@stll/anonymize-wasm 1.4.11 → 1.5.0

package/dist/wasm.mjs

@@ -1,6 +1,6 @@
 import { DEFAULT_ENTITY_LABELS, DETECTION_SOURCES, DETECTOR_PRIORITY, OPERATOR_TYPES } from "./constants.mjs";
 import { TextSearch } from "@stll/text-search-wasm";
-import { at, au, be, bg, br, ch, cn, cy, cz, de, dk, ee, es, fi, fr, gb, gr, hr, hu, ie, it, lt, lu, lv, mt, nl, no, pl, pt, ro, se, si, sk, us } from "@stll/stdnum";
+import { at, au, be, bg, br, ch, cn, crypto, cy, cz, de, dk, ee, es, fi, fr, gb, gr, hr, hu, ie, it, lt, lu, lv, mt, nl, no, pl, pt, ro, se, si, sk, us } from "@stll/stdnum";
 import { toRegex } from "@stll/stdnum/patterns";
 //#region \0rolldown/runtime.js
 var __defProp = Object.defineProperty;
@@ -7959,6 +7959,15 @@ const processCountryMatches = (allMatches, sliceStart, sliceEnd, fullText, data)
 	return results;
 };
 //#endregion
+//#region src/util/entity-source.ts
+/**
+* True when the entity came from a caller-supplied detector — a
+* custom deny-list entry or a custom regex. These spans carry
+* caller-locked boundaries: the pipeline does not sanitise, trim,
+* or boundary-adjust them, since the caller chose the exact text.
+*/
+const isCallerOwnedEntity = (entity) => entity.sourceDetail === "custom-deny-list" || entity.sourceDetail === "custom-regex";
+//#endregion
 //#region src/config/titles.ts
 /**
 * Academic and professional title prefixes.
@@ -8405,6 +8414,7 @@ const initNameCorpus = (ctx = defaultContext, dictionaries, languages) => {
 				titleTokens: Object.freeze(new Set(dedupTitles)),
 				titleAbbreviations: Object.freeze(titleAbbrSet),
 				excludedWords: Object.freeze(new Set(exclusions)),
+				commonWords: Object.freeze(commonWords),
 				nonWesternNames: Object.freeze(new Set(dedupNonWestern)),
 				excludedAllCaps: Object.freeze(new Set(dedupExcludedAllCaps)),
 				firstNamesList: Object.freeze(dedupFirst),
@@ -9119,13 +9129,14 @@ const detectNameCorpus = (fullText, ctx = defaultContext, options = {}) => {
 		const corpusCount = chain.filter((t) => isCorpusMatch(t.type)).length;
 		const capitalizedCount = chain.filter((t) => t.type === TOKEN_TYPE.CAPITALIZED).length;
 		const nonWesternCount = chain.filter((t) => t.nonWestern).length;
+		const chainAllCommonWords = chain.every((t) => corpus.commonWords.has(t.text.toLowerCase()));
 		let score;
 		if (hasNonWestern) {
 			if (hasTitle && (nonWesternCount > 0 || capitalizedCount > 0)) score = .95;
 			else if (hasJaSuffix && (capitalizedCount > 0 || nonWesternCount > 0)) score = .9;
 			else if (hasArabicConnector && nonWesternCount > 0) score = .9;
 			else if (nonWesternCount >= 2) score = .9;
-			else if (nonWesternCount > 0 && (capitalizedCount > 0 || hasAbbreviation)) score = .9;
+			else if (nonWesternCount > 0 && (capitalizedCount > 0 || hasAbbreviation) && !chainAllCommonWords) score = .9;
 			else if (nonWesternCount === 1 && chain.length === 1 && !isSentenceStart(fullText, token.start)) score = .5;
 			else continue;
 			if (supplementalMode && score < .9) continue;
@@ -9754,6 +9765,7 @@ var amount_words_default = {
 //#region src/detectors/regex.ts
 const MIN_PHONE_LENGTH = 7;
 const MIN_MONTH_NAME_LENGTH = 3;
+const US_STATE_CODE = "(?:(?i:A[KLZR]|C[AOT]|D[CE]|FL|GA|HI|I[ADL]|K[SY]|LA|M[ADEHINOPST]|N[CDEHJMVY]|O[HK]|PA|RI|S[CD]|T[NX]|UT|V[AIT]|W[AIVY])|I[Nn]|iN|O[Rr]|oR)";
 const escapeTitle = (title) => title.replace(/[.*+?^${}()|[\]\\]/g, "\\$&").replace(/\s+/g, "\\s*");
 /** Escape for use inside a regex alternation. */
 const escapeRegex$2 = (s) => s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
@@ -9991,6 +10003,58 @@ const ES_CIF = {
 	score: .95,
 	validator: es.cif
 };
+const NHS_NUMBER_CONTEXT = {
+	pattern: "\\b(?:(?i:NHS)(?:[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|#))?|(?i:National)[^\\S\\n]+(?i:Health)[^\\S\\n]+(?i:Service)[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|#))[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}\\d{3}[^\\S\\n]?\\d{3}[^\\S\\n]?\\d{4}\\b",
+	label: "national identification number",
+	score: .95,
+	validator: gb.nhs,
+	validatorInput: (text) => text.replace(/\D/g, "")
+};
+const PASSPORT_CONTEXT = {
+	pattern: "\\b(?:(?:(?i:U\\.?S\\.?)|(?i:USA)|(?i:United)[^\\S\\n]+(?i:States)|(?i:UK)|(?i:U\\.?K\\.?)|(?i:GBR)|(?i:British)|(?i:EU)|(?i:European)|(?i:France)|(?i:French))[^\\S\\n]{1,4})?(?i:passports?)(?:[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|#))?[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}(?:[A-Za-z]{1,2}\\d{6,8}|\\d{2}[A-Za-z]{2}\\d{5}|\\d{7,9})\\b",
+	label: "passport number",
+	score: .96
+};
+const FR_CNI_CONTEXT = {
+	pattern: "\\b(?:(?i:CNI)|(?i:carte)[^\\S\\n]+(?i:nationale)[^\\S\\n]+(?i:d)[’'](?i:identit[ée])|(?i:French)[^\\S\\n]+(?i:national)[^\\S\\n]+(?i:identity)[^\\S\\n]+(?i:card))(?:[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|(?i:n[°ºo]\\.?))[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}[A-Za-z0-9]{9,12}|[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}(?=[A-Za-z0-9]{0,11}\\d)[A-Za-z0-9]{9,12})\\b",
+	label: "identity card number",
+	score: .96
+};
+const CY_TIC_CONTEXT = {
+	pattern: "\\b(?:(?:(?i:Cyprus)|(?i:Cypriot))[^\\S\\n]{1,4})?(?:(?i:TIC)|(?i:tax)[^\\S\\n]+(?i:identification)[^\\S\\n]+(?i:code))(?:[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|#))?[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}\\d{8}[A-Za-z]\\b",
+	label: "tax identification number",
+	score: .96
+};
+const CY_ID_CARD_CONTEXT = {
+	pattern: "\\b(?:(?i:Cyprus)|(?i:Cypriot))[^\\S\\n]+(?:(?i:identity)[^\\S\\n]+(?i:card)|(?i:ID)[^\\S\\n]+(?i:card))(?:[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|#))?[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}\\d{6,8}\\b",
+	label: "identity card number",
+	score: .96
+};
+const UK_DRIVING_LICENCE_CONTEXT = {
+	pattern: "\\b(?:(?:(?i:UK)|(?i:U\\.?K\\.?)|(?i:British))[^\\S\\n]{1,4})?(?i:driving)[^\\S\\n]+(?i:licen[cs]e)(?:[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|#))?[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}[A-Za-z9]{5}\\d{6}[A-Za-z0-9]{2}\\d[A-Za-z]{2}\\b",
+	label: "identity card number",
+	score: .96
+};
+const US_DRIVER_LICENSE_CONTEXT = {
+	pattern: `\\b(?:(?:(?i:U\\.?S\\.?)|(?i:USA)|(?i:United)[^\\S\\n]+(?i:States)|${US_STATE_CODE})[^\\S\\n]{1,4})?(?:(?i:driver['’]?s?)|(?i:driving))[^\\S\\n]+(?i:licen[cs]e)(?:[^\\S\\n]+(?:(?i:number)|(?i:no\\.?)|#)[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}[A-Za-z0-9]{5,15}|[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}(?=[A-Za-z0-9]{0,14}\\d)[A-Za-z0-9]{5,15})\\b`,
+	label: "identity card number",
+	score: .8
+};
+const MEDICAL_LICENSE_CONTEXT = {
+	pattern: "\\b(?:(?:(?i:GMC)|(?i:NMC))(?:[^\\S\\n]+(?:(?i:licen[cs]e)|(?i:registration)|(?i:reg\\.?)|(?i:pin)|(?i:number)|(?i:no\\.?)))*|(?:(?i:medical)|(?i:physician)|(?i:doctor)|(?i:surgeon)|(?i:nursing)|(?i:nurse))(?:[^\\S\\n]+(?:(?i:licen[cs]e)|(?i:registration)|(?i:reg\\.?)|(?i:pin)|(?i:number)|(?i:no\\.?)))+)[^\\S\\n]{0,4}:?[^\\S\\n]{0,4}(?:[A-Za-z]{0,3}\\d{5,8}|\\d{2}[A-Za-z]\\d{4}[A-Za-z])\\b",
+	label: "registration number",
+	score: .85
+};
+const CRYPTO_WALLET_CANDIDATE = crypto.wallet.candidatePattern ?? "(?!)";
+const CRYPTO_WALLET_CANDIDATE_REGEX = new RegExp(CRYPTO_WALLET_CANDIDATE);
+const getCryptoWalletCandidate = (text) => CRYPTO_WALLET_CANDIDATE_REGEX.exec(text)?.[0] ?? text;
+const CRYPTO_WALLET_ADDRESS = {
+	pattern: "\\b(?:0x[0-9A-Fa-f]{40}|bc1[ac-hj-np-z02-9]{11,71}|BC1[AC-HJ-NP-Z02-9]{11,71})\\b|\\b(?:(?i:BTC|Bitcoin|crypto|wallet|address)[^\\S\\n]{0,4}:?[^\\S\\n]{1,8}){1,4}[13][a-km-zA-HJ-NP-Z1-9]{25,34}\\b",
+	label: "crypto",
+	score: .85,
+	validator: crypto.wallet,
+	validatorInput: getCryptoWalletCandidate
+};
 const AU_ABN_FORMATTED = {
 	pattern: `\\b\\d{2}[^\\S\\n]\\d{3}[^\\S\\n]\\d{3}[^\\S\\n]\\d{3}\\b`,
 	label: "tax identification number",
@@ -10137,6 +10201,15 @@ const ALL_REGEX_DEFS = [
 	ES_DNI,
 	ES_NIE,
 	ES_CIF,
+	NHS_NUMBER_CONTEXT,
+	PASSPORT_CONTEXT,
+	FR_CNI_CONTEXT,
+	CY_TIC_CONTEXT,
+	CY_ID_CARD_CONTEXT,
+	UK_DRIVING_LICENCE_CONTEXT,
+	US_DRIVER_LICENSE_CONTEXT,
+	MEDICAL_LICENSE_CONTEXT,
+	CRYPTO_WALLET_ADDRESS,
 	AU_ABN_FORMATTED,
 	NO_ORGNR_FORMATTED,
 	NO_MVA_FORMATTED,
@@ -10168,6 +10241,7 @@ const REGEX_META = ALL_REGEX_DEFS.map((d) => {
 		score: d.score
 	};
 	if (d.validator) meta.validator = d.validator;
+	if (d.validatorInput) meta.validatorInput = d.validatorInput;
 	return meta;
 });
 /**
@@ -10439,7 +10513,8 @@ const processRegexMatches = (allMatches, sliceStart, sliceEnd, meta_) => {
 		if (!meta) continue;
 		if (meta.sourceDetail !== "custom-regex" && meta.label === "phone number" && match.text.length < MIN_PHONE_LENGTH) continue;
 		if (meta.validator) {
-			const compacted = meta.validator.compact(match.text);
+			const validatorText = meta.validatorInput ? meta.validatorInput(match.text) : match.text;
+			const compacted = meta.validator.compact(validatorText);
 			if (!meta.validator.validate(compacted).valid) continue;
 		}
 		const entity = {
@@ -11895,7 +11970,40 @@ const trimTrailingAddressProse = (text) => {
 	}
 	return text;
 };
-const normalizeEntity = (entity) => {
+const collapseEntityWhitespace = (text) => text.replace(/\s*\n\s*/g, " ").replace(/\s{2,}/g, " ");
+const normalizeEntityFromRaw = (entity, fullText) => {
+	let start = entity.start;
+	let end = entity.end;
+	let text = fullText.slice(start, end);
+	const trimLeading = (re) => {
+		const match = re.exec(text);
+		if (!match) return;
+		start += match[0].length;
+		text = text.slice(match[0].length);
+	};
+	trimLeading(LEADING_ARTIFACT_RE);
+	trimLeading(/^\s+/u);
+	if (entity.label === "address") {
+		trimLeading(ADDRESS_ROLE_PREFIX_RE);
+		const beforeTrim = text;
+		text = trimTrailingAddressProse(text);
+		end -= beforeTrim.length - text.length;
+	}
+	const trailingMatch = /[,\s]+$/u.exec(text);
+	if (trailingMatch) {
+		end -= trailingMatch[0].length;
+		text = text.slice(0, text.length - trailingMatch[0].length);
+	}
+	if (text.length === 0) return null;
+	return {
+		...entity,
+		start,
+		end,
+		text: collapseEntityWhitespace(text)
+	};
+};
+const normalizeEntity = (entity, fullText) => {
+	if (fullText !== void 0) return normalizeEntityFromRaw(entity, fullText);
 	let start = entity.start;
 	let text = entity.text;
 	const trimLeading = (re) => {
@@ -11913,10 +12021,13 @@ const normalizeEntity = (entity) => {
 	const trailingMatch = /[,\s]+$/u.exec(text);
 	if (trailingMatch) text = text.slice(0, text.length - trailingMatch[0].length);
 	if (text.length === 0) return null;
+	const removedFromFront = start - entity.start;
+	const removedFromBack = entity.text.length - removedFromFront - text.length;
+	const end = entity.end - removedFromBack;
 	return {
 		...entity,
 		start,
-		end: start + text.length,
+		end,
 		text
 	};
 };
@@ -12021,7 +12132,6 @@ const initAddressComponents = () => {
 * Czech house number form.
 */
 const hasAddressComponent = (text) => _streetTypesRe.test(text) || ADDRESS_COMPONENT_EXTRA_RE.test(text);
-const isCallerOwnedEntity$3 = (entity) => entity.sourceDetail === "custom-deny-list" || entity.sourceDetail === "custom-regex";
 /**
 * Filter out entities that are likely false positives:
 * template placeholders, clause/section numbers,
@@ -12034,14 +12144,14 @@ const filterFalsePositives = (entities, ctx = defaultContext, fullText) => {
 	const filtered = [];
 	const roles = getGenericRoles(ctx);
 	for (const entity of entities) {
-		if (isCallerOwnedEntity$3(entity)) {
+		if (isCallerOwnedEntity(entity)) {
 			filtered.push(entity);
 			continue;
 		}
-		const normalized = normalizeEntity(entity);
+		const normalized = normalizeEntity(entity, fullText);
 		if (!normalized) continue;
 		const trimmed = normalized.text;
-		if (isCallerOwnedEntity$3(normalized)) {
+		if (isCallerOwnedEntity(normalized)) {
 			filtered.push(normalized);
 			continue;
 		}
@@ -12120,6 +12230,32 @@ const loadCommonWords = () => {
 	})();
 	return commonWordsPromise;
 };
+let monthNamesPromise = null;
+let monthNamesCache = null;
+/**
+* English month names, lowercased. City gazetteers carry a
+* handful of entries that collide with month words (e.g. a
+* place named "August"); in prose these surface as address
+* false positives where the token is plainly a date, so they
+* are dropped from address patterns at build time.
+*/
+const loadMonthNames = () => {
+	if (monthNamesCache) return Promise.resolve(monthNamesCache);
+	if (monthNamesPromise) return monthNamesPromise;
+	monthNamesPromise = (async () => {
+		try {
+			const mod = await import("./date-months.mjs");
+			const set = new Set((mod.default?.en ?? []).map((month) => month.toLowerCase()));
+			monthNamesCache = set;
+			return set;
+		} catch {
+			const empty = /* @__PURE__ */ new Set();
+			monthNamesCache = empty;
+			return empty;
+		}
+	})();
+	return monthNamesPromise;
+};
 /**
 * Curated dictionary entries that are pure dotted
 * single-letter acronyms (e.g. `S.C.`, `D.N.J.`, `C.E.C.`)
@@ -12433,10 +12569,12 @@ const buildDenyList = async (config, ctx = defaultContext) => {
 		loadPersonStopwords(ctx),
 		loadAddressStopwords(ctx),
 		loadCommonWords(),
+		loadMonthNames(),
 		loadStreetTypeRe(),
 		loadGenericRoles(ctx)
 	]);
 	const commonWords = await loadCommonWords();
+	const monthNames = await loadMonthNames();
 	const dictionaries = config.dictionaries;
 	const hasDenyList = dictionaries?.denyList && dictionaries?.denyListMeta;
 	const hasCustomDenyList = config.customDenyList !== void 0 && config.customDenyList.length > 0;
@@ -12455,9 +12593,11 @@ const buildDenyList = async (config, ctx = defaultContext) => {
 		const normalized = source === "custom-deny-list" ? normalizeForSearch(entry) : stripCuratedPatternSyntax(normalizeForSearch(entry));
 		if (normalized.length === 0) return;
 		const lower = normalized.toLowerCase();
-		if (source !== "custom-deny-list" && label !== "address") {
-			if (SINGLE_WORD_RE.test(normalized) && commonWords.has(lower)) return;
-			if (isShortCuratedNoiseAcronym(normalized)) return;
+		if (source !== "custom-deny-list") {
+			if (label !== "address") {
+				if (SINGLE_WORD_RE.test(normalized) && commonWords.has(lower)) return;
+				if (isShortCuratedNoiseAcronym(normalized)) return;
+			} else if (monthNames.has(lower)) return;
 		}
 		const existing = patternIndex.get(lower);
 		if (existing !== void 0) {
@@ -13336,7 +13476,6 @@ const TRAILING_SEP = /[,\s]+$/;
 const WORD_CHAR_RE = /[\p{L}\p{N}]/u;
 const ORG_PROPAGATION_SCORE = .9;
 const ORG_DETERMINER_RE = /(?<![\p{L}\p{N}])(společnost(?:i|í|em|u)?|spolecnost(?:i|em|u)?|the\s+(?:company|corporation|firm)|die\s+(?:gesellschaft|firma)|la\s+(?:société|empresa|sociedad)|el\s+(?:empresa|sociedad))\s+$/iu;
-const isCallerOwnedEntity$2 = (entity) => entity.sourceDetail === "custom-deny-list" || entity.sourceDetail === "custom-regex";
 /**
 * After the main detection pass, collect organization
 * entities with a legal form suffix, strip the suffix
@@ -13354,7 +13493,7 @@ const propagateOrgNames = (entities, fullText) => {
 	const seedByBase = /* @__PURE__ */ new Map();
 	for (const e of entities) {
 		if (e.label !== "organization") continue;
-		if (isCallerOwnedEntity$2(e)) continue;
+		if (isCallerOwnedEntity(e)) continue;
 		for (const suffix of LEGAL_SUFFIXES) if (e.text.endsWith(suffix)) {
 			const base = e.text.slice(0, -suffix.length).replace(TRAILING_SEP, "").trim();
 			if (base.length >= 3) {
@@ -13724,7 +13863,6 @@ const NEAR_MISS_BAND = .15;
 const BOOST_PER_NEIGHBOUR = .05;
 const CONTEXT_WINDOW_CHARS = 150;
 const HIGH_CONFIDENCE_FLOOR = .9;
-const isCallerOwnedEntity$1 = (entity) => entity.sourceDetail === "custom-deny-list" || entity.sourceDetail === "custom-regex";
 /**
 * Boost confidence of near-miss NER entities that appear
 * near high-confidence detections (regex, trigger phrase).
@@ -13756,7 +13894,7 @@ const boostNearMissEntities = (entities, threshold) => {
 		const boostedScore = entity.score + neighbourCount * BOOST_PER_NEIGHBOUR;
 		if (boostedScore >= threshold) boosted.push({
 			...entity,
-			score: boostedScore
+			score: Math.min(1, boostedScore)
 		});
 	}
 	return boosted;
@@ -13835,7 +13973,7 @@ const getStreetAbbrevs = () => _streetAbbrevs ?? /* @__PURE__ */ new Set();
 */
 const detectStreetPatternsNearAddresses = (fullText, existingEntities) => {
 	const results = [];
-	const addressEntities = existingEntities.filter((entity) => !(entity.label === "address" && isCallerOwnedEntity$1(entity))).filter((e) => e.label === "address");
+	const addressEntities = existingEntities.filter((entity) => !(entity.label === "address" && isCallerOwnedEntity(entity))).filter((e) => e.label === "address");
 	const headerEnd = Math.floor(fullText.length * HEADER_ZONE_FRACTION);
 	const houseNumRe = /\b(?:\d{1,4}\/\d+[a-zA-Z]\b|\d{3,4}\/\d+\b|(?:1[3-9]|[2-9]\d)\/\d{3,}\b)/g;
 	houseNumRe.lastIndex = 0;
@@ -13934,7 +14072,7 @@ const ORPHAN_STREET_RE = /^[^\S\n]*(\p{Lu}[\p{Ll}\p{Lu}]+(?:[^\S\n]+[\p{Lu}\p{Ll
 */
 const detectOrphanStreetLines = (fullText, existingEntities) => {
 	const headerEnd = Math.floor(fullText.length * HEADER_ZONE_FRACTION);
-	const contextEntities = existingEntities.filter((entity) => !(entity.label === "address" && isCallerOwnedEntity$1(entity)));
+	const contextEntities = existingEntities.filter((entity) => !(entity.label === "address" && isCallerOwnedEntity(entity)));
 	const results = [];
 	ORPHAN_STREET_RE.lastIndex = 0;
 	for (let m = ORPHAN_STREET_RE.exec(fullText); m !== null; m = ORPHAN_STREET_RE.exec(fullText)) {
@@ -14292,7 +14430,7 @@ const applyHotwordRules = (entities, fullText) => {
 //#region src/filters/boundary-consistency.ts
 /** Max gap (in chars) between entities to merge. */
 const MAX_GAP = 3;
-const hasLockedBoundary$1 = (entity) => entity.sourceDetail === "custom-deny-list" || entity.sourceDetail === "custom-regex";
+const hasLockedBoundary$1 = (entity) => isCallerOwnedEntity(entity);
 const hasDetectorLockedBoundary = (entity) => entity.label === "phone number" && entity.source === DETECTION_SOURCES.TRIGGER;
 /**
 * Characters allowed in the gap between two adjacent
@@ -14875,7 +15013,6 @@ const unmaskNerEntities = (nerEntities, maskResult, fullText) => {
 * so the containment rule trusts their length.
 */
 const LITERAL_SOURCES = new Set(["deny-list", "gazetteer"]);
-const isCallerOwnedEntity = (entity) => entity.sourceDetail === "custom-deny-list" || entity.sourceDetail === "custom-regex";
 const hasLockedBoundary = (entity) => isCallerOwnedEntity(entity);
 const isCoveredByDenyListEntity = (entity, denyListEntity) => entity.label === denyListEntity.label && denyListEntity.start <= entity.start && denyListEntity.end >= entity.end;
 const filterNameCorpusCoveredByDenyList = (nameCorpusEntities, denyListEntities) => nameCorpusEntities.filter((entity) => !denyListEntities.some((denyListEntity) => isCoveredByDenyListEntity(entity, denyListEntity)));
@@ -15099,7 +15236,7 @@ const sanitizeEntities = (entities) => entities.flatMap((e) => {
 	return [{
 		...e,
 		start: e.start + lead,
-		end: e.start + lead + collapsed.length,
+		end: e.start + lead + cleaned.length,
 		text: collapsed
 	}];
 });
@@ -15215,7 +15352,7 @@ const filterAllowedLabels = (entities, allowedLabels) => {
 	return entities.filter((e) => allowedLabels.has(e.label));
 };
 const labelIsAllowed = (label, allowedLabels) => !allowedLabels || allowedLabels.has(label);
-const NON_NER_LABELS = new Set(["misc"]);
+const NON_NER_LABELS = new Set(["crypto", "misc"]);
 const getRequestedNerLabels = (config, expandForHotwords = false) => {
 	const labels = config.labels.length > 0 ? config.labels : DEFAULT_ENTITY_LABELS;
 	return (expandForHotwords ? expandLabelsForHotwordRules(labels) : labels).filter((label) => !NON_NER_LABELS.has(label));
@@ -15545,7 +15682,23 @@ const resolveOperator = (config, label) => config.operators[label] ?? "replace";
 //#region src/redact.ts
 const WHITESPACE_RE = /\s+/g;
 const PHONE_NOISE_RE = /[()\s-]/g;
+const ETHEREUM_ADDRESS_RE = /0x[0-9A-Fa-f]{40}/;
+const BECH32_ADDRESS_RE = /\bbc1[ac-hj-np-z02-9]{11,71}\b/i;
+const BASE58_ADDRESS_RE = /\b[13][a-km-zA-HJ-NP-Z1-9]{25,34}\b/;
+const NHS_NUMBER_CUE_RE = /\b(?:NHS|National\s+Health\s+Service)\b/i;
+const PASSPORT_IDENTIFIER_RE = /\b(?:[A-Za-z]{1,2}\d{6,8}|\d{2}[A-Za-z]{2}\d{5}|\d{7,9})\b/;
 const ID_SEPARATOR_RE = /[\s\-/.]/g;
+const normalizeCryptoText = (text) => {
+	const trimmed = text.trim();
+	const ethereumAddress = ETHEREUM_ADDRESS_RE.exec(trimmed)?.[0];
+	if (ethereumAddress) return ethereumAddress.toLowerCase();
+	const bech32Address = BECH32_ADDRESS_RE.exec(trimmed)?.[0];
+	if (bech32Address) return bech32Address.toLowerCase();
+	return BASE58_ADDRESS_RE.exec(trimmed)?.[0] ?? trimmed;
+};
+const normalizePassportText = (text) => {
+	return (PASSPORT_IDENTIFIER_RE.exec(text)?.[0] ?? text).replace(ID_SEPARATOR_RE, "").toUpperCase();
+};
 /**
 * Normalize entity text so that surface-form variations
 * of the same real-world value map to a single canonical
@@ -15555,7 +15708,10 @@ const normalizeEntityText = (label, text) => {
 	const upper = label.toUpperCase().replace(WHITESPACE_RE, "_");
 	if (upper === "EMAIL_ADDRESS" || upper === "EMAIL") return text.toLowerCase().trim();
 	if (upper === "PHONE_NUMBER" || upper === "PHONE") return text.replace(PHONE_NOISE_RE, "");
-	if (upper === "IBAN" || upper === "BANK_ACCOUNT_NUMBER" || upper === "TAX_IDENTIFICATION_NUMBER" || upper === "REGISTRATION_NUMBER" || upper === "NATIONAL_IDENTIFICATION_NUMBER" || upper === "SOCIAL_SECURITY_NUMBER" || upper === "BIRTH_NUMBER" || upper === "IDENTITY_CARD_NUMBER" || upper === "PASSPORT_NUMBER" || upper === "CREDIT_CARD_NUMBER") return text.replace(ID_SEPARATOR_RE, "").toUpperCase();
+	if (upper === "CRYPTO") return normalizeCryptoText(text);
+	if (upper === "NATIONAL_IDENTIFICATION_NUMBER" && NHS_NUMBER_CUE_RE.test(text)) return text.replace(/\D/g, "");
+	if (upper === "IBAN" || upper === "BANK_ACCOUNT_NUMBER" || upper === "TAX_IDENTIFICATION_NUMBER" || upper === "REGISTRATION_NUMBER" || upper === "NATIONAL_IDENTIFICATION_NUMBER" || upper === "SOCIAL_SECURITY_NUMBER" || upper === "BIRTH_NUMBER" || upper === "IDENTITY_CARD_NUMBER" || upper === "CREDIT_CARD_NUMBER") return text.replace(ID_SEPARATOR_RE, "").toUpperCase();
+	if (upper === "PASSPORT_NUMBER") return normalizePassportText(text);
 	if (upper === "PERSON" || upper === "ORGANIZATION" || upper === "ADDRESS" || upper === "LAND_PARCEL" || upper === "MISC") return text.replace(WHITESPACE_RE, " ").toLowerCase().trim();
 	return text.trim();
 };
@@ -15688,7 +15844,7 @@ const hasOverlapping = (a, b, multiLabel) => {
 	const b0 = b[0] ?? 0;
 	const b1 = b[1] ?? 0;
 	if (a0 === b0 && a1 === b1) return !multiLabel;
-	return !(a0 > b1 || b0 > a1);
+	return a0 < b1 && b0 < a1;
 };
 /**
 * Greedy non-overlapping span selection.
@@ -15900,7 +16056,12 @@ const encodeInputs = (tokenizer, texts, promptLengths) => {
 		allWordsMasks
 	];
 };
-/** Build span index pairs and masks for the model. */
+/**
+* Build span index pairs and masks for the model.
+*
+* Exported for unit testing the span-mask invariant; not part
+* of the public package barrel.
+*/
 const prepareSpans = (batchTokens, maxWidth) => {
 	const spanIdxs = [];
 	const spanMasks = [];
@@ -15911,7 +16072,7 @@ const prepareSpans = (batchTokens, maxWidth) => {
 		for (let i = 0; i < len; i++) for (let j = 0; j < maxWidth; j++) {
 			const endIdx = Math.min(i + j, len - 1);
 			idx.push([i, endIdx]);
-			mask.push(endIdx < len);
+			mask.push(i + j < len);
 		}
 		spanIdxs.push(idx);
 		spanMasks.push(mask);
@@ -15977,13 +16138,20 @@ const MAX_CHUNK_CHARS = 1500;
 const OVERLAP_CHARS = 50;
 const MIN_CHUNK_LENGTH = 10;
 /**
-* Split text into overlapping chunks for GLiNER's
-* ~512 token context window. Character-based splitting
-* (rough approximation of token limits).
+* Split text into overlapping chunks, each paired with its
+* exact start offset in the source text.
+*
+* Carrying the offset out of the splitter is the robust way to
+* map chunk-local entity offsets back to document offsets:
+* downstream code never has to re-locate a chunk by content
+* search (which mis-locates when boilerplate repeats; see
+* computeChunkOffsets).
 *
-* Tries to break at sentence boundaries when possible.
+* Character-based splitting (rough token approximation for
+* GLiNER's ~512 token window); breaks at sentence boundaries
+* when possible.
 */
-const chunkText = (text) => {
+const chunkTextWithOffsets = (text) => {
 	const chunks = [];
 	let offset = 0;
 	while (offset < text.length) {
@@ -15993,14 +16161,34 @@ const chunkText = (text) => {
 			if (lastPeriod > MAX_CHUNK_CHARS * .5) end = offset + lastPeriod + 2;
 		}
 		const chunk = text.slice(offset, end);
-		if (chunk.trim().length > MIN_CHUNK_LENGTH) chunks.push(chunk);
+		if (chunk.trim().length > MIN_CHUNK_LENGTH) chunks.push({
+			text: chunk,
+			offset
+		});
+		if (end === text.length) break;
 		offset = Math.max(offset + 1, end - OVERLAP_CHARS);
 	}
 	return chunks;
 };
 /**
-* Compute the byte offset of each chunk within the
-* original document text.
+* Split text into overlapping chunks for GLiNER's ~512 token
+* context window. Character-based splitting (rough token
+* approximation); breaks at sentence boundaries when possible.
+*
+* Prefer chunkTextWithOffsets when you also need each chunk's
+* document offset.
+*/
+const chunkText = (text) => chunkTextWithOffsets(text).map((chunk) => chunk.text);
+/**
+* Compute the start offset of each chunk within the original
+* document text by content search.
+*
+* @deprecated Re-locates each chunk with `indexOf`, which can
+*   match the wrong position when identical content repeats in
+*   the document (common in boilerplate-heavy legal text) and
+*   then desyncs every subsequent offset. Use
+*   `chunkTextWithOffsets`, which carries exact offsets out of
+*   the splitter.
 */
 const computeChunkOffsets = (fullText, chunks) => {
 	const offsets = [];
@@ -16101,6 +16289,6 @@ const levenshtein = (rawA, rawB) => {
 //#region src/wasm.ts
 initTextSearch(TextSearch);
 //#endregion
-export { CURRENCY_PATTERN_META, DATE_PATTERN_META, DEFAULT_ENTITY_LABELS, DEFAULT_OPERATOR_CONFIG, DETECTION_SOURCES, DETECTOR_PRIORITY, OPERATOR_REGISTRY, OPERATOR_TYPES, REGEX_META, REGEX_PATTERNS, REGIONS, ZONE_SCORE_ADJUSTMENTS, applyHotwordRules, applyZoneAdjustments, boostNearMissEntities, buildDenyList, buildGazetteerPatterns, buildLegalFormPatterns, buildPlaceholderMap, buildStreetTypePatterns, buildTriggerPatterns, buildUnifiedSearch, chunkText, classifyZones, computeChunkOffsets, corefKey, createPipelineContext, deanonymise, decodeSpans, decodeTokenSpans, detectNameCorpus, ensureDenyListData, exportRedactionKey, extractDefinedTerms, filterFalsePositives, findCoreferenceSpans, getCurrencyPatterns, getDatePatterns, getNameCorpusNonWesternNames, initAddressComponents, initHotwordRules, initNameCorpus, initZoneClassifier, levenshtein, mergeAndDedup, mergeChunkEntities, normalizeForSearch, prepareBatch, preparePipelineSearch, processAddressSeeds, processDenyListMatches, processGazetteerMatches, processLegalFormMatches, processRegexMatches, processTriggerMatches, propagateOrgNames, redactText, resolveCountries, resolveOperator, runPipeline, runUnifiedSearch, sanitizeEntities, tokenizeText, warmLegalRoleHeads };
+export { CURRENCY_PATTERN_META, DATE_PATTERN_META, DEFAULT_ENTITY_LABELS, DEFAULT_OPERATOR_CONFIG, DETECTION_SOURCES, DETECTOR_PRIORITY, OPERATOR_REGISTRY, OPERATOR_TYPES, REGEX_META, REGEX_PATTERNS, REGIONS, ZONE_SCORE_ADJUSTMENTS, applyHotwordRules, applyZoneAdjustments, boostNearMissEntities, buildDenyList, buildGazetteerPatterns, buildLegalFormPatterns, buildPlaceholderMap, buildStreetTypePatterns, buildTriggerPatterns, buildUnifiedSearch, chunkText, chunkTextWithOffsets, classifyZones, computeChunkOffsets, corefKey, createPipelineContext, deanonymise, decodeSpans, decodeTokenSpans, detectNameCorpus, ensureDenyListData, exportRedactionKey, extractDefinedTerms, filterFalsePositives, findCoreferenceSpans, getCurrencyPatterns, getDatePatterns, getNameCorpusNonWesternNames, initAddressComponents, initHotwordRules, initNameCorpus, initZoneClassifier, levenshtein, mergeAndDedup, mergeChunkEntities, normalizeForSearch, prepareBatch, preparePipelineSearch, processAddressSeeds, processDenyListMatches, processGazetteerMatches, processLegalFormMatches, processRegexMatches, processTriggerMatches, propagateOrgNames, redactText, resolveCountries, resolveOperator, runPipeline, runUnifiedSearch, sanitizeEntities, tokenizeText, warmLegalRoleHeads };
 
 //# sourceMappingURL=wasm.mjs.map