chaffjs 0.33.0 → 0.34.0

package/src/name-variants.ts

@@ -2,6 +2,7 @@ import type { Token } from "./plugin.ts";
 import { escapeRegExp } from "./orthography.ts";
 import { nameCueAt, type NameCue, type NameCues } from "./name-cue.ts";
 import { isNearSurname } from "./surname-near.ts";
+import { isEndWordSlip, NO_WORD_USE, type WordUse } from "./name-end-letter.ts";
 import { foldedKeysOf, NO_SPELLING, type SpellingInput } from "./name-spelling-chars.ts";
 import { NO_CHAR_READINGS, readingOfWords } from "./name-char-reading.ts";
 import { isPersonRun } from "./name-person.ts";
@@ -15,7 +16,8 @@ import { heavierFirst, outweighs, type NameWeight } from "./name-weight.ts";
  * words は名前の語（記号を除く）。読みが同じ二つの名前が、どの語で違うかを見るため。person は人の名前と読めるもの（解析器が
  * 人名と読むか、敬称が付く）。cue は前後の語から見た名前らしさ（name-cue.ts）。ordinary は、ふつうの語でもある一語の名前で、
  * 文書がその形をどこでも名前に使わないもの（name-ordinary-word.ts）。labelled は、人を書く欄（氏名：、Contractor:）に書いた名前
- * （person-field.ts）。
+ * （person-field.ts）。givenNameReading は、空白を挟んだ姓と名の名を、名の中での字の読みで読み直した読み（name-char-reading.ts の
+ * givenNameReadingOf）。解析器が名を別の読みに切ったとき（建太 を タテフトシ）のもう一通りの読み。
  */
 export type NameMention = {
   readonly surface: string;
@@ -26,6 +28,7 @@ export type NameMention = {
   readonly cue?: NameCue;
   readonly ordinary?: boolean;
   readonly labelled?: boolean;
+  readonly givenNameReading?: string | undefined;
 };
 
 export type NameVariant = {
@@ -163,8 +166,12 @@ export const nameKey = (surface: string): string =>
  */
 export type NameEvidence = { readonly person: boolean; readonly named: boolean; readonly nameLike: boolean };
 
-/** readings は書き方の現れの読みのすべて。解析器は同じ字を所によって別に読む（佐伯 を サエキ と サイキ）。 */
-type Tally = NameEvidence & NameWeight & { readonly surface: string; readonly first: NameMention; readonly readings: readonly string[] };
+/**
+ * readings は書き方の現れの読みのすべて。解析器は同じ字を所によって別に読む（佐伯 を サエキ と サイキ）。givenNameReadings は
+ * 現れの givenNameReading のすべて。
+ */
+type Tally = NameEvidence &
+  NameWeight & { readonly surface: string; readonly first: NameMention; readonly readings: readonly string[]; readonly givenNameReadings: readonly string[] };
 
 const readingsWith = (readings: readonly string[], reading: string | undefined): readonly string[] =>
   reading === undefined || readings.includes(reading) ? readings : [...readings, reading];
@@ -184,11 +191,20 @@ const talliesOf = (mentions: readonly NameMention[]): Tally[] => {
     tallies.set(
       mention.surface,
       tally === undefined
-        ? { surface: mention.surface, first: mention, count: 1, readings: readingsWith([], mention.reading), labelled: mention.labelled === true, ...evidence }
+        ? {
+            surface: mention.surface,
+            first: mention,
+            count: 1,
+            readings: readingsWith([], mention.reading),
+            givenNameReadings: readingsWith([], mention.givenNameReading),
+            labelled: mention.labelled === true,
+            ...evidence,
+          }
         : {
             ...tally,
             count: tally.count + 1,
             readings: readingsWith(tally.readings, mention.reading),
+            givenNameReadings: readingsWith(tally.givenNameReadings, mention.givenNameReading),
             person: tally.person || evidence.person,
             named: tally.named || evidence.named,
             nameLike: tally.nameLike || evidence.nameLike,
@@ -268,14 +284,17 @@ const isSwapOrReplace = (left: string, right: string): boolean => {
 const isInnerDrop = (longer: string, shorter: string): boolean =>
   [...longer].some((_char, index) => index > 0 && index < longer.length - 1 && longer.slice(0, index) + longer.slice(index + 1) === shorter);
 
-/** 二つの語が一字違いか。英字だけの、長さ 5 以上の語で。 */
-export const isNearWord = (left: string, right: string): boolean => {
-  if (left === right || !LATIN_WORD.test(left) || !LATIN_WORD.test(right) || Math.min(left.length, right.length) < MIN_NEAR_LENGTH) return false;
+/** 二つの英字だけの語が一字違いか（置き換え・隣どうしの入れ替え・中の抜け）。長さは問わない。 */
+const isOneLetterApart = (left: string, right: string): boolean => {
+  if (left === right || !LATIN_WORD.test(left) || !LATIN_WORD.test(right)) return false;
   if (left.length === right.length) return isSwapOrReplace(left, right);
   if (left.length === right.length + 1) return isInnerDrop(left, right);
   return right.length === left.length + 1 && isInnerDrop(right, left);
 };
 
+/** 二つの語が一字違いか。英字だけの、長さ 5 以上の語で。 */
+export const isNearWord = (left: string, right: string): boolean => Math.min(left.length, right.length) >= MIN_NEAR_LENGTH && isOneLetterApart(left, right);
+
 /** 語の一つを伏せた形。ほかの語が同じ二つの名前は、同じ形を一つ持つ。 */
 const maskedKeys = (tally: Tally): string[] => {
   const words = wordsOf(tally.surface);
@@ -297,7 +316,24 @@ const isSlipOf = (tally: Tally, other: Tally): boolean => {
   return words !== undefined && isNearWord(...words);
 };
 
-const nearVariants = (tallies: readonly Tally[]): NameVariant[] => variantsIn(groupBy(tallies, maskedKeys), "near", isSlipOf, true);
+/**
+ * 二語以上の名前の一語だけが、終わりの一字の抜けか足し（Rachel Whitmore と Rachel Whitmor）か、最後の語（姓）が四字でも一字違い
+ * （Daniel Reyes と Daniel Reys）で、相手のほうが重い（outweighs）。違う語のどちらかがふつうの語なら語の形の違い。
+ */
+const isEndSlipOf =
+  (use: WordUse): Alike =>
+  (tally, other) => {
+    const [words, others] = [wordsOf(tally.surface), wordsOf(other.surface)];
+    const at = words.findIndex((word, index) => word !== others[index]);
+    const [word, usual] = [words[at], others[at]];
+    if (!outweighs(other, tally) || words.length < 2 || words.length !== others.length || word === undefined || usual === undefined) return false;
+    return !use.ordinary.has(word) && !use.ordinary.has(usual) && isEndWordSlip(words, at, usual, use, isOneLetterApart);
+  };
+
+const nearVariants = (tallies: readonly Tally[], use: WordUse): NameVariant[] => {
+  const isEndSlip = isEndSlipOf(use);
+  return variantsIn(groupBy(tallies, maskedKeys), "near", (tally, other) => isSlipOf(tally, other) || isEndSlip(tally, other), true);
+};
 
 /**
  * 名が同じで姓だけが二字ほど違う人の名前（Rachel Whitford と Rachel Whitfield）。一字違い（near）より遠いので、名と姓の二語以上
@@ -343,7 +379,7 @@ const oneCharApart = (left: string, right: string): boolean => {
 
 /**
  * 読みが同じ人の名前が、一字だけ違う（松本 と 松元）。読みの同じ別の姓（伊藤 と 伊東）もあるので、相手のほうが重い（outweighs）
- * ときに限る。
+ * ときに限る。名を名の中での字の読みで読み直した読み（健太 と 建太 の ケンタ）でも比べる。
  */
 const isPersonSlipOf = (tally: Tally, other: Tally): boolean => outweighs(other, tally) && oneCharApart(tally.surface, other.surface);
 
@@ -351,7 +387,7 @@ const personReadingVariants = (tallies: readonly Tally[]): NameVariant[] =>
   variantsIn(
     groupBy(
       tallies.filter((tally) => tally.person),
-      (tally) => tally.readings,
+      (tally) => [...tally.readings, ...tally.givenNameReadings],
     ),
     "reading",
     isPersonSlipOf,
@@ -360,9 +396,14 @@ const personReadingVariants = (tallies: readonly Tally[]): NameVariant[] =>
 
 /**
  * 同じ名前の、少ないほうの書き方。書き方ごとに最初の現れを一つ。同じ現れを二つの見方が言えば、先の見方だけ。chars は字体の
- * 違う同じ字の組（語彙表 name-variant-char）。
+ * 違う同じ字の組（語彙表 name-variant-char）。use は文書での英字の語の使い方。
  */
-export const nameVariants = (mentions: readonly NameMention[], chars: VariantChars = new Map(), spelling: SpellingInput = NO_SPELLING): NameVariant[] => {
+export const nameVariants = (
+  mentions: readonly NameMention[],
+  chars: VariantChars = new Map(),
+  spelling: SpellingInput = NO_SPELLING,
+  use: WordUse = NO_WORD_USE,
+): NameVariant[] => {
   const tallies = talliesOf(mentions);
   const found = [
     ...variantsIn(
@@ -378,7 +419,7 @@ export const nameVariants = (mentions: readonly NameMention[], chars: VariantCha
       "reading",
       oneWordApart,
     ),
-    ...nearVariants(tallies),
+    ...nearVariants(tallies, use),
     ...surnameVariants(tallies),
     ...characterVariants(tallies, chars),
     ...personReadingVariants(tallies),

package/src/number-name.ts

@@ -6,7 +6,7 @@ import type { Span, Token } from "./plugin.ts";
  * 「〒100-8916 東京都」の郵便番号、「紀尾井町1-3 東京ガーデンテラス」の番地は、後ろの空白が番号と題・項目の区切りで、「3回」「3 回」のような空け方の好みではない。
  * 空け方を数える材料にしない。
  *
- * 形で名前と読むのは六つだけ。0 で始まる組を含むハイフンつなぎの番号、※・〒や鉤括弧のすぐ後ろの番号、市・区・町まで下りた地名のすぐ後ろのハイフンつなぎの番号、
+ * 形で名前と読むのは七つだけ。桁の区切りの無い 6 桁以上の番号（口座番号 3056172）、0 で始まる組を含むハイフンつなぎの番号、※・〒や鉤括弧のすぐ後ろの番号、市・区・町まで下りた地名のすぐ後ろのハイフンつなぎの番号、
  * 文頭の階層つきの節番号（2.1、3.1.2）、続き番号になった行や番号の並ぶ箇条書きの頭の番号（白書の注、関連コード）、「問3」のように番号の前に書く語のすぐ後ろの番号。
  * 文頭でも「223 言語」「1ターン」は数量なので、階層の無い数は文頭というだけでは名前と読まない。
  * そのうえで、すぐ後ろ（空白 1 つまで）の語が数につく語（助数詞・数・助詞・助動詞）なら数量として数え続ける（「3-5 日」「26.7 万行」）。
@@ -53,8 +53,16 @@ const wordAfter = (tokens: readonly Token[], text: string, run: Span, base: numb
   return tokens.find((token) => token.span.start === base + next);
 };
 
-const isBoundToNumber = (token: Token): boolean =>
-  token.features?.["NounType"] === "Class" || token.features?.["NumType"] === "Card" || BOUND_TO_NUMBER.has(token.pos);
+/** 助数詞（NounType=Class）か数（NumType=Card）。 */
+const isCounterOrNumber = (token: Token | undefined): boolean => token?.features?.["NounType"] === "Class" || token?.features?.["NumType"] === "Card";
+
+const isBoundToNumber = (token: Token): boolean => isCounterOrNumber(token) || BOUND_TO_NUMBER.has(token.pos);
+
+/**
+ * 桁の区切りの無い、6 桁以上の数字だけの番号（口座番号 3056172 を、普通 1234567）。この大きさの数量は桁を区切るか単位を付けて書くので、
+ * 助数詞か数が続くとき（1234567円、250000 人）だけ数量と読む。助詞が続いても番号のまま。
+ */
+const PLAIN_CODE = /^\d{6,}$/u;
 
 /** 「1 又は 2」の接続詞は数どうしをつなぐが、続き番号の注の本文は「31 ただし、…」のように接続詞でも始まる。 */
 const CONJUNCTIONS = new Set(["CCONJ", "SCONJ"]);
@@ -221,6 +229,7 @@ export type NameContext = {
 export const isNumberName = (text: string, run: Span, tokens: readonly Token[] | undefined, base: number, context: NameContext = {}): boolean => {
   const inSequence = context.sequence?.has(base + run.start) ?? false;
   if (tokens === undefined) return false;
+  if (PLAIN_CODE.test(text.slice(run.start, run.end))) return !isCounterOrNumber(wordAfter(tokens, text, run, base));
   const shaped =
     inSequence ||
     isNameShaped(text, run) ||

package/src/profile/cue-lines.ts

@@ -0,0 +1,14 @@
+/**
+ * 行頭の箇条の印。Markdown の印（- * + と 1. 1)）は後ろに空白が要る。「・」は日本語で手で打つ箇条の印で、空白を置かずに続けて書く。
+ * 「**賃料**」の太字や「-5」は印ではない。
+ */
+const LIST_MARKER = /^([ \t]*)(?:(?:[-*+]|\d{1,9}[.)])[ \t]+|・[ \t\u3000]*)/gmu;
+
+/** 行頭で項目の名だけを包む太字（「**賃料**：」）。名の後ろにすぐ区切りが来るときだけで、文の中の太字は外さない。 */
+const BOLD_LABEL = /^([ \t]*)\*\*([^*:：|｜\n]+)\*\*(?=[ \t]*[|｜:：])/gmu;
+
+/**
+ * 内容から種類を選ぶ行を読むために、行頭の箇条の印と、項目の名を包む太字を外す。字下げは残す。
+ * 「- 賃料：85,000円」「・賃料：85,000円」「**賃料**：85,000円」も「賃料：85,000円」と同じ項目の行として数えるため。
+ */
+export const withoutListMarkers = (source: string): string => source.replace(LIST_MARKER, "$1").replace(BOLD_LABEL, "$1$2");

package/src/profile/parse.ts

@@ -1,4 +1,5 @@
 import type { DocumentProfile, RelativeVocabulary } from "../plugin.ts";
+import { withoutListMarkers } from "./cue-lines.ts";
 
 /** profiles/*.yaml を一つ読んだもの。言語ごとの中身と、内容から選ぶための形を持つ。 */
 export type ProfileDefinition = {
@@ -137,10 +138,10 @@ export const parseProfile = (raw: unknown): ProfileDefinition | undefined => {
   return { id, name: texts(raw["name"]), detect, languages: languagesOf(id, raw) };
 };
 
-/** 行頭から数える。source 全体に m フラグで当てれば、行ごとに分けずに数えられる。 */
+/** 行頭から数える。source 全体に m フラグで当てれば、行ごとに分けずに数えられる。箇条の印は行頭に数えない。 */
 export const detected = (definition: ProfileDefinition, source: string, language: string): boolean => {
   const detect = definition.detect[language];
   if (detect === undefined) return false;
-  const count = [...source.matchAll(new RegExp(detect.line, "gmu"))].length;
+  const count = [...withoutListMarkers(source).matchAll(new RegExp(detect.line, "gmu"))].length;
   return count >= detect.minLines;
 };

package/src/reference-numbers.ts

@@ -18,7 +18,7 @@ export const MAX_CHARACTERS = 24;
 const DIGIT = /\d/gu;
 const SEPARATOR = /[-‐−－]/gu;
 /** Only spaces and a colon, comma or opening bracket between the label and its number (「予約番号：」, "booking number, "); not a sentence's end. */
-const LABEL_TO_NUMBER = /^[\s:：,，、#＃(（「『[［]{0,4}$/u;
+export const LABEL_TO_NUMBER = /^[\s:：,，、#＃(（「『[［]{0,4}$/u;
 
 export const charactersOf = (written: string): string => written.normalize("NFKC").replaceAll(SEPARATOR, "").replaceAll(" ", "").toUpperCase();
 

package/src/spaced-person-name.ts

@@ -4,7 +4,7 @@ import { isInPersonField as isFieldValue } from "./person-field.ts";
 // 空白を挟んで書いた姓と名（高瀬 誠、佐伯 未咲）を、解析器の品詞に頼らず、前後の語で人の名前と読む。解析器は姓を地名やふつうの
 // 名詞と読み、名を二語に切ることがあるので、姓と名を一つの名前にできない。後ろに敬称か肩書き（様、部長）が付くか、人を書く欄
 // （氏名：、承認者：）の行の終わりにあれば、人の名前と読む。欄に書いた名と同じ名を空白の後ろに書いたもの（受託者：西村 亮 の
-// 文書の 西邑 亮が）も、人の名前と読む。
+// 文書の 西邑 亮が）も、人の名前と読む。名はかなで書いてもよい（松本 あかり 様）。
 
 /** suffixes は敬称（様、さん）、titles は名前のすぐ後ろの肩書き（部長）、labels は人を書く欄の語（氏名、承認者）。語彙表から。 */
 export type SpacedNameCues = { readonly suffixes: readonly string[]; readonly titles: readonly string[]; readonly labels: readonly string[] };
@@ -31,6 +31,13 @@ const LINE_MARKS = /^[\s>*#+-]*/u;
 const FIELD_END = /^[\u0020\u3000]*(?:[（(][^（()）\n]*[）)][\u0020\u3000]*)?(?:\r?\n|$)/u;
 
 const isHan = (letter: string | undefined): boolean => letter !== undefined && HAN.test(letter);
+const KANA = /^[\p{Script=Hiragana}\p{Script=Katakana}]$/u;
+const LEADING_KANA = /^[\p{Script=Hiragana}\p{Script=Katakana}ー]+/u;
+/** かなで書いた名の字数（あかり、エミリー）。一字のかな（ご担当 の ご）は名でない。 */
+const MIN_KANA_GIVEN = 2;
+const MAX_KANA_GIVEN = 5;
+
+const isKana = (letter: string | undefined): boolean => letter !== undefined && KANA.test(letter);
 
 /** 人を書く欄の行（承認者：営業部長 高瀬 誠）で、姓が欄の語の後ろにあるか。 */
 const isInPersonField = (source: string, surnameStart: number, labels: readonly string[]): boolean => {
@@ -53,8 +60,23 @@ const givenBeforeSuffix = (run: string, after: string, words: readonly string[])
     });
 };
 
+/**
+ * かなで書いた名（松本 あかり 様）。かなの連なりまるごとを名と読み、後ろに空白一つまでを挟んで敬称か肩書きが付くか、人を書く欄の
+ * 行の終わりにあるときだけ。かなの続きの中で敬称を切り出さない（みなさま を みな と さま にしない）。
+ */
+const kanaGivenAt = (source: string, start: number, cues: SpacedNameCues, inField: boolean, fieldGivens: ReadonlySet<string>): string | undefined => {
+  const run = LEADING_KANA.exec(source.slice(start, start + LOOKBACK))?.[0];
+  const length = [...(run ?? "")].length;
+  if (run === undefined || length < MIN_KANA_GIVEN || length > MAX_KANA_GIVEN) return undefined;
+  const after = source.slice(start + run.length, start + run.length + LOOKBACK);
+  const rest = after.replace(ONE_SPACE, "");
+  if ([...cues.suffixes, ...cues.titles].some((word) => rest.startsWith(word))) return run;
+  return fieldGivens.has(run) || (inField && FIELD_END.test(after)) ? run : undefined;
+};
+
 /** fieldGivens は、文書が人を書く欄に書いた名（亮）。その名と同じ漢字の連なりは、敬称が無くても名と読む。 */
 const givenAt = (source: string, start: number, cues: SpacedNameCues, inField: boolean, fieldGivens: ReadonlySet<string>): string | undefined => {
+  if (isKana(source.charAt(start))) return kanaGivenAt(source, start, cues, inField, fieldGivens);
   const run = LEADING_HAN.exec(source.slice(start, start + LOOKBACK))?.[0];
   if (run === undefined) return undefined;
   const after = source.slice(start + run.length, start + run.length + LOOKBACK);
@@ -73,10 +95,13 @@ const surnameBefore = (source: string, space: number, cues: SpacedNameCues): str
   return [...cues.titles, ...cues.suffixes].some((word) => run.endsWith(word)) ? undefined : run;
 };
 
+/** 名の頭の字。漢字か、かな（松本 あかり）。 */
+const isNameLetter = (letter: string | undefined): boolean => isHan(letter) || isKana(letter);
+
 const namesIn = (source: string, cues: SpacedNameCues, fieldGivens: ReadonlySet<string>): SpacedName[] =>
   [...source.matchAll(NAME_SPACE)].flatMap((match) => {
     const space = match.index;
-    if (!isHan(LAST_LETTER.exec(source.slice(Math.max(0, space - 2), space))?.[0]) || !isHan(FIRST_LETTER.exec(source.slice(space + 1, space + 3))?.[0]))
+    if (!isHan(LAST_LETTER.exec(source.slice(Math.max(0, space - 2), space))?.[0]) || !isNameLetter(FIRST_LETTER.exec(source.slice(space + 1, space + 3))?.[0]))
       return [];
     const surname = surnameBefore(source, space, cues);
     if (surname === undefined) return [];

package/src/spacing-code.ts

@@ -28,6 +28,15 @@ const HYPHENATED = new RegExp(
   "gu",
 );
 const CAPITALS_SERIAL = new RegExp(`(?<![A-Za-z0-9.${HYPHEN_CHARS}])[A-Z]+\\d{4,}(?![A-Za-z0-9${HYPHEN_CHARS}]|\\.[A-Za-z0-9])`, "gu");
+/**
+ * A Japanese number plate after its place name: a class number of two or three digits, one hiragana and a serial
+ * (330 ね 48-15, 500み・・12). A one-digit class number, or a serial that a kanji or katakana word follows, is a reference
+ * such as 「表3の12-15行」「表30の12-34行」, not a plate. The spaces inside a plate are its layout, and the one before it ends the place name.
+ */
+const PLATE = new RegExp(
+  `(?<![\\d${HYPHEN_CHARS}])\\d{2,3}[ \\u3000]?\\p{Script=Hiragana}[ \\u3000]?(?:\\d{1,2}[${HYPHEN_CHARS}]\\d{2}|[・･]{1,3}\\d{1,3})(?![\\d${HYPHEN_CHARS}\\p{Script=Han}\\p{Script=Katakana}])`,
+  "gu",
+);
 const HAS_LETTER = /[A-Za-z]/u;
 const SERIAL_DIGITS = /\d{3}/u;
 
@@ -40,8 +49,10 @@ export const addressSpans = (text: string): Span[] => spansOf(text, ASCII_RUN).f
 
 const isIdentifier = (word: string): boolean => HAS_LETTER.test(word) && SERIAL_DIGITS.test(word);
 
-export const identifierSpans = (text: string): Span[] =>
-  [...spansOf(text, HYPHENATED), ...spansOf(text, CAPITALS_SERIAL)].filter((span) => isIdentifier(text.slice(span.start, span.end)));
+export const identifierSpans = (text: string): Span[] => [
+  ...[...spansOf(text, HYPHENATED), ...spansOf(text, CAPITALS_SERIAL)].filter((span) => isIdentifier(text.slice(span.start, span.end))),
+  ...spansOf(text, PLATE),
+];
 
 /** Whether the character at `at` is inside one of the spans. */
 export const isInsideSpan = (spans: readonly Span[], at: number): boolean => spans.some((span) => span.start <= at && at < span.end);

package/src/spacing-minority.ts

@@ -25,3 +25,20 @@ export const minorityReport = <T extends { readonly spaced: boolean }>(ofKind: r
   const spaced = ofKind.filter((entry) => entry.spaced).length;
   return { mode: "mixed", first, odd: odd.length, spaced, touching: ofKind.length - spaced };
 };
+
+/** 同数のとき、書き方が二通りあると 1 件で言う最少の数。1 対 1 は一つずつの書き違いとも区別がつかないので言わない。 */
+const MIN_TIE = 2;
+
+/**
+ * 二通りが同じ数なら少ないほうは無く、1 箇所ずつ「違う」と指すとどちらかの側に立つ。MIN_TIE 対 MIN_TIE 以上なら書き方が二通りあることを 1 件で言い、
+ * それより少なければ言わない。同数でなければ report のまま。
+ */
+export const withoutSideOnTie = <T extends { readonly spaced: boolean }>(
+  report: MinorityReport<T> | undefined,
+  ofKind: readonly T[],
+): MinorityReport<T> | undefined => {
+  if (report?.mode !== "each" || report.odd.length * 2 !== ofKind.length) return report;
+  const first = ofKind[0];
+  if (report.odd.length < MIN_TIE || first === undefined) return undefined;
+  return { mode: "mixed", first, odd: report.odd.length, spaced: report.odd.length, touching: report.odd.length };
+};

package/src/structure/closing-balance.ts

@@ -11,7 +11,11 @@ import { amountOf, isExact, shown, type LabelledLine, type NetAmount, type NetCe
 export const BALANCE_ROLES = ["opening", "inflow", "outflow", "closing"] as const;
 export type BalanceRole = (typeof BALANCE_ROLES)[number];
 
-export type BalanceWords = Pick<NetWords, "magnitudes" | "percentUnits"> & { readonly labels: Readonly<Record<BalanceRole, readonly string[]>> };
+export type BalanceWords = Pick<NetWords, "magnitudes" | "percentUnits"> & {
+  readonly labels: Readonly<Record<BalanceRole, readonly string[]>>;
+  /** Words that make a label's note the date the balance is taken on (時点, as of): such a note is not a unit. */
+  readonly asOf?: readonly string[] | undefined;
+};
 
 export type BalanceAmount = NetAmount & { readonly role: BalanceRole };
 
@@ -54,12 +58,46 @@ export const closingMismatch = (amounts: readonly BalanceAmount[], words: Balanc
   return { offset: closing.start, values: { written: closing.written, expected: shown(expected, closing, decimals) } };
 };
 
-const labelledCells = (labels: readonly NetCell[], values: readonly NetCell[], words: BalanceWords): BalanceAmount[] =>
+const DIGIT = /\d/u;
+
+/** A note naming the date a balance is taken on (（2026年6月30日時点）, (as of June 30, 2026)): a date, not a unit. */
+export const isDatedNote = (note: string, asOf: readonly string[]): boolean => {
+  const key = keyOf(note);
+  return DIGIT.test(key) && asOf.some((word) => key.includes(keyOf(word)));
+};
+
+const CURRENCY_CODE = /\b[A-Z]{3}\b/u;
+
+/** A note that names a scale or a currency besides the date ((thousand yen, as of June 30, 2026)) still names a unit. */
+const namesUnitToo = (note: string, words: BalanceWords): boolean =>
+  CURRENCY_CODE.test(note.normalize("NFKC")) || words.magnitudes.some((word) => keyOf(note).includes(keyOf(word)));
+
+/** The note of a label that may give its amount a unit: a note that only dates it gives none, so dated and undated rows compare. */
+const unitNoteOf = (label: string, words: BalanceWords): string => {
+  const note = noteOf(label);
+  return isDatedNote(note, words.asOf ?? []) && !namesUnitToo(note, words) ? "" : note;
+};
+
+/** A column heading that is a word with no digit or bracket (ポイント, Points): the unit of the bare numbers under it. */
+const UNIT_HEADING = /^[^\d()[\]（）]+$/u;
+const BARE_MARKS = "||";
+
+/**
+ * A bare number under a heading naming its unit compares with the cells that write the unit out (14,310 under ポイント,
+ * against 4,960ポイント). Its marks stay empty, so it is still read as possibly rounded.
+ */
+export const withHeadingUnit = (amount: NetAmount, heading: string | undefined): NetAmount => {
+  const unit = keyOf(heading ?? "");
+  if (amount.prefix + amount.suffix !== "" || !UNIT_HEADING.test(unit)) return amount;
+  return { ...amount, unit: `|${unit}|${amount.unit.slice(BARE_MARKS.length)}` };
+};
+
+const labelledCells = (labels: readonly NetCell[], values: readonly NetCell[], words: BalanceWords, heading?: string): BalanceAmount[] =>
   labels.flatMap((label, column) => {
     const role = balanceRoleOf(label.text, words);
     const value = values[column];
-    const amount = role === undefined || value === undefined ? undefined : amountOf(value.text, value.start, noteOf(label.text), words);
-    return role === undefined || amount === undefined ? [] : [{ ...amount, role }];
+    const amount = role === undefined || value === undefined ? undefined : amountOf(value.text, value.start, unitNoteOf(label.text, words), words);
+    return role === undefined || amount === undefined ? [] : [{ ...withHeadingUnit(amount, heading), role }];
   });
 
 const issuesOf = (scopes: readonly (readonly BalanceAmount[])[], words: BalanceWords): NetIssue[] =>
@@ -75,14 +113,14 @@ const byColumns = (header: readonly NetCell[], rows: readonly (readonly NetCell[
     words,
   );
 
-/** A table whose rows are labelled in their first cell: each column after the labels is one scope. */
-const byRows = (rows: readonly (readonly NetCell[])[], words: BalanceWords): NetIssue[] => {
+/** A table whose rows are labelled in their first cell: each column after the labels is one scope, under its heading. */
+const byRows = (header: readonly NetCell[], rows: readonly (readonly NetCell[])[], words: BalanceWords): NetIssue[] => {
   const labelled = rows.filter((row) => row[0] !== undefined && balanceRoleOf(row[0].text, words) !== undefined);
   const width = labelled[0]?.length ?? 0;
   if (labelled.length === 0 || labelled.some((row) => row.length !== width)) return [];
   const columns = Array.from({ length: width - 1 }, (_, index) => index + 1);
   return issuesOf(
-    columns.map((column) => labelled.flatMap((row) => labelledCells(row.slice(0, 1), row.slice(column, column + 1), words))),
+    columns.map((column) => labelled.flatMap((row) => labelledCells(row.slice(0, 1), row.slice(column, column + 1), words, header[column]?.text))),
     words,
   );
 };
@@ -92,7 +130,7 @@ const byRows = (rows: readonly (readonly NetCell[])[], words: BalanceWords): Net
  * four totals in its header or its first column, so it gives no scope; running-balance-mismatch reads it row by row.
  */
 export const tableClosingMismatches = (header: readonly NetCell[], rows: readonly (readonly NetCell[])[], words: BalanceWords): NetIssue[] => {
-  const issues = [...byColumns(header, rows, words), ...byRows(rows, words)];
+  const issues = [...byColumns(header, rows, words), ...byRows(header, rows, words)];
   return issues.filter((issue, index) => issues.findIndex((other) => other.offset === issue.offset) === index);
 };
 

package/src/structure/definition-use.ts

@@ -200,6 +200,27 @@ const echoesLongName = (texts: readonly BodyText[], defined: DefinedTerm, offset
   return name.endsWith(defined.term) && sharedTail(name, written) >= defined.term.length + MIN_ECHO;
 };
 
+/** What may follow a word inside a name up to a definition's bracket: capitalised words (Services Agreement), or kanji and katakana (本サービス利用契約). */
+const NAME_RUN = /^(?:['’]?s)?(?:\s+\p{Lu}[\p{L}\p{N}&.'’-]*)*$|^[\p{Script=Han}\p{Script=Katakana}ー・･]*$/u;
+
+/**
+ * 別の語を定義する括弧の前の名前の中の現れ（This Services Agreement (this "Agreement") の Services、
+ * 本サービス利用契約（以下「本契約」という。）の本サービス）。文書が自分に付けた名前で、後で定義する語の使用ではない。
+ */
+const insideDefiningName = (texts: readonly BodyText[], inline: readonly DefinedTerm[], defined: DefinedTerm, offset: number): boolean => {
+  const body = sentenceAt(texts, offset);
+  if (body === undefined) return false;
+  const from = offset + defined.term.length;
+  const bodyEnd = body.start + body.text.length;
+  return inline.some(
+    (other) =>
+      other.term !== defined.term &&
+      other.span.start >= from &&
+      other.span.start <= bodyEnd &&
+      NAME_RUN.test(nameBefore(body.text.slice(from - body.start, other.span.start - body.start))),
+  );
+};
+
 const CAPITAL_OR_NON_LATIN = /\p{Lu}|[^\p{Script=Latin}\p{N}\s\p{P}]/u;
 
 /**
@@ -226,6 +247,7 @@ export const usesBeforeDefinition = (
   groupsOf(terms)
     .filter((group) => group.first.inline && isDistinctive(group.first.term))
     .flatMap(({ first: defined, spans }) => {
+      const inline = terms.filter((term) => term.inline);
       const defining = sentenceAt(texts, defined.span.start)?.start ?? defined.span.start;
       const uses = usesOf(defined.term, texts, spans, "exact", defined.bare === true ? mentioned : NEVER_MENTIONED);
       const first = uses.find(
@@ -234,7 +256,8 @@ export const usesBeforeDefinition = (
           !isTitle(offset) &&
           !continuesInHan(texts, defined.term, offset) &&
           !insideKatakanaName(texts, defined.term, offset) &&
-          !echoesLongName(texts, defined, offset),
+          !echoesLongName(texts, defined, offset) &&
+          !insideDefiningName(texts, inline, defined, offset),
       );
       return first !== undefined && first < defining ? [{ term: defined, offset: first }] : [];
     });

package/src/structure/document-context.ts

@@ -0,0 +1,46 @@
+import { linesOf } from "./lines.ts";
+import { withoutEdgeMarks } from "../facts/trim-marks.ts";
+import { leadingText } from "./total.ts";
+
+/**
+ * Whether a document says what kind of document it is with one of the words: its title contains one (会議室ご利用御見積書,
+ * Meeting Room Quote), or a field's label is one (見積番号：MR-0733, Quote number: Q-1). The title is the first line with
+ * text after any front matter, and the first heading; a table row is neither. A label only starting with a word does not
+ * count (Withholding tax: on an invoice), nor does a word in a sentence. Width and case are ignored. Pure.
+ */
+
+const LABEL_SEPARATOR = /[:：|]/u;
+const HEADING = /^[ \t]*#+[ \t]*/u;
+const TABLE_ROW = /^[ \t]*\|/u;
+const FRONT_MATTER_FENCE = /^---[ \t]*$/u;
+
+const keyOf = (text: string): string => text.normalize("NFKC").toLowerCase().replace(/\s+/gu, " ").trim();
+
+/** The lines after a front matter block (--- … --- at the very top); all of them when there is none. */
+const afterFrontMatter = (lines: readonly string[]): readonly string[] => {
+  if (!FRONT_MATTER_FENCE.test(lines[0] ?? "")) return lines;
+  const close = lines.findIndex((line, index) => index > 0 && FRONT_MATTER_FENCE.test(line));
+  return close === -1 ? lines : lines.slice(close + 1);
+};
+
+const titlesOf = (lines: readonly string[]): string[] => {
+  const first = lines.find((line) => line.trim().length > 0 && !TABLE_ROW.test(line));
+  const heading = lines.find((line) => HEADING.test(line));
+  return [first, heading].flatMap((line) => (line === undefined ? [] : [keyOf(line.replace(HEADING, ""))]));
+};
+
+const labelOf = (line: string): string | undefined => {
+  const lead = leadingText(line);
+  const separator = LABEL_SEPARATOR.exec(lead);
+  return separator === null ? undefined : keyOf(withoutEdgeMarks(lead.slice(0, separator.index)));
+};
+
+export const namesDocumentKind = (source: string, words: readonly string[]): boolean => {
+  const keys = words.map(keyOf);
+  const lines = afterFrontMatter(linesOf(source).map((line) => line.text));
+  if (titlesOf(lines).some((title) => keys.some((key) => title.includes(key)))) return true;
+  return lines.some((line) => {
+    const label = labelOf(line);
+    return label !== undefined && keys.includes(label);
+  });
+};

package/src/structure/fenced-lines.ts

@@ -0,0 +1,28 @@
+import { codeFences } from "../detectors/code-fences.ts";
+
+const FENCE_MARK = /`{3}|~{3}/u;
+
+const NONE: ReadonlySet<number> = new Set();
+
+/** The index of the line holding an offset of the joined lines. */
+const lineIndexAt = (starts: readonly number[], offset: number): number => starts.findLastIndex((start) => start <= offset);
+
+/**
+ * The indexes of the lines of fenced code blocks (``` or ~~~), the fences included. A table or a list written in one is an
+ * example, not the document's own values. The blocks are the ones the Markdown parser reads (code-fences.ts): a fence in
+ * a list item is one, a line indented as code or a fence in a quotation is not, and an unclosed fence runs to the end.
+ */
+export const fencedLineIndexes = (texts: readonly string[]): ReadonlySet<number> => {
+  if (!texts.some((text) => FENCE_MARK.test(text))) return NONE;
+  const starts: number[] = [];
+  texts.reduce((start, text) => {
+    starts.push(start);
+    return start + text.length + 1;
+  }, 0);
+  const fenced = new Set<number>();
+  codeFences(texts.join("\n")).forEach((fence) => {
+    const last = lineIndexAt(starts, Math.max(fence.start, fence.end - 1));
+    for (let index = lineIndexAt(starts, fence.start); index <= last; index += 1) fenced.add(index);
+  });
+  return fenced;
+};