the assay office / record

The Company Orange Keeps

Written 2026-08-02. Claims re-read against their sources on 2026-09-29: 39 checked, 1 confirmed, 33 wrong, 0 unverifiable, 5 true when written and overtaken since (dated, not corrected). By codex worker directed by claude-assaying-codex-0929, fixing the source-first pilot's adjudicated findings (research/source-first-pilot/adjudication/the-company-orange-keeps.md).

This pass fixed the adjudicated findings of the source-first pilot. It is not a fresh full audit. Every acted-on row was re-read against the cited source, repository code or data, or a recreated computation before editing. C30 was re-read and left unchanged. C24, C34, C37 and C42 remain outside this pass because their adjudication is unclear. Source access on 2026-09-29: curl fetched Wikipedia, Wiktionary, Language Log, the COCA spreadsheet and Thewlis source successfully. Merriam-Webster returned HTTP 403 to curl; browser retrieval then exposed the cited dictionary entries and article. No failed fetch was treated as evidence of absence. The exact spreadsheet cells were read from the downloaded XLSX using Python's zipfile and XML parser. Scratch downloads were temporary. Original computed observations were compared with commit dfc965e9c3 (2026-08-02) and the August 6 correction. A statement about an old table or measured sample can be STALE even where that computation used the since-corrected engine. The already-retracted claims about which words actually rhyme remain corrections. Historical audit decisions are explicitly marked where superseded, rather than rewriting the recorded 19/6/2 tally as a new observation.

Claims

• MINOR C1: Nothing rhymes with have, including the dek, hero, plain card and descriptions
  https://www.merriam-webster.com/dictionary/calve ; https://www.merriam-webster.com/dictionary/salve : Re-read 2026-09-29: calve has ˈkav and salve has ˈsav. The local CMUdict omits calve and has only S AA1 V for salve. The census claim holds; the unrestricted claim does not.
• WRONG C2: With one exception nobody has put the two-sound orphans on a list
  https://en.wikipedia.org/wiki/List_of_English_words_without_rhymes ; research/rhyme-orphans/data/census.json : Re-read the list: gouge(d), kirsch, beige, fugue and oblige appear; all five are in shortRimeOrphans[2]. More than one suffices to contradict the claim.
• MINOR C3: Orange and 18,746 others had no rhyme
  research/rhyme-orphans/data/census.json : headline.orphans is 10,142, hence 10,141 others. The original total was retracted as false under the intended rhyme definition in the August 6 correction.
• MINOR C4: This page and the scansion bench both name General American as their limit
  public/strata/prosody-scansion-bench/index.html ; src/content/strata/prosody-scansion-bench.md : Both files have zero occurrences of General American; visible prose has no accent/American/dialect. Positive control: both files exist and discuss CMUdict; this layer explicitly names General American.
• MINOR C5: Nobody has counted the famous language claims
  https://languagelog.ldc.upenn.edu/nll/?p=1946 : Re-read Liberman: 50,344 rhyme equivalence classes, 30,905 singletons, December 8, 2009. The page itself cites this count.
• MINOR C6: Film rhymes with sovexportfilm in the raw dictionary
  research/prosody-workshop/cmudict.dict ; research/rhyme-orphans/census.mjs : Recreated with loadCmudict, rimeKeysOf and buildIndex: both end IH L M with onset F; film has no partners even across the raw dictionary. Chimney has partner zimny.
• WRONG C7: Markdown retains 18,747/38.6%, the old law rates, algorithm not rhyming with rhythm, a 70% hit rate and a range near 33%
  research/rhyme-orphans/data/census.json ; research/rhyme-orphans/data/orphans.txt : Current data: 10,142/48,614 = 20.9%; length 1 rate 0.1%, length 5 rate 33.7%, 100% from length 10; algorithm/rhythm share IH DH AH M with different onsets. Current audit 12/18 = 66.7%; estimated rates 13.6 to 20.9%. The original 70.4% audit was a true observation, separately dated in the update.
• MINOR C8: After the engine repair every number on this page was recomputed
  research/rhyme-orphans/page.html ; research/rhyme-orphans/data/census.json : Before this pass the prose still said 1,599, 887, 712, two points and 126,053 while the data and table said otherwise.
• STALE C9: The share stays near two words in five whichever vocabulary is chosen
  git show dfc965e9c3:research/rhyme-orphans/data/census.json ; research/rhyme-orphans/data/census.json : The August 2 computed bands were 36.1% to 40.4%, supporting the observation then. After the August 6 engine repair the range is 20.0% to 34.9%. This is a description of the displayed computation, not evidence that English changed.
• WRONG C10: 1,599 orphans, 887 rescued, 712 remaining
  git show dfc965e9c3:research/rhyme-orphans/page.html ; git show dfc965e9c3:research/rhyme-orphans/data/census.json ; research/rhyme-orphans/page.html ; research/rhyme-orphans/data/census.json : The August 2 page already defined rhyme from the last stressed vowel. CMUdict and all 63 SCOWL files are byte-identical to that commit. The August correction explicitly says the old results were wrong under the same definition and sources, so the figures were false when written. Current pools 10/70/95 have 1,383/546/520: 837 find partners through band 70, or 863 through band 95.
• MINOR C11: Have is two rows above orange
  research/rhyme-orphans/data/census.json : Law rows run consecutively from 1 to 11; orange has 5 phonemes and have 2, a difference of 3.
• MINOR C12: Trio has no rhyme
  research/rhyme-orphans/data/census.json ; research/prosody-workshop/cmudict.dict : Trio is T R IY1 OW2 and shares the last-stress OW class with partners; it is absent from orphans.txt. This is one of the false rhyme denials inherited from the primary-stress rule.
• MINOR C13: Sporange -anj belongs with range and hinge
  research/prosody-workshop/cmudict.dict ; https://www.merriam-webster.com/dictionary/sporange : Re-read dictionary phones: range EY N JH, hinge IH N JH, flange AE N JH. Merriam-Webster gives sporange spəˈranj and ˈspōrˌanj.
• WRONG C14: Angel is the one word on the famous list with a partner
  git show dfc965e9c3:research/rhyme-orphans/page.html ; git show dfc965e9c3:research/rhyme-orphans/data/census.json ; research/rhyme-orphans/page.html ; research/rhyme-orphans/data/census.json : The August 2 page already defined rhyme from the last stressed vowel, and CMUdict plus all 63 SCOWL files are unchanged. The August correction says the old results were wrong under that same definition and those sources. Angel, elbow, rhythm and marathon have partners under the stated rule; the claim that angel was the only one was false when written.
• MINOR C15: The rhyme-lab correction note is in the apparatus at the foot of this page
  git show HEAD:research/rhyme-orphans/page.html ; git show dfc965e9c3:research/rhyme-orphans/page.html : Read the full Apparatus and the earlier template: the promised correction note is absent. The unrelated Choices paragraph mentions the rhyme engine, so a search could find a real reference.
• MINOR C16: Six clusters were refused and every one is coronal plus yod
  research/rhyme-orphans/onset.mjs:130 ; git show dfc965e9c3:research/rhyme-orphans/onset.mjs : Both original and current REFUSED_BY_THE_DICTIONARY have five entries: S Y, Z Y, TH Y, S T Y, SH P. Four are yod clusters; SH P is not.
• MINOR C17: The audit selects every Nth item of the sorted list
  research/rhyme-orphans/audit.mjs:40 ; git show dfc965e9c3:research/rhyme-orphans/audit.mjs : The original committed sampler already used hash32(orphan) modulo 160. The positional sampler is documented as an earlier version, not the shipped one.
• MINOR C18: Every rescue count is discounted before use
  research/rhyme-orphans/data/census.json ; research/rhyme-orphans/build-data.mjs : The dek and table use raw 2,665/10,142 = 26.3%; discounted[70] is 1,777 (17.5% of orphans), which feeds the 17.2% estimated orphan rate.
• STALE C19: The two draws land within two percentage points
  git show dfc965e9c3:research/rhyme-orphans/data/census.json ; research/rhyme-orphans/data/census.json : August 2 audit rates 19/27 = 70.4% and 17/24 = 70.8% are within two points. Since August 6 the live rate is 12/18 = 66.7%; the gap rounds to 4.1 points.
• MINOR C20: Inconveniency is not in current Merriam-Webster
  https://www.merriam-webster.com/dictionary/inconveniency : Re-read entry: noun, plural inconveniencies, definition inconvenience. Curl returned 403; browser retrieval exposed the actual entry.
• WRONG C21: Liberman used the same rhyme definition
  https://languagelog.ldc.upenn.edu/nll/?p=1946 : Re-read the definition: identical pronunciation from the main-stressed vowel to the end. It lacks this census's different-onset condition as well as its later last-stress anchor, so same definition was already too strong.
• MINOR C22: SCOWL was vendored whole
  research/rhyme-orphans/lexicon.mjs:111 ; research/rhyme-orphans/data/scowl-raw/ : Re-read the documented selection and counted exactly 63 vendored files; the module reads English and American lists, not the whole Debian package.
• STALE C23: The rime domain uses only the last primary-stressed vowel, following the shipped engine
  research/rhyme-orphans/census.mjs:13 ; git show dfc965e9c3:research/rhyme-orphans/data/census.json : The original code used last primary stress. The August 6 repair uses anchorCandidates: last primary plus last stress of any grade.
• MINOR C25: SEO title says a fifth of English
  research/rhyme-orphans/data/census.json ; research/rhyme-orphans/page.html : 20.9% is the fraction of the named 48,614-word lexicon; the page explicitly disclaims knowing the orphan rate of all English.
• MINOR C26: The dictionary has 126,053 headwords
  research/prosody-workshop/cmudict.dict ; research/rhyme-orphans/data/census.json : Recounted with loadCmudict: 126,052 headwords and 135,166 lines.
• STALE C27: The section contains three answers that are not answers
  git show dfc965e9c3:research/rhyme-orphans/page.html ; research/rhyme-orphans/page.html : It contained three rejections on August 2; the August 6 edit withdrew algorithm/rhythm while leaving the old heading. This is section bookkeeping, not confirmation of the withdrawn rejection.
• MINOR C28: Both Wiktionary accent entries stress only sporange's second syllable
  https://en.wiktionary.org/wiki/sporange : Re-read: RP /spəˈɹænd͡ʒ/, /ˈspəʊɹ.ænd͡ʒ/; US /spəˈɹænd͡ʒ/, /ˈspoʊɹ.ænd͡ʒ/. Both list both stress positions.
• MINOR C29: Merriam-Webster's objection to sporange as a poetic word is its verdict against the rhyme
  https://www.merriam-webster.com/grammar/what-rhymes-with-orange : Re-read the article: it calls sporange near-perfect and the closest to a perfect rhyme; its stated objection to poetic usefulness is botanical character. The quotation itself was accurate.
• CONFIRMED C30: The whole legal-onset inventory is stated cluster by cluster
  research/rhyme-orphans/data/census.json ; research/rhyme-orphans/page.html : The embedded AW_CENSUS data contains singletons, clusters2, clusters3 and their example words; the verifier reads that inventory. Adjudication says not false.
• MINOR C31: Lieu is simply L UW
  research/prosody-workshop/cmudict.dict ; research/rhyme-orphans/onset.mjs:120 : Re-read and computed: lieu L UW1 and lieu(2) L Y UW1; both onsets survive.
• MINOR C32: The surviving proper nouns, including Thewlis, are all borrowed
  https://surnamedb.com/Surname/Thewlis ; research/rhyme-orphans/onset.mjs : Fetched source: this is an English surname. Onset data admits L Y via lieu, so Leeuwen is not an example of a refused cluster.
• MINOR C33: Merriam-Webster's hirple pronunciation rhymes with purple
  https://www.merriam-webster.com/dictionary/hirple ; https://www.merriam-webster.com/dictionary/curple ; https://en.wiktionary.org/wiki/hirple : Re-read: Merriam-Webster has hirple ˈhirpəl versus curple ˈkərpəl. Wiktionary supplies UK /ˈhɜː.pəl/ with purple's rhyme.
• MINOR C35: Adnate and stagnate are examples of eSpeak rhyme failures
  https://www.merriam-webster.com/dictionary/adnate ; https://www.merriam-webster.com/dictionary/stagnate ; research/prosody-workshop/cmudict.dict : Both dictionary entries mark secondary stress on -nate. Recreated stagnated rime EY T IH D; the old primary-only reasons do not establish a non-rhyme. Both audited pairs are outside the live 18.
• MINOR C36: Wikipedia's list uses the same definition
  https://en.wikipedia.org/wiki/List_of_English_words_without_rhymes : Re-read lead: main stressed syllable, compiled from Received Pronunciation with a few General American exceptions. Its definition section says last stressed vowel.
• MINOR C38: Knosp's gloss and pronunciation are cited in place
  git show HEAD:research/rhyme-orphans/page.html : Before editing, knosp occurred only in that Sources sentence in visible authored prose; the separate embedded named-gap data is not an in-place gloss. Other named entries do have prose discussions.
• MINOR C39: The placard says Last verified 2 August 2026, when it was written
  scripts/check-placard.mjs:253 ; public/strata/the-company-orange-keeps/index.html : Read the generator: a dossier with reverified true selects when its claims were re-read against their sources. The current placard is preserved byte for byte; the directing instance owns regeneration.
• MINOR C40: Have and halve are the first two short-orphan entries
  research/rhyme-orphans/data/census.json : shortRimeOrphans[2] begins employee, have, halve, in both original and current data.
• STALE C41: The audit table's rows all belong to the stated sample
  git show dfc965e9c3:research/rhyme-orphans/data/census.json ; research/rhyme-orphans/data/census.json : August 2: table and tally both had 27 pairs. Since August 6, the live tally has 18 pairs but all 27 original records were still rendered. Filtering with the regenerated sample gives 12 hold, 4 fail, 2 unresolved.
• MINOR C43: Have is eighth by COCA, without an edition qualification
  https://en.wikipedia.org/wiki/Most_common_words_in_English ; https://www.wordfrequency.info/samples/wordFrequency.xlsx : Fetched both. Wikipedia gives have rank 8 and cites the older 450-million-word list (accessed April 11, 2018). The spreadsheet's introductory sheet identifies one billion words; sheets 2 and 3 rank have 11. No date for the transition since this page was written was established, so this repairs attribution, not an asserted later change in language.

What was done

• [fixed] C1: Restricted every copy to this census; social descriptions now include orange and have within 10,142 rather than adding 10,142 other words.
• [fixed] C2: Named the existing list and examples; removed nobody names from both markdown copies.
• [fixed] C3: Changed the relates note to 10,141 others in this census.
• [fixed] C4: The relates note now attributes that limit only to this census.
• [fixed] C5: Removed nobody has counted from the relates note and its generated nearby copy.
• [fixed] C6: Used the independently reproduced chimney/zimny example in the page and markdown.
• [fixed] C7: Updated all markdown figures and the false non-rhyme; the correction distinguishes the current figures and the update dates the changed audit.
• [fixed] C8: Narrowed the old correction to the census data was recomputed.
• [dated] C9: Dated the original table observation and changed the prose to 34.9% falling to 20.0%.
• [fixed] C10: Moved the false figures from Updated to Corrected in the page and markdown mirror, stating the corrected band-95 counts as 1,383/863/520. The revised example explicitly uses band 70, with 1,383 minus 546 = 837 rescued, all filled from census.json.
• [fixed] C11: Changed the script-built chart label to three rows.
• [fixed] C12: Removed trio from the no-partner sentence.
• [fixed] C13: Replaced range and hinge with flange.
• [fixed] C14: Moved the false only-partner claim from Updated to Corrected in the page and markdown mirror; angel, elbow, rhythm and marathon have partners, and the script says one of four.
• [fixed] C15: Removed the false pointer; retained the supported statement that the rhyme lab was corrected.
• [fixed] C16: Changed six to five and every one to four; explained SH P and corrected the verifier comment, whose assertion already loops over the actual list.
• [fixed] C17: Described selection by a hash of the orphan word.
• [fixed] C18: Said explicitly which values are raw and which estimated rates apply the discount; corrected the same assertion in audit.mjs comments.
• [dated] C19: Dated the old comparison and compute the current gap in the page script; dated the old comparison in the audit comment too.
• [fixed] C20: Corrected the authored gloss in audit.mjs and regenerated census.json and the embedded data.
• [fixed] C21: Named his main-stress anchor and absence of an onset condition.
• [fixed] C22: Said vendored in part and identified the 63 lists.
• [dated] C23: Dated the former implementation and described both candidate domains.
• [fixed] C25: Qualified seoTitle, HTML title and JSON-LD name as this census and its words.
• [fixed] C26: Replaced the denominator with the existing CMUHEADS data placeholder.
• [dated] C27: Changed the page and markdown headings to Three proposed answers and dated the section change.
• [fixed] C28: Described both readings and their full second-syllable vowel.
• [fixed] C29: Kept the verbatim quotation and supplied the article's actual rhyme assessment and reason.
• [fixed] C31: Gave both readings.
• [fixed] C32: Removed borrowed and Leeuwen from the list, specified the four refused yod clusters, and corrected the verifier comment.
• [fixed] C33: Attributed the hirple rhyme to Wiktionary's UK reading and displayed the two sources' pronunciations.
• [fixed] C35: Removed these failure examples from prose; retained the original audit booleans as historical decisions with superseded flags and explicit retractions, without changing the observed original tally.
• [fixed] C36: Stated the accent and the lead's main-stress wording without claiming equivalence.
• [fixed] C38: Removed knosp from that attribution list.
• [fixed] the placard line is generated by scripts/check-placard.mjs and becomes 'when its claims were re-read' now this record exists (C39; directing-instance regeneration pending.)
• [fixed] C40: Changed to second and third.
• [dated] C41: Stored the actual regenerated sample in census.json, render only those 18 rows, and dated the change; all 27 historical records remain available.
• [fixed] C43: Qualified eighth as the older list and gave eleventh for the current billion-word list in page and markdown.