@mailwoman/neural 2.2.0 → 4.0.0

package/out/postcode-repair.js

@@ -0,0 +1,171 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Postcode regex repair pass — v0.7 task #35 ("postcode regex pre-pass").
+ *
+ *   The 2026-05-29 postcode diagnostic showed the neural model fragments alphanumeric postcodes at
+ *   the SentencePiece layer (GB/CA/NL at 0%, US 80.5%, FR 70.1%). Three failure modes were visible
+ *   in the data:
+ *
+ *   1. Total miss — "London SW1A 1AA" → (no postcode label)
+ *   2. Truncation — "M5V 2T6" → "2T6"; "B12 8QX" → "B12"
+ *   3. Char-drift — "75008" → "5008"; "62701" → "2701" (and smear: "1200-030 Lisboa" → "200-030 Lis")
+ *
+ *   This pass runs AFTER the model's per-token BIO labels are decoded but BEFORE `buildAddressTree`.
+ *   It detects postcode-shaped substrings with per-country regexes and repairs the label sequence
+ *   so the postcode span matches the detected shape. The model is untouched — this is a
+ *   deterministic decoder-side correction, the "lowest risk" lever in the v0.7 plan (vs. #36's soft
+ *   FST shallow-fusion or #41's char-level encoder).
+ *
+ *   PRECISION GUARDS (so we never regress the countries already passing):
+ *
+ *   - Alphanumeric shapes (GB/CA/NL/DE-prefixed) are high-confidence "this IS a postcode" patterns →
+ *       eligible to ADD a span where the model emitted none, but only over non-structural labels
+ *       (never over house_number/street/etc.).
+ *   - Numeric shapes (\d{5}, ZIP+4, JP, PT, PL) are ambiguous (a bare 5-digit could be a house number)
+ *       → SNAP-only: they expand/clip an EXISTING postcode span, never create one from scratch.
+ *   - Smear cleanup is LOCAL: only postcode tokens immediately flanking a snapped span are cleared. We
+ *       never globally clear unmatched postcode tokens — that would regress shapes we don't
+ *       pattern-match (AU 4-digit, IN 6-digit, …).
+ */
+/**
+ * Per-country postcode shape patterns, ordered most-specific → least. Alphanumeric patterns require
+ * uppercase letters (postcodes are conventionally uppercase, and the eval data has them uppercase)
+ * — this keeps them from matching ordinary lowercase prose.
+ */
+export const POSTCODE_PATTERNS = [
+    // --- Alphanumeric (eligible to ADD) ---
+    // GB: outward + space + inward, e.g. SW1A 1AA, EH8 9YL, W1J 9PN, IP13 6SU, B12 8QX
+    { label: "GB", kind: "alnum", re: /\b[A-Z]{1,2}\d[A-Z\d]?\s+\d[A-Z]{2}\b/g },
+    // CA: A1A 1A1 (space optional), e.g. M5V 2T6, H2X 2T6, H3B 1A3
+    { label: "CA", kind: "alnum", re: /\b[A-Z]\d[A-Z]\s?\d[A-Z]\d\b/g },
+    // DE-prefixed: D-68161
+    { label: "DE", kind: "alnum", re: /\bD-\d{5}\b/g },
+    // NL: 1234 AB / 1234AB — space optional (glued is common). The US "2737 CA" (ZIP+4 tail +
+    // state) collision is resolved by longest-match-wins below, which lets the ZIP+4 claim it.
+    { label: "NL", kind: "alnum", re: /\b\d{4}\s?[A-Z]{2}\b/g },
+    // --- Numeric (SNAP-only) ---
+    { label: "ZIP4", kind: "numeric", re: /\b\d{5}-\d{4}\b/g }, // US ZIP+4
+    { label: "JP", kind: "numeric", re: /\b\d{3}-\d{4}\b/g }, // 100-0001
+    { label: "PT", kind: "numeric", re: /\b\d{4}-\d{3}\b/g }, // 3060-187
+    { label: "PL", kind: "numeric", re: /\b\d{2}-\d{3}\b/g }, // 47-400
+    { label: "NUM5", kind: "numeric", re: /\b\d{5}\b/g }, // US/FR/DE/ES 5-digit
+];
+/**
+ * Labels a postcode span is allowed to overwrite when the model emitted no postcode at all (ADD
+ * path). These are the geographic-container tags postcodes get confused with per the diagnostic
+ * ("often labeled as locality or O"). Structural tags (house_number, street*, unit, po_box, venue,
+ * …) are intentionally absent so we never clobber a confidently-labeled street/number with a false
+ * postcode.
+ */
+const ADD_OVER_TAGS = new Set(["locality", "dependent_locality", "region", "subregion", "country"]);
+const POSTCODE_B = "B-postcode";
+const POSTCODE_I = "I-postcode";
+const LOCALITY_B = "B-locality";
+const LOCALITY_I = "I-locality";
+const OUTSIDE = "O";
+function isPostcodeLabel(label) {
+    return label === "B-postcode" || label === "I-postcode";
+}
+/** Extract the bare tag from a BIO label ("B-locality" → "locality", "O" → null). */
+function tagOf(label) {
+    return label === "O" ? null : label.slice(2);
+}
+/** Collect non-overlapping postcode matches, preferring more-specific (earlier) patterns. */
+export function collectMatches(text) {
+    const candidates = [];
+    POSTCODE_PATTERNS.forEach((pat, priority) => {
+        pat.re.lastIndex = 0;
+        for (let m = pat.re.exec(text); m; m = pat.re.exec(text)) {
+            candidates.push({ start: m.index, end: m.index + m[0].length, kind: pat.kind, priority });
+        }
+    });
+    // Greedy longest-match-wins: accept by (length desc, then priority asc); reject anything
+    // overlapping an accepted match. Longest-first lets a US ZIP+4 ("94610-2737") claim its span
+    // before the shorter NL-shaped false positive in its tail ("2737 CA") can.
+    candidates.sort((a, b) => b.end - b.start - (a.end - a.start) || a.priority - b.priority);
+    const accepted = [];
+    for (const c of candidates) {
+        if (accepted.some((a) => c.start < a.end && a.start < c.end))
+            continue;
+        accepted.push(c);
+    }
+    return accepted;
+}
+/**
+ * Repair postcode label spans in a decoded token sequence using per-country regexes. Returns a NEW
+ * token array (inputs are not mutated) plus a change count.
+ */
+export function repairPostcodeLabels(text, input) {
+    const matches = collectMatches(text);
+    const tokens = input.map((t) => ({ ...t }));
+    if (matches.length === 0)
+        return { tokens, changed: 0 };
+    let changed = 0;
+    const setLabel = (i, label) => {
+        if (tokens[i].label !== label) {
+            tokens[i].label = label;
+            changed++;
+        }
+    };
+    for (const m of matches) {
+        // Tokens whose char span intersects the match.
+        const overlap = [];
+        for (let i = 0; i < tokens.length; i++) {
+            const t = tokens[i];
+            if (t.start < m.end && m.start < t.end)
+                overlap.push(i);
+        }
+        if (overlap.length === 0)
+            continue;
+        const hasPostcode = overlap.some((i) => isPostcodeLabel(tokens[i].label));
+        if (!hasPostcode) {
+            // ADD path — only for high-confidence alphanumeric shapes, only over safe labels.
+            if (m.kind !== "alnum")
+                continue;
+            const safe = overlap.every((i) => {
+                const tag = tagOf(tokens[i].label);
+                return tag === null || ADD_OVER_TAGS.has(tag);
+            });
+            if (!safe)
+                continue;
+        }
+        // SNAP/ADD: relabel the matched run as a single postcode span.
+        overlap.forEach((i, k) => setLabel(i, k === 0 ? POSTCODE_B : POSTCODE_I));
+        // Leading smear clip: postcode tokens immediately BEFORE the snapped run are noise (e.g. a
+        // house-number digit the model over-labeled) — clear to O as before.
+        for (let j = overlap[0] - 1; j >= 0 && isPostcodeLabel(tokens[j].label); j--)
+            setLabel(j, OUTSIDE);
+        // Trailing smear: the model over-extended the postcode to the RIGHT. In postcode-before-city
+        // locales (DE/FR/ES/IT, "08523 Plauen") this swallows the leading characters of the city, which
+        // the historical clip-to-O then DISCARDED ("08523 Pl|auen Vogtl" → postcode "08523" + O +
+        // locality "auen Vogtl", dropping the "Pl"). When the smear connects to a following locality run,
+        // hand those characters BACK to the city — reassign them to locality and demote the city's
+        // leading B so the prefix + city form ONE span ("Pl"+"auen"+"Vogtl" → "Plauen Vogtl"). A
+        // standalone neighbour with no following locality (a country, "Paris 75008 France") keeps the
+        // historical clip-to-O. This is the decoder-side repair for the cross-tag postcode→city
+        // absorption diagnosed in the PR3 Pilot A postmortem (+36pp DE exact-locality, no-op on US,
+        // where the postcode sits at the end with nothing to trim).
+        const trailing = [];
+        for (let j = overlap[overlap.length - 1] + 1; j < tokens.length && isPostcodeLabel(tokens[j].label); j++) {
+            trailing.push(j);
+        }
+        if (trailing.length > 0) {
+            const after = trailing[trailing.length - 1] + 1;
+            const connectsToCity = after < tokens.length && tagOf(tokens[after].label) === "locality";
+            if (connectsToCity) {
+                trailing.forEach((j, k) => setLabel(j, k === 0 ? LOCALITY_B : LOCALITY_I));
+                if (tokens[after].label === "B-locality")
+                    setLabel(after, LOCALITY_I);
+            }
+            else {
+                for (const j of trailing)
+                    setLabel(j, OUTSIDE);
+            }
+        }
+    }
+    return { tokens, changed };
+}
+//# sourceMappingURL=postcode-repair.js.map

package/out/postcode-repair.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"postcode-repair.js","sourceRoot":"","sources":["../postcode-repair.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAcH;;;;GAIG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAoE;IACjG,yCAAyC;IACzC,mFAAmF;IACnF,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,EAAE,EAAE,wCAAwC,EAAE;IAC5E,+DAA+D;IAC/D,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,EAAE,EAAE,+BAA+B,EAAE;IACnE,uBAAuB;IACvB,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,EAAE,EAAE,cAAc,EAAE;IAClD,0FAA0F;IAC1F,2FAA2F;IAC3F,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,EAAE,EAAE,uBAAuB,EAAE;IAC3D,8BAA8B;IAC9B,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,EAAE,kBAAkB,EAAE,EAAE,WAAW;IACvE,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,EAAE,kBAAkB,EAAE,EAAE,WAAW;IACrE,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,EAAE,kBAAkB,EAAE,EAAE,WAAW;IACrE,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,EAAE,kBAAkB,EAAE,EAAE,SAAS;IACnE,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,EAAE,YAAY,EAAE,EAAE,sBAAsB;CAC5E,CAAA;AAED;;;;;;GAMG;AACH,MAAM,aAAa,GAAG,IAAI,GAAG,CAAS,CAAC,UAAU,EAAE,oBAAoB,EAAE,QAAQ,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC,CAAA;AAE3G,MAAM,UAAU,GAAG,YAAqC,CAAA;AACxD,MAAM,UAAU,GAAG,YAAqC,CAAA;AACxD,MAAM,UAAU,GAAG,YAAqC,CAAA;AACxD,MAAM,UAAU,GAAG,YAAqC,CAAA;AACxD,MAAM,OAAO,GAAG,GAA4B,CAAA;AAE5C,SAAS,eAAe,CAAC,KAAa;IACrC,OAAO,KAAK,KAAK,YAAY,IAAI,KAAK,KAAK,YAAY,CAAA;AACxD,CAAC;AAED,qFAAqF;AACrF,SAAS,KAAK,CAAC,KAAa;IAC3B,OAAO,KAAK,KAAK,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAA;AAC7C,CAAC;AAED,6FAA6F;AAC7F,MAAM,UAAU,cAAc,CAAC,IAAY;IAC1C,MAAM,UAAU,GAAoB,EAAE,CAAA;IACtC,iBAAiB,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,QAAQ,EAAE,EAAE;QAC3C,GAAG,CAAC,EAAE,CAAC,SAAS,GAAG,CAAC,CAAA;QACpB,KAAK,IAAI,CAAC,GAAG,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,CAAC,GAAG,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;YAC1D,UAAU,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,KAAK,EAAE,GAAG,EAAE,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,EAAE,IAAI,EAAE,GAAG,CAAC,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAA;QAC1F,CAAC;IACF,CAAC,CAAC,CAAA;IACF,yFAAyF;IACzF,6FAA6F;IAC7F,2EAA2E;IAC3E,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,QAAQ,GAAG,CAAC,CAAC,QAAQ,CAAC,CAAA;IACzF,MAAM,QAAQ,GAAoB,EAAE,CAAA;IACpC,KAAK,MAAM,CAAC,IAAI,UAAU,EAAE,CAAC;QAC5B,IAAI,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,CAAC;YAAE,SAAQ;QACtE,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;IACjB,CAAC;IACD,OAAO,QAAQ,CAAA;AAChB,CAAC;AAQD;;;GAGG;AACH,MAAM,UAAU,oBAAoB,CAAC,IAAY,EAAE,KAA8B;IAChF,MAAM,OAAO,GAAG,cAAc,CAAC,IAAI,CAAC,CAAA;IACpC,MAAM,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC,CAAA;IAC3C,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC,EAAE,CAAA;IAEvD,IAAI,OAAO,GAAG,CAAC,CAAA;IACf,MAAM,QAAQ,GAAG,CAAC,CAAS,EAAE,KAA4B,EAAQ,EAAE;QAClE,IAAI,MAAM,CAAC,CAAC,CAAE,CAAC,KAAK,KAAK,KAAK,EAAE,CAAC;YAChC,MAAM,CAAC,CAAC,CAAE,CAAC,KAAK,GAAG,KAAK,CAAA;YACxB,OAAO,EAAE,CAAA;QACV,CAAC;IACF,CAAC,CAAA;IAED,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACzB,+CAA+C;QAC/C,MAAM,OAAO,GAAa,EAAE,CAAA;QAC5B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACxC,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAE,CAAA;YACpB,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG;gBAAE,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;QACxD,CAAC;QACD,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,SAAQ;QAElC,MAAM,WAAW,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,eAAe,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,KAAK,CAAC,CAAC,CAAA;QAC1E,IAAI,CAAC,WAAW,EAAE,CAAC;YAClB,kFAAkF;YAClF,IAAI,CAAC,CAAC,IAAI,KAAK,OAAO;gBAAE,SAAQ;YAChC,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE;gBAChC,MAAM,GAAG,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,KAAK,CAAC,CAAA;gBACnC,OAAO,GAAG,KAAK,IAAI,IAAI,aAAa,CAAC,GAAG,CAAC,GAAG,CAAC,CAAA;YAC9C,CAAC,CAAC,CAAA;YACF,IAAI,CAAC,IAAI;gBAAE,SAAQ;QACpB,CAAC;QAED,+DAA+D;QAC/D,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAA;QAEzE,2FAA2F;QAC3F,qEAAqE;QACrE,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,CAAC,CAAE,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,eAAe,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,KAAK,CAAC,EAAE,CAAC,EAAE;YAAE,QAAQ,CAAC,CAAC,EAAE,OAAO,CAAC,CAAA;QAEpG,6FAA6F;QAC7F,gGAAgG;QAChG,0FAA0F;QAC1F,kGAAkG;QAClG,2FAA2F;QAC3F,yFAAyF;QACzF,8FAA8F;QAC9F,wFAAwF;QACxF,4FAA4F;QAC5F,4DAA4D;QAC5D,MAAM,QAAQ,GAAa,EAAE,CAAA;QAC7B,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAE,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,IAAI,eAAe,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,KAAK,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5G,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;QACjB,CAAC;QACD,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACzB,MAAM,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAE,GAAG,CAAC,CAAA;YAChD,MAAM,cAAc,GAAG,KAAK,GAAG,MAAM,CAAC,MAAM,IAAI,KAAK,CAAC,MAAM,CAAC,KAAK,CAAE,CAAC,KAAK,CAAC,KAAK,UAAU,CAAA;YAC1F,IAAI,cAAc,EAAE,CAAC;gBACpB,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAA;gBAC1E,IAAI,MAAM,CAAC,KAAK,CAAE,CAAC,KAAK,KAAK,YAAY;oBAAE,QAAQ,CAAC,KAAK,EAAE,UAAU,CAAC,CAAA;YACvE,CAAC;iBAAM,CAAC;gBACP,KAAK,MAAM,CAAC,IAAI,QAAQ;oBAAE,QAAQ,CAAC,CAAC,EAAE,OAAO,CAAC,CAAA;YAC/C,CAAC;QACF,CAAC;IACF,CAAC;IAED,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,CAAA;AAC3B,CAAC"}

package/out/proposal-classifier.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"proposal-classifier.d.ts","sourceRoot":"","sources":["../proposal-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAIH,OAAO,KAAK,EAGX,YAAY,EACZ,kBAAkB,EAElB,MAAM,uBAAuB,CAAA;AAC9B,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,iBAAiB,CAAA;AAG9D,MAAM,WAAW,8BAA8B;IAC9C,wFAAwF;IACxF,EAAE,EAAE,MAAM,CAAA;IACV,iDAAiD;IACjD,UAAU,EAAE,uBAAuB,CAAA;IACnC;;;;OAIG;IACH,KAAK,CAAC,EAAE,SAAS,YAAY,EAAE,CAAA;IAC/B,mFAAmF;IACnF,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,GAAG,GAAG,CAAC,EAAE,CAAA;IACnC,+DAA+D;IAC/D,OAAO,CAAC,EAAE,MAAM,CAAA;CAChB;AAED,0EAA0E;AAC1E,wBAAgB,8BAA8B,CAAC,GAAG,EAAE,8BAA8B,GAAG,kBAAkB,CA6CtG"}
+{"version":3,"file":"proposal-classifier.d.ts","sourceRoot":"","sources":["../proposal-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAIH,OAAO,KAAK,EAGX,YAAY,EACZ,kBAAkB,EAElB,MAAM,uBAAuB,CAAA;AAC9B,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,iBAAiB,CAAA;AAG9D,MAAM,WAAW,8BAA8B;IAC9C,wFAAwF;IACxF,EAAE,EAAE,MAAM,CAAA;IACV,iDAAiD;IACjD,UAAU,EAAE,uBAAuB,CAAA;IACnC;;;;OAIG;IACH,KAAK,CAAC,EAAE,SAAS,YAAY,EAAE,CAAA;IAC/B,mFAAmF;IACnF,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,GAAG,GAAG,CAAC,EAAE,CAAA;IACnC,+DAA+D;IAC/D,OAAO,CAAC,EAAE,MAAM,CAAA;CAChB;AAED,0EAA0E;AAC1E,wBAAgB,8BAA8B,CAAC,GAAG,EAAE,8BAA8B,GAAG,kBAAkB,CA+CtG"}

package/out/proposal-classifier.js

@@ -24,7 +24,9 @@ export function createNeuralProposalClassifier(cfg) {
     const emitsSet = new Set(emits);
     const penalty = cfg.penalty ?? 0;
     async function classify(section, _ctx) {
-        const tree = await cfg.classifier.parse(section.body);
+        // Postcode regex repair on by default (v0.7 #35, operator-signed): +135/0 on the postcode
+        // harness, model-independent. Fixes the SentencePiece-fragmentation misses (GB/CA/NL/…).
+        const tree = await cfg.classifier.parse(section.body, { postcodeRepair: true });
         const proposals = [];
         const sectionOffset = section.start;
         const visit = (node) => {

package/out/proposal-classifier.js.map

@@ -1 +1 @@
-{"version":3,"file":"proposal-classifier.js","sourceRoot":"","sources":["../proposal-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAYH,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAA;AAmBzC,0EAA0E;AAC1E,MAAM,UAAU,8BAA8B,CAAC,GAAmC;IACjF,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,IAAI,WAAW,CAAA;IACtC,MAAM,QAAQ,GAAG,IAAI,GAAG,CAAe,KAAgC,CAAC,CAAA;IACxE,MAAM,OAAO,GAAG,GAAG,CAAC,OAAO,IAAI,CAAC,CAAA;IAEhC,KAAK,UAAU,QAAQ,CAAC,OAAgB,EAAE,IAAuB;QAChE,MAAM,IAAI,GAAG,MAAM,GAAG,CAAC,UAAU,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,CAAA;QACrD,MAAM,SAAS,GAA6B,EAAE,CAAA;QAC9C,MAAM,aAAa,GAAG,OAAO,CAAC,KAAK,CAAA;QAEnC,MAAM,KAAK,GAAG,CAAC,IAAiB,EAAQ,EAAE;YACzC,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC5B,wFAAwF;gBACxF,oFAAoF;gBACpF,uFAAuF;gBACvF,iFAAiF;gBACjF,mFAAmF;gBACnF,2EAA2E;gBAC3E,MAAM,IAAI,GAAG;oBACZ,KAAK,EAAE,aAAa,GAAG,IAAI,CAAC,KAAK;oBACjC,GAAG,EAAE,aAAa,GAAG,IAAI,CAAC,GAAG;oBAC7B,IAAI,EAAE,IAAI,CAAC,KAAK;iBACG,CAAA;gBACpB,SAAS,CAAC,IAAI,CAAC;oBACd,IAAI;oBACJ,SAAS,EAAE,IAAI,CAAC,GAAG;oBACnB,UAAU,EAAE,IAAI,CAAC,UAAU;oBAC3B,MAAM,EAAE,QAAQ;oBAChB,SAAS,EAAE,GAAG,CAAC,EAAE;oBACjB,OAAO;iBACP,CAAC,CAAA;YACH,CAAC;YACD,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,QAAQ;gBAAE,KAAK,CAAC,KAAK,CAAC,CAAA;QAChD,CAAC,CAAA;QAED,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,KAAK;YAAE,KAAK,CAAC,IAAI,CAAC,CAAA;QAC1C,OAAO,SAAS,CAAA;IACjB,CAAC;IAED,OAAO;QACN,EAAE,EAAE,GAAG,CAAC,EAAE;QACV,KAAK;QACL,OAAO,EAAE,GAAG,CAAC,OAAO,IAAI,CAAC,GAAG,CAAC;QAC7B,QAAQ;KACR,CAAA;AACF,CAAC"}
+{"version":3,"file":"proposal-classifier.js","sourceRoot":"","sources":["../proposal-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAYH,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAA;AAmBzC,0EAA0E;AAC1E,MAAM,UAAU,8BAA8B,CAAC,GAAmC;IACjF,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,IAAI,WAAW,CAAA;IACtC,MAAM,QAAQ,GAAG,IAAI,GAAG,CAAe,KAAgC,CAAC,CAAA;IACxE,MAAM,OAAO,GAAG,GAAG,CAAC,OAAO,IAAI,CAAC,CAAA;IAEhC,KAAK,UAAU,QAAQ,CAAC,OAAgB,EAAE,IAAuB;QAChE,0FAA0F;QAC1F,yFAAyF;QACzF,MAAM,IAAI,GAAG,MAAM,GAAG,CAAC,UAAU,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,EAAE,EAAE,cAAc,EAAE,IAAI,EAAE,CAAC,CAAA;QAC/E,MAAM,SAAS,GAA6B,EAAE,CAAA;QAC9C,MAAM,aAAa,GAAG,OAAO,CAAC,KAAK,CAAA;QAEnC,MAAM,KAAK,GAAG,CAAC,IAAiB,EAAQ,EAAE;YACzC,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC5B,wFAAwF;gBACxF,oFAAoF;gBACpF,uFAAuF;gBACvF,iFAAiF;gBACjF,mFAAmF;gBACnF,2EAA2E;gBAC3E,MAAM,IAAI,GAAG;oBACZ,KAAK,EAAE,aAAa,GAAG,IAAI,CAAC,KAAK;oBACjC,GAAG,EAAE,aAAa,GAAG,IAAI,CAAC,GAAG;oBAC7B,IAAI,EAAE,IAAI,CAAC,KAAK;iBACG,CAAA;gBACpB,SAAS,CAAC,IAAI,CAAC;oBACd,IAAI;oBACJ,SAAS,EAAE,IAAI,CAAC,GAAG;oBACnB,UAAU,EAAE,IAAI,CAAC,UAAU;oBAC3B,MAAM,EAAE,QAAQ;oBAChB,SAAS,EAAE,GAAG,CAAC,EAAE;oBACjB,OAAO;iBACP,CAAC,CAAA;YACH,CAAC;YACD,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,QAAQ;gBAAE,KAAK,CAAC,KAAK,CAAC,CAAA;QAChD,CAAC,CAAA;QAED,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,KAAK;YAAE,KAAK,CAAC,IAAI,CAAC,CAAA;QAC1C,OAAO,SAAS,CAAA;IACjB,CAAC;IAED,OAAO;QACN,EAAE,EAAE,GAAG,CAAC,EAAE;QACV,KAAK;QACL,OAAO,EAAE,GAAG,CAAC,OAAO,IAAI,CAAC,GAAG,CAAC;QAC7B,QAAQ;KACR,CAAA;AACF,CAAC"}

package/out/query-shape-prior.d.ts

@@ -24,6 +24,11 @@
  */
 export interface QueryShapeLike {
     knownFormats: ReadonlyArray<KnownFormatHitLike>;
+    regionAbbreviations?: ReadonlyArray<RegionAbbreviationHitLike>;
+}
+export interface RegionAbbreviationHitLike {
+    start: number;
+    span: string;
 }
 export interface KnownFormatHitLike {
     format: string;
@@ -45,6 +50,13 @@ export interface BuildPriorsOpts {
      * favored label. Confidence-scaled, so a 0.6-confidence format hit gets +0.6 max bias.
      */
     biasScale?: number;
+    /**
+     * Bias magnitude for the locality soft prior (in log-odds units). Default 2.0 — adds ~e^2 ≈ 7.4×
+     * odds to B-locality / I-locality for tokens preceding a detected region abbreviation.
+     */
+    localityBiasScale?: number;
+    /** Raw input text for region-name matching in the locality bias guard. */
+    inputText?: string;
 }
 /**
  * Build a `[seqLen][numLabels]` matrix of additive log-bias to be added to encoder emissions before

package/out/query-shape-prior.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"query-shape-prior.d.ts","sourceRoot":"","sources":["../query-shape-prior.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH;;;GAGG;AACH,MAAM,WAAW,cAAc;IAC9B,YAAY,EAAE,aAAa,CAAC,kBAAkB,CAAC,CAAA;CAC/C;AAED,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAA;IACd,IAAI,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAA;IACpC,4EAA4E;IAC5E,UAAU,EAAE,MAAM,CAAA;CAClB;AAED,+DAA+D;AAC/D,MAAM,WAAW,SAAS;IACzB,KAAK,EAAE,MAAM,CAAA;IACb,GAAG,EAAE,MAAM,CAAA;CACX;AAiBD,MAAM,WAAW,eAAe;IAC/B;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAA;CAClB;AAED;;;;;;;;;GASG;AACH,wBAAgB,mBAAmB,CAClC,KAAK,EAAE,cAAc,EACrB,MAAM,EAAE,aAAa,CAAC,SAAS,CAAC,EAChC,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC,EAC7B,IAAI,GAAE,eAAoB,GACxB,MAAM,EAAE,EAAE,CA4BZ;AAMD,0EAA0E;AAC1E,wBAAgB,iBAAiB,CAAC,SAAS,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE,GAAG,MAAM,EAAE,EAAE,CAWvF"}
+{"version":3,"file":"query-shape-prior.d.ts","sourceRoot":"","sources":["../query-shape-prior.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH;;;GAGG;AACH,MAAM,WAAW,cAAc;IAC9B,YAAY,EAAE,aAAa,CAAC,kBAAkB,CAAC,CAAA;IAC/C,mBAAmB,CAAC,EAAE,aAAa,CAAC,yBAAyB,CAAC,CAAA;CAC9D;AAED,MAAM,WAAW,yBAAyB;IACzC,KAAK,EAAE,MAAM,CAAA;IACb,IAAI,EAAE,MAAM,CAAA;CACZ;AAED,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAA;IACd,IAAI,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAA;IACpC,4EAA4E;IAC5E,UAAU,EAAE,MAAM,CAAA;CAClB;AAED,+DAA+D;AAC/D,MAAM,WAAW,SAAS;IACzB,KAAK,EAAE,MAAM,CAAA;IACb,GAAG,EAAE,MAAM,CAAA;CACX;AAiBD,MAAM,WAAW,eAAe;IAC/B;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAA;IAElB;;;OAGG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAA;IAC1B,0EAA0E;IAC1E,SAAS,CAAC,EAAE,MAAM,CAAA;CAClB;AAED;;;;;;;;;GASG;AACH,wBAAgB,mBAAmB,CAClC,KAAK,EAAE,cAAc,EACrB,MAAM,EAAE,aAAa,CAAC,SAAS,CAAC,EAChC,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC,EAC7B,IAAI,GAAE,eAAoB,GACxB,MAAM,EAAE,EAAE,CAoCZ;AA2ID,0EAA0E;AAC1E,wBAAgB,iBAAiB,CAAC,SAAS,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE,GAAG,MAAM,EAAE,EAAE,CAWvF"}

package/out/query-shape-prior.js

@@ -49,12 +49,13 @@ export function buildEmissionPriors(shape, tokens, labels, opts = {}) {
     const matrix = [];
     for (let t = 0; t < T; t++)
         matrix.push(new Array(L).fill(0));
-    if (shape.knownFormats.length === 0)
-        return matrix;
     // Index label → column for fast lookup.
     const labelToCol = new Map();
     for (let k = 0; k < labels.length; k++)
         labelToCol.set(labels[k], k);
+    if (shape.knownFormats.length === 0 && (!shape.regionAbbreviations || shape.regionAbbreviations.length === 0)) {
+        return matrix;
+    }
     for (const hit of shape.knownFormats) {
         const targetLabel = FORMAT_TO_LABEL.get(hit.format);
         if (!targetLabel)
@@ -70,8 +71,137 @@ export function buildEmissionPriors(shape, tokens, labels, opts = {}) {
             }
         }
     }
+    // Locality soft prior: when a region abbreviation is detected (e.g., "DC", "NY"), bias
+    // preceding alphabetic tokens toward B-locality / I-locality. This counters the WOF
+    // bare-name frequency dominance that makes the model over-emit B-region on ambiguous
+    // place names like "Washington" or "New York".
+    applyLocalityBias(matrix, shape, tokens, labelToCol, opts.localityBiasScale ?? 2.0, opts.inputText);
     return matrix;
 }
+/**
+ * Apply locality bias to tokens preceding a detected region abbreviation.
+ *
+ * For "Washington, DC" — "DC" is the region abbreviation; "Washington" gets biased toward
+ * B-locality. For "New York, NY" — "New" gets B-locality and "York" gets I-locality.
+ *
+ * Guard: if the preceding text matches the full name of the region that the abbreviation represents
+ * (e.g., "Washington" before "WA"), the locality bias is NOT applied — the text IS the region, not
+ * a locality within it.
+ *
+ * Constraint: only bias tokens that appear BEFORE the abbreviation's character offset and are
+ * alphabetic (start with uppercase). Tokens that are part of a known postcode format or are
+ * themselves region abbreviations are skipped.
+ */
+function applyLocalityBias(matrix, shape, tokens, labelToCol, localityBias, inputText) {
+    const abbrevs = shape.regionAbbreviations;
+    if (!abbrevs || abbrevs.length === 0)
+        return;
+    const bLocCol = labelToCol.get("B-locality");
+    const iLocCol = labelToCol.get("I-locality");
+    if (bLocCol === undefined)
+        return;
+    for (const abbrev of abbrevs) {
+        const candidates = [];
+        let prevStart = abbrev.start;
+        for (let t = tokens.length - 1; t >= 0; t--) {
+            const tok = tokens[t];
+            if (tok.end > abbrev.start)
+                continue;
+            const gap = prevStart - tok.end;
+            if (candidates.length === 0 && gap > 4)
+                break;
+            if (candidates.length > 0 && gap > 2)
+                break;
+            let isPostcode = false;
+            for (const fmt of shape.knownFormats) {
+                if (overlaps(tok, fmt.span)) {
+                    isPostcode = true;
+                    break;
+                }
+            }
+            if (isPostcode)
+                break;
+            candidates.push(t);
+            prevStart = tok.start;
+        }
+        if (candidates.length === 0)
+            continue;
+        candidates.reverse();
+        if (inputText) {
+            const firstTok = tokens[candidates[0]];
+            const lastTok = tokens[candidates[candidates.length - 1]];
+            const candidateText = inputText.slice(firstTok.start, lastTok.end).toLowerCase();
+            const regionNames = ABBREV_TO_REGION.get(abbrev.span);
+            if (regionNames?.some((name) => candidateText === name))
+                continue;
+        }
+        for (let i = 0; i < candidates.length; i++) {
+            const t = candidates[i];
+            const col = i === 0 ? bLocCol : iLocCol;
+            if (col === undefined)
+                continue;
+            matrix[t][col] = Math.max(matrix[t][col], localityBias);
+        }
+    }
+}
+const ABBREV_TO_REGION = new Map([
+    ["AL", ["alabama"]],
+    ["AK", ["alaska"]],
+    ["AZ", ["arizona"]],
+    ["AR", ["arkansas"]],
+    ["CA", ["california"]],
+    ["CO", ["colorado"]],
+    ["CT", ["connecticut"]],
+    ["DE", ["delaware"]],
+    ["DC", ["district of columbia"]],
+    ["FL", ["florida"]],
+    ["GA", ["georgia"]],
+    ["HI", ["hawaii"]],
+    ["ID", ["idaho"]],
+    ["IL", ["illinois"]],
+    ["IN", ["indiana"]],
+    ["IA", ["iowa"]],
+    ["KS", ["kansas"]],
+    ["KY", ["kentucky"]],
+    ["LA", ["louisiana"]],
+    ["ME", ["maine"]],
+    ["MD", ["maryland"]],
+    ["MA", ["massachusetts"]],
+    ["MI", ["michigan"]],
+    ["MN", ["minnesota"]],
+    ["MS", ["mississippi"]],
+    ["MO", ["missouri"]],
+    ["MT", ["montana"]],
+    ["NE", ["nebraska"]],
+    ["NV", ["nevada"]],
+    ["NH", ["new hampshire"]],
+    ["NJ", ["new jersey"]],
+    ["NM", ["new mexico"]],
+    ["NY", ["new york"]],
+    ["NC", ["north carolina"]],
+    ["ND", ["north dakota"]],
+    ["OH", ["ohio"]],
+    ["OK", ["oklahoma"]],
+    ["OR", ["oregon"]],
+    ["PA", ["pennsylvania"]],
+    ["RI", ["rhode island"]],
+    ["SC", ["south carolina"]],
+    ["SD", ["south dakota"]],
+    ["TN", ["tennessee"]],
+    ["TX", ["texas"]],
+    ["UT", ["utah"]],
+    ["VT", ["vermont"]],
+    ["VA", ["virginia"]],
+    ["WA", ["washington"]],
+    ["WV", ["west virginia"]],
+    ["WI", ["wisconsin"]],
+    ["WY", ["wyoming"]],
+    ["AS", ["american samoa"]],
+    ["GU", ["guam"]],
+    ["MP", ["northern mariana islands"]],
+    ["PR", ["puerto rico"]],
+    ["VI", ["virgin islands"]],
+]);
 function overlaps(a, b) {
     return a.start < b.end && b.start < a.end;
 }

package/out/query-shape-prior.js.map

@@ -1 +1 @@
-{"version":3,"file":"query-shape-prior.js","sourceRoot":"","sources":["../query-shape-prior.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAuBH;;;GAGG;AACH,MAAM,eAAe,GAAgC,IAAI,GAAG,CAAC;IAC5D,CAAC,QAAQ,EAAE,YAAY,CAAC;IACxB,CAAC,SAAS,EAAE,YAAY,CAAC;IACzB,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,QAAQ,EAAE,UAAU,CAAC;CACtB,CAAC,CAAA;AAUF;;;;;;;;;GASG;AACH,MAAM,UAAU,mBAAmB,CAClC,KAAqB,EACrB,MAAgC,EAChC,MAA6B,EAC7B,OAAwB,EAAE;IAE1B,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,CAAA;IACvB,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,CAAA;IACvB,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,GAAG,CAAA;IACvC,MAAM,MAAM,GAAe,EAAE,CAAA;IAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,MAAM,CAAC,IAAI,CAAC,IAAI,KAAK,CAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAA;IAErE,IAAI,KAAK,CAAC,YAAY,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,MAAM,CAAA;IAElD,wCAAwC;IACxC,MAAM,UAAU,GAAG,IAAI,GAAG,EAAkB,CAAA;IAC5C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE;QAAE,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC,CAAA;IAErE,KAAK,MAAM,GAAG,IAAI,KAAK,CAAC,YAAY,EAAE,CAAC;QACtC,MAAM,WAAW,GAAG,eAAe,CAAC,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,CAAA;QACnD,IAAI,CAAC,WAAW;YAAE,SAAQ;QAC1B,MAAM,GAAG,GAAG,UAAU,CAAC,GAAG,CAAC,WAAW,CAAC,CAAA;QACvC,IAAI,GAAG,KAAK,SAAS;YAAE,SAAQ;QAC/B,MAAM,IAAI,GAAG,GAAG,CAAC,UAAU,GAAG,SAAS,CAAA;QACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5B,MAAM,GAAG,GAAG,MAAM,CAAC,CAAC,CAAE,CAAA;YACtB,IAAI,QAAQ,CAAC,GAAG,EAAE,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC7B,MAAM,CAAC,CAAC,CAAE,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,GAAG,CAAE,EAAE,IAAI,CAAC,CAAA;YACnD,CAAC;QACF,CAAC;IACF,CAAC;IAED,OAAO,MAAM,CAAA;AACd,CAAC;AAED,SAAS,QAAQ,CAAC,CAAiC,EAAE,CAAiC;IACrF,OAAO,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,CAAA;AAC1C,CAAC;AAED,0EAA0E;AAC1E,MAAM,UAAU,iBAAiB,CAAC,SAAqB,EAAE,MAAkB;IAC1E,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,KAAK,EAAE,CAAC,CAAA;IACnE,MAAM,GAAG,GAAe,EAAE,CAAA;IAC1B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3C,MAAM,CAAC,GAAG,SAAS,CAAC,CAAC,CAAE,CAAA;QACvB,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,IAAI,IAAI,KAAK,CAAS,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;QAC1D,MAAM,GAAG,GAAG,IAAI,KAAK,CAAS,CAAC,CAAC,MAAM,CAAC,CAAA;QACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;YAAE,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAA;QAC/D,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACd,CAAC;IACD,OAAO,GAAG,CAAA;AACX,CAAC"}
+{"version":3,"file":"query-shape-prior.js","sourceRoot":"","sources":["../query-shape-prior.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AA6BH;;;GAGG;AACH,MAAM,eAAe,GAAgC,IAAI,GAAG,CAAC;IAC5D,CAAC,QAAQ,EAAE,YAAY,CAAC;IACxB,CAAC,SAAS,EAAE,YAAY,CAAC;IACzB,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,aAAa,EAAE,YAAY,CAAC;IAC7B,CAAC,QAAQ,EAAE,UAAU,CAAC;CACtB,CAAC,CAAA;AAkBF;;;;;;;;;GASG;AACH,MAAM,UAAU,mBAAmB,CAClC,KAAqB,EACrB,MAAgC,EAChC,MAA6B,EAC7B,OAAwB,EAAE;IAE1B,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,CAAA;IACvB,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,CAAA;IACvB,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,GAAG,CAAA;IACvC,MAAM,MAAM,GAAe,EAAE,CAAA;IAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,MAAM,CAAC,IAAI,CAAC,IAAI,KAAK,CAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAA;IAErE,wCAAwC;IACxC,MAAM,UAAU,GAAG,IAAI,GAAG,EAAkB,CAAA;IAC5C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE;QAAE,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC,CAAA;IAErE,IAAI,KAAK,CAAC,YAAY,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,mBAAmB,IAAI,KAAK,CAAC,mBAAmB,CAAC,MAAM,KAAK,CAAC,CAAC,EAAE,CAAC;QAC/G,OAAO,MAAM,CAAA;IACd,CAAC;IAED,KAAK,MAAM,GAAG,IAAI,KAAK,CAAC,YAAY,EAAE,CAAC;QACtC,MAAM,WAAW,GAAG,eAAe,CAAC,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,CAAA;QACnD,IAAI,CAAC,WAAW;YAAE,SAAQ;QAC1B,MAAM,GAAG,GAAG,UAAU,CAAC,GAAG,CAAC,WAAW,CAAC,CAAA;QACvC,IAAI,GAAG,KAAK,SAAS;YAAE,SAAQ;QAC/B,MAAM,IAAI,GAAG,GAAG,CAAC,UAAU,GAAG,SAAS,CAAA;QACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5B,MAAM,GAAG,GAAG,MAAM,CAAC,CAAC,CAAE,CAAA;YACtB,IAAI,QAAQ,CAAC,GAAG,EAAE,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC7B,MAAM,CAAC,CAAC,CAAE,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,GAAG,CAAE,EAAE,IAAI,CAAC,CAAA;YACnD,CAAC;QACF,CAAC;IACF,CAAC;IAED,uFAAuF;IACvF,oFAAoF;IACpF,qFAAqF;IACrF,+CAA+C;IAC/C,iBAAiB,CAAC,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,UAAU,EAAE,IAAI,CAAC,iBAAiB,IAAI,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,CAAA;IAEnG,OAAO,MAAM,CAAA;AACd,CAAC;AAED;;;;;;;;;;;;;GAaG;AACH,SAAS,iBAAiB,CACzB,MAAkB,EAClB,KAAqB,EACrB,MAAqD,EACrD,UAA+B,EAC/B,YAAoB,EACpB,SAAkB;IAElB,MAAM,OAAO,GAAG,KAAK,CAAC,mBAAmB,CAAA;IACzC,IAAI,CAAC,OAAO,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAM;IAE5C,MAAM,OAAO,GAAG,UAAU,CAAC,GAAG,CAAC,YAAY,CAAC,CAAA;IAC5C,MAAM,OAAO,GAAG,UAAU,CAAC,GAAG,CAAC,YAAY,CAAC,CAAA;IAC5C,IAAI,OAAO,KAAK,SAAS;QAAE,OAAM;IAEjC,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;QAC9B,MAAM,UAAU,GAAa,EAAE,CAAA;QAC/B,IAAI,SAAS,GAAG,MAAM,CAAC,KAAK,CAAA;QAE5B,KAAK,IAAI,CAAC,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7C,MAAM,GAAG,GAAG,MAAM,CAAC,CAAC,CAAE,CAAA;YACtB,IAAI,GAAG,CAAC,GAAG,GAAG,MAAM,CAAC,KAAK;gBAAE,SAAQ;YAEpC,MAAM,GAAG,GAAG,SAAS,GAAG,GAAG,CAAC,GAAG,CAAA;YAC/B,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC,IAAI,GAAG,GAAG,CAAC;gBAAE,MAAK;YAC7C,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,IAAI,GAAG,GAAG,CAAC;gBAAE,MAAK;YAE3C,IAAI,UAAU,GAAG,KAAK,CAAA;YACtB,KAAK,MAAM,GAAG,IAAI,KAAK,CAAC,YAAY,EAAE,CAAC;gBACtC,IAAI,QAAQ,CAAC,GAAG,EAAE,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;oBAC7B,UAAU,GAAG,IAAI,CAAA;oBACjB,MAAK;gBACN,CAAC;YACF,CAAC;YACD,IAAI,UAAU;gBAAE,MAAK;YAErB,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;YAClB,SAAS,GAAG,GAAG,CAAC,KAAK,CAAA;QACtB,CAAC;QAED,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;YAAE,SAAQ;QACrC,UAAU,CAAC,OAAO,EAAE,CAAA;QAEpB,IAAI,SAAS,EAAE,CAAC;YACf,MAAM,QAAQ,GAAG,MAAM,CAAC,UAAU,CAAC,CAAC,CAAE,CAAE,CAAA;YACxC,MAAM,OAAO,GAAG,MAAM,CAAC,UAAU,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAE,CAAE,CAAA;YAC3D,MAAM,aAAa,GAAG,SAAS,CAAC,KAAK,CAAC,QAAQ,CAAC,KAAK,EAAE,OAAO,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,CAAA;YAChF,MAAM,WAAW,GAAG,gBAAgB,CAAC,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,CAAA;YACrD,IAAI,WAAW,EAAE,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,aAAa,KAAK,IAAI,CAAC;gBAAE,SAAQ;QAClE,CAAC;QAED,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,UAAU,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5C,MAAM,CAAC,GAAG,UAAU,CAAC,CAAC,CAAE,CAAA;YACxB,MAAM,GAAG,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAA;YACvC,IAAI,GAAG,KAAK,SAAS;gBAAE,SAAQ;YAC/B,MAAM,CAAC,CAAC,CAAE,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,GAAG,CAAE,EAAE,YAAY,CAAC,CAAA;QAC3D,CAAC;IACF,CAAC;AACF,CAAC;AAED,MAAM,gBAAgB,GAAkC,IAAI,GAAG,CAAC;IAC/D,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,CAAC;IAClB,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,YAAY,CAAC,CAAC;IACtB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,aAAa,CAAC,CAAC;IACvB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,sBAAsB,CAAC,CAAC;IAChC,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,CAAC;IAClB,CAAC,IAAI,EAAE,CAAC,OAAO,CAAC,CAAC;IACjB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,MAAM,CAAC,CAAC;IAChB,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,CAAC;IAClB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,WAAW,CAAC,CAAC;IACrB,CAAC,IAAI,EAAE,CAAC,OAAO,CAAC,CAAC;IACjB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,eAAe,CAAC,CAAC;IACzB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,WAAW,CAAC,CAAC;IACrB,CAAC,IAAI,EAAE,CAAC,aAAa,CAAC,CAAC;IACvB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,CAAC;IAClB,CAAC,IAAI,EAAE,CAAC,eAAe,CAAC,CAAC;IACzB,CAAC,IAAI,EAAE,CAAC,YAAY,CAAC,CAAC;IACtB,CAAC,IAAI,EAAE,CAAC,YAAY,CAAC,CAAC;IACtB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,gBAAgB,CAAC,CAAC;IAC1B,CAAC,IAAI,EAAE,CAAC,cAAc,CAAC,CAAC;IACxB,CAAC,IAAI,EAAE,CAAC,MAAM,CAAC,CAAC;IAChB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,CAAC;IAClB,CAAC,IAAI,EAAE,CAAC,cAAc,CAAC,CAAC;IACxB,CAAC,IAAI,EAAE,CAAC,cAAc,CAAC,CAAC;IACxB,CAAC,IAAI,EAAE,CAAC,gBAAgB,CAAC,CAAC;IAC1B,CAAC,IAAI,EAAE,CAAC,cAAc,CAAC,CAAC;IACxB,CAAC,IAAI,EAAE,CAAC,WAAW,CAAC,CAAC;IACrB,CAAC,IAAI,EAAE,CAAC,OAAO,CAAC,CAAC;IACjB,CAAC,IAAI,EAAE,CAAC,MAAM,CAAC,CAAC;IAChB,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC;IACpB,CAAC,IAAI,EAAE,CAAC,YAAY,CAAC,CAAC;IACtB,CAAC,IAAI,EAAE,CAAC,eAAe,CAAC,CAAC;IACzB,CAAC,IAAI,EAAE,CAAC,WAAW,CAAC,CAAC;IACrB,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC;IACnB,CAAC,IAAI,EAAE,CAAC,gBAAgB,CAAC,CAAC;IAC1B,CAAC,IAAI,EAAE,CAAC,MAAM,CAAC,CAAC;IAChB,CAAC,IAAI,EAAE,CAAC,0BAA0B,CAAC,CAAC;IACpC,CAAC,IAAI,EAAE,CAAC,aAAa,CAAC,CAAC;IACvB,CAAC,IAAI,EAAE,CAAC,gBAAgB,CAAC,CAAC;CAC1B,CAAC,CAAA;AAEF,SAAS,QAAQ,CAAC,CAAiC,EAAE,CAAiC;IACrF,OAAO,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,CAAA;AAC1C,CAAC;AAED,0EAA0E;AAC1E,MAAM,UAAU,iBAAiB,CAAC,SAAqB,EAAE,MAAkB;IAC1E,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,KAAK,EAAE,CAAC,CAAA;IACnE,MAAM,GAAG,GAAe,EAAE,CAAA;IAC1B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3C,MAAM,CAAC,GAAG,SAAS,CAAC,CAAC,CAAE,CAAA;QACvB,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,IAAI,IAAI,KAAK,CAAS,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;QAC1D,MAAM,GAAG,GAAG,IAAI,KAAK,CAAS,CAAC,CAAC,MAAM,CAAC,CAAA;QACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;YAAE,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAA;QAC/D,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACd,CAAC;IACD,OAAO,GAAG,CAAA;AACX,CAAC"}

package/out/street-morphology-prior.d.ts

@@ -0,0 +1,56 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Street-morphology emission bias — Layer 1 of the four-layer street-supplement architecture (see
+ *   `docs/articles/concepts/street-supplement-architecture.md`).
+ *
+ *   This module composes with {@linkcode buildFstEmissionPriors} (admin FST) and the QueryShape prior
+ *   via {@linkcode addEmissionMatrix} — same shape, same additive semantics. Where the admin FST
+ *   biases admin BIO labels (`B/I-locality`, `B/I-region`, ...), the morphology FST biases:
+ *
+ *   - **Affix-token (the matched span):** toward `B/I-street_prefix` AND `B/I-street_suffix` (position
+ *       unknown — let the model + context disambiguate).
+ *   - **Adjacent token (one before AND one after each match):** toward `B/I-street`, AWAY from
+ *       `B/I-dependent_locality`. The negative bias on `dependent_locality` is the load-bearing
+ *       piece — it closes the inference-time vacuum that caused v0.6.1's 1066 dep_locality
+ *       hallucinations (see [[project-v061-failure-mechanism]]).
+ *
+ *   The morphology FST itself is built by `resolver-wof-sqlite/street-morphology-fst-builder.ts` and
+ *   ships as a separate binary (`fst-street-morphology.bin`) loaded into a second `FstMatcher`
+ *   instance.
+ */
+import { type FstMatcherLike } from "./fst-prior.js";
+import type { TokenLike } from "./query-shape-prior.js";
+export interface StreetMorphologyPriorOpts {
+    /** Multiplier on the base bias before {@linkcode maxBias} is applied. Default 1.0. */
+    biasScale?: number;
+    /**
+     * Maximum bias magnitude (logits) on the affix span itself. Default 3.0 — same as the admin FST.
+     * The morphology signal is structurally less ambiguous than admin names (`Avenue` is almost never
+     * anything but street-typing), so equal magnitude is justified.
+     */
+    maxAffixBias?: number;
+    /**
+     * Maximum bias magnitude (logits) on the adjacent (neighbour) tokens for the `street` label.
+     * Default 2.0 — a touch weaker than the affix bias because the neighbour is inferred from
+     * adjacency, not direct match.
+     */
+    maxNeighbourStreetBias?: number;
+    /**
+     * Magnitude of the negative bias applied to `dependent_locality` BIO labels on the adjacent
+     * tokens. Default 2.0. This is the load-bearing piece.
+     */
+    dependentLocalityPenalty?: number;
+}
+/**
+ * Build a `[seqLen][numLabels]` bias matrix from street-morphology FST matches.
+ *
+ * The output composes with the admin FST bias matrix via {@linkcode addEmissionMatrix} — same
+ * `addEmissionMatrix(emissions, fstBias) → biasedEmissions` pattern as the existing admin prior.
+ */
+export declare function buildStreetMorphologyEmissionPriors(fst: FstMatcherLike, pieces: ReadonlyArray<TokenLike & {
+    piece: string;
+}>, labels: ReadonlyArray<string>, opts?: StreetMorphologyPriorOpts): number[][];
+//# sourceMappingURL=street-morphology-prior.d.ts.map

package/out/street-morphology-prior.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"street-morphology-prior.d.ts","sourceRoot":"","sources":["../street-morphology-prior.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAwB,KAAK,cAAc,EAAkB,MAAM,gBAAgB,CAAA;AAC1F,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,wBAAwB,CAAA;AAEvD,MAAM,WAAW,yBAAyB;IACzC,sFAAsF;IACtF,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB;;;;OAIG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,MAAM,CAAA;IAC/B;;;OAGG;IACH,wBAAwB,CAAC,EAAE,MAAM,CAAA;CACjC;AAED;;;;;GAKG;AACH,wBAAgB,mCAAmC,CAClD,GAAG,EAAE,cAAc,EACnB,MAAM,EAAE,aAAa,CAAC,SAAS,GAAG;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC,EACpD,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC,EAC7B,IAAI,GAAE,yBAA8B,GAClC,MAAM,EAAE,EAAE,CAgIZ"}

package/out/street-morphology-prior.js

@@ -0,0 +1,159 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Street-morphology emission bias — Layer 1 of the four-layer street-supplement architecture (see
+ *   `docs/articles/concepts/street-supplement-architecture.md`).
+ *
+ *   This module composes with {@linkcode buildFstEmissionPriors} (admin FST) and the QueryShape prior
+ *   via {@linkcode addEmissionMatrix} — same shape, same additive semantics. Where the admin FST
+ *   biases admin BIO labels (`B/I-locality`, `B/I-region`, ...), the morphology FST biases:
+ *
+ *   - **Affix-token (the matched span):** toward `B/I-street_prefix` AND `B/I-street_suffix` (position
+ *       unknown — let the model + context disambiguate).
+ *   - **Adjacent token (one before AND one after each match):** toward `B/I-street`, AWAY from
+ *       `B/I-dependent_locality`. The negative bias on `dependent_locality` is the load-bearing
+ *       piece — it closes the inference-time vacuum that caused v0.6.1's 1066 dep_locality
+ *       hallucinations (see [[project-v061-failure-mechanism]]).
+ *
+ *   The morphology FST itself is built by `resolver-wof-sqlite/street-morphology-fst-builder.ts` and
+ *   ships as a separate binary (`fst-street-morphology.bin`) loaded into a second `FstMatcher`
+ *   instance.
+ */
+import { groupPiecesIntoWords } from "./fst-prior.js";
+/**
+ * Build a `[seqLen][numLabels]` bias matrix from street-morphology FST matches.
+ *
+ * The output composes with the admin FST bias matrix via {@linkcode addEmissionMatrix} — same
+ * `addEmissionMatrix(emissions, fstBias) → biasedEmissions` pattern as the existing admin prior.
+ */
+export function buildStreetMorphologyEmissionPriors(fst, pieces, labels, opts = {}) {
+    const T = pieces.length;
+    const L = labels.length;
+    const biasScale = opts.biasScale ?? 1.0;
+    const maxAffixBias = opts.maxAffixBias ?? 3.0;
+    const maxNeighbourStreetBias = opts.maxNeighbourStreetBias ?? 2.0;
+    const dependentLocalityPenalty = opts.dependentLocalityPenalty ?? 2.0;
+    const matrix = [];
+    for (let t = 0; t < T; t++)
+        matrix.push(new Array(L).fill(0));
+    const labelToCol = new Map();
+    for (let k = 0; k < labels.length; k++)
+        labelToCol.set(labels[k], k);
+    const bStreetPrefix = labelToCol.get("B-street_prefix");
+    const iStreetPrefix = labelToCol.get("I-street_prefix");
+    const bStreetSuffix = labelToCol.get("B-street_suffix");
+    const iStreetSuffix = labelToCol.get("I-street_suffix");
+    const bStreet = labelToCol.get("B-street");
+    const iStreet = labelToCol.get("I-street");
+    const bDepLoc = labelToCol.get("B-dependent_locality");
+    const iDepLoc = labelToCol.get("I-dependent_locality");
+    // If the label vocabulary doesn't include street tags at all (e.g. a Stage 1 model), there's
+    // nothing to bias toward. Return zero-matrix and let the additive pipeline no-op.
+    if (bStreet === undefined || bStreetPrefix === undefined || bStreetSuffix === undefined) {
+        return matrix;
+    }
+    const wordGroups = groupPiecesIntoWords(pieces);
+    if (wordGroups.length === 0)
+        return matrix;
+    const affixMatches = [];
+    // Pass 1 — walk every contiguous subpath, collect accepting morphology matches, and apply
+    // the affix bias to matched tokens.
+    for (let start = 0; start < wordGroups.length; start++) {
+        const group = wordGroups[start];
+        if (group.fstToken === "")
+            continue;
+        const initial = fst.walk([group.fstToken]);
+        if (!initial)
+            continue;
+        let bestEnd = -1;
+        let bestStateId = -1;
+        if (initial.accepted) {
+            bestEnd = start;
+            bestStateId = initial.stateId;
+        }
+        let current = initial;
+        for (let end = start + 1; end < wordGroups.length; end++) {
+            const nextGroup = wordGroups[end];
+            if (nextGroup.fstToken === "")
+                continue;
+            const next = fst.walkFrom(current, nextGroup.fstToken);
+            if (!next)
+                break;
+            if (next.accepted) {
+                bestEnd = end;
+                bestStateId = next.stateId;
+            }
+            current = next;
+        }
+        if (bestEnd === -1)
+            continue;
+        // Verify the accepting entries are street_affix (the morphology FST may eventually contain
+        // other placetypes if the binary format is reused for related priors).
+        const entries = fst.accepting(bestStateId);
+        const hasAffix = entries.some((e) => e.placetype === "street_affix");
+        if (!hasAffix)
+            continue;
+        affixMatches.push({ startGroupIdx: start, endGroupIdx: bestEnd });
+        // Collect piece indices for the matched span.
+        const affixPieceIndices = [];
+        for (let g = start; g <= bestEnd; g++) {
+            const wg = wordGroups[g];
+            if (wg.fstToken === "")
+                continue;
+            for (const pi of wg.pieceIndices)
+                affixPieceIndices.push(pi);
+        }
+        // Apply affix bias: positive bias toward both prefix and suffix BIO labels on the matched
+        // tokens. The model's existing logits + the QueryShape prior + the adjacent context (via
+        // pass 2) determine which of {prefix, suffix} actually wins. We don't pre-commit to one.
+        const affixBias = biasScale * maxAffixBias;
+        for (let k = 0; k < affixPieceIndices.length; k++) {
+            const pi = affixPieceIndices[k];
+            const prefixCol = k === 0 ? bStreetPrefix : (iStreetPrefix ?? bStreetPrefix);
+            const suffixCol = k === 0 ? bStreetSuffix : (iStreetSuffix ?? bStreetSuffix);
+            matrix[pi][prefixCol] = Math.max(matrix[pi][prefixCol], affixBias);
+            matrix[pi][suffixCol] = Math.max(matrix[pi][suffixCol], affixBias);
+        }
+    }
+    if (affixMatches.length === 0)
+        return matrix;
+    // Pass 2 — for each affix match, identify the immediately-adjacent word groups (skipping
+    // empty/punctuation groups) on either side and bias them toward street, away from
+    // dependent_locality.
+    const neighbourStreetBias = biasScale * maxNeighbourStreetBias;
+    for (const match of affixMatches) {
+        const before = findNeighbour(wordGroups, match.startGroupIdx, -1);
+        const after = findNeighbour(wordGroups, match.endGroupIdx, +1);
+        for (const neighbour of [before, after]) {
+            if (!neighbour)
+                continue;
+            const indices = neighbour.pieceIndices;
+            for (let k = 0; k < indices.length; k++) {
+                const pi = indices[k];
+                const streetCol = k === 0 ? bStreet : (iStreet ?? bStreet);
+                matrix[pi][streetCol] = Math.max(matrix[pi][streetCol], neighbourStreetBias);
+                if (bDepLoc !== undefined) {
+                    const depLocCol = k === 0 ? bDepLoc : (iDepLoc ?? bDepLoc);
+                    matrix[pi][depLocCol] = Math.min(matrix[pi][depLocCol], -dependentLocalityPenalty);
+                }
+            }
+        }
+    }
+    return matrix;
+}
+/**
+ * Walk word groups outward from `fromGroupIdx` in `direction` (+1 or -1), skipping empty groups
+ * (whitespace / punctuation), and return the first non-empty group encountered — or `null` if no
+ * such neighbour exists.
+ */
+function findNeighbour(groups, fromGroupIdx, direction) {
+    for (let i = fromGroupIdx + direction; i >= 0 && i < groups.length; i += direction) {
+        const g = groups[i];
+        if (g.fstToken !== "")
+            return g;
+    }
+    return null;
+}
+//# sourceMappingURL=street-morphology-prior.js.map