@ai-sdlc/orchestrator 0.10.0 → 0.14.0

package/dist/signal-ingestion/classifier.js

@@ -0,0 +1,494 @@
+/**
+ * RFC-0030 Phase 2 — Signal classifier.
+ *
+ * Classifies raw signals on three deterministic axes and applies the
+ * English-only language gate (OQ-13.2 resolution):
+ *
+ *   1. **Tier** — enterprise / mid / smb / free / churned (metadata-driven).
+ *      Inference order: adapter-provided `customerTier` → customerId lookup →
+ *      adapter's `defaultTier` (mapped to smb for Tier 1, free for Tier 2).
+ *
+ *   2. **ICP resonance** — strong / partial / weak (BM25 default; embedding
+ *      when RFC-0019 adapter is configured — currently always BM25 in v1).
+ *      For v1 the classifier receives an explicit `icpSegments` list from the
+ *      caller. BM25 matching measures how closely the signal payload matches
+ *      the declared ICP segments.
+ *
+ *   3. **Recency decay** — exponential decay: exp(-age_days × ln(2) / half_life).
+ *      Applied at scoring time so the pipeline doesn't need to re-compute as
+ *      time passes (age is passed as input). Half-life is read from config
+ *      (default 30 days).
+ *
+ * **Language gate** (OQ-13.2 v0.3 re-walkthrough): non-accepted-language signal
+ * payloads are detected via the `franc` library (deterministic; <10ms per
+ * signal; JS-native; MIT-licensed; 95%+ accuracy on text >50 chars; returns
+ * ISO 639-3 codes). Signals in unsupported languages are dropped and logged as
+ * `Decision: signal-language-unsupported`. Per-org `acceptedLanguages` config
+ * is respected (default `['en']`); multi-language opt-in (e.g. `['en', 'fr',
+ * 'es']`) accepts those languages knowingly — adopters take on documented BM25
+ * quality degradation (~15-30% precision drop without per-language stopwords)
+ * in exchange for non-English signal coverage.
+ *
+ * All multipliers and weights are read from `SignalIngestionConfig`; they are
+ * NOT hardcoded, satisfying AC #6.
+ *
+ * @module signal-ingestion/classifier
+ */
+import { franc, francAll } from 'franc';
+import { DEFAULT_SIGNAL_INGESTION_CONFIG } from './config.js';
+// ── Main entry point ────────────────────────────────────────────────────────
+/**
+ * Classify a batch of raw signals.
+ *
+ * Per RFC-0030 §6 inference order for tier:
+ *   1. `signal.customerTier` (adapter-provided structured field)
+ *   2. `tierRegistry.resolve(signal.customerId)` (registry lookup)
+ *   3. Adapter `defaultTier` (mapped: Tier 1 → smb, Tier 2 → free)
+ *
+ * Per RFC-0030 §13.2: non-accepted-language signals are dropped and
+ * returned in `languageDecisions`.
+ */
+export function classifySignals(signals, options = {}) {
+    const config = options.config ?? DEFAULT_SIGNAL_INGESTION_CONFIG;
+    const asOf = options.asOf ?? new Date();
+    const icpSegments = options.icpSegments ?? [];
+    const tierRegistry = options.tierRegistry;
+    const adapterTiers = options.adapterTiers ?? new Map();
+    const acceptedLanguages = config.acceptedLanguages;
+    const languageDetection = config.languageDetection;
+    const classified = [];
+    const languageDecisions = [];
+    for (const signal of signals) {
+        // Language gate — drop non-accepted-language signals
+        const langCheck = checkLanguage(signal.payload, acceptedLanguages, languageDetection);
+        if (!langCheck.accepted) {
+            languageDecisions.push({
+                type: 'Decision',
+                decision: 'signal-language-unsupported',
+                sourceId: signal.sourceId,
+                detectedLanguage: langCheck.detectedLanguage,
+                detectedScript: langCheck.detectedScript,
+                acceptedLanguages,
+                message: `Signal ${signal.sourceId} dropped: detected language '${langCheck.detectedLanguage}' (script '${langCheck.detectedScript}') not in accepted languages ${JSON.stringify(acceptedLanguages)}`,
+            });
+            continue;
+        }
+        const customerTier = resolveCustomerTier(signal, tierRegistry, adapterTiers);
+        const icpResonance = resolveIcpResonance(signal.payload, icpSegments);
+        const recencyDecay = computeRecencyDecay(signal.sourceTimestamp, asOf, config.recencyHalfLifeDays);
+        const tierMultiplier = config.tierMultipliers[customerTier];
+        const icpResonanceWeight = config.icpResonanceWeights[icpResonance];
+        // Resolve adapter tier for base weight (Phase 4 will refine with significance threshold)
+        const sourceAdapterTier = adapterTiers.get(signal.metadata?.['adapterName']) ?? 1;
+        const baseWeight = sourceAdapterTier === 1 ? 1.0 : 0.3;
+        classified.push({
+            signal,
+            customerTier,
+            icpResonance,
+            recencyDecay,
+            tierMultiplier,
+            icpResonanceWeight,
+            baseWeight,
+        });
+    }
+    return { classified, languageDecisions };
+}
+// ── Tier resolution ─────────────────────────────────────────────────────────
+/**
+ * Resolve customer tier per RFC-0030 §6.1 inference order:
+ *   1. `signal.customerTier`
+ *   2. `tierRegistry.resolve(signal.customerId)`
+ *   3. Adapter defaultTier → smb (Tier 1) or free (Tier 2)
+ */
+export function resolveCustomerTier(signal, tierRegistry, adapterTiers) {
+    // Step 1: adapter-provided structured tier
+    if (signal.customerTier !== undefined) {
+        return signal.customerTier;
+    }
+    // Step 2: customerId registry lookup
+    if (tierRegistry && signal.customerId !== undefined) {
+        const resolved = tierRegistry.resolve(signal.customerId);
+        if (resolved !== undefined)
+            return resolved;
+    }
+    // Step 3: adapter defaultTier mapping
+    const adapterName = signal.metadata?.['adapterName'];
+    const signalTier = adapterName ? (adapterTiers?.get(adapterName) ?? 1) : 1;
+    // Tier 1 adapters (support tickets, CRM) → smb as baseline customer tier
+    // Tier 2 adapters (community, competitive) → free as baseline
+    return signalTier === 1 ? 'smb' : 'free';
+}
+// ── ICP resonance (BM25) ────────────────────────────────────────────────────
+/**
+ * Classify ICP resonance for a signal payload against the declared ICP
+ * segment descriptors using BM25 similarity.
+ *
+ * When `icpSegments` is empty, returns `partial` (no basis for strong/weak).
+ *
+ * BM25 parameters (k1=1.5, b=0.75) are standard.
+ * Score mapping:
+ *   - ≥ 0.5  → strong
+ *   - ≥ 0.15 → partial
+ *   - < 0.15 → weak
+ */
+export function resolveIcpResonance(payload, icpSegments) {
+    if (icpSegments.length === 0)
+        return 'partial';
+    const score = bm25MaxScore(payload, icpSegments);
+    if (score >= 0.5)
+        return 'strong';
+    if (score >= 0.15)
+        return 'partial';
+    return 'weak';
+}
+/**
+ * Compute the BM25-based ICP match score between a signal payload and the
+ * ICP segment corpus.
+ *
+ * **Scoring direction**: ICP segments are the "documents"; the payload tokens
+ * are the "query". We score each ICP segment against the payload terms, then
+ * take the maximum across segments. This measures how closely the signal
+ * payload covers the ICP segment vocabulary — the deterministic-first
+ * approach per RFC-0029 Principle 2.
+ *
+ * **Normalisation**: we compute the self-score of each ICP segment (score
+ * when the document perfectly matches the query — i.e., we pretend the
+ * segment itself is the payload) and use the maximum self-score as the
+ * denominator. This produces a well-bounded [0, 1] score.
+ *
+ * Implementation: simplified BM25 without external deps so the orchestrator
+ * stays self-contained.
+ */
+function bm25MaxScore(payload, corpus) {
+    const K1 = 1.5;
+    const B = 0.75;
+    const payloadTokens = tokenize(payload);
+    if (payloadTokens.length === 0)
+        return 0;
+    // Build term frequency maps for both the corpus docs and the payload
+    const corpusTfs = corpus.map((doc) => termFrequency(tokenize(doc)));
+    const avgDocLen = corpus.reduce((sum, doc) => sum + tokenize(doc).length, 0) / corpus.length;
+    const N = corpus.length;
+    /**
+     * IDF of a term over the ICP corpus.
+     * Smoothed to avoid negatives: log((N - df + 0.5) / (df + 0.5) + 1)
+     */
+    const idf = (term) => {
+        const df = corpusTfs.filter((tf) => (tf.get(term) ?? 0) > 0).length;
+        return Math.log((N - df + 0.5) / (df + 0.5) + 1);
+    };
+    /**
+     * BM25 score for one ICP segment against a set of query terms.
+     * query terms come from the entity we're scoring (payload OR the segment
+     * itself for self-score normalisation).
+     */
+    const scoreSegment = (segTf, queryTerms) => {
+        const docLen = Array.from(segTf.values()).reduce((a, b) => a + b, 0);
+        let score = 0;
+        for (const term of new Set(queryTerms)) {
+            const freq = segTf.get(term) ?? 0;
+            if (freq === 0)
+                continue;
+            const termIdf = idf(term);
+            const numerator = freq * (K1 + 1);
+            const denominator = freq + K1 * (1 - B + B * (docLen / Math.max(avgDocLen, 1)));
+            score += termIdf * (numerator / denominator);
+        }
+        return score;
+    };
+    // Score each ICP segment using payload tokens as the query
+    let maxScore = 0;
+    let maxSelfScore = 0;
+    for (let i = 0; i < corpus.length; i++) {
+        const segTf = corpusTfs[i];
+        const segTokens = tokenize(corpus[i]);
+        // Self-score: how much would this segment score if queried by its own tokens?
+        const selfScore = scoreSegment(segTf, segTokens);
+        if (selfScore > maxSelfScore)
+            maxSelfScore = selfScore;
+        // Payload score: score this segment using the payload tokens
+        const payloadScore = scoreSegment(segTf, payloadTokens);
+        if (payloadScore > maxScore)
+            maxScore = payloadScore;
+    }
+    if (maxSelfScore <= 0)
+        return 0;
+    // Normalise by the maximum achievable score (self-match score)
+    return Math.min(1, maxScore / maxSelfScore);
+}
+/** Tokenize a string to lowercase alphanumeric tokens, minimum length 2. */
+export function tokenize(text) {
+    return text.toLowerCase().match(/[a-z0-9]{2,}/g) ?? [];
+}
+/** Count term frequencies in a token list. */
+function termFrequency(tokens) {
+    const tf = new Map();
+    for (const t of tokens)
+        tf.set(t, (tf.get(t) ?? 0) + 1);
+    return tf;
+}
+// ── Recency decay ───────────────────────────────────────────────────────────
+/**
+ * Compute exponential recency decay for a signal.
+ *
+ * `decay = exp(-age_days * ln(2) / halfLifeDays)`
+ *
+ * At age = halfLifeDays, decay = 0.5.
+ * At age = 0, decay = 1.0.
+ * Future-dated signals (age < 0) clamp to 1.0.
+ */
+export function computeRecencyDecay(sourceTimestamp, asOf, halfLifeDays) {
+    if (halfLifeDays <= 0)
+        return 1.0;
+    const ageDays = (asOf.getTime() - sourceTimestamp.getTime()) / (1000 * 60 * 60 * 24);
+    if (ageDays <= 0)
+        return 1.0; // future or now → no decay
+    return Math.exp((-ageDays * Math.LN2) / halfLifeDays);
+}
+/**
+ * Language gate using the `franc` trigram-based detector (RFC-0030 OQ-13.2
+ * v0.3 re-walkthrough resolution).
+ *
+ * `franc` returns ISO 639-3 three-letter codes (`'eng'`, `'cmn'`, `'fra'`,
+ * `'spa'`, etc.) or `'und'` for undetermined text (too short / no script
+ * match). It is deterministic, MIT-licensed, JS-native (no model download),
+ * runs in <10ms per signal, and is 95%+ accurate on text >50 characters.
+ *
+ * Behaviour:
+ *  - The language-tag matcher accepts BCP-47 / ISO 639-1 / ISO 639-3 forms in
+ *    `acceptedLanguages` (e.g. `'en'`, `'eng'`, `'en-US'` all match English).
+ *  - Empty payloads return `accepted: true` (nothing to gate on; downstream
+ *    classifier handles empty-payload signals).
+ *  - When `franc` returns `'und'` (text too short, no script matched, or
+ *    `library: 'none'`), behaviour is controlled by
+ *    `languageDetection.onUndetermined` — `'accept'` (default) is conservative
+ *    because the dominant short-payload case is legitimate (e.g. "Crash on
+ *    save", "BUG: 500"); `'drop'` is strict.
+ *  - When `library: 'none'`, the gate is disabled entirely — all signals
+ *    accepted regardless of `acceptedLanguages`. Useful for testing or for
+ *    adopters that pre-filter signals upstream.
+ *  - **Fuzzy acceptance via francAll top-N**: short technical English (under
+ *    ~50 chars; e.g. "auth login failure when SAML callback returns no nonce")
+ *    can misdetect as Nordic / Latin-script neighbours (nno, nld, dan) due to
+ *    uncommon trigrams. To avoid false-drops on legitimate English, we accept
+ *    when ANY accepted language appears in the top-N candidates with a relative
+ *    score ≥ `TOP_N_ACCEPT_THRESHOLD` of the top hit. This is the standard
+ *    franc-misclassification workaround documented in
+ *    https://github.com/wooorm/franc#data
+ */
+function checkLanguage(payload, acceptedLanguages, languageDetection) {
+    // Empty payload — nothing to gate on.
+    if (!payload || payload.length === 0) {
+        return { accepted: true, detectedLanguage: 'und', detectedScript: 'unknown' };
+    }
+    // `library: 'none'` disables the gate entirely.
+    if (languageDetection.library === 'none') {
+        return { accepted: true, detectedLanguage: 'und', detectedScript: 'unknown' };
+    }
+    // Top language for reporting purposes (Decision payload + script hint).
+    const detectedLanguage = franc(payload, { minLength: languageDetection.minDetectionLength });
+    const detectedScript = scriptHintFor(detectedLanguage);
+    // 'und' — text too short OR no script matched — handle per config policy.
+    if (detectedLanguage === 'und') {
+        if (languageDetection.onUndetermined === 'accept') {
+            return { accepted: true, detectedLanguage, detectedScript };
+        }
+        return { accepted: false, detectedLanguage, detectedScript };
+    }
+    // Top-of-list match — fast path.
+    if (languageMatchesAccepted(detectedLanguage, acceptedLanguages)) {
+        return { accepted: true, detectedLanguage, detectedScript };
+    }
+    // Fuzzy acceptance via script-prefiltered candidate list. `francAll` only
+    // returns languages whose SCRIPT matches the input — Chinese text returns
+    // only CJK languages; Cyrillic text returns only Cyrillic languages. So if
+    // any accepted language appears at all in the candidate list (above a low
+    // relative-score floor), the input shares a script family with an accepted
+    // language — which is the load-bearing signal we care about. Short technical
+    // English may misclassify the TOP slot as nno/dan/nld, but `eng` will still
+    // appear in the list with a high relative score because the script
+    // (Latin-script) matches. Genuinely-foreign signals (Chinese, Arabic, Hindi)
+    // exclude `eng` from the list entirely via the script prefilter.
+    const allCandidates = francAll(payload, {
+        minLength: languageDetection.minDetectionLength,
+    });
+    for (const [lang, score] of allCandidates) {
+        if (score < FUZZY_ACCEPT_SCORE_FLOOR)
+            break;
+        if (languageMatchesAccepted(lang, acceptedLanguages)) {
+            return { accepted: true, detectedLanguage: lang, detectedScript: scriptHintFor(lang) };
+        }
+    }
+    return { accepted: false, detectedLanguage, detectedScript };
+}
+/**
+ * Minimum relative score (normalised against the top hit's 1.0) for a
+ * candidate to be considered for fuzzy acceptance. 0.80 is permissive enough
+ * to catch English misclassified as Nordic in short technical bug reports
+ * (typical eng-score for sub-60-char English is 0.85-0.98) while still
+ * excluding genuinely-distant guesses (franc rarely puts distant-script
+ * languages above 0.5 because the script prefilter already excludes them).
+ *
+ * The load-bearing protection against false-accepts of foreign-script text is
+ * franc's built-in SCRIPT PREFILTER: Chinese text returns only CJK
+ * candidates, Arabic text returns only Arabic candidates, etc. The fuzzy
+ * threshold here only matters for Latin-script close-runners-up.
+ */
+const FUZZY_ACCEPT_SCORE_FLOOR = 0.8;
+/**
+ * Check whether a franc-detected ISO 639-3 code matches any of the
+ * adopter-configured accepted-language tags.
+ *
+ * Accepts the common forms in `acceptedLanguages`:
+ *  - ISO 639-3 three-letter codes: `'eng'`, `'fra'`, `'spa'`, `'cmn'`, …
+ *  - ISO 639-1 two-letter codes: `'en'`, `'fr'`, `'es'`, `'zh'`, …
+ *  - BCP-47 regional tags: `'en-US'`, `'fr-CA'`, `'zh-Hans'`, … (prefix matched
+ *    against the two-letter base code)
+ *
+ * The two-letter ↔ three-letter mapping covers the most common languages
+ * adopters opt into; uncommon languages can be specified directly in their
+ * ISO 639-3 form to bypass the mapping.
+ */
+function languageMatchesAccepted(detectedIso6393, acceptedLanguages) {
+    for (const raw of acceptedLanguages) {
+        const tag = raw.toLowerCase();
+        // Exact ISO 639-3 match (e.g. 'eng' === 'eng').
+        if (tag === detectedIso6393)
+            return true;
+        // Two-letter ISO 639-1 (or BCP-47 prefix) match via the mapping table.
+        const base = tag.split('-')[0]; // 'en' from 'en-US', 'en' from 'en'
+        const mapped = ISO_639_1_TO_639_3[base];
+        if (mapped && mapped === detectedIso6393)
+            return true;
+    }
+    return false;
+}
+/**
+ * Two-letter ISO 639-1 → three-letter ISO 639-3 mapping for the most common
+ * languages adopters opt into. Not exhaustive — uncommon languages should be
+ * configured directly in their ISO 639-3 form (e.g. `'tgl'` for Tagalog).
+ *
+ * Source: ISO 639 standard cross-reference. The list intentionally includes
+ * the languages most likely to appear in multi-language customer-support
+ * pipelines (top ~30 languages by speaker count plus EU + East-Asian
+ * majors).
+ */
+const ISO_639_1_TO_639_3 = {
+    en: 'eng',
+    fr: 'fra',
+    es: 'spa',
+    de: 'deu',
+    it: 'ita',
+    pt: 'por',
+    nl: 'nld',
+    ru: 'rus',
+    pl: 'pol',
+    uk: 'ukr',
+    cs: 'ces',
+    sk: 'slk',
+    hu: 'hun',
+    ro: 'ron',
+    bg: 'bul',
+    el: 'ell',
+    tr: 'tur',
+    sv: 'swe',
+    no: 'nob',
+    da: 'dan',
+    fi: 'fin',
+    is: 'isl',
+    ga: 'gle',
+    zh: 'cmn',
+    ja: 'jpn',
+    ko: 'kor',
+    vi: 'vie',
+    th: 'tha',
+    id: 'ind',
+    ms: 'zlm',
+    tl: 'tgl',
+    hi: 'hin',
+    bn: 'ben',
+    ta: 'tam',
+    te: 'tel',
+    ur: 'urd',
+    fa: 'pes',
+    ar: 'arb',
+    he: 'heb',
+    sw: 'swh',
+    am: 'amh',
+};
+/**
+ * Coarse script-family hint for a franc-detected ISO 639-3 code. Returned in
+ * `SignalLanguageUnsupportedDecision.detectedScript` for backwards compat with
+ * v0.2 dashboards that key on script family rather than language.
+ *
+ * This is a small lookup table — only covers the languages we expect to
+ * surface in customer-signal pipelines. Unknown codes get `'unknown'`.
+ */
+function scriptHintFor(iso6393) {
+    if (iso6393 === 'und')
+        return 'unknown';
+    if (CJK_LANGUAGES.has(iso6393))
+        return 'cjk';
+    if (CYRILLIC_LANGUAGES.has(iso6393))
+        return 'cyrillic';
+    if (ARABIC_LANGUAGES.has(iso6393))
+        return 'arabic';
+    if (iso6393 === 'heb' || iso6393 === 'ydd')
+        return 'hebrew';
+    if (DEVANAGARI_LANGUAGES.has(iso6393))
+        return 'devanagari';
+    if (iso6393 === 'tha')
+        return 'thai';
+    if (iso6393 === 'ell')
+        return 'greek';
+    if (iso6393 === 'hye')
+        return 'armenian';
+    if (iso6393 === 'kat')
+        return 'georgian';
+    // Default: assume Latin-script for everything else (most languages franc
+    // detects in our coverage are Latin-script — fr/es/de/pt/it/nl/etc.).
+    return 'latin';
+}
+const CJK_LANGUAGES = new Set(['cmn', 'jpn', 'kor', 'yue', 'wuu', 'nan', 'hak']);
+const CYRILLIC_LANGUAGES = new Set([
+    'rus',
+    'ukr',
+    'bel',
+    'bul',
+    'mkd',
+    'srp',
+    'tat',
+    'kaz',
+    'kir',
+    'tgk',
+    'mon',
+    'khk',
+]);
+const ARABIC_LANGUAGES = new Set(['arb', 'pes', 'urd', 'pbu', 'prs', 'ckb', 'uig']);
+const DEVANAGARI_LANGUAGES = new Set([
+    'hin',
+    'mar',
+    'nep',
+    'npi',
+    'mai',
+    'bho',
+    'awa',
+    'mag',
+    'san',
+]);
+// ── Weight computation (convenience) ─────────────────────────────────────────
+/**
+ * Compute the composite signal weight from a `ClassifiedSignal`.
+ *
+ * `weight = baseWeight × tierMultiplier × icpResonanceWeight × recencyDecay`
+ *
+ * Note: SA resonance filter (RFC-0030 §9) is applied at the cluster level in
+ * Phase 5, not at the signal level here.
+ */
+export function computeSignalWeight(classified) {
+    return (classified.baseWeight *
+        classified.tierMultiplier *
+        classified.icpResonanceWeight *
+        classified.recencyDecay);
+}
+//# sourceMappingURL=classifier.js.map

package/dist/signal-ingestion/clustering-types.d.ts

@@ -0,0 +1,36 @@
+/**
+ * Input shape consumed by `clusterSignals()` per RFC-0030 Phase 3.
+ *
+ * `ClusteredSignalInput` is a subset of `ClassifiedSignal` (Phase 2 output)
+ * carried through the clusterer plus optional adapter-tier metadata used
+ * during cluster aggregation. Kept in a separate file so consumers can
+ * import the input shape without pulling in the clusterer's full surface
+ * (algorithm types, options) — and so the classifier doesn't depend on
+ * clustering.ts (avoid circular imports).
+ *
+ * @module signal-ingestion/clustering-types
+ */
+import type { ICPResonance } from './classifier.js';
+import type { CustomerTier, RawSignal, SignalTier } from './types.js';
+/**
+ * Per-signal input to the clusterer.
+ *
+ * A subset of `ClassifiedSignal` (Phase 2) chosen to keep clustering
+ * decoupled from Phase 2's full output shape:
+ *   - `signal`, `customerTier`, `icpResonance`, `recencyDecay` are carried
+ *     through and surfaced on the resulting `DemandCluster.members` so
+ *     Phase 4/5 consumers retain access without re-running classification.
+ *   - `adapterTier` is an OPTIONAL hint that lets Phase 4 sum `tier1` /
+ *     `tier2` member counts without re-deriving from `signal.metadata`.
+ *     When absent, the clusterer falls back to `signal.metadata.adapterTier`
+ *     then to Tier 1 (matches classifier convention).
+ */
+export interface ClusteredSignalInput {
+    signal: RawSignal;
+    customerTier: CustomerTier;
+    icpResonance: ICPResonance;
+    recencyDecay: number;
+    /** Optional adapter tier hint (Tier 1 default when absent). */
+    adapterTier?: SignalTier;
+}
+//# sourceMappingURL=clustering-types.d.ts.map

package/dist/signal-ingestion/clustering-types.js

@@ -0,0 +1,14 @@
+/**
+ * Input shape consumed by `clusterSignals()` per RFC-0030 Phase 3.
+ *
+ * `ClusteredSignalInput` is a subset of `ClassifiedSignal` (Phase 2 output)
+ * carried through the clusterer plus optional adapter-tier metadata used
+ * during cluster aggregation. Kept in a separate file so consumers can
+ * import the input shape without pulling in the clusterer's full surface
+ * (algorithm types, options) — and so the classifier doesn't depend on
+ * clustering.ts (avoid circular imports).
+ *
+ * @module signal-ingestion/clustering-types
+ */
+export {};
+//# sourceMappingURL=clustering-types.js.map