply("Bäh", lang="de", deleted="a", fold_diacritics=False) returned "Bh" — deleting a letter through the very fold the caller had switched off. It now raises DegenerateOutput, because with folding off ä is ä, the text holds no a, and the slenderizing of it is the text itself — which is also the answer _check gives, so the refusal is the two agreeing for the first time. ADR 0035, Consequences.
• Twenty-five fillers in data/devices/poesieautomat_2000.yaml — published data, CC BY 4.0 — replaced, because assembled they read as camp and confinement imagery. Twenty-four were of that family: der Zaun, am Zaun, Die Mauer, Die Sperre, auf Posten, die Wache, die Rampe, das Gleis, Das Lager, im Lager, im Graben, im Moor, in Ketten, in Trümmern, in Flammen, unter Zwang, aus Draht, mit Draht and Der Draht, plus Das Gitter (hinter Gittern), in Deckung, die Kammer, die Sirene and brennt, the last five found by sweeping rather than listed. Each is innocent read alone — a loading ramp, a warehouse, a railway track, a watch, a fence — and the defect was never one flap but what a board of 10^36 readings eventually sets beside what. That is countable exactly rather than by sampling, because a line is six modules of ten and the per-module counts convolve: of the 6,000,000 lines the board can assemble, 1,633,120 (27.2%) carried at least one of these words and 244,620 (4.08%) carried two or more, which put 22.686% of all 10^36 poems — better than one poem in five — on the wrong side of it. Over the replaced file both figures are 0. Nobody can read 10^36 combinations, so the guard is to keep the material out of the modules rather than to look for it in the output, and the device header now carries that as a seventh constraint, marked as resting on judgement rather than on a test — as, it also now records, the fourth rule already did, there being no test in this repository for the ADV prohibition either.
• The twenty-fifth, ohne Zeugen, went on a neighbouring ground: beside Die Anzeige, which is a notice and a display and also a criminal complaint, it read as a denunciation, in 1.00% of that line's 1,000,000 readings — 10,000 of them. It is now ohne Strom. Everything about the machine is unchanged: every module still holds ten alternatives, Device.combinations is still exactly 10^36, no filler exceeds eleven characters, no line's assembled width rose, all 360 stay distinct case-folded, and the six schemas, the three grammaticality rules and the ADV rule are untouched. kippt, which replaced brennt, is the only one of the thirty verbs that is ambitransitive; rule 3 holds for all 10^36 readings anyway, because no schema offers a verb an object, and the header records the exception. The second positive golden fixture spun four of the replaced flaps and was rewritten to the same seed's new reading, which leaves both properties its source claims — disjointness from the first fixture and no repeated word stem — true and re-derived on every run.
• Breaking: seed is a parameter on SeedParams rather than a keyword in the Constructive.apply signature, carried only by the ten procedures that draw at random. A keyword named in the signature bound before **params, so seed was never validated and all 27 generators accepted seed="not-an-int". Passing seed to a procedure that does not draw now raises InvalidParams instead of being silently ignored. lang remains a reserved keyword on the same signature for the same reason — no params field may be named lang either, and ApplyParams now says so.
• Breaking: spoonerism raises InputTooShort rather than NoCandidateWord for a text of fewer than two words, and ideenwuerfeln raises it rather than MalformedCorpus for a corpus holding fewer entries than a throw needs. Both were the shape recombination already raised InputTooShort for, so the code a caller retries on depended on which procedure they had called. NoCandidateWord keeps its narrower meaning — enough units, none of them suitable — and MalformedCorpus keeps corpus.parse's: unreadable, not merely small. InputTooShort's message now ends on a remedy, as the other errors do, and counts its units in English rather than reading "0 line".
• Breaking: ConstructiveProcedure.apply_params_model() is abstract. It used to synthesise the checker's model widened by ApplyParams; all 27 generators declare their own, so the default had no caller and carried an MRO trap.
• boustrophedon, cent_mille_milliards, recombination and wechselsatz raise InputTooShort where they used to return their input unchanged. wechselsatz raises it twice over: once for a frame offering no choice at all, and once for a frame whose alternatives the tokenizer used to check the result cannot read back as single words — a gap _apply's drawable filter now closes by simply never drawing such an alternative, which is itself a known, accepted cost documented in the code and deferred to chapter 2.
• apply_procedure (MCP) returns texts and truncated alongside text. Additive: text keeps its meaning and value as the first of texts.
• paragram returns every candidate it scores, best first, where it scored them all and returned one. apply's result is unchanged.
• apply_procedure (MCP)'s message for a non-constructive procedure, from a hand-built sentence to NotConstructive's own wording, which additionally points a caller at constructive in describe_procedure. code and detail are unchanged.
• Breaking for anyone who has written a generator: _produce returns Produced rather than list[str]. All 27 generators moved with it; denckring.core was already outside the stability statement, and this is what that clause is for.

• Breaking: the identity guard compares casefolded text, so output differing from its input only in capitalisation is refused. word_ladder.apply("Cat", target="cat") now raises DegenerateOutput.IDENTICAL where it returned "cat". apply("cat", target="cat") already raised, so the change makes that row consistent with itself rather than taking anything away.

• get_pack("fr") returns a pack instead of raising. Breaking for anyone catching UnknownLanguage on "fr": check(..., lang="fr") on a row needing only core capabilities now runs, and on a row needing a lexicon it raises MissingCapability naming the capability rather than UnknownLanguage naming the language. Lang has been Literal["en", "de", "fr"] since the first release and every catalogue row carries a French name and definition, so the old error pointed at denckring[fr] — a distribution that does not exist and was never planned. The failure moves to the right layer; it does not go away.

• MissingCapability's remedy is conditional on a data distribution existing. It interpolated the language into pip install denckring[{lang}], which was harmless while every language reaching it had a package and became a false promise the moment French did not: the 49 rows a French caller cannot run were each pointing at denckring[fr], which does not exist. English and German keep the install hint; anything else now reads "No data distribution supplies it for 'fr'". The code (missing_capability) and detail are unchanged, so a caller matching on either is unaffected; one matching on message text for a language with no extra is not. Untouched and older than this: naming a language's extra still assumes that extra supplies the capability, and denckring[de] carries no stress.
• anagram ranks covers by letters_used descending before ADR 0028's three keys, because under allow_subset every single word that fits the source is a valid transposal — 373 of them for astronomer — and word count first would bury every cover worth reading. With the flag off the new key is constant across covers and the order is byte-identical to ADR 0028's, held by a regression test rather than by the argument. The flag costs no search: astronomer visits 747,769 nodes either way, because a cover is recorded at nodes the walk already visits. What it multiplies is results — astronomer 1,421 covers to 15,185, dormitory 48 to 742, one fewer out of produce in each case, which drops the identity cover — which makes Production.truncated true on nearly every subset call against the default max_results of 10. That field has meant both "the search abandoned its budget" and "max_results capped the list" since it existed; allow_subset makes the second the common case, and a caller still cannot tell them apart.
• tests/test_round_trip.py draws blank-line paragraphs and terminated sentences by construction instead of waiting for a flat draw to offer them, and max_examples falls from 1,000 to 600. The old budget was never measured: over twelve fixed seeds the flat-only strategy reached every constructive row on 3 of 12, missing recombination on nine and every_nth_word on three. The new composition reaches every reachable row on 60 of 60 replayed seeds at 300 examples, and 600 is twice that floor. The cost is a narrower flat draw — st.one_of dedupes branches by identity and does not sample them uniformly, measured at about 20% for a flat branch against 45% for the prose one — which starves spoonerism, the row that wants flat text and the first to go missing when the budget is cut.

Fixed

• German lexicon.graded_words is no longer reported as a permanent ceiling. It was in _PERMANENTLY_MISSING citing ADR 0028, whose own comment anticipated the removal; it now names denckring[de-frequency]. French stress remains permanent (ADR 0034 D2).
• missing_capability no longer names an extra that cannot supply the missing capability. German is the one language with two: stress, phonemes, syllables.dictionary and lexicon.glosses come only from denckring[de-wiktionary], and the remedy named denckring[de], which the caller may already have installed.
• Six catalogue definitions promised what their checkers structurally cannot do, in all three languages: alexandrine's caesura, limerick's metre and line lengths, alliterative_verse's caesura and stress, homoteleuton's "or syllable", dactylic_hexameter's reading of the German spondee, and blank_verse withholding the strict-stress caveat iambic_pentameter already published.
• apply_procedure's docstring sent callers to constructive, which does not depend on lang and so cannot answer whether a call will run. It now names apply_missing.
• The word ladder's letter box had been sizing every tile on the board. Two unscoped .tile rules sat in one stylesheet; the ladder's, declared later at equal specificity, forced all 156 board tiles to a 38px square with their contents overflowing across their neighbours. Both landed on 2026-08-20 and nothing caught it, because the explorer's tests drive Starlette's TestClient, which renders no CSS.

• The source distribution shipped docs/superpowers/ and docs/seed/ — 1.22 MB of specs and plans that the documentation site had always refused on the same argument. The two exclusion lists are now one list with two consumers.

• s_plus_7/n_plus_7 inverted a verdict on a word carrying an elided proclitic — worse than refusing. noun_index looked an apostrophe-bearing token like l'île up whole and never found it, since no noun carries an internal apostrophe, so the correct S+7 l'îlot failed as changed_a_non_noun while retyping the source unchanged passed, hiding exactly what ambiguous_nouns: "strict" exists to catch. d', qu' and the typographic ’ behaved the same; both rows share the bug, since s_plus_7 delegates entirely to n_plus_7's displace/displacement_report. A new split_elision splits at the last apostrophe; the checker and generator both work on the tail and treat the proclitic as a separate unit that must match exactly, under a new changed_proclitic violation. The proclitic is reattached exactly as written rather than re-elided against the displaced noun — l'aïoli displaced by one entry becomes l'b, a known, honestly pinned limitation rather than a silent one; choosing l'/le/la correctly needs the noun's gender, which the list does not carry. ADR 0036.

• identical_rhyme could not tell l'amour from amour. The comparison read each rhyme scheme line's final word as written, so a minimal pair differing only by a leading l' matched by does_rhyme and by neither identical_rhyme nor does_not_rhyme, scoring 1.0 where the bare pair correctly failed — hemeling's own allow_identical publishes "Permit a word to rhyme with itself, as French rime riche does," and this was the dial bypassed by the commonest orthographic fact in the language. Fixed with the same split_elision, moved to core/text.py so core/prosody.py can share it; hemeling, limerick and rhyme_scheme all inherit the fix through scheme_violations. ADR 0036.
• kangaroo_word's own lookup broke on any accented French synonym under default folding. synonym: "école" failed not_a_word, because is_word is not fold-aware — French keeps its accents on purpose (ADR 0009) — and this row folded the synonym before asking it, so a diacritic-folded ecole could never match the table's école. Every other is_word caller (paragram, word_ladder, semordnilap) already checks membership on the word as written; this row now does too, folding only for its own distinctness and ordering checks. ADR 0036.
• missing_capability advised a remedy that could not work. The remedy was picked from lang in _EXTRAS alone, never asking whether that extra supplies the missing capability: all 18 French stress refusals said pip install denckring[fr], where French has no lexical stress at all, permanently (ADR 0034 D2); German anagram said denckring[de] for lexicon.graded_words, which only denckring-en-data ships (ADR 0028) — measured with denckring-de-data already installed, so the advice was to reinstall what was already there. A hand-maintained set of (lang, capability) pairs no extra will ever supply is checked first. The comment claiming French carries no syllables or phonemes was also false since ADR 0034 and is corrected in the same place.

• quenina accepted every text put to it whenever n was left out — which is the default. The size inference asked for the first n whose first n end-words are all distinct, and that is 1 for every text there is, so no stanza was ever compared against the rotation, total came back 0, and _report scored that 1.0 as vacuously satisfied. A whole sestina with an end-word wrong passed. The checker itself was correct all along: with n given it always graded properly, which is why the fault was in the one path nothing exercised — every golden case, every strategy and every unit test passed n explicitly. The inference now reads the size off the longest all-distinct run of end-words, and two golden cases cover the unparametrised path.

• assonance_constraint and spoonerism identified a vowel phoneme by CMUdict's stress digit — a convention no other source uses. Both rows declare only phonemes, so the moment German had phonemes they ran in German and found no vowels in any German word: assonance_constraint failed every German text and spoonerism read every word as pure onset. LanguagePack.is_vowel_phoneme puts the question on the pack, where the answer lives.

• denckring-en-data declared the syllables capability and never implemented it, so get_pack("en").syllables("table") raised MissingCapability naming a capability the pack declared. No row requires it, so nothing was broken; but runs_in computes from capabilities, so the first row to require it would have been reported as running in en and would then have raised. The claim is removed: a pronouncing dictionary is not a hyphenation dictionary.
• device.load and device.load_figure built a filename by interpolating the caller's id directly — directory / f"{item_id}.yaml" — which Path does not make safe: Path.__truediv__ silently discards the left operand when the right is absolute, and a .. segment is never rejected. An id of /etc/passwd or ../../etc/passwd read that file and returned it as though it were a device. This predates DENCKRING_DEVICE_PATH — the same interpolation was already present for both loaders when they only ever read the packaged directory — but the cartridge is what turns a latent gap into a routine one: it multiplies the directories an id can escape from and documents "point this at your own files" as a supported feature. It is reachable from outside the process, not only from Python: DeviceParams.device is a bare str, an MCP client supplies it directly via mcp/server.py's **(params or {}), and the CLI reaches it through --param device=.... Both loaders now validate an id as a bare filename stem — no separators, no .., never absolute — through one shared helper, and independently confirm the resolved path is inside the directory it was joined to, so the rule and the filesystem agree. An id that fails either check is treated exactly as an id no directory offers: UnknownDevice/UnknownFigure, not a different error.
• A cartridge file that failed to parse as YAML, or parsed but did not fit the Device/ Figure schema, let the raw yaml.YAMLError or Pydantic ValidationError escape. Both exceptions quote content from the file in their default message — a YAML parse error echoes a snippet of the source around the failure, and a Pydantic error echoes each offending value back — which is a content-disclosure path once a cartridge is by design arbitrary user-authored YAML, and it broke the contract mcp/server.py states in its own docstring: only a DenckringError is ever converted to data for a model to act on. Both loaders now raise MalformedDevice/MalformedFigure — new DenckringError subclasses alongside MalformedTable and MalformedCorpus — naming the path and a short, content-free reason (an exception's class name, or a count of validation errors) instead.
• Two language packs claiming one language were resolved silently by load order. They now raise DuplicatePack naming both, since answers that depend on installation order are the failure ADR 0004 exists to prevent.
• A report could say satisfied and still list violations, when a violation was appended without being counted in the score. The invariant suite asserted satisfied == (score == 1.0) from the start but never that a satisfied report is silent; it does now, and the Arca's tone check was the case that showed the gap.
• apply let Pydantic's validation error escape where check wrapped it in InvalidParams, so one kind of mistake raised two kinds of failure depending on which method you called. Both now go through one door.
• Three catalogue edits in earlier releases silently did nothing, because a string replacement that matches nothing succeeds quietly. The Denckring, Wechselsatz and Proteus-verse rows never received the text they were reported as having. All three are now written, and catalogue edits assert that they applied.
• The Denckring row overstated Harsdörffer, who labels the rings but states no total, and understated the case against the 97,209,600 the literature repeats. That figure factors as 2^8 x 3 x 5^2 x 61 x 83, and neither 61 nor 83 divides any ring size on any count — a stronger refutation than the divisibility argument it replaces. The row now also records that two independent counts of the parts disagree about one boundary while both coming to 264, and cites the Erquickstunden rather than the Poetischer Trichter, where the device is often wrongly placed.
• denckring, cent_mille_milliards and wechselsatz were filed as having no computable acceptance criterion. Asking whether a word is producible by five rings is a segmentation question, and asking whether a poem is one of Queneau's is a selection check; both are decidable. line_permutation, stanza_permutation and word_permutation keep their classification, because their definitions claim every ordering reads as a finished text and a permutation check does not verify that.
• Thirteen catalogue rows declared lexicon.nouns when they wanted word membership, synonyms, antonyms or glosses. Each now names what it actually needs.
• Entry-point language packs were shadowed by the built-in English default, so an installed data package had no effect. Precedence is now explicit registration, then entry point, then built-in default.
• Unknown parameters passed to check were silently dropped, so a mistyped --param frobidden=e looked like a constraint being applied when it was not. They now raise InvalidParams naming the parameters the procedure actually accepts.
• Four catalogue rows misdescribed what they need: boustrophedon is decidable only against its source, and semordnilap, kangaroo_word, word_square and charade cannot be decided without a lexicon.
• prisoners_constraint compared folded letters, so it read ß as ss and ä as a and wrongly accepted groß, Maße and Mädchen — and English naïve with them. An ascender is a property of the written glyph, and folding destroys it. The checker now reads raw characters. This changes existing English results: text containing accented letters that previously satisfied the constraint no longer does.
• Six form rows published a capability contract their own code broke. sonnet, ottava_rima, ballade, curtal_sonnet and spenserian_stanza declared [tokens, phonemes, syllables]; run against a pack carrying exactly that, all five raised MissingCapability('stress'). Scanning metre reaches stress (prosody.word_stress → pack.stress_patterns) and reaches syllable counting only on the out-of-dictionary fallback, which is syllables.heuristic; nothing in this codebase calls pack.syllables() at all. All five now declare [tokens, phonemes, stress, syllables.heuristic], the shape double_dactyl already used, and englyn — which is syllabic, not accentual — declares [tokens, syllables.heuristic, phonemes].
• The catalogue got more honest in both directions this batch. Rows that understated what they need: word_ladder and tmesis gained lexicon.words. Six rows that reach pack.vowels(), most of them through letter_class_report, gained alphabet: univocalic, bivocalic, monoconsonantal, homoconsonantism, homovocalism and univocalic_translation. dactylic_hexameter gained syllables.heuristic, which any metre scan falls back on for a word the pronouncing dictionary does not carry. Rows that declared what they never used: limerick dropped stress — it checks the rhyme scheme only, never the anapestic metre or shortened couplet its definition names; prisoners_constraint dropped tokens and fold_diacritics — it scans raw characters and asks only about x-height, and folding first would destroy the ascender the check needs to see; spoonerism dropped fold_diacritics — its onset comparison lives entirely in phonemes, where accents do not apply — and gained alphabet, which its written-onset guess does call. haikuization is the interesting one: it gained phonemes early in the batch on the assumption that a rhyme-word reading and a line-end reading could be told apart, then lost it again, because this codebase's rhyme machinery always resolves a line to its final word regardless of whether it pronounces as a rhyme with anything — the two readings are not constructible here, so the capability did no work and only made the row refuse unusual vocabulary.
• back_translation, transduction and intralingual_translation reclassified to checkability: none, each recording in notes why no acceptance criterion exists rather than merely asserting it: a translated-onward-and-back rendering, a chain of translations whose drift is the point, and a rewording into another register have no independent text to check the result against. implementable falls from 129 to 126 as a result — the fall is the point, not a regression.
• Four catalogue rows are blocked on a missing capability and now say so in notes: homosyntaxism and verbless_prose need a part-of-speech capability no pack provides; homophonic_translation needs phonemes in two languages at once; perverb needs a proverb corpus, which ADR 0020 already rules this project does not ship.
• The English lexicon is now derived from Open English WordNet 2024 rather than Princeton WordNet 3.1, which is unmaintained. The noun list grows from 55,239 to 56,468 entries and 78,865 glosses arrive with it; LICENSE-WORDNET is replaced with the CC BY 4.0 notice OEWN carries, and the data package's README stops naming the old release. This changes published output: n_plus_7 and s_plus_7 both count forward through that list, so a displacement that landed on one word can now land on another — cat displaces to catacomb, where it displaced to catafalque before. N+7's golden fixture, strategy and tests are updated to the new list; S+7's fixture words were not affected, but the same shift applies to any input.
• Four catalogue rows named capabilities they never needed, each corrected against what the row actually does: kangaroo_word needs a word list (lexicon.words), not a thesaurus, since whether two words are synonyms is the writer's claim; chimera substitutes by part of speech and is blocked on pos alongside homosyntaxism and verbless_prose; definitional_translation needs glosses in the target language (lexicon.glosses.bilingual), which English ones do not serve and no shipped pack provides, named the way homophonic_translation's phonemes.bilingual already is.
• glosses' inflection fallback tried stripping ed and ing, which resolved cared to car and poled to pol — unrelated headwords, returned with the confidence of a real answer, and enough to make the definitional rows demand the wrong sense of a correctly expanded word. Measured over a sample sentence, s/es stripping earns its place and ed/ing bought two extra resolutions of which both were wrong, so only s/es is kept. The docstring no longer claims resolution never guesses: it is best-effort, and only the empty sequence is a guarantee.
• The three distributions are version-locked to each other. denckring-en-data subclasses the core English pack, and release.yml publishes all three from one workflow for exactly that reason, but each declared a bare denckring dependency and each extra a bare data-package one — so pip install denckring-en-data was free to pair a subclass with a core it was never built against. All four declarations now pin ==0.1.0; [tool.uv.sources] keeps local development on the workspace copies.
• project.urls gained Documentation and Changelog entries, which are what PyPI's sidebar shows.
• cent_mille_milliards silently dropped a sheet line offering no alternatives, drawing one fewer line than the sheet has positions — the shape its own missing_line violation exists to catch. apply('a|a\n|') returned one line where check required two. It now raises InputTooShort for a sheet with an empty position rather than drawing from it.
• recombination shuffled an unterminated trailing fragment into the draw, where " ".join could merge it with its new neighbour and re-split into a different multiset than the one shuffled — exactly what check compares against. It now pins the tail out of the shuffle and permutes only the terminated sentences. Both were made reachable by the previous chapter's wider round-trip input alphabet and surfaced only intermittently, because tests/test_round_trip.py::test_apply_output_satisfies_check is not derandomized.
• apply anagram "astronomer" raised DegenerateOutput, and dormitory with it. The generator walked pack.nouns() greedily — a source word that is itself a noun is the longest cover of its own letters — and emitted whatever letters it could not spend as a run of single letters, which kept check satisfied and said nothing. It now searches the graded lexicon depth-first for sets of words whose letters are exactly the source's, bounded by max_words, min_word_length, max_size and a budget of 1,000,000 nodes, and ranks what it finds by fewest words, then by the SCOWL band of the cover's least common word, then alphabetically. dormitory gives dirty room, one of 47 covers; astronomer gives arrest moon, one of 1,420. A walk that cannot spend its remaining letters is not a cover and is abandoned rather than padded. The catalogue row declares lexicon.graded_words under apply_requires and not requires, since check has always run on core alone.
• The identity guard compared produced.strip() == text.strip(), so a generator whose output is casefolded never compared equal to a capitalised input and slipped it: apply anagram "Dormitory" returned dormitory, the reported defect still live for anyone who types a word the way people write one. Fixed at the spine rather than in anagram, because the rule is the spine's and belongs in one place. See Changed for the one existing result this moves.
• The diacritic fold applied to the text but not to the parameter compared against it, so vowel="ä", vowels="äö" and a ß acrostic target were unsatisfiable in German under the default fold_diacritics: true, and the violation mixed a folded found with an unfolded expected. Worse, slenderizing's generated output failed its own checker: _produce kept a character when fold_diacritics(ch).lower() != deleted, and ß folds to ss, so ß always survived a deletion _check then required — apply gave Die traße war groß. and its own check scored that 0.412 with seven violations. core.text.fold_letter is the parameter-side twin of letter_spans and is now applied at the seven rows this decision covers — univocalic, bivocalic, monoconsonantal, slenderizing and the acrostic family — not at every such comparison; five more were named under Known remaining in ADR 0035 and were fixed the same day, below. A target that is a phrase is flattened by core.text.fold_target, so a ß claims the two units its two letters need. The acrostic family is three rows, not two: telestich inherits from Acrostic, double_acrostic kept its own copy of the expression and now shares the one function. ADR 0035 D3 and D6.
• slenderizing's generator failed its own checker in French too, and on one letter it had been made worse rather than better. Folding deleted was necessary and not sufficient: dropping a source character whenever any letter it folds to matched agrees with _check only for a uniform fold, and œ → oe is not uniform. apply("Le cœur et la sœur", lang="fr", deleted="e") returned L cur t la sur, scoring 0.167; deleted="o" returned Le cur et la sur, scoring 0.214, where the code before this release raised DegenerateOutput — an honest refusal turned into text failing its own checker. The rule is now three-way: keep the source character untouched when none of its letters is deleted, drop it when all are, emit the survivors when only some are. Both round-trip at 1.0 (L cour t la sour, Le ceur et la seur), with three golden cases and French unit tests. ADR 0035 D6.
• Five more rows had the identical seam, unfiled by the sweep because it walks fixtures and these needed a parameter no fixture supplies. lipogram(forbidden="ä") on "Rätsel" returned satisfied=True — an inverted verdict, the text plainly contained the letter — with pangrammatic_lipogram the same shape behind it. tautogram and homoteleuton failed closed instead: an unfolded initial/final never matched a folded word boundary. abecedarian's start="ä" was vacuous rather than wrong — alphabet.index fell back to 0 and agreed with the right answer only because ä happens to fold to the alphabet's first letter; start="ö" (index 14) showed the real defect, disagreeing with the correct answer on every line. serial_lipogram had its own copy of the bug, the opposite way: it refused start="ä" outright with InvalidParams, even under default folding, stricter than its own text-side comparison. All six now fold via single_letter(); abecedarian and serial_lipogram raise InvalidParams when a folded start still is not a letter of the row's own flat alphabet, rather than silently defaulting. Ten golden cases, one per row plus a negative for four of them. ADR 0035, Known remaining.
• A single-letter parameter that folds to several characters is now refused by name. consonant="ß" compared a two-character expected against a one-character got and could never be satisfied; single_letter() raises InvalidParams saying what it folded to and naming fold_diacritics: false, which genuinely works. This replaces slenderizing's degenerate_output advising allow_identity=true, which named neither the cause nor a remedy and would have returned the untouched text as a slenderizing. ADR 0035 D4.
• supervocalic required every accented vowel letter its pack names — eight in German, twenty-one in French — against a published definition that says five. Folded text can never contain an umlaut, so ä ö ü were permanently missing while inflating a o u into repeats: no German or French text could satisfy the row at all. vowel_inventory() is the reading it wanted. ADR 0035 D1.
• slenderizing was unexercised twice over — in PARAMETER_GATED in tests/test_round_trip.py and pinning lang: en at file level in its golden file — so the row that broke the project's thesis was the one row neither net could see. A structural guard now holds those two facts together for every gated row that declares fold_diacritics, and tests/test_fold_symmetry.py states the invariant the whole seam violated once rather than per row: a verdict on folded input with folding on must agree with the verdict on pre-folded input with folding off.