Spaces:

NagaYu
/

dasumae

Running

dasumae / src /detect.js

Upload 6 files

c4413a6 verified about 5 hours ago

21 kB

// DasuMae detection module — dependency-free.
//
// Input:  words = [{ text, bbox:{x,y,w,h}, confidence? }]  (OCR output, pixels)
// Output: detections = [{ class, text, bbox:{x,y,w,h}, confidence }]
//
// Approach: reconstruct line-level text from words, run freshly-written patterns
// + validators (Luhn, My Number check digit, octet/range checks, length/charset
// rules) against each line, then map the matched character span back to the
// words it covers and merge their boxes to cover the whole planted item.
//
// Patterns are written from the public shape of each format (RFC email shape,
// Luhn, the My Number check-digit algorithm, well-known API-key prefixes). No
// rule file, pattern list, or code is copied from any other project.
//
// ESM module (package.json "type": "module").

// ---------------------------------------------------------------------------
// Normalisation — only length-preserving 1:1 replacements, so character offsets
// stay aligned with the word-owner map.
// ---------------------------------------------------------------------------
const DASHES = /[‐‑‒–—―−－]/g; // ‐‑‒–—―−－ -> -
function normalizeChar(ch) {
  const code = ch.codePointAt(0);
  // full-width digits ０-９ -> 0-9
  if (code >= 0xff10 && code <= 0xff19) return String.fromCharCode(code - 0xff10 + 0x30);
  // full-width Latin letters Ａ-Ｚ ａ-ｚ -> ASCII
  if (code >= 0xff21 && code <= 0xff3a) return String.fromCharCode(code - 0xff21 + 0x41);
  if (code >= 0xff41 && code <= 0xff5a) return String.fromCharCode(code - 0xff41 + 0x61);
  if (code === 0xff20) return '@'; // ＠
  if (code === 0xff0e) return '.'; // ．
  return ch;
}
function normalize(s) {
  let out = '';
  for (const ch of String(s)) out += normalizeChar(ch);
  return out.replace(DASHES, '-');
}

// ---------------------------------------------------------------------------
// Validators (exported for unit tests)
// ---------------------------------------------------------------------------
function luhn(numStr) {
  const s = String(numStr).replace(/\D/g, '');
  if (s.length === 0) return false;
  let sum = 0, alt = false;
  for (let i = s.length - 1; i >= 0; i--) {
    let d = s.charCodeAt(i) - 48;
    if (alt) { d *= 2; if (d > 9) d -= 9; }
    sum += d; alt = !alt;
  }
  return sum % 10 === 0;
}

function myNumberCheckDigit(payload11) {
  let sum = 0;
  for (let n = 1; n <= 11; n++) {
    const p = payload11.charCodeAt(11 - n) - 48;
    const q = n <= 6 ? n + 1 : n - 5;
    sum += p * q;
  }
  const r = sum % 11;
  return r <= 1 ? 0 : 11 - r;
}
function myNumberValid(twelveDigits) {
  const s = String(twelveDigits).replace(/\D/g, '');
  if (!/^\d{12}$/.test(s)) return false;
  return myNumberCheckDigit(s.slice(0, 11)) === s.charCodeAt(11) - 48;
}

function octetsValid(ip) {
  const parts = ip.split('.');
  if (parts.length !== 4) return false;
  return parts.every((p) => /^\d{1,3}$/.test(p) && Number(p) <= 255 &&
    (p.length === 1 || p[0] !== '0')); // reject 01, 00 style
}

// ---------------------------------------------------------------------------
// Detectors. Each returns matches [{ class, text, start, end, confidence }]
// against a normalised line string. start/end are character offsets.
// Confidence heuristic (documented):
//   0.95  exact validator passed (Luhn, My Number check digit, IBAN mod-97) or a
//         distinctive vendor key prefix / 〒 marker matched
//   0.90  strong structural pattern with a range/shape check (email, IP)
//   0.80  pattern-only match (phone) — shape matched, no external validator
//   0.50  signal-only class (address_jp) — surfaced, never auto-gated
// ---------------------------------------------------------------------------

function pushMatch(out, cls, m, conf, textOverride) {
  out.push({ class: cls, text: textOverride ?? m[0], start: m.index, end: m.index + m[0].length, confidence: conf });
}

// Each detector receives the line object { text, owners, compact } and pushes
// raw matches with character offsets into line.text.

// email — OCR often drops the tiny dot in ".jp", so a dot-less domain is still
// accepted when it ends in a known TLD. (No '@' appears in the clean set, so FP
// stays 0 despite the looser rule.)
// ',' is in the domain charset because OCR frequently reads a domain '.' as ','
// (e.g. "example,jp"); it is normalised back to '.' below. The '@' is required
// and never appears in the clean set, so FP stays 0 despite the looser rule.
const RE_EMAIL = /[A-Za-z0-9._%+\-]+@[A-Za-z0-9.,\-]{2,}/g;
const EMAIL_TLD = /(?:com|jp|org|net|co|io|dev|info|edu|gov|ne)$/i;
// Space-tolerant recovery pass: OCR often splits a domain with a stray space
// ("example. com", "examp le. com"), which breaks a contiguous match. On the
// space-stripped view we match up to a KNOWN TLD boundary (\b), so the match
// can't run on into the next word. '@' is required (absent from the clean set).
const RE_EMAIL_COMPACT = /[A-Za-z0-9._%+\-]+@[A-Za-z0-9.,\-]+?[.,](?:com|jp|org|net|co|io|dev|info|edu|gov|ne)\b/gi;
function detectEmail(line, out) {
  const s = line.text;
  for (const m of s.matchAll(RE_EMAIL)) {
    let tok = m[0].replace(/[.,\-]+$/, ''); // trim trailing punctuation
    tok = tok.replace(/,/g, '.'); // OCR comma -> dot (length-preserving)
    const at = tok.indexOf('@');
    const local = tok.slice(0, at), domain = tok.slice(at + 1);
    if (!local || domain.length < 2 || /\.\./.test(tok)) continue;
    const hasDot = domain.includes('.');
    if (hasDot ? !/\.[A-Za-z]{2,}$/.test(tok) : !EMAIL_TLD.test(domain)) continue;
    out.push({ class: 'email', text: tok, start: m.index, end: m.index + tok.length, confidence: hasDot ? 0.9 : 0.75 });
  }
  // recovery pass on the compact view (offsets mapped back to the raw line)
  const { text: ctext, map } = line.compact;
  for (const m of ctext.matchAll(RE_EMAIL_COMPACT)) {
    const tok = m[0].replace(/,/g, '.');
    if (tok.indexOf('@') < 1 || /\.\./.test(tok)) continue;
    const os = map[m.index], oe = map[m.index + m[0].length - 1] + 1;
    out.push({ class: 'email', text: tok, start: os, end: oe, confidence: 0.85 });
  }
}

// credit card — digit runs (spaces/dashes allowed), 13–19 digits, Luhn.
const RE_CARD = /\d(?:[ -]?\d){11,18}/g;
function detectCard(line, out) {
  const s = line.text;
  for (const m of s.matchAll(RE_CARD)) {
    const digits = m[0].replace(/\D/g, '');
    if (digits.length < 13 || digits.length > 19) continue;
    const before = s[m.index - 1], after = s[m.index + m[0].length];
    if (before && /[A-Za-z]/.test(before)) continue;
    if (after && /[A-Za-z]/.test(after)) continue;
    if (!luhn(digits)) continue;
    pushMatch(out, 'credit_card', m, 0.95);
  }
}

// My Number — exactly 12 digits (grouped or contiguous), valid check digit.
const RE_MYNUM = /\d(?:[ -]?\d){11}/g;
function detectMyNumber(line, out) {
  const s = line.text;
  for (const m of s.matchAll(RE_MYNUM)) {
    const digits = m[0].replace(/\D/g, '');
    if (digits.length !== 12) continue;
    const before = s[m.index - 1], after = s[m.index + m[0].length];
    if (before && /[0-9A-Za-z]/.test(before)) continue;
    if (after && /[0-9A-Za-z]/.test(after)) continue;
    if (!myNumberValid(digits)) continue;
    pushMatch(out, 'my_number', m, 0.95);
  }
}

// IP address — 4 dotted octets, each 0–255.
const RE_IP = /\d{1,3}(?:\.\d{1,3}){3}/g;
function detectIp(line, out) {
  const s = line.text;
  for (const m of s.matchAll(RE_IP)) {
    if (!octetsValid(m[0])) continue;
    const before = s[m.index - 1], after = s[m.index + m[0].length];
    if (before && /[0-9.]/.test(before)) continue; // part of a longer dotted string
    if (after && /[0-9.]/.test(after)) continue;
    pushMatch(out, 'ip_address', m, 0.9);
  }
}

// API keys — prefix-anchored, run on a SPACE-STRIPPED view of the line because
// OCR scatters spaces inside a long key and doubles dashes ("sk-" -> "sk--").
// Once a vendor prefix is seen the whole token is secret, so the body is matched
// loosely (length-tolerant, allows the stray @/./- that OCR substitutes in). The
// prefixes don't occur in ordinary UI text, so clean-set FP stays 0.
// Key body charset: the real key alphabet (base64/base62 + _ @ . - + / =) plus
// the stray '[' ']' OCR sometimes injects mid-token (e.g. "sk-pro]j-..."). Once a
// vendor prefix is seen the whole run is secret, so the body is matched loosely.
const BODY = '[A-Za-z0-9_@.+/=\\-\\]\\[]';
const JWTBODY = '[A-Za-z0-9_@\\-]'; // JWT part body, tolerating OCR '0'->'@'
const API_PATTERNS = [
  /gh[opusr][_\-][A-Za-z0-9@\]\[]{18,}/g,
  /github[_\-]pat[_\-][A-Za-z0-9_@\]\[]{25,}/g,
  /AKIA[A-Z0-9]{12,}/g,
  new RegExp(`x[o0]x[bpars]${BODY}{16,}`, 'g'), // Slack xox*-; tolerate OCR o->0
  new RegExp(`eyJ${JWTBODY}{6,}\\.${JWTBODY}{6,}\\.${JWTBODY}{4,}`, 'g'), // JWT
  new RegExp(`sk-+${BODY}{16,}`, 'g'), // sk-ant / sk-proj / any sk- key
];
function detectApiKey(line, out) {
  const { text: ctext, map } = line.compact;
  const taken = [];
  for (const re of API_PATTERNS) {
    for (const m of ctext.matchAll(re)) {
      const s = m.index, e = m.index + m[0].length;
      if (taken.some(([ts, te]) => s < te && e > ts)) continue;
      taken.push([s, e]);
      // map compact offsets back to original line.text offsets for box lookup
      const os = map[s], oe = map[e - 1] + 1;
      out.push({ class: 'api_key', text: m[0], start: os, end: oe, confidence: 0.95 });
    }
  }
}

// phone (JP) — leading 0, dashed groups, <=11 digits total.
const RE_PHONE_JP = /0\d{0,3}-\d{2,4}-\d{3,4}/g;
function detectPhoneJp(line, out) {
  const s = line.text;
  for (const m of s.matchAll(RE_PHONE_JP)) {
    const digits = m[0].replace(/\D/g, '');
    if (digits.length < 9 || digits.length > 11) continue;
    const before = s[m.index - 1], after = s[m.index + m[0].length];
    if (before && /[0-9]/.test(before)) continue;
    if (after && /[0-9-]/.test(after)) continue;
    pushMatch(out, 'phone_jp', m, 0.8);
  }
}

// phone (international) — leading +, 8–15 digits.
const RE_PHONE_INTL = /\+\d{1,3}[\s-]?\(?\d{1,4}\)?[\s-]?\d{2,4}[\s-]?\d{2,4}/g;
function detectPhoneIntl(line, out) {
  const s = line.text;
  for (const m of s.matchAll(RE_PHONE_INTL)) {
    const digits = m[0].replace(/\D/g, '');
    if (digits.length < 8 || digits.length > 15) continue;
    pushMatch(out, 'phone_intl', m, 0.8);
  }
}

// postal (JP) — the 〒 marker is required, but OCR reads 〒 as "T"/"TT"
// (look-alike), so an isolated leading T before NNN-NNNN is accepted too. Bare
// NNN-NNNN with no marker is left to manual to avoid flagging generic numbers.
// The 〒 marker (or its "T"/"TT" OCR look-alike) is required. The first group is
// \d{3,4}: a true code is 〒NNN-NNNN, but OCR often injects a phantom digit right
// after 〒 (e.g. "〒7080-7118"), so a 4-digit first group is accepted too.
const RE_POSTAL = /(?:〒|\bT{1,2})\s?\d{3,4}\s?-\s?\d{4}/g;
function detectPostal(line, out) {
  const s = line.text;
  for (const m of s.matchAll(RE_POSTAL)) {
    const after = s[m.index + m[0].length];
    if (after && /\d/.test(after)) continue; // part of a longer number
    pushMatch(out, 'postal_jp', m, 0.85);
  }
}

// address (JP) — signal only, never auto-gated.
const RE_ADDRESS = /(?:東京都|北海道|(?:京都|大阪)府|[一-鿿]{2,3}県)[一-鿿0-9ー\-]{2,}[0-9]/g;
function detectAddress(line, out) {
  for (const m of line.text.matchAll(RE_ADDRESS)) pushMatch(out, 'address_jp', m, 0.5);
}

// ===========================================================================
// EU LOCALE PACK — IBAN (ISO 13616). Self-contained block: country length
// table, ISO 7064 mod-97 checksum, candidate matching, and the detector. Kept
// in one clearly-delimited section so a future locale pack can be added without
// tangling with the universal core above. A candidate is accepted ONLY when its
// length matches the country table AND mod-97 == 1. The checksum is the point —
// like Luhn for cards, it is what keeps false positives at zero: an IBAN-shaped
// string with a broken checksum, or a wrong length for its country code, is
// never boxed.
// ===========================================================================

// ISO 13616 IBAN length by 2-letter country code (full published registry).
const IBAN_LEN = {
  AD: 24, AE: 23, AL: 28, AT: 20, AZ: 28, BA: 20, BE: 16, BG: 22, BH: 22,
  BR: 29, BY: 28, CH: 21, CR: 22, CY: 28, CZ: 24, DE: 22, DK: 18, DO: 28,
  EE: 20, EG: 29, ES: 24, FI: 18, FO: 18, FR: 27, GB: 22, GE: 22, GI: 23,
  GL: 18, GR: 27, GT: 28, HR: 21, HU: 28, IE: 22, IL: 23, IQ: 23, IS: 26,
  IT: 27, JO: 30, KW: 30, KZ: 20, LB: 28, LC: 32, LI: 21, LT: 20, LU: 20,
  LV: 21, LY: 25, MC: 27, MD: 24, ME: 22, MK: 19, MR: 27, MT: 31, MU: 30,
  NL: 18, NO: 15, PK: 24, PL: 28, PS: 29, PT: 25, QA: 29, RO: 24, RS: 22,
  SA: 24, SC: 31, SD: 18, SE: 24, SI: 19, SK: 24, SM: 27, ST: 25, SV: 28,
  TL: 23, TN: 24, TR: 26, UA: 29, VA: 22, VG: 24, XK: 20,
};

// ISO 7064 mod-97-10: move the first 4 chars to the end, map A-Z -> 10..35, and
// reduce the resulting big number mod 97 piecewise. A valid IBAN yields exactly 1.
function ibanMod97(compact) {
  const s = compact.slice(4) + compact.slice(0, 4);
  let rem = 0;
  for (let i = 0; i < s.length; i++) {
    const code = s.charCodeAt(i);
    if (code >= 48 && code <= 57) {        // 0-9
      rem = (rem * 10 + (code - 48)) % 97;
    } else if (code >= 65 && code <= 90) { // A-Z -> 10..35 (two decimal digits)
      rem = (rem * 100 + (code - 55)) % 97;
    } else {
      return false;                        // non-alphanumeric: not an IBAN body
    }
  }
  return rem === 1;
}

// Normalise a candidate: drop spaces, uppercase. (Line text is already
// character-normalised upstream, so letters are ASCII here.)
function ibanCompact(tok) { return tok.replace(/\s+/g, '').toUpperCase(); }

// Validate a compact token as a FULL IBAN: known country, exact country length,
// well-formed body, mod-97 == 1. No prefix trimming — a run longer than the
// country length is a wrong-length lookalike, not an IBAN.
function ibanValid(tok) {
  const c = ibanCompact(tok);
  if (c.length < 15 || c.length > 34) return false;
  const L = IBAN_LEN[c.slice(0, 2)];
  if (!L || c.length !== L) return false;
  if (!/^[A-Z]{2}[0-9]{2}[A-Z0-9]+$/.test(c)) return false;
  return ibanMod97(c);
}

// Candidate shapes on the reconstructed line text:
//   grouped    — CC + 2 check digits, then space-separated groups (of 4, final
//                group 1-4). Each group is bounded by a non-alphanumeric via the
//                lookahead, so a following word is never glued onto the run.
//   contiguous — CC + 2 check digits + 11-30 alphanumerics, no spaces.
// The whole matched run (including its internal spaces) becomes the box, so the
// redaction covers every space-separated group, never a partial group.
const RE_IBAN_GROUPED = /[A-Za-z]{2}[0-9]{2}(?:\s[A-Za-z0-9]{4}(?![A-Za-z0-9]))*(?:\s[A-Za-z0-9]{1,4}(?![A-Za-z0-9]))?/g;
const RE_IBAN_CONTIG = /[A-Za-z]{2}[0-9]{2}[A-Za-z0-9]{11,30}/g;
function detectIban(line, out) {
  const s = line.text;
  const claimed = [];
  const consider = (st, en, tok) => {
    if (!ibanValid(tok)) return;
    if (claimed.some(([a, b]) => st < b && en > a)) return;
    claimed.push([st, en]);
    out.push({ class: 'iban', text: ibanCompact(tok), start: st, end: en, confidence: 0.95 });
  };
  // grouped pass — must contain at least one space, and not start mid-token.
  for (const m of s.matchAll(RE_IBAN_GROUPED)) {
    if (!/\s/.test(m[0])) continue;
    const before = s[m.index - 1];
    if (before && /[A-Za-z0-9]/.test(before)) continue;
    consider(m.index, m.index + m[0].length, m[0]);
  }
  // contiguous pass — reject if glued to adjacent alphanumerics (wrong-length run).
  for (const m of s.matchAll(RE_IBAN_CONTIG)) {
    const before = s[m.index - 1], after = s[m.index + m[0].length];
    if (before && /[A-Za-z0-9]/.test(before)) continue;
    if (after && /[A-Za-z0-9]/.test(after)) continue;
    consider(m.index, m.index + m[0].length, m[0]);
  }
}

const DETECTORS = [detectEmail, detectCard, detectMyNumber, detectIp, detectApiKey,
  detectPhoneJp, detectPhoneIntl, detectPostal, detectAddress, detectIban];

// ---------------------------------------------------------------------------
// Line reconstruction from words + char->word owner map.
// ---------------------------------------------------------------------------
function buildLines(words) {
  const ws = words
    .filter((w) => w && typeof w.text === 'string' && w.text.trim() !== '' && w.bbox)
    .map((w) => ({ ...w, _yc: w.bbox.y + w.bbox.h / 2 }))
    .sort((a, b) => a._yc - b._yc || a.bbox.x - b.bbox.x);

  const lines = [];
  for (const w of ws) {
    let line = lines.find((l) => Math.abs(l.yc - w._yc) <= Math.max(6, w.bbox.h * 0.6));
    if (!line) { line = { yc: w._yc, words: [] }; lines.push(line); }
    line.words.push(w);
    // running mean of the line's vertical centre keeps drift small
    line.yc = line.words.reduce((s, x) => s + x._yc, 0) / line.words.length;
  }

  for (const line of lines) {
    line.words.sort((a, b) => a.bbox.x - b.bbox.x);
    let text = '';
    const owners = []; // char index -> word index (or -1 for inserted spaces)
    line.words.forEach((w, wi) => {
      if (text.length) { text += ' '; owners.push(-1); }
      const t = normalize(w.text);
      for (let i = 0; i < t.length; i++) owners.push(wi);
      text += t;
    });
    line.text = text;
    line.owners = owners;
    // space-stripped view with a map back to original offsets, for key scanning
    let ctext = '';
    const map = [];
    for (let i = 0; i < text.length; i++) {
      if (/\s/.test(text[i])) continue;
      ctext += text[i];
      map.push(i);
    }
    line.compact = { text: ctext, map };
  }
  return lines;
}

function mergeBoxes(boxes) {
  const x1 = Math.min(...boxes.map((b) => b.x));
  const y1 = Math.min(...boxes.map((b) => b.y));
  const x2 = Math.max(...boxes.map((b) => b.x + b.w));
  const y2 = Math.max(...boxes.map((b) => b.y + b.h));
  return { x: x1, y: y1, w: x2 - x1, h: y2 - y1 };
}

// Map a [start,end) character span in a line to the merged bbox of the words
// it overlaps.
function spanBox(line, start, end) {
  const wi = new Set();
  for (let i = start; i < end; i++) {
    const o = line.owners[i];
    if (o >= 0) wi.add(o);
  }
  if (wi.size === 0) return null;
  const boxes = [...wi].map((i) => line.words[i].bbox);
  return mergeBoxes(boxes);
}

// Numeric classes whose value is written as several digit groups. OCR can mangle
// the middle of a long number so that only part of it matches a validator (e.g.
// a garbled card whose first 12 clean digits coincidentally pass the My Number
// check digit), which would box only some of the groups and leave the rest of
// the number readable. For these classes we expand the matched span outward
// over the full CONTIGUOUS run of digit groups it belongs to — digits, ASCII
// spaces, and hyphens — so a partially-read number still hides the whole thing.
// Real text (letters, '@', '.', ':', '(', currency, etc.) breaks the run, so a
// neighbouring unrelated number separated by text is never swallowed.
const NUMERIC_EXPAND = new Set(['credit_card', 'my_number', 'phone_jp', 'phone_intl']);
const isRunChar = (ch) => ch === ' ' || ch === '-' || (ch >= '0' && ch <= '9');
function expandNumericSpan(text, start, end) {
  let s = start, e = end;
  while (s > 0 && isRunChar(text[s - 1])) s--;
  while (e < text.length && isRunChar(text[e])) e++;
  // Trim pure separators at the edges so the box hugs the digit groups.
  while (s < e && (text[s] === ' ' || text[s] === '-')) s++;
  while (e > s && (text[e - 1] === ' ' || text[e - 1] === '-')) e--;
  return [s, e];
}

// ---------------------------------------------------------------------------
// Public API
// ---------------------------------------------------------------------------
function detect(words) {
  const lines = buildLines(words);
  const detections = [];
  for (const line of lines) {
    const raw = [];
    for (const d of DETECTORS) d(line, raw);
    // resolve overlaps: a span already claimed by a higher-priority/validated
    // class wins (card/my_number/api_key/email before phone/address).
    raw.sort((a, b) => b.confidence - a.confidence || (b.end - b.start) - (a.end - a.start));
    const claimed = [];
    for (const r of raw) {
      if (claimed.some(([s, e]) => r.start < e && r.end > s)) continue;
      claimed.push([r.start, r.end]);
      let bs = r.start, be = r.end;
      if (NUMERIC_EXPAND.has(r.class)) [bs, be] = expandNumericSpan(line.text, r.start, r.end);
      const bbox = spanBox(line, bs, be);
      if (!bbox) continue;
      detections.push({ class: r.class, text: r.text, bbox, confidence: r.confidence });
    }
  }
  return detections;
}

export {
  detect, buildLines, normalize, luhn, myNumberCheckDigit, myNumberValid, octetsValid,
  ibanMod97, ibanValid, IBAN_LEN,
};

	// DasuMae detection module — dependency-free.
	//
	// Input: words = [{ text, bbox:{x,y,w,h}, confidence? }] (OCR output, pixels)
	// Output: detections = [{ class, text, bbox:{x,y,w,h}, confidence }]
	//
	// Approach: reconstruct line-level text from words, run freshly-written patterns
	// + validators (Luhn, My Number check digit, octet/range checks, length/charset
	// rules) against each line, then map the matched character span back to the
	// words it covers and merge their boxes to cover the whole planted item.
	//
	// Patterns are written from the public shape of each format (RFC email shape,
	// Luhn, the My Number check-digit algorithm, well-known API-key prefixes). No
	// rule file, pattern list, or code is copied from any other project.
	//
	// ESM module (package.json "type": "module").

	// ---------------------------------------------------------------------------
	// Normalisation — only length-preserving 1:1 replacements, so character offsets
	// stay aligned with the word-owner map.
	// ---------------------------------------------------------------------------
	const DASHES = /[‐‑‒–—―−－]/g; // ‐‑‒–—―−－ -> -
	function normalizeChar(ch) {
	const code = ch.codePointAt(0);
	// full-width digits ０-９ -> 0-9
	if (code >= 0xff10 && code <= 0xff19) return String.fromCharCode(code - 0xff10 + 0x30);
	// full-width Latin letters Ａ-Ｚａ-ｚ -> ASCII
	if (code >= 0xff21 && code <= 0xff3a) return String.fromCharCode(code - 0xff21 + 0x41);
	if (code >= 0xff41 && code <= 0xff5a) return String.fromCharCode(code - 0xff41 + 0x61);
	if (code === 0xff20) return '@'; // ＠
	if (code === 0xff0e) return '.'; // ．
	return ch;
	}
	function normalize(s) {
	let out = '';
	for (const ch of String(s)) out += normalizeChar(ch);
	return out.replace(DASHES, '-');
	}

	// ---------------------------------------------------------------------------
	// Validators (exported for unit tests)
	// ---------------------------------------------------------------------------
	function luhn(numStr) {
	const s = String(numStr).replace(/\D/g, '');
	if (s.length === 0) return false;
	let sum = 0, alt = false;
	for (let i = s.length - 1; i >= 0; i--) {
	let d = s.charCodeAt(i) - 48;
	if (alt) { d *= 2; if (d > 9) d -= 9; }
	sum += d; alt = !alt;
	}
	return sum % 10 === 0;
	}

	function myNumberCheckDigit(payload11) {
	let sum = 0;
	for (let n = 1; n <= 11; n++) {
	const p = payload11.charCodeAt(11 - n) - 48;
	const q = n <= 6 ? n + 1 : n - 5;
	sum += p * q;
	}
	const r = sum % 11;
	return r <= 1 ? 0 : 11 - r;
	}
	function myNumberValid(twelveDigits) {
	const s = String(twelveDigits).replace(/\D/g, '');
	if (!/^\d{12}$/.test(s)) return false;
	return myNumberCheckDigit(s.slice(0, 11)) === s.charCodeAt(11) - 48;
	}

	function octetsValid(ip) {
	const parts = ip.split('.');
	if (parts.length !== 4) return false;
	return parts.every((p) => /^\d{1,3}$/.test(p) && Number(p) <= 255 &&
	(p.length === 1 \|\| p[0] !== '0')); // reject 01, 00 style
	}

	// ---------------------------------------------------------------------------
	// Detectors. Each returns matches [{ class, text, start, end, confidence }]
	// against a normalised line string. start/end are character offsets.
	// Confidence heuristic (documented):
	// 0.95 exact validator passed (Luhn, My Number check digit, IBAN mod-97) or a
	// distinctive vendor key prefix / 〒 marker matched
	// 0.90 strong structural pattern with a range/shape check (email, IP)
	// 0.80 pattern-only match (phone) — shape matched, no external validator
	// 0.50 signal-only class (address_jp) — surfaced, never auto-gated
	// ---------------------------------------------------------------------------

	function pushMatch(out, cls, m, conf, textOverride) {
	out.push({ class: cls, text: textOverride ?? m[0], start: m.index, end: m.index + m[0].length, confidence: conf });
	}

	// Each detector receives the line object { text, owners, compact } and pushes
	// raw matches with character offsets into line.text.

	// email — OCR often drops the tiny dot in ".jp", so a dot-less domain is still
	// accepted when it ends in a known TLD. (No '@' appears in the clean set, so FP
	// stays 0 despite the looser rule.)
	// ',' is in the domain charset because OCR frequently reads a domain '.' as ','
	// (e.g. "example,jp"); it is normalised back to '.' below. The '@' is required
	// and never appears in the clean set, so FP stays 0 despite the looser rule.
	const RE_EMAIL = /[A-Za-z0-9._%+\-]+@[A-Za-z0-9.,\-]{2,}/g;
	const EMAIL_TLD = /(?:com\|jp\|org\|net\|co\|io\|dev\|info\|edu\|gov\|ne)$/i;
	// Space-tolerant recovery pass: OCR often splits a domain with a stray space
	// ("example. com", "examp le. com"), which breaks a contiguous match. On the
	// space-stripped view we match up to a KNOWN TLD boundary (\b), so the match
	// can't run on into the next word. '@' is required (absent from the clean set).
	const RE_EMAIL_COMPACT = /[A-Za-z0-9._%+\-]+@[A-Za-z0-9.,\-]+?[.,](?:com\|jp\|org\|net\|co\|io\|dev\|info\|edu\|gov\|ne)\b/gi;
	function detectEmail(line, out) {
	const s = line.text;
	for (const m of s.matchAll(RE_EMAIL)) {
	let tok = m[0].replace(/[.,\-]+$/, ''); // trim trailing punctuation
	tok = tok.replace(/,/g, '.'); // OCR comma -> dot (length-preserving)
	const at = tok.indexOf('@');
	const local = tok.slice(0, at), domain = tok.slice(at + 1);
	if (!local \|\| domain.length < 2 \|\| /\.\./.test(tok)) continue;
	const hasDot = domain.includes('.');
	if (hasDot ? !/\.[A-Za-z]{2,}$/.test(tok) : !EMAIL_TLD.test(domain)) continue;
	out.push({ class: 'email', text: tok, start: m.index, end: m.index + tok.length, confidence: hasDot ? 0.9 : 0.75 });
	}
	// recovery pass on the compact view (offsets mapped back to the raw line)
	const { text: ctext, map } = line.compact;
	for (const m of ctext.matchAll(RE_EMAIL_COMPACT)) {
	const tok = m[0].replace(/,/g, '.');
	if (tok.indexOf('@') < 1 \|\| /\.\./.test(tok)) continue;
	const os = map[m.index], oe = map[m.index + m[0].length - 1] + 1;
	out.push({ class: 'email', text: tok, start: os, end: oe, confidence: 0.85 });
	}
	}

	// credit card — digit runs (spaces/dashes allowed), 13–19 digits, Luhn.
	const RE_CARD = /\d(?:[ -]?\d){11,18}/g;
	function detectCard(line, out) {
	const s = line.text;
	for (const m of s.matchAll(RE_CARD)) {
	const digits = m[0].replace(/\D/g, '');
	if (digits.length < 13 \|\| digits.length > 19) continue;
	const before = s[m.index - 1], after = s[m.index + m[0].length];
	if (before && /[A-Za-z]/.test(before)) continue;
	if (after && /[A-Za-z]/.test(after)) continue;
	if (!luhn(digits)) continue;
	pushMatch(out, 'credit_card', m, 0.95);
	}
	}

	// My Number — exactly 12 digits (grouped or contiguous), valid check digit.
	const RE_MYNUM = /\d(?:[ -]?\d){11}/g;
	function detectMyNumber(line, out) {
	const s = line.text;
	for (const m of s.matchAll(RE_MYNUM)) {
	const digits = m[0].replace(/\D/g, '');
	if (digits.length !== 12) continue;
	const before = s[m.index - 1], after = s[m.index + m[0].length];
	if (before && /[0-9A-Za-z]/.test(before)) continue;
	if (after && /[0-9A-Za-z]/.test(after)) continue;
	if (!myNumberValid(digits)) continue;
	pushMatch(out, 'my_number', m, 0.95);
	}
	}

	// IP address — 4 dotted octets, each 0–255.
	const RE_IP = /\d{1,3}(?:\.\d{1,3}){3}/g;
	function detectIp(line, out) {
	const s = line.text;
	for (const m of s.matchAll(RE_IP)) {
	if (!octetsValid(m[0])) continue;
	const before = s[m.index - 1], after = s[m.index + m[0].length];
	if (before && /[0-9.]/.test(before)) continue; // part of a longer dotted string
	if (after && /[0-9.]/.test(after)) continue;
	pushMatch(out, 'ip_address', m, 0.9);
	}
	}

	// API keys — prefix-anchored, run on a SPACE-STRIPPED view of the line because
	// OCR scatters spaces inside a long key and doubles dashes ("sk-" -> "sk--").
	// Once a vendor prefix is seen the whole token is secret, so the body is matched
	// loosely (length-tolerant, allows the stray @/./- that OCR substitutes in). The
	// prefixes don't occur in ordinary UI text, so clean-set FP stays 0.
	// Key body charset: the real key alphabet (base64/base62 + _ @ . - + / =) plus
	// the stray '[' ']' OCR sometimes injects mid-token (e.g. "sk-pro]j-..."). Once a
	// vendor prefix is seen the whole run is secret, so the body is matched loosely.
	const BODY = '[A-Za-z0-9_@.+/=\\-\\]\\[]';
	const JWTBODY = '[A-Za-z0-9_@\\-]'; // JWT part body, tolerating OCR '0'->'@'
	const API_PATTERNS = [
	/gh[opusr][_\-][A-Za-z0-9@\]\[]{18,}/g,
	/github[_\-]pat[_\-][A-Za-z0-9_@\]\[]{25,}/g,
	/AKIA[A-Z0-9]{12,}/g,
	new RegExp(`x[o0]x[bpars]${BODY}{16,}`, 'g'), // Slack xox*-; tolerate OCR o->0
	new RegExp(`eyJ${JWTBODY}{6,}\\.${JWTBODY}{6,}\\.${JWTBODY}{4,}`, 'g'), // JWT
	new RegExp(`sk-+${BODY}{16,}`, 'g'), // sk-ant / sk-proj / any sk- key
	];
	function detectApiKey(line, out) {
	const { text: ctext, map } = line.compact;
	const taken = [];
	for (const re of API_PATTERNS) {
	for (const m of ctext.matchAll(re)) {
	const s = m.index, e = m.index + m[0].length;
	if (taken.some(([ts, te]) => s < te && e > ts)) continue;
	taken.push([s, e]);
	// map compact offsets back to original line.text offsets for box lookup
	const os = map[s], oe = map[e - 1] + 1;
	out.push({ class: 'api_key', text: m[0], start: os, end: oe, confidence: 0.95 });
	}
	}
	}

	// phone (JP) — leading 0, dashed groups, <=11 digits total.
	const RE_PHONE_JP = /0\d{0,3}-\d{2,4}-\d{3,4}/g;
	function detectPhoneJp(line, out) {
	const s = line.text;
	for (const m of s.matchAll(RE_PHONE_JP)) {
	const digits = m[0].replace(/\D/g, '');
	if (digits.length < 9 \|\| digits.length > 11) continue;
	const before = s[m.index - 1], after = s[m.index + m[0].length];
	if (before && /[0-9]/.test(before)) continue;
	if (after && /[0-9-]/.test(after)) continue;
	pushMatch(out, 'phone_jp', m, 0.8);
	}
	}

	// phone (international) — leading +, 8–15 digits.
	const RE_PHONE_INTL = /\+\d{1,3}[\s-]?\(?\d{1,4}\)?[\s-]?\d{2,4}[\s-]?\d{2,4}/g;
	function detectPhoneIntl(line, out) {
	const s = line.text;
	for (const m of s.matchAll(RE_PHONE_INTL)) {
	const digits = m[0].replace(/\D/g, '');
	if (digits.length < 8 \|\| digits.length > 15) continue;
	pushMatch(out, 'phone_intl', m, 0.8);
	}
	}

	// postal (JP) — the 〒 marker is required, but OCR reads 〒 as "T"/"TT"
	// (look-alike), so an isolated leading T before NNN-NNNN is accepted too. Bare
	// NNN-NNNN with no marker is left to manual to avoid flagging generic numbers.
	// The 〒 marker (or its "T"/"TT" OCR look-alike) is required. The first group is
	// \d{3,4}: a true code is 〒NNN-NNNN, but OCR often injects a phantom digit right
	// after 〒 (e.g. "〒7080-7118"), so a 4-digit first group is accepted too.
	const RE_POSTAL = /(?:〒\|\bT{1,2})\s?\d{3,4}\s?-\s?\d{4}/g;
	function detectPostal(line, out) {
	const s = line.text;
	for (const m of s.matchAll(RE_POSTAL)) {
	const after = s[m.index + m[0].length];
	if (after && /\d/.test(after)) continue; // part of a longer number
	pushMatch(out, 'postal_jp', m, 0.85);
	}
	}

	// address (JP) — signal only, never auto-gated.
	const RE_ADDRESS = /(?:東京都\|北海道\|(?:京都\|大阪)府\|[一-鿿]{2,3}県)[一-鿿0-9ー\-]{2,}[0-9]/g;
	function detectAddress(line, out) {
	for (const m of line.text.matchAll(RE_ADDRESS)) pushMatch(out, 'address_jp', m, 0.5);
	}

	// ===========================================================================
	// EU LOCALE PACK — IBAN (ISO 13616). Self-contained block: country length
	// table, ISO 7064 mod-97 checksum, candidate matching, and the detector. Kept
	// in one clearly-delimited section so a future locale pack can be added without
	// tangling with the universal core above. A candidate is accepted ONLY when its
	// length matches the country table AND mod-97 == 1. The checksum is the point —
	// like Luhn for cards, it is what keeps false positives at zero: an IBAN-shaped
	// string with a broken checksum, or a wrong length for its country code, is
	// never boxed.
	// ===========================================================================

	// ISO 13616 IBAN length by 2-letter country code (full published registry).
	const IBAN_LEN = {
	AD: 24, AE: 23, AL: 28, AT: 20, AZ: 28, BA: 20, BE: 16, BG: 22, BH: 22,
	BR: 29, BY: 28, CH: 21, CR: 22, CY: 28, CZ: 24, DE: 22, DK: 18, DO: 28,
	EE: 20, EG: 29, ES: 24, FI: 18, FO: 18, FR: 27, GB: 22, GE: 22, GI: 23,
	GL: 18, GR: 27, GT: 28, HR: 21, HU: 28, IE: 22, IL: 23, IQ: 23, IS: 26,
	IT: 27, JO: 30, KW: 30, KZ: 20, LB: 28, LC: 32, LI: 21, LT: 20, LU: 20,
	LV: 21, LY: 25, MC: 27, MD: 24, ME: 22, MK: 19, MR: 27, MT: 31, MU: 30,
	NL: 18, NO: 15, PK: 24, PL: 28, PS: 29, PT: 25, QA: 29, RO: 24, RS: 22,
	SA: 24, SC: 31, SD: 18, SE: 24, SI: 19, SK: 24, SM: 27, ST: 25, SV: 28,
	TL: 23, TN: 24, TR: 26, UA: 29, VA: 22, VG: 24, XK: 20,
	};

	// ISO 7064 mod-97-10: move the first 4 chars to the end, map A-Z -> 10..35, and
	// reduce the resulting big number mod 97 piecewise. A valid IBAN yields exactly 1.
	function ibanMod97(compact) {
	const s = compact.slice(4) + compact.slice(0, 4);
	let rem = 0;
	for (let i = 0; i < s.length; i++) {
	const code = s.charCodeAt(i);
	if (code >= 48 && code <= 57) { // 0-9
	rem = (rem * 10 + (code - 48)) % 97;
	} else if (code >= 65 && code <= 90) { // A-Z -> 10..35 (two decimal digits)
	rem = (rem * 100 + (code - 55)) % 97;
	} else {
	return false; // non-alphanumeric: not an IBAN body
	}
	}
	return rem === 1;
	}

	// Normalise a candidate: drop spaces, uppercase. (Line text is already
	// character-normalised upstream, so letters are ASCII here.)
	function ibanCompact(tok) { return tok.replace(/\s+/g, '').toUpperCase(); }

	// Validate a compact token as a FULL IBAN: known country, exact country length,
	// well-formed body, mod-97 == 1. No prefix trimming — a run longer than the
	// country length is a wrong-length lookalike, not an IBAN.
	function ibanValid(tok) {
	const c = ibanCompact(tok);
	if (c.length < 15 \|\| c.length > 34) return false;
	const L = IBAN_LEN[c.slice(0, 2)];
	if (!L \|\| c.length !== L) return false;
	if (!/^[A-Z]{2}[0-9]{2}[A-Z0-9]+$/.test(c)) return false;
	return ibanMod97(c);
	}

	// Candidate shapes on the reconstructed line text:
	// grouped — CC + 2 check digits, then space-separated groups (of 4, final
	// group 1-4). Each group is bounded by a non-alphanumeric via the
	// lookahead, so a following word is never glued onto the run.
	// contiguous — CC + 2 check digits + 11-30 alphanumerics, no spaces.
	// The whole matched run (including its internal spaces) becomes the box, so the
	// redaction covers every space-separated group, never a partial group.
	const RE_IBAN_GROUPED = /[A-Za-z]{2}[0-9]{2}(?:\s[A-Za-z0-9]{4}(?![A-Za-z0-9]))*(?:\s[A-Za-z0-9]{1,4}(?![A-Za-z0-9]))?/g;
	const RE_IBAN_CONTIG = /[A-Za-z]{2}[0-9]{2}[A-Za-z0-9]{11,30}/g;
	function detectIban(line, out) {
	const s = line.text;
	const claimed = [];
	const consider = (st, en, tok) => {
	if (!ibanValid(tok)) return;
	if (claimed.some(([a, b]) => st < b && en > a)) return;
	claimed.push([st, en]);
	out.push({ class: 'iban', text: ibanCompact(tok), start: st, end: en, confidence: 0.95 });
	};
	// grouped pass — must contain at least one space, and not start mid-token.
	for (const m of s.matchAll(RE_IBAN_GROUPED)) {
	if (!/\s/.test(m[0])) continue;
	const before = s[m.index - 1];
	if (before && /[A-Za-z0-9]/.test(before)) continue;
	consider(m.index, m.index + m[0].length, m[0]);
	}
	// contiguous pass — reject if glued to adjacent alphanumerics (wrong-length run).
	for (const m of s.matchAll(RE_IBAN_CONTIG)) {
	const before = s[m.index - 1], after = s[m.index + m[0].length];
	if (before && /[A-Za-z0-9]/.test(before)) continue;
	if (after && /[A-Za-z0-9]/.test(after)) continue;
	consider(m.index, m.index + m[0].length, m[0]);
	}
	}

	const DETECTORS = [detectEmail, detectCard, detectMyNumber, detectIp, detectApiKey,
	detectPhoneJp, detectPhoneIntl, detectPostal, detectAddress, detectIban];

	// ---------------------------------------------------------------------------
	// Line reconstruction from words + char->word owner map.
	// ---------------------------------------------------------------------------
	function buildLines(words) {
	const ws = words
	.filter((w) => w && typeof w.text === 'string' && w.text.trim() !== '' && w.bbox)
	.map((w) => ({ ...w, _yc: w.bbox.y + w.bbox.h / 2 }))
	.sort((a, b) => a._yc - b._yc \|\| a.bbox.x - b.bbox.x);

	const lines = [];
	for (const w of ws) {
	let line = lines.find((l) => Math.abs(l.yc - w._yc) <= Math.max(6, w.bbox.h * 0.6));
	if (!line) { line = { yc: w._yc, words: [] }; lines.push(line); }
	line.words.push(w);
	// running mean of the line's vertical centre keeps drift small
	line.yc = line.words.reduce((s, x) => s + x._yc, 0) / line.words.length;
	}

	for (const line of lines) {
	line.words.sort((a, b) => a.bbox.x - b.bbox.x);
	let text = '';
	const owners = []; // char index -> word index (or -1 for inserted spaces)
	line.words.forEach((w, wi) => {
	if (text.length) { text += ' '; owners.push(-1); }
	const t = normalize(w.text);
	for (let i = 0; i < t.length; i++) owners.push(wi);
	text += t;
	});
	line.text = text;
	line.owners = owners;
	// space-stripped view with a map back to original offsets, for key scanning
	let ctext = '';
	const map = [];
	for (let i = 0; i < text.length; i++) {
	if (/\s/.test(text[i])) continue;
	ctext += text[i];
	map.push(i);
	}
	line.compact = { text: ctext, map };
	}
	return lines;
	}

	function mergeBoxes(boxes) {
	const x1 = Math.min(...boxes.map((b) => b.x));
	const y1 = Math.min(...boxes.map((b) => b.y));
	const x2 = Math.max(...boxes.map((b) => b.x + b.w));
	const y2 = Math.max(...boxes.map((b) => b.y + b.h));
	return { x: x1, y: y1, w: x2 - x1, h: y2 - y1 };
	}

	// Map a [start,end) character span in a line to the merged bbox of the words
	// it overlaps.
	function spanBox(line, start, end) {
	const wi = new Set();
	for (let i = start; i < end; i++) {
	const o = line.owners[i];
	if (o >= 0) wi.add(o);
	}
	if (wi.size === 0) return null;
	const boxes = [...wi].map((i) => line.words[i].bbox);
	return mergeBoxes(boxes);
	}

	// Numeric classes whose value is written as several digit groups. OCR can mangle
	// the middle of a long number so that only part of it matches a validator (e.g.
	// a garbled card whose first 12 clean digits coincidentally pass the My Number
	// check digit), which would box only some of the groups and leave the rest of
	// the number readable. For these classes we expand the matched span outward
	// over the full CONTIGUOUS run of digit groups it belongs to — digits, ASCII
	// spaces, and hyphens — so a partially-read number still hides the whole thing.
	// Real text (letters, '@', '.', ':', '(', currency, etc.) breaks the run, so a
	// neighbouring unrelated number separated by text is never swallowed.
	const NUMERIC_EXPAND = new Set(['credit_card', 'my_number', 'phone_jp', 'phone_intl']);
	const isRunChar = (ch) => ch === ' ' \|\| ch === '-' \|\| (ch >= '0' && ch <= '9');
	function expandNumericSpan(text, start, end) {
	let s = start, e = end;
	while (s > 0 && isRunChar(text[s - 1])) s--;
	while (e < text.length && isRunChar(text[e])) e++;
	// Trim pure separators at the edges so the box hugs the digit groups.
	while (s < e && (text[s] === ' ' \|\| text[s] === '-')) s++;
	while (e > s && (text[e - 1] === ' ' \|\| text[e - 1] === '-')) e--;
	return [s, e];
	}

	// ---------------------------------------------------------------------------
	// Public API
	// ---------------------------------------------------------------------------
	function detect(words) {
	const lines = buildLines(words);
	const detections = [];
	for (const line of lines) {
	const raw = [];
	for (const d of DETECTORS) d(line, raw);
	// resolve overlaps: a span already claimed by a higher-priority/validated
	// class wins (card/my_number/api_key/email before phone/address).
	raw.sort((a, b) => b.confidence - a.confidence \|\| (b.end - b.start) - (a.end - a.start));
	const claimed = [];
	for (const r of raw) {
	if (claimed.some(([s, e]) => r.start < e && r.end > s)) continue;
	claimed.push([r.start, r.end]);
	let bs = r.start, be = r.end;
	if (NUMERIC_EXPAND.has(r.class)) [bs, be] = expandNumericSpan(line.text, r.start, r.end);
	const bbox = spanBox(line, bs, be);
	if (!bbox) continue;
	detections.push({ class: r.class, text: r.text, bbox, confidence: r.confidence });
	}
	}
	return detections;
	}

	export {
	detect, buildLines, normalize, luhn, myNumberCheckDigit, myNumberValid, octetsValid,
	ibanMod97, ibanValid, IBAN_LEN,
	};