chaffjs 0.31.0 → 0.32.0

package/rules/variance-mismatch.yaml

@@ -0,0 +1,67 @@
+id: variance-mismatch
+layer: L3
+status: stable
+how_to_find: variance-mismatch
+# The column headings are lexicon variance-column (groups budget, actual, variance, rate); words of magnitude are amount-multiplier.
+severity: warning
+
+name:
+  ja: 予算と決算の差が、増減の列と合わない
+  en: A variance that is not the difference of the budget and the actual
+
+why:
+  ja: 予算額・決算額・増減を並べた表では、読み手は増減の列だけを見て、予算をどれだけ超えたか、どれだけ残ったかを判断します。決算額を直したのに増減を直し忘れたとき、桁を写し間違えたときに、増減だけが食い違います。見出しの語で予算の列（予算額、予算）、決算の列（決算額、実績）、増減の列（増減、差異、差額、予算比）が一つずつ分かる表では、行ごとに決算額から予算額を引いた額と、書いた増減を比べます。△・▲・−・- の付いた額と括弧に入れた額は負の数と読みます。増減を「予算額から決算額を引いた額」で書く表もあるので、どちらの向きかは表のほかの行（どの行も向きを言わない表では文書のほかの表）が多く従うほうに決め、決まらなければ言いません。
+  en: In a table of budget, actual and variance, readers look at the variance column alone to see how far over or under budget each line went. When an actual figure is corrected and its variance is not, or a digit is copied wrong, the variance alone is off. When the headings name one budget column (Budget), one actual column (Actual) and one variance column (Variance, Difference, Over (under)), each row's actual less its budget is compared with the variance written. An amount marked -, −, ▲ or △, or written in brackets, is negative. Some tables write the variance the other way, as the budget less the actual, so the direction is the one most of the table's other rows follow (or, when no row of the table tells, the document's other tables); when the rows split evenly, nothing is said.
+
+message:
+  ja: 「{item}」の増減は「{written}」ですが、予算と決算の差からは「{expected}」です
+  en: '"{item}" shows a variance of "{written}", but its budget and actual give "{expected}"'
+
+how_to_fix:
+  ja: 予算額・決算額・増減のどれが正しいかを確かめて、ほかを直してください。
+  en: Check which of the budget, the actual and the variance is right, and correct the others.
+
+levels: { strict: error, normal: warning, relaxed: info }
+
+use_for: [business, technical, blog, legal, docs, academic, literature, speech]
+off_for:
+  literature: "Length, rhythm, repetition and register are the author's choices in literature, and dialogue mixes styles. Only what one work can hold itself to runs: consistent spacing and commas, and numbering without gaps."
+
+group: facts
+summary:
+  ja: 予算額と決算額の差が、書いた増減と合わない行
+  en: A table row whose variance is not the difference of its budget and actual
+example:
+  ja:
+    before: |
+      | 項目 | 予算額 | 決算額 | 増減 |
+      | --- | --- | --- | --- |
+      | 会費 | 510,000円 | 504,000円 | △6,000円 |
+      | 防災用品の購入費 | 80,000円 | 86,900円 | 6,090円 |
+      | 会議・事務費 | 40,000円 | 35,280円 | △4,720円 |
+    after: |
+      | 項目 | 予算額 | 決算額 | 増減 |
+      | --- | --- | --- | --- |
+      | 会費 | 510,000円 | 504,000円 | △6,000円 |
+      | 防災用品の購入費 | 80,000円 | 86,900円 | 6,900円 |
+      | 会議・事務費 | 40,000円 | 35,280円 | △4,720円 |
+  en:
+    before: |
+      | Item | Budget | Actual | Variance |
+      | --- | --- | --- | --- |
+      | Street lighting | $480.00 | $469.20 | ($10.80) |
+      | Emergency supplies | $400.00 | $434.50 | $43.50 |
+      | Seniors' lunch | $600.00 | $590.00 | ($10.00) |
+    after: |
+      | Item | Budget | Actual | Variance |
+      | --- | --- | --- | --- |
+      | Street lighting | $480.00 | $469.20 | ($10.80) |
+      | Emergency supplies | $400.00 | $434.50 | $34.50 |
+      | Seniors' lunch | $600.00 | $590.00 | ($10.00) |
+
+not_flagged:
+  ja: 見出しから予算・決算・増減の列が一つずつ決まらない表（当初予算と補正予算が並ぶ表など）、率の列（増減率、執行率、% を書いた列）、升が金額でない行（—、98.8%）、予算額と決算額の単位の書き方が違う行は比べません。増減の向きが、表の行で同じ数ずつに分かれる表は言いません。もう一方の向きに従う行が二つ以上ある表では、それらを書き手が選んだ向きと読み、向きの違いだけでは言いません。千円・百万円や単位の無い数は丸めた額と読み、一桁の差は言いません。
+  en: "A table whose headings do not name one budget, one actual and one variance column (an original and a revised budget side by side), a column of rates (Variance %, Rate, Ratio), a row whose cells are not amounts (—, 98.8%), and a row whose budget and actual are written in different units are not compared. A table whose rows split evenly between the two directions is not reported, and when two or more rows follow the other direction, the difference in direction alone is not reported. Amounts in thousands or millions, and bare numbers, are read as rounded, so a difference of one in the last digit is not reported."
+level_meaning:
+  ja: 予算と決算の差が増減の列と合わない行を指摘
+  en: reported when a row's variance is not its actual less its budget (or the reverse, as the table's rows do)

package/rules/very-adjective.yaml

@@ -34,7 +34,7 @@ off_for:
 group: wording
 summary:
   ja: 「とても便利」「非常に危険」（very important）のような強め + 形容詞が、文書の長さのわりに多い所
-  en: '"Very" + adjective, such as "very important" (とても便利), dense for the document''s length'
+  en: '"Very" + adjective, such as "very important", used too often'
 example:
   ja:
     pad: true

package/rules/walk-time-distance-mismatch.yaml

@@ -29,7 +29,7 @@ off_for:
 group: facts
 summary:
   ja: 「徒歩3分（約600m）」のように、徒歩の分数が書いた距離と合わない所
-  en: A walking time that does not match the distance beside it, at the rate the document states
+  en: A walking time that does not match the distance beside it
 example:
   ja:
     before: |

package/rules/warning-without-action.yaml

@@ -36,7 +36,7 @@ off_for:
 group: wording
 summary:
   ja: 「注意」「警告」で始まる段落で、危険だけを書き、読み手のすること（〜してください、〜しないこと）が無いもの
-  en: A paragraph opening with Warning or Caution that names a risk and no action for the reader
+  en: A Warning that names a risk but not what the reader should do
 example:
   ja:
     config: { genre: docs/manual }

package/rules/weasel-word.yaml

@@ -33,8 +33,8 @@ off_for:
 
 group: wording
 summary:
-  ja: 「〜と言われています」「ある研究によると」（many experts say, studies show）のように、主張を名前の無い誰かに預ける言い回し
-  en: Phrases such as "many experts say" and "studies show" (と言われています, ある研究によると) that credit a claim to no one in particular
+  ja: 「〜と言われています」「ある研究によると」のように、主張を名前の無い誰かに預ける言い方
+  en: Phrases like "many experts say" or "studies show" that credit no one
 example:
   ja:
     before: |

package/rules/wordy-phrase.yaml

@@ -34,8 +34,8 @@ off_for:
 
 group: wording
 summary:
-  ja: 「することが可能」「現時点において」（in order to、due to the fact that）のように、一語で言えることを何語もかけて言う所
-  en: Phrases such as "in order to" and "due to the fact that" (することが可能, 現時点において) that take several words to say what one says
+  ja: 「することが可能」「現時点において」のように、一語で言えることに何語もかける所
+  en: Phrases like "in order to" or "due to the fact that" where one word does
 example:
   ja:
     before: |

package/rules/yori-as-from.yaml

@@ -35,7 +35,7 @@ sources: [koyobun-2022]
 group: wording
 summary:
   ja: 「4月1日より開始」のように、起点を「から」でなく「より」で書いた所
-  en: より written for a starting point (4月1日より開始) where Japanese official style writes から
+  en: より written for a starting point (4月1日より) where official style writes から
 example:
   ja:
     before: |

package/skills/chaff/SKILL.md

@@ -1,12 +1,14 @@
 ---
 name: chaff
-description: "Check prose with chaff (npx chaffjs): pick the genre preset for the kind of document (contract, statute, manual, FAQ, paper, fiction, transcript…), run it on Markdown or text, read each finding, fix the text or record why not, and tune chaff.yaml with rules --json and relax --why. Use when asked to lint, proofread or tighten documents, or to set up a team's writing rules. Also use when asked to make text sound less AI-generated (「AIっぽさをなくして」「AI臭さを消して」「人間らしく書き直して」, 'make this sound human'): Light, Bold or Full mode, with Full (全面書き直し) when the request says 「全面的に」 or 'from scratch'."
+description: "Check prose with chaff (npx chaffjs): pick the genre preset for the kind of document (contract, statute, manual, FAQ, paper, fiction, transcript…), run it on Markdown or text, read each finding, fix the text or record why not, and tune chaff.yaml with rules --json and relax --why. Use when asked to lint, proofread or tighten documents, to check that a document's totals, dates and references agree, or to set up a team's writing rules. Also use when asked to make text sound less AI-generated (「AIっぽさをなくして」「AI臭さを消して」「人間らしく書き直して」, 'make this sound human'): Light, Bold or Full mode, with Full (全面書き直し) when the request says 「全面的に」 or 'from scratch'."
 ---
 
 # Checking writing with chaff
 
 chaff finds what makes writing hard to read — long sentences, padding, mixed styles, broken references in a
-contract — and never rewrites anything. Reading the findings and changing the text is your part. Japanese and
+contract. It also finds where a document's own numbers and dates disagree: a total that is not the sum of its
+items, a date and its weekday that do not match, an end date before the start. It never rewrites anything, and the
+same text gives the same result every time. Reading the findings and changing the text is your part. Japanese and
 English documents both work; the language is detected per file.
 
 ## Pick the genre first
@@ -26,7 +28,7 @@ read. A contract checked as a blog post gets wrong findings, so choose the genre
 5. With the genre set, the report opens with the genre's guide: what a good document of that kind does, and the
    rules that matter most for it. Read it first, before any finding, and hold the draft (or your rewrite) to each
    line; most of it is what no rule can check. The team may have changed it with `guide:` in `chaff.yaml`. A
-   guessed or default genre shows no guide.
+   guessed or default genre shows no guide, and `--no-guide` leaves it out.
 
 A rule the genre turns off is listed under "did not run" with the genre as the reason. That is the preset's
 choice, not a gap; turn it back on (`rules:` in `chaff.yaml`) only when the person asks.
@@ -43,11 +45,25 @@ A run exits 1 when a finding is an `error`. It also exits 1 when it could not ch
 check, an unknown `--genre`, no package for the language. So a clean exit means the files were checked.
 
 After the findings comes the list of rules that did not run, and why. Only the default output shows it;
-`--compact` gives just their number. Read it before saying a document is clean.
+`--compact` gives just their number. Read it before saying a document is clean. Two reasons there are not about
+the genre: a rule that compares documents runs only when chaff is given two or more files or a folder, and a rule
+that reads meaning runs only under `npx chaffjs test` (below).
 
 A run over several files prints only the files that have findings. To call one file clean, run chaff on that
 file alone and read its list.
 
+The default output ends each file with an `AI-likeness:` line (low, medium or high, or why it was not scored).
+`npx chaffjs ai-score <file>` gives the same on its own. It counts signs common in generated text against human
+documents of the genre; it does not say who wrote the text.
+
+## Checks that read meaning
+
+A few checks need a reader: does the proposal state its risks, does the conclusion only restate the body, does a
+number claiming an effect have a source. chaff never runs a model in a plain check. `npx chaffjs test <file|dir>`
+narrows the candidates by machine and sends only those passages to an AI (Anthropic or OpenAI, set by
+`ai_backend` in `chaff.yaml`; it needs an API key). Run it with `--dry-run` first: it shows what would be sent and
+calls nothing. Ask the person before a real run; the passages leave the machine.
+
 ## For each finding, choose one of three
 
 | | When | How |
@@ -59,6 +75,18 @@ file alone and read its list.
 `npx chaffjs explain <rule-id>` shows why a rule exists, its levels and the current one. Read it before
 relaxing a rule.
 
+## When numbers or dates disagree
+
+The group "Facts that disagree" in `npx chaffjs rules` checks what a sum or a calendar can settle. Among them:
+totals and their items, quantity times price, tax and net amounts, and percentages that should add up to 100%.
+Also a date and its weekday, dates and periods out of order, a reference number written two ways, and one fee or
+deadline given two values. Most of these are an `error`.
+
+Such a finding says two places in the document disagree. It does not say which one is wrong. Check the source
+(the quote, the ledger, the calendar, the booking, or the person) before changing anything, then fix the wrong one.
+Never "fix" it by changing the figure that was right, and never adjust both so they merely agree. If you cannot
+tell which is right, leave both and ask.
+
 ## Tune chaff.yaml
 
 - `npx chaffjs init --genre <genre>` writes a commented `chaff.yaml` for that genre when there is none (an
@@ -67,7 +95,8 @@ relaxing a rule.
   the genre's preset included) and why a rule is off. Read it instead of guessing rule names or limits.
   `npx chaffjs rules` shows the same as a table, grouped, with the level each rule runs at now.
 - A level is `strict` / `normal` / `relaxed` / `off`; when even `relaxed` is too tight, a positive number is the
-  limit itself (`max-sentence-length: 260`).
+  limit itself (`max-sentence-length: 260`). A rule with nothing to count takes fewer levels, and no number;
+  `levels_you_can_set` in `rules --json` lists them.
 - `prefer:` maps spellings to the team's (`e-mail: email`); `preferred-term: normal` turns the rule on.
 - Unknown rule names and unreadable values are reported on stderr — read stderr after editing `chaff.yaml`.
 - `genre:` sets the kind of document. `by_path` sets it per path, as a list of `files` globs, each with `genre`,
@@ -82,6 +111,10 @@ relaxing a rule.
   A file's front matter `genre:` (or `type:`) sets that one file, when neither `genre:` nor a matching `by_path`
   entry does. Strongest first: `--genre` for one run, then `by_path`, `genre:`, front matter, the path, the content,
   and `blog/tech`. `rules:` wins over the genre's preset.
+- `profile:` names a bundled kind of document with fixed conventions (`statute`). Leave it out: the genre or the
+  content chooses it, and the first line of `npx chaffjs tree` shows the choice. `none` stops the choice.
+- `names:` lists the team's names (organisations, products), each read as one name. `include:` adds files besides
+  Markdown when a folder is checked (`include: ["*.yaml"]`).
 
 ## Turn a team's style note into chaff.yaml
 
@@ -109,7 +142,7 @@ Pick the easiest way that can say the requirement:
 | A fixed phrase or pattern, with the team's own message | `custom_rules` of type `words` or `pattern` |
 | Decided by part of speech or inflection | `custom_rules` of type `tokens` |
 | Counting or comparing that the above cannot express | a Node function, `type: module` |
-| The same rules shared across repositories | a rule pack of YAML files: `npx chaffjs init --plugin <name>`, then `plugins:` in `chaff.yaml` |
+| The same rules shared across repositories | a rule pack of YAML files: `npx chaffjs init --plugin <name>`, `npx chaffjs plugin-test` in its folder, then `plugins:` in `chaff.yaml` |
 
 A `words` rule names the words, the team's text for each field `explain` shows, and one example:
 
@@ -137,11 +170,13 @@ Use this section when asked to make a text sound less generated: 「AIっぽさ
 
 chaff marks the shapes common in generated text (group "Signs of generated text" in `npx chaffjs rules`: `ai-tell`,
 `contrast-framing`, `stock-transition`, `announcing-opener`, `colon-lead-in` (Japanese only), `assistant-residue`, and the signals
-`ai-generated-composite` reads). They run by default outside legal, manual, literary and spoken genres. None of them says the text was
-generated. chaff never rewrites; the rewriting is yours.
+`ai-generated-composite` reads, among others). Most of them do not run in `technical/spec` or in the legal, docs, literature and
+speech genres; `assistant-residue` (words left from a pasted chat reply) runs in almost every genre. `npx chaffjs rules --json`
+says per rule and genre. None of them says the text was generated. chaff never rewrites; the rewriting is yours.
 
 **Start with the plan.** Run `npx chaffjs fix-plan <file>` (with the genre's `--genre`; `--json` to
-read it as data). It prints, in the document's language:
+read it as data; `--depth light|structure|register` when how deep the rewrite may go is already decided). It
+prints, in the document's language:
 
 - the genre's guide first, when the genre is set: what a good document of the kind does, the lines your rewrite
   is held to;
@@ -167,6 +202,11 @@ Pick one of three modes and say which.
 | **Bold** | each section's prose, keeping the outline | the outline is fixed (a report template, a manual, required sections) |
 | **Full** | the whole document, from scratch | the request says 「全面的に」「一から」「全部書き直して」 / "from scratch" / "rewrite the whole thing"; or the genre is `blog/*` or an essay (`literature/essay`); or `ai-generated-composite` fires; or the plan's structure score reaches its limit |
 
+The plan also names a fourth way, Register: a Full rewrite that also converts the style (polite to plain endings,
+です/ます to だ/である, and so on). chaff never recommends it; use it only when the person asks for the style to change
+(`--depth register`). With `--depth light`, a rule whose direction needs a deeper rewrite is listed by name and
+spot count only.
+
 For a blog or an essay, and whenever `ai-generated-composite` fires, recommend Full. A Light or Bold pass leaves the
 skeleton of generated text: a heading every few paragraphs, bold lead-ins, symmetric sections, a closing まとめ that
 repeats the body. The writer usually wants the structure changed, not only the sentences.
@@ -286,6 +326,16 @@ output before and after.
   terms defined twice.
 - `npx chaffjs cite <source> <quotes.json>` checks that quotations (`[{ "address", "quote" }]`) are really in
   the source; exit 1 if any is not. Use it to back every quotation you put in an answer or a summary.
+  `cite --url <URL> <quotes.json>` checks against a web page; `cite --scaffold <file>` prints the quotations in a
+  file that give no source, as JSON to fill in and pass to `cite`.
+- `tree --format json` and `cite --format json` give the same as data.
+
+## Grading model outputs
+
+`npx chaffjs grade <items.jsonl>` grades a file of model outputs, one JSON object per line (`id`, `output`, and
+`reference`, `sources` and `citations` when there are any): findings, dropped or added facts, quotations, pass or
+fail. It exits 1 when an output fails (with `--baseline`, when the run got worse) and sends nothing.
+`chaffjs/grade` does the same in code. The guide page "Using chaff for AI evals" covers baselines and variants.
 
 ## When chaff is wrong
 

package/src/company-kana-spelling.ts

@@ -0,0 +1,105 @@
+import type { CompanyMention } from "./company-names.ts";
+import { toKatakana } from "./kana-spelling.ts";
+import { escapeRegExp } from "./orthography.ts";
+
+// 会社の名前の漢字一字を、その読みのかなで書いた所（みどり野生命保険株式会社 と みどりの生命保険株式会社）。読みは品詞解析が言う。
+
+/** 名前の中の語。start と end は source の上の位置。reading はカタカナ。 */
+export type ReadWord = { readonly start: number; readonly end: number; readonly surface: string; readonly reading: string | undefined };
+
+/** 漢字の名前を、かなで書いた所。usual は多いほうの書き方。 */
+export type KanaSpelled = { readonly surface: string; readonly offset: number; readonly usual: string };
+
+const HAN = /^\p{Script=Han}$/u;
+const KANA = /^[\p{Script=Hiragana}\p{Script=Katakana}ー]$/u;
+const KATAKANA = /[ァ-ヶ]/gu;
+const KATAKANA_TO_HIRAGANA = 0x60;
+/** かなで書いた所の外側に続くと、名前の続き（別の名前）になる字。 */
+const NAME_CHAR = /[\p{Script=Katakana}\p{Script=Han}\p{Script=Latin}\p{N}ー]/u;
+/** 多いほうの書き方の数。一度だけの名前は、どちらが正しいとも言えない。 */
+const MIN_USUAL_COUNT = 2;
+
+const toHiragana = (text: string): string => text.replace(KATAKANA, (char) => String.fromCodePoint((char.codePointAt(0) ?? 0) - KATAKANA_TO_HIRAGANA));
+
+/**
+ * かなと漢字一字の語（みどり野 ミドリノ、野 ノ）の、漢字の読みと、語の中の位置（UTF-16 の単位）。かなの部分が読みの頭と終わりに
+ * 合い、残りがあるときだけ。漢字二字以上の語（生命 セイメイ）は、どの字がどの音か分からないので読まない。
+ */
+export const kanjiReadingOf = (
+  surface: string,
+  reading: string | undefined,
+): { readonly index: number; readonly kanji: string; readonly reading: string } | undefined => {
+  const chars = [...surface];
+  const at = chars.findIndex((char) => HAN.test(char));
+  const kanji = chars[at];
+  if (reading === undefined || kanji === undefined || !chars.every((char, index) => index === at || KANA.test(char))) return undefined;
+  const [head, tail] = [toKatakana(chars.slice(0, at).join("")), toKatakana(chars.slice(at + 1).join(""))];
+  const read = toKatakana(reading);
+  if (read.length <= head.length + tail.length || !read.startsWith(head) || !read.endsWith(tail)) return undefined;
+  return { index: chars.slice(0, at).join("").length, kanji, reading: read.slice(head.length, read.length - tail.length) };
+};
+
+const baseStartOf = (mention: CompanyMention): number =>
+  mention.position === "after" ? mention.offset : mention.offset + mention.surface.length - mention.base.length;
+
+/** 会社の名前の、漢字一字をその読みのひらがなかカタカナで書いた形（みどりの生命保険株式会社、みどりノ生命保険株式会社）。 */
+export const kanaSpellingsOf = (mention: CompanyMention, words: readonly ReadWord[]): string[] => {
+  const baseStart = baseStartOf(mention);
+  const baseEnd = baseStart + mention.base.length;
+  return words
+    .filter((word) => word.start >= baseStart && word.end <= baseEnd)
+    .flatMap((word) => {
+      const read = kanjiReadingOf(word.surface, word.reading);
+      if (read === undefined) return [];
+      const at = word.start - mention.offset + read.index;
+      const [before, after] = [mention.surface.slice(0, at), mention.surface.slice(at + read.kanji.length)];
+      return [toHiragana(read.reading), read.reading].map((kana) => `${before}${kana}${after}`);
+    });
+};
+
+const HIRAGANA = /^\p{Script=Hiragana}$/u;
+const HIRAGANA_AFTER = /^\p{Script=Hiragana}+/u;
+/** 助詞を探す、前のひらがなの字数。語彙表の一番長い助詞より長い。 */
+const PARTICLE_LOOKBACK = 10;
+
+const hiraganaBefore = (source: string, at: number): string => {
+  const chars = Array.from(source.slice(Math.max(0, at - PARTICLE_LOOKBACK), at));
+  return chars.slice(chars.findLastIndex((char) => !HIRAGANA.test(char)) + 1).join("");
+};
+
+/**
+ * 名前の側に名前の字が続かない現れ。続けば、別の長い名前の一部（あおいみどりの生命保険）。ひらがなが続くときは、助詞（当社は、
+ * と、における）で切れるときだけ名前の外と見る。particles は語彙表 company-name-kana の particle。
+ */
+const standsAlone = (source: string, at: number, length: number, position: CompanyMention["position"], particles: readonly string[]): boolean => {
+  const neighbour = position === "after" ? source.charAt(at - 1) : source.charAt(at + length);
+  if (NAME_CHAR.test(neighbour)) return false;
+  const kana = position === "after" ? hiraganaBefore(source, at) : (HIRAGANA_AFTER.exec(source.slice(at + length))?.[0] ?? "");
+  if (kana === "") return true;
+  return particles.some((particle) => (position === "after" ? kana.endsWith(particle) : kana.startsWith(particle)));
+};
+
+const occurrencesOf = (source: string, text: string): number[] => [...source.matchAll(new RegExp(escapeRegExp(text), "gu"))].map((match) => match.index);
+
+/**
+ * 二度以上書いた会社の名前の漢字一字を、その読みのかなで書いた所が一度だけある所。名前のほかの所はそのまま同じもの。
+ * 語の読みは words が言う（品詞解析の語）。前後のひらがなを切る助詞は particles。
+ */
+export const kanaSpelledCompanies = (
+  source: string,
+  mentions: readonly CompanyMention[],
+  words: readonly ReadWord[],
+  particles: readonly string[],
+): KanaSpelled[] => {
+  const counts = new Map<string, number>();
+  mentions.forEach((mention) => counts.set(mention.surface, (counts.get(mention.surface) ?? 0) + 1));
+  const usuals = mentions.filter(
+    (mention, index) => (counts.get(mention.surface) ?? 0) >= MIN_USUAL_COUNT && mentions.findIndex((other) => other.surface === mention.surface) === index,
+  );
+  return usuals.flatMap((usual) =>
+    [...new Set(kanaSpellingsOf(usual, words))].flatMap((spelled) => {
+      const found = occurrencesOf(source, spelled).filter((at) => standsAlone(source, at, spelled.length, usual.position, particles));
+      return found.length === 1 && found[0] !== undefined ? [{ surface: spelled, offset: found[0], usual: usual.surface }] : [];
+    }),
+  );
+};