@mailwoman/neural 2.2.0 → 4.1.0

package/out/fst-prior.js

@@ -0,0 +1,173 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Soft-prior emission biases derived from the FST gazetteer. When the FST finds that a token
+ *   sequence matches a known place name (e.g., "New York" → locality + region), this module
+ *   produces additive biases that nudge the Viterbi decoder toward the matching BIO labels.
+ *
+ *   Composes with the QueryShape prior via addEmissionMatrix — same integration point, same additive
+ *   semantics.
+ *
+ *   SentencePiece ↔ FST bridge: SentencePiece pieces are grouped into whitespace words (by the ▁
+ *   sentinel), normalized through the same pipeline as FST edges (NFKC, lowercase, strip
+ *   non-alnum), and walked through the FST as contiguous subpaths.
+ *
+ *   Uses structural typing for the FST input so this module has zero dependencies on
+ *   `@mailwoman/resolver-wof-sqlite` — consumers pass an FstMatcher instance, but this file only
+ *   consumes the shape.
+ */
+const SPACE_SENTINEL = "▁";
+// ---------------------------------------------------------------------------
+// Placetype → BIO label mapping
+// ---------------------------------------------------------------------------
+const PLACETYPE_TO_BIO = new Map([
+    ["country", "country"],
+    ["region", "region"],
+    ["locality", "locality"],
+    ["postalcode", "postcode"],
+]);
+const SUPPRESS_WHEN_PLACE = ["B-street", "I-street", "B-house_number", "I-house_number", "B-venue"];
+/**
+ * Build a `[seqLen][numLabels]` bias matrix from FST gazetteer matches.
+ *
+ * Walks all contiguous subpaths of the reconstructed whitespace-token sequence through the FST. For
+ * each accepting state, biases the corresponding BIO labels on the matched pieces.
+ */
+export function buildFstEmissionPriors(fst, pieces, labels, opts = {}) {
+    const T = pieces.length;
+    const L = labels.length;
+    const biasScale = opts.biasScale ?? 1.0;
+    const seenWOFIDs = new Set();
+    const maxBias = opts.maxBias ?? 3.0;
+    const suppressionScale = opts.suppressionScale ?? 1.5;
+    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 wordGroups = groupPiecesIntoWords(pieces);
+    if (wordGroups.length === 0)
+        return matrix;
+    for (let start = 0; start < wordGroups.length; start++) {
+        const group = wordGroups[start];
+        if (group.fstToken === "")
+            continue;
+        const match = fst.walk([group.fstToken]);
+        if (!match)
+            continue;
+        if (match.accepted) {
+            applyBias(matrix, labelToCol, fst.accepting(match.stateId), [group], biasScale, maxBias, suppressionScale, seenWOFIDs);
+        }
+        let current = match;
+        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) {
+                const matchedGroups = wordGroups.slice(start, end + 1).filter((g) => g.fstToken !== "");
+                applyBias(matrix, labelToCol, fst.accepting(next.stateId), matchedGroups, biasScale, maxBias, suppressionScale, seenWOFIDs);
+            }
+            current = next;
+        }
+    }
+    return matrix;
+}
+/**
+ * Group SentencePiece pieces into whitespace-delimited words. Each word's literal text is
+ * reconstructed by concatenating pieces (minus leading ▁), then normalized through the same
+ * pipeline the FST builder uses.
+ *
+ * Exported (alongside {@linkcode normalizeFstToken} and the {@linkcode WordGroup} type) so the
+ * street-morphology prior can reuse the same piece-grouping/normalization pipeline without
+ * duplication. Internal helper signature; not part of the public neural API.
+ */
+export function groupPiecesIntoWords(pieces) {
+    const groups = [];
+    let current = null;
+    for (let i = 0; i < pieces.length; i++) {
+        const p = pieces[i];
+        const hasAlnum = /[\p{L}\p{N}]/u.test(p.piece);
+        if (p.piece.startsWith(SPACE_SENTINEL) || i === 0 || !hasAlnum) {
+            if (current)
+                groups.push(current);
+            if (!hasAlnum) {
+                groups.push({ fstToken: "", pieceIndices: [i] });
+                current = null;
+                continue;
+            }
+            const literal = p.piece.startsWith(SPACE_SENTINEL) ? p.piece.slice(SPACE_SENTINEL.length) : p.piece;
+            current = { fstToken: literal, pieceIndices: [i] };
+        }
+        else {
+            if (current) {
+                current.pieceIndices.push(i);
+                current.fstToken += p.piece;
+            }
+        }
+    }
+    if (current)
+        groups.push(current);
+    for (const g of groups) {
+        if (g.fstToken !== "") {
+            g.fstToken = normalizeFstToken(g.fstToken);
+        }
+    }
+    return groups;
+}
+function normalizeFstToken(s) {
+    const cleaned = s
+        .normalize("NFKC")
+        .toLowerCase()
+        .replace(/[\p{P}\p{S}]/gu, "");
+    return cleaned.length > 0 ? cleaned : "";
+}
+function applyBias(matrix, labelToCol, entries, groups, biasScale, maxBias, suppressionScale, seenWOFIDs) {
+    const seenTags = new Map();
+    for (const entry of entries) {
+        if (seenWOFIDs.has(entry.wofID))
+            continue;
+        seenWOFIDs.add(entry.wofID);
+        const bioTag = PLACETYPE_TO_BIO.get(entry.placetype);
+        if (!bioTag)
+            continue;
+        const impBias = entry.importance * biasScale * maxBias;
+        const existing = seenTags.get(bioTag) ?? 0;
+        if (impBias > existing)
+            seenTags.set(bioTag, impBias);
+    }
+    if (seenTags.size === 0)
+        return;
+    const allPieceIndices = [];
+    for (const group of groups) {
+        for (const pi of group.pieceIndices)
+            allPieceIndices.push(pi);
+    }
+    for (const [bioTag, bias] of seenTags) {
+        const bCol = labelToCol.get(`B-${bioTag}`);
+        const iCol = labelToCol.get(`I-${bioTag}`);
+        if (bCol === undefined)
+            continue;
+        for (let k = 0; k < allPieceIndices.length; k++) {
+            const pi = allPieceIndices[k];
+            const col = k === 0 ? bCol : (iCol ?? bCol);
+            matrix[pi][col] = Math.max(matrix[pi][col], bias);
+        }
+    }
+    if (suppressionScale > 0) {
+        for (const pi of allPieceIndices) {
+            for (const label of SUPPRESS_WHEN_PLACE) {
+                const col = labelToCol.get(label);
+                if (col !== undefined) {
+                    matrix[pi][col] = Math.min(matrix[pi][col], -suppressionScale);
+                }
+            }
+        }
+    }
+}
+//# sourceMappingURL=fst-prior.js.map

package/out/fst-prior.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fst-prior.js","sourceRoot":"","sources":["../fst-prior.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAIH,MAAM,cAAc,GAAG,GAAG,CAAA;AAwB1B,8EAA8E;AAC9E,gCAAgC;AAChC,8EAA8E;AAE9E,MAAM,gBAAgB,GAAgC,IAAI,GAAG,CAAC;IAC7D,CAAC,SAAS,EAAE,SAAS,CAAC;IACtB,CAAC,QAAQ,EAAE,QAAQ,CAAC;IACpB,CAAC,UAAU,EAAE,UAAU,CAAC;IACxB,CAAC,YAAY,EAAE,UAAU,CAAC;CAC1B,CAAC,CAAA;AAWF,MAAM,mBAAmB,GAAsB,CAAC,UAAU,EAAE,UAAU,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,SAAS,CAAC,CAAA;AAYtH;;;;;GAKG;AACH,MAAM,UAAU,sBAAsB,CACrC,GAAmB,EACnB,MAAoD,EACpD,MAA6B,EAC7B,OAAqB,EAAE;IAEvB,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,UAAU,GAAG,IAAI,GAAG,EAAU,CAAA;IACpC,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,IAAI,GAAG,CAAA;IACnC,MAAM,gBAAgB,GAAG,IAAI,CAAC,gBAAgB,IAAI,GAAG,CAAA;IACrD,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,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,MAAM,UAAU,GAAG,oBAAoB,CAAC,MAAM,CAAC,CAAA;IAC/C,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,MAAM,CAAA;IAE1C,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,UAAU,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC;QACxD,MAAM,KAAK,GAAG,UAAU,CAAC,KAAK,CAAE,CAAA;QAChC,IAAI,KAAK,CAAC,QAAQ,KAAK,EAAE;YAAE,SAAQ;QAEnC,MAAM,KAAK,GAAG,GAAG,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAA;QACxC,IAAI,CAAC,KAAK;YAAE,SAAQ;QAEpB,IAAI,KAAK,CAAC,QAAQ,EAAE,CAAC;YACpB,SAAS,CACR,MAAM,EACN,UAAU,EACV,GAAG,CAAC,SAAS,CAAC,KAAK,CAAC,OAAO,CAAC,EAC5B,CAAC,KAAK,CAAC,EACP,SAAS,EACT,OAAO,EACP,gBAAgB,EAChB,UAAU,CACV,CAAA;QACF,CAAC;QAED,IAAI,OAAO,GAAG,KAAK,CAAA;QACnB,KAAK,IAAI,GAAG,GAAG,KAAK,GAAG,CAAC,EAAE,GAAG,GAAG,UAAU,CAAC,MAAM,EAAE,GAAG,EAAE,EAAE,CAAC;YAC1D,MAAM,SAAS,GAAG,UAAU,CAAC,GAAG,CAAE,CAAA;YAClC,IAAI,SAAS,CAAC,QAAQ,KAAK,EAAE;gBAAE,SAAQ;YAEvC,MAAM,IAAI,GAAG,GAAG,CAAC,QAAQ,CAAC,OAAO,EAAE,SAAS,CAAC,QAAQ,CAAC,CAAA;YACtD,IAAI,CAAC,IAAI;gBAAE,MAAK;YAEhB,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACnB,MAAM,aAAa,GAAG,UAAU,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,KAAK,EAAE,CAAC,CAAA;gBACvF,SAAS,CACR,MAAM,EACN,UAAU,EACV,GAAG,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,CAAC,EAC3B,aAAa,EACb,SAAS,EACT,OAAO,EACP,gBAAgB,EAChB,UAAU,CACV,CAAA;YACF,CAAC;YAED,OAAO,GAAG,IAAI,CAAA;QACf,CAAC;IACF,CAAC;IAED,OAAO,MAAM,CAAA;AACd,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,oBAAoB,CAAC,MAAwC;IAC5E,MAAM,MAAM,GAAgB,EAAE,CAAA;IAC9B,IAAI,OAAO,GAAqB,IAAI,CAAA;IAEpC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACxC,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAE,CAAA;QACpB,MAAM,QAAQ,GAAG,eAAe,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAA;QAE9C,IAAI,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC,cAAc,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;YAChE,IAAI,OAAO;gBAAE,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;YACjC,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACf,MAAM,CAAC,IAAI,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;gBAChD,OAAO,GAAG,IAAI,CAAA;gBACd,SAAQ;YACT,CAAC;YACD,MAAM,OAAO,GAAG,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAA;YACnG,OAAO,GAAG,EAAE,QAAQ,EAAE,OAAO,EAAE,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAA;QACnD,CAAC;aAAM,CAAC;YACP,IAAI,OAAO,EAAE,CAAC;gBACb,OAAO,CAAC,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;gBAC5B,OAAO,CAAC,QAAQ,IAAI,CAAC,CAAC,KAAK,CAAA;YAC5B,CAAC;QACF,CAAC;IACF,CAAC;IACD,IAAI,OAAO;QAAE,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;IAEjC,KAAK,MAAM,CAAC,IAAI,MAAM,EAAE,CAAC;QACxB,IAAI,CAAC,CAAC,QAAQ,KAAK,EAAE,EAAE,CAAC;YACvB,CAAC,CAAC,QAAQ,GAAG,iBAAiB,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAA;QAC3C,CAAC;IACF,CAAC;IAED,OAAO,MAAM,CAAA;AACd,CAAC;AAED,SAAS,iBAAiB,CAAC,CAAS;IACnC,MAAM,OAAO,GAAG,CAAC;SACf,SAAS,CAAC,MAAM,CAAC;SACjB,WAAW,EAAE;SACb,OAAO,CAAC,gBAAgB,EAAE,EAAE,CAAC,CAAA;IAC/B,OAAO,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAA;AACzC,CAAC;AAED,SAAS,SAAS,CACjB,MAAkB,EAClB,UAA+B,EAC/B,OAAyC,EACzC,MAAmB,EACnB,SAAiB,EACjB,OAAe,EACf,gBAAwB,EACxB,UAAuB;IAEvB,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAkB,CAAA;IAE1C,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,IAAI,UAAU,CAAC,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC;YAAE,SAAQ;QACzC,UAAU,CAAC,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,CAAA;QAC3B,MAAM,MAAM,GAAG,gBAAgB,CAAC,GAAG,CAAC,KAAK,CAAC,SAAS,CAAC,CAAA;QACpD,IAAI,CAAC,MAAM;YAAE,SAAQ;QACrB,MAAM,OAAO,GAAG,KAAK,CAAC,UAAU,GAAG,SAAS,GAAG,OAAO,CAAA;QACtD,MAAM,QAAQ,GAAG,QAAQ,CAAC,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,CAAA;QAC1C,IAAI,OAAO,GAAG,QAAQ;YAAE,QAAQ,CAAC,GAAG,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;IACtD,CAAC;IAED,IAAI,QAAQ,CAAC,IAAI,KAAK,CAAC;QAAE,OAAM;IAE/B,MAAM,eAAe,GAAa,EAAE,CAAA;IACpC,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QAC5B,KAAK,MAAM,EAAE,IAAI,KAAK,CAAC,YAAY;YAAE,eAAe,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;IAC9D,CAAC;IAED,KAAK,MAAM,CAAC,MAAM,EAAE,IAAI,CAAC,IAAI,QAAQ,EAAE,CAAC;QACvC,MAAM,IAAI,GAAG,UAAU,CAAC,GAAG,CAAC,KAAK,MAAM,EAAE,CAAC,CAAA;QAC1C,MAAM,IAAI,GAAG,UAAU,CAAC,GAAG,CAAC,KAAK,MAAM,EAAE,CAAC,CAAA;QAC1C,IAAI,IAAI,KAAK,SAAS;YAAE,SAAQ;QAEhC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,eAAe,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjD,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,CAAE,CAAA;YAC9B,MAAM,GAAG,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,IAAI,IAAI,CAAC,CAAA;YAC3C,MAAM,CAAC,EAAE,CAAE,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAE,CAAC,GAAG,CAAE,EAAE,IAAI,CAAC,CAAA;QACrD,CAAC;IACF,CAAC;IAED,IAAI,gBAAgB,GAAG,CAAC,EAAE,CAAC;QAC1B,KAAK,MAAM,EAAE,IAAI,eAAe,EAAE,CAAC;YAClC,KAAK,MAAM,KAAK,IAAI,mBAAmB,EAAE,CAAC;gBACzC,MAAM,GAAG,GAAG,UAAU,CAAC,GAAG,CAAC,KAAK,CAAC,CAAA;gBACjC,IAAI,GAAG,KAAK,SAAS,EAAE,CAAC;oBACvB,MAAM,CAAC,EAAE,CAAE,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAE,CAAC,GAAG,CAAE,EAAE,CAAC,gBAAgB,CAAC,CAAA;gBAClE,CAAC;YACF,CAAC;QACF,CAAC;IACF,CAAC;AACF,CAAC"}

package/out/index.d.ts

@@ -3,9 +3,12 @@
  * @license AGPL-3.0
  * @author Teffen Ellis, et al.
  */
+export * from "./anchor-inference.js";
 export * from "./classifier.js";
 export * from "./labels.js";
 export * from "./onnx-runner.js";
+export * from "./postcode-anchor.js";
+export * from "./postcode-binary-resolver.js";
 export * from "./proposal-classifier.js";
 export { addEmissionMatrix, buildEmissionPriors } from "./query-shape-prior.js";
 export type { BuildPriorsOpts, KnownFormatHitLike, QueryShapeLike, TokenLike } from "./query-shape-prior.js";

package/out/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,cAAc,iBAAiB,CAAA;AAC/B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,0BAA0B,CAAA;AACxC,OAAO,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,wBAAwB,CAAA;AAC/E,YAAY,EAAE,eAAe,EAAE,kBAAkB,EAAE,cAAc,EAAE,SAAS,EAAE,MAAM,wBAAwB,CAAA;AAC5G,cAAc,gBAAgB,CAAA;AAC9B,OAAO,EACN,eAAe,EACf,iBAAiB,EACjB,sBAAsB,EACtB,cAAc,EACd,OAAO,EACP,OAAO,GACP,MAAM,cAAc,CAAA;AACrB,YAAY,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,cAAc,CAAA;AAC/D,cAAc,cAAc,CAAA"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,cAAc,uBAAuB,CAAA;AACrC,cAAc,iBAAiB,CAAA;AAC/B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,sBAAsB,CAAA;AACpC,cAAc,+BAA+B,CAAA;AAC7C,cAAc,0BAA0B,CAAA;AACxC,OAAO,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,wBAAwB,CAAA;AAC/E,YAAY,EAAE,eAAe,EAAE,kBAAkB,EAAE,cAAc,EAAE,SAAS,EAAE,MAAM,wBAAwB,CAAA;AAC5G,cAAc,gBAAgB,CAAA;AAC9B,OAAO,EACN,eAAe,EACf,iBAAiB,EACjB,sBAAsB,EACtB,cAAc,EACd,OAAO,EACP,OAAO,GACP,MAAM,cAAc,CAAA;AACrB,YAAY,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,cAAc,CAAA;AAC/D,cAAc,cAAc,CAAA"}

package/out/index.js

@@ -3,9 +3,12 @@
  * @license AGPL-3.0
  * @author Teffen Ellis, et al.
  */
+export * from "./anchor-inference.js";
 export * from "./classifier.js";
 export * from "./labels.js";
 export * from "./onnx-runner.js";
+export * from "./postcode-anchor.js";
+export * from "./postcode-binary-resolver.js";
 export * from "./proposal-classifier.js";
 export { addEmissionMatrix, buildEmissionPriors } from "./query-shape-prior.js";
 export * from "./tokenizer.js";

package/out/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,cAAc,iBAAiB,CAAA;AAC/B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,0BAA0B,CAAA;AACxC,OAAO,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,wBAAwB,CAAA;AAE/E,cAAc,gBAAgB,CAAA;AAC9B,OAAO,EACN,eAAe,EACf,iBAAiB,EACjB,sBAAsB,EACtB,cAAc,EACd,OAAO,EACP,OAAO,GACP,MAAM,cAAc,CAAA;AAErB,cAAc,cAAc,CAAA"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,cAAc,uBAAuB,CAAA;AACrC,cAAc,iBAAiB,CAAA;AAC/B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,sBAAsB,CAAA;AACpC,cAAc,+BAA+B,CAAA;AAC7C,cAAc,0BAA0B,CAAA;AACxC,OAAO,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,wBAAwB,CAAA;AAE/E,cAAc,gBAAgB,CAAA;AAC9B,OAAO,EACN,eAAe,EACf,iBAAiB,EACjB,sBAAsB,EACtB,cAAc,EACd,OAAO,EACP,OAAO,GACP,MAAM,cAAc,CAAA;AAErB,cAAc,cAAc,CAAA"}

package/out/labels.d.ts

@@ -38,4 +38,7 @@ export declare const STAGE2_TAGS: readonly ["country", "region", "locality", "de
  * never gets argmax'd because Stage 1 only emits 15 logits.
  */
 export declare const STAGE2_BIO_LABELS: readonly BioLabel[];
+export declare const STAGE3_FINE_TAGS: readonly ["street_prefix", "street_suffix", "unit", "po_box", "intersection_a", "intersection_b"];
+export declare const STAGE3_TAGS: readonly ["country", "region", "locality", "dependent_locality", "postcode", "subregion", "cedex", "venue", "street", "house_number", "street_prefix", "street_suffix", "unit", "po_box", "intersection_a", "intersection_b"];
+export declare const STAGE3_BIO_LABELS: readonly BioLabel[];
 //# sourceMappingURL=labels.d.ts.map

package/out/labels.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"labels.d.ts","sourceRoot":"","sources":["../labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,yBAAyB,CAAA;AAEvD,0EAA0E;AAC1E,eAAO,MAAM,kBAAkB,oGAQrB,CAAA;AAEV,yFAAyF;AACzF,eAAO,MAAM,iBAAiB,EAAE,SAAS,QAAQ,EAG/C,CAAA;AAEF;;;;GAIG;AACH,eAAO,MAAM,gBAAgB,8CAA+C,CAAA;AAE5E,8FAA8F;AAC9F,eAAO,MAAM,WAAW,uIAAwD,CAAA;AAEhF;;;;;;GAMG;AACH,eAAO,MAAM,iBAAiB,EAAE,SAAS,QAAQ,EAG/C,CAAA"}
+{"version":3,"file":"labels.d.ts","sourceRoot":"","sources":["../labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,yBAAyB,CAAA;AAEvD,0EAA0E;AAC1E,eAAO,MAAM,kBAAkB,oGAQrB,CAAA;AAEV,yFAAyF;AACzF,eAAO,MAAM,iBAAiB,EAAE,SAAS,QAAQ,EAG/C,CAAA;AAEF;;;;GAIG;AACH,eAAO,MAAM,gBAAgB,8CAA+C,CAAA;AAE5E,8FAA8F;AAC9F,eAAO,MAAM,WAAW,uIAAwD,CAAA;AAEhF;;;;;;GAMG;AACH,eAAO,MAAM,iBAAiB,EAAE,SAAS,QAAQ,EAG/C,CAAA;AAEF,eAAO,MAAM,gBAAgB,mGAOnB,CAAA;AAEV,eAAO,MAAM,WAAW,+NAAiD,CAAA;AAEzE,eAAO,MAAM,iBAAiB,EAAE,SAAS,QAAQ,EAG/C,CAAA"}

package/out/labels.js

@@ -51,4 +51,17 @@ export const STAGE2_BIO_LABELS = Object.freeze([
     "O",
     ...STAGE2_TAGS.flatMap((tag) => [`B-${tag}`, `I-${tag}`]),
 ]);
+export const STAGE3_FINE_TAGS = [
+    "street_prefix",
+    "street_suffix",
+    "unit",
+    "po_box",
+    "intersection_a",
+    "intersection_b",
+];
+export const STAGE3_TAGS = [...STAGE2_TAGS, ...STAGE3_FINE_TAGS];
+export const STAGE3_BIO_LABELS = Object.freeze([
+    "O",
+    ...STAGE3_TAGS.flatMap((tag) => [`B-${tag}`, `I-${tag}`]),
+]);
 //# sourceMappingURL=labels.js.map

package/out/labels.js.map

@@ -1 +1 @@
-{"version":3,"file":"labels.js","sourceRoot":"","sources":["../labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAIH,0EAA0E;AAC1E,MAAM,CAAC,MAAM,kBAAkB,GAAG;IACjC,SAAS;IACT,QAAQ;IACR,UAAU;IACV,oBAAoB;IACpB,UAAU;IACV,WAAW;IACX,OAAO;CACE,CAAA;AAEV,yFAAyF;AACzF,MAAM,CAAC,MAAM,iBAAiB,GAAwB,MAAM,CAAC,MAAM,CAAC;IACnE,GAAe;IACf,GAAG,kBAAkB,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,KAAK,GAAG,EAAc,EAAE,KAAK,GAAG,EAAc,CAAC,CAAC;CACxF,CAAC,CAAA;AAEF;;;;GAIG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAG,CAAC,OAAO,EAAE,QAAQ,EAAE,cAAc,CAAU,CAAA;AAE5E,8FAA8F;AAC9F,MAAM,CAAC,MAAM,WAAW,GAAG,CAAC,GAAG,kBAAkB,EAAE,GAAG,gBAAgB,CAAU,CAAA;AAEhF;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAwB,MAAM,CAAC,MAAM,CAAC;IACnE,GAAe;IACf,GAAG,WAAW,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,KAAK,GAAG,EAAc,EAAE,KAAK,GAAG,EAAc,CAAC,CAAC;CACjF,CAAC,CAAA"}
+{"version":3,"file":"labels.js","sourceRoot":"","sources":["../labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAIH,0EAA0E;AAC1E,MAAM,CAAC,MAAM,kBAAkB,GAAG;IACjC,SAAS;IACT,QAAQ;IACR,UAAU;IACV,oBAAoB;IACpB,UAAU;IACV,WAAW;IACX,OAAO;CACE,CAAA;AAEV,yFAAyF;AACzF,MAAM,CAAC,MAAM,iBAAiB,GAAwB,MAAM,CAAC,MAAM,CAAC;IACnE,GAAe;IACf,GAAG,kBAAkB,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,KAAK,GAAG,EAAc,EAAE,KAAK,GAAG,EAAc,CAAC,CAAC;CACxF,CAAC,CAAA;AAEF;;;;GAIG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAG,CAAC,OAAO,EAAE,QAAQ,EAAE,cAAc,CAAU,CAAA;AAE5E,8FAA8F;AAC9F,MAAM,CAAC,MAAM,WAAW,GAAG,CAAC,GAAG,kBAAkB,EAAE,GAAG,gBAAgB,CAAU,CAAA;AAEhF;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAwB,MAAM,CAAC,MAAM,CAAC;IACnE,GAAe;IACf,GAAG,WAAW,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,KAAK,GAAG,EAAc,EAAE,KAAK,GAAG,EAAc,CAAC,CAAC;CACjF,CAAC,CAAA;AAEF,MAAM,CAAC,MAAM,gBAAgB,GAAG;IAC/B,eAAe;IACf,eAAe;IACf,MAAM;IACN,QAAQ;IACR,gBAAgB;IAChB,gBAAgB;CACP,CAAA;AAEV,MAAM,CAAC,MAAM,WAAW,GAAG,CAAC,GAAG,WAAW,EAAE,GAAG,gBAAgB,CAAU,CAAA;AAEzE,MAAM,CAAC,MAAM,iBAAiB,GAAwB,MAAM,CAAC,MAAM,CAAC;IACnE,GAAe;IACf,GAAG,WAAW,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,KAAK,GAAG,EAAc,EAAE,KAAK,GAAG,EAAc,CAAC,CAAC;CACjF,CAAC,CAAA"}

package/out/onnx-runner.d.ts

@@ -50,7 +50,14 @@ export declare class OnnxRunner {
      * back to the actual input length.
      *
      * @param tokenIds The id sequence produced by the tokenizer (no special tokens added).
+     * @param anchor Optional postcode-anchor channel (#239/#240). When supplied (only for anchor
+     *   models — exported with the `anchor_features`/`anchor_confidence` inputs), per-piece features
+     *   `(seqLen × dim)` + confidence `(seqLen,)` are fed, zero-padded to `fixedSeqLen`. Omit for
+     *   plain models, whose ONNX has no anchor inputs.
      */
-    infer(tokenIds: number[]): Promise<InferResult>;
+    infer(tokenIds: number[], anchor?: {
+        features: ReadonlyArray<ReadonlyArray<number>>;
+        confidence: ReadonlyArray<number>;
+    }): Promise<InferResult>;
 }
 //# sourceMappingURL=onnx-runner.d.ts.map

package/out/onnx-runner.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"onnx-runner.d.ts","sourceRoot":"","sources":["../onnx-runner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAKH,MAAM,WAAW,cAAc;IAC9B,wEAAwE;IACxE,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB;;;;;OAKG;IACH,WAAW,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,8FAA8F;AAC9F,eAAO,MAAM,qBAAqB,MAAM,CAAA;AAExC,MAAM,WAAW,WAAW;IAC3B,2EAA2E;IAC3E,MAAM,EAAE,MAAM,EAAE,EAAE,CAAA;IAClB,oEAAoE;IACpE,SAAS,EAAE,MAAM,CAAA;CACjB;AAED,qBAAa,UAAU;IAMrB,OAAO,CAAC,QAAQ,CAAC,SAAS;IAC1B,OAAO,CAAC,QAAQ,CAAC,UAAU;IAN5B,OAAO,CAAC,OAAO,CAAoC;IACnD,OAAO,CAAC,WAAW,CAA6C;IAChE,SAAgB,WAAW,EAAE,MAAM,CAAA;IAEnC,OAAO;IAQP,oEAAoE;WACvD,MAAM,CAAC,SAAS,EAAE,MAAM,EAAE,IAAI,GAAE,cAAmB,GAAG,OAAO,CAAC,UAAU,CAAC;IAMtF,6CAA6C;WAChC,SAAS,CAAC,UAAU,EAAE,UAAU,EAAE,IAAI,GAAE,cAAmB,GAAG,OAAO,CAAC,UAAU,CAAC;YAMhF,aAAa;IAgB3B;;;;;;;OAOG;IACG,KAAK,CAAC,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,WAAW,CAAC;CA8BrD"}
+{"version":3,"file":"onnx-runner.d.ts","sourceRoot":"","sources":["../onnx-runner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAOH,MAAM,WAAW,cAAc;IAC9B,wEAAwE;IACxE,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB;;;;;OAKG;IACH,WAAW,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,8FAA8F;AAC9F,eAAO,MAAM,qBAAqB,MAAM,CAAA;AAExC,MAAM,WAAW,WAAW;IAC3B,2EAA2E;IAC3E,MAAM,EAAE,MAAM,EAAE,EAAE,CAAA;IAClB,oEAAoE;IACpE,SAAS,EAAE,MAAM,CAAA;CACjB;AAED,qBAAa,UAAU;IAMrB,OAAO,CAAC,QAAQ,CAAC,SAAS;IAC1B,OAAO,CAAC,QAAQ,CAAC,UAAU;IAN5B,OAAO,CAAC,OAAO,CAAoC;IACnD,OAAO,CAAC,WAAW,CAA6C;IAChE,SAAgB,WAAW,EAAE,MAAM,CAAA;IAEnC,OAAO;IAQP,oEAAoE;WACvD,MAAM,CAAC,SAAS,EAAE,MAAM,EAAE,IAAI,GAAE,cAAmB,GAAG,OAAO,CAAC,UAAU,CAAC;IAMtF,6CAA6C;WAChC,SAAS,CAAC,UAAU,EAAE,UAAU,EAAE,IAAI,GAAE,cAAmB,GAAG,OAAO,CAAC,UAAU,CAAC;YAMhF,aAAa;IAgB3B;;;;;;;;;;;OAWG;IACG,KAAK,CACV,QAAQ,EAAE,MAAM,EAAE,EAClB,MAAM,CAAC,EAAE;QAAE,QAAQ,EAAE,aAAa,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC;QAAC,UAAU,EAAE,aAAa,CAAC,MAAM,CAAC,CAAA;KAAE,GAC5F,OAAO,CAAC,WAAW,CAAC;CAqDvB"}

package/out/onnx-runner.js

@@ -14,6 +14,7 @@
  */
 import { promises as fs } from "node:fs";
 import ort from "onnxruntime-node";
+import { ANCHOR_FEATURE_DIM } from "./anchor-inference.js";
 /** Default sequence length for v0.1.0 / v0.2.0 (BertConfig max_position_embeddings = 128). */
 export const DEFAULT_FIXED_SEQ_LEN = 128;
 export class OnnxRunner {
@@ -64,8 +65,12 @@ export class OnnxRunner {
      * back to the actual input length.
      *
      * @param tokenIds The id sequence produced by the tokenizer (no special tokens added).
+     * @param anchor Optional postcode-anchor channel (#239/#240). When supplied (only for anchor
+     *   models — exported with the `anchor_features`/`anchor_confidence` inputs), per-piece features
+     *   `(seqLen × dim)` + confidence `(seqLen,)` are fed, zero-padded to `fixedSeqLen`. Omit for
+     *   plain models, whose ONNX has no anchor inputs.
      */
-    async infer(tokenIds) {
+    async infer(tokenIds, anchor) {
         const session = await this.ensureSession();
         const seqLen = Math.min(tokenIds.length, this.fixedSeqLen);
         const padded = new BigInt64Array(this.fixedSeqLen);
@@ -78,6 +83,31 @@ export class OnnxRunner {
             input_ids: new ort.Tensor("int64", padded, [1, this.fixedSeqLen]),
             attention_mask: new ort.Tensor("int64", mask, [1, this.fixedSeqLen]),
         };
+        if (anchor) {
+            const dim = anchor.features[0]?.length ?? 0;
+            const af = new Float32Array(this.fixedSeqLen * dim);
+            const ac = new Float32Array(this.fixedSeqLen);
+            for (let i = 0; i < seqLen; i++) {
+                ac[i] = anchor.confidence[i] ?? 0;
+                const row = anchor.features[i];
+                if (row)
+                    for (let d = 0; d < dim; d++)
+                        af[i * dim + d] = row[d] ?? 0;
+            }
+            feeds.anchor_features = new ort.Tensor("float32", af, [1, this.fixedSeqLen, dim]);
+            feeds.anchor_confidence = new ort.Tensor("float32", ac, [1, this.fixedSeqLen]);
+        }
+        else if (session.inputNames.includes("anchor_features")) {
+            // Anchor-trained model (its ONNX declares the anchor inputs as mandatory) but no anchor data
+            // was supplied: feed zeros. That's the `confidence = 0` identity — the model's anchor-off
+            // behavior. Without it the session throws on the missing required inputs.
+            feeds.anchor_features = new ort.Tensor("float32", new Float32Array(this.fixedSeqLen * ANCHOR_FEATURE_DIM), [
+                1,
+                this.fixedSeqLen,
+                ANCHOR_FEATURE_DIM,
+            ]);
+            feeds.anchor_confidence = new ort.Tensor("float32", new Float32Array(this.fixedSeqLen), [1, this.fixedSeqLen]);
+        }
         const output = await session.run(feeds);
         const logitsTensor = output.logits;
         if (!logitsTensor)

package/out/onnx-runner.js.map

@@ -1 +1 @@
-{"version":3,"file":"onnx-runner.js","sourceRoot":"","sources":["../onnx-runner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,QAAQ,IAAI,EAAE,EAAE,MAAM,SAAS,CAAA;AACxC,OAAO,GAAG,MAAM,kBAAkB,CAAA;AAclC,8FAA8F;AAC9F,MAAM,CAAC,MAAM,qBAAqB,GAAG,GAAG,CAAA;AASxC,MAAM,OAAO,UAAU;IAMJ;IACA;IANV,OAAO,GAAgC,IAAI,CAAA;IAC3C,WAAW,GAAyC,IAAI,CAAA;IAChD,WAAW,CAAQ;IAEnC,YACkB,SAAiB,EACjB,UAA6B,EAC9C,IAAoB;QAFH,cAAS,GAAT,SAAS,CAAQ;QACjB,eAAU,GAAV,UAAU,CAAmB;QAG9C,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,WAAW,IAAI,qBAAqB,CAAA;IAC7D,CAAC;IAED,oEAAoE;IACpE,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,SAAiB,EAAE,OAAuB,EAAE;QAC/D,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,SAAS,EAAE,IAAI,EAAE,IAAI,CAAC,CAAA;QACpD,IAAI,IAAI,CAAC,MAAM;YAAE,MAAM,MAAM,CAAC,aAAa,EAAE,CAAA;QAC7C,OAAO,MAAM,CAAA;IACd,CAAC;IAED,6CAA6C;IAC7C,MAAM,CAAC,KAAK,CAAC,SAAS,CAAC,UAAsB,EAAE,OAAuB,EAAE;QACvE,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,SAAS,EAAE,UAAU,EAAE,IAAI,CAAC,CAAA;QAC1D,IAAI,IAAI,CAAC,MAAM;YAAE,MAAM,MAAM,CAAC,aAAa,EAAE,CAAA;QAC7C,OAAO,MAAM,CAAA;IACd,CAAC;IAEO,KAAK,CAAC,aAAa;QAC1B,IAAI,IAAI,CAAC,OAAO;YAAE,OAAO,IAAI,CAAC,OAAO,CAAA;QACrC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC;YACvB,IAAI,CAAC,WAAW,GAAG,CAAC,KAAK,IAAI,EAAE;gBAC9B,MAAM,KAAK,GAAG,IAAI,CAAC,UAAU,IAAI,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAA;gBAClF,MAAM,OAAO,GAAG,MAAM,GAAG,CAAC,gBAAgB,CAAC,MAAM,CAAC,KAAK,EAAE;oBACxD,kBAAkB,EAAE,CAAC,KAAK,CAAC;oBAC3B,sBAAsB,EAAE,KAAK;iBAC7B,CAAC,CAAA;gBACF,IAAI,CAAC,OAAO,GAAG,OAAO,CAAA;gBACtB,OAAO,OAAO,CAAA;YACf,CAAC,CAAC,EAAE,CAAA;QACL,CAAC;QACD,OAAO,IAAI,CAAC,WAAW,CAAA;IACxB,CAAC;IAED;;;;;;;OAOG;IACH,KAAK,CAAC,KAAK,CAAC,QAAkB;QAC7B,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,aAAa,EAAE,CAAA;QAC1C,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,CAAC,CAAA;QAC1D,MAAM,MAAM,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,CAAA;QAClD,MAAM,IAAI,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,CAAA;QAChD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjC,MAAM,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAA;YAChC,IAAI,CAAC,CAAC,CAAC,GAAG,EAAE,CAAA;QACb,CAAC;QAED,MAAM,KAAK,GAA+B;YACzC,SAAS,EAAE,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;YACjE,cAAc,EAAE,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;SACpE,CAAA;QAED,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,CAAA;QACvC,MAAM,YAAY,GAAG,MAAM,CAAC,MAAM,CAAA;QAClC,IAAI,CAAC,YAAY;YAAE,MAAM,IAAI,KAAK,CAAC,6CAA6C,CAAC,CAAA;QACjF,MAAM,IAAI,GAAG,YAAY,CAAC,IAAoB,CAAA;QAC9C,MAAM,CAAC,EAAE,AAAD,EAAG,SAAS,CAAC,GAAG,YAAY,CAAC,IAAyC,CAAA;QAE9E,MAAM,MAAM,GAAe,EAAE,CAAA;QAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjC,MAAM,GAAG,GAAa,IAAI,KAAK,CAAC,SAAS,CAAC,CAAA;YAC1C,MAAM,IAAI,GAAG,CAAC,GAAG,SAAS,CAAA;YAC1B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE;gBAAE,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,GAAG,CAAC,CAAE,CAAA;YAC5D,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;QACjB,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,CAAA;IAC7B,CAAC;CACD"}
+{"version":3,"file":"onnx-runner.js","sourceRoot":"","sources":["../onnx-runner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,QAAQ,IAAI,EAAE,EAAE,MAAM,SAAS,CAAA;AACxC,OAAO,GAAG,MAAM,kBAAkB,CAAA;AAElC,OAAO,EAAE,kBAAkB,EAAE,MAAM,uBAAuB,CAAA;AAc1D,8FAA8F;AAC9F,MAAM,CAAC,MAAM,qBAAqB,GAAG,GAAG,CAAA;AASxC,MAAM,OAAO,UAAU;IAMJ;IACA;IANV,OAAO,GAAgC,IAAI,CAAA;IAC3C,WAAW,GAAyC,IAAI,CAAA;IAChD,WAAW,CAAQ;IAEnC,YACkB,SAAiB,EACjB,UAA6B,EAC9C,IAAoB;QAFH,cAAS,GAAT,SAAS,CAAQ;QACjB,eAAU,GAAV,UAAU,CAAmB;QAG9C,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,WAAW,IAAI,qBAAqB,CAAA;IAC7D,CAAC;IAED,oEAAoE;IACpE,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,SAAiB,EAAE,OAAuB,EAAE;QAC/D,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,SAAS,EAAE,IAAI,EAAE,IAAI,CAAC,CAAA;QACpD,IAAI,IAAI,CAAC,MAAM;YAAE,MAAM,MAAM,CAAC,aAAa,EAAE,CAAA;QAC7C,OAAO,MAAM,CAAA;IACd,CAAC;IAED,6CAA6C;IAC7C,MAAM,CAAC,KAAK,CAAC,SAAS,CAAC,UAAsB,EAAE,OAAuB,EAAE;QACvE,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,SAAS,EAAE,UAAU,EAAE,IAAI,CAAC,CAAA;QAC1D,IAAI,IAAI,CAAC,MAAM;YAAE,MAAM,MAAM,CAAC,aAAa,EAAE,CAAA;QAC7C,OAAO,MAAM,CAAA;IACd,CAAC;IAEO,KAAK,CAAC,aAAa;QAC1B,IAAI,IAAI,CAAC,OAAO;YAAE,OAAO,IAAI,CAAC,OAAO,CAAA;QACrC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC;YACvB,IAAI,CAAC,WAAW,GAAG,CAAC,KAAK,IAAI,EAAE;gBAC9B,MAAM,KAAK,GAAG,IAAI,CAAC,UAAU,IAAI,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAA;gBAClF,MAAM,OAAO,GAAG,MAAM,GAAG,CAAC,gBAAgB,CAAC,MAAM,CAAC,KAAK,EAAE;oBACxD,kBAAkB,EAAE,CAAC,KAAK,CAAC;oBAC3B,sBAAsB,EAAE,KAAK;iBAC7B,CAAC,CAAA;gBACF,IAAI,CAAC,OAAO,GAAG,OAAO,CAAA;gBACtB,OAAO,OAAO,CAAA;YACf,CAAC,CAAC,EAAE,CAAA;QACL,CAAC;QACD,OAAO,IAAI,CAAC,WAAW,CAAA;IACxB,CAAC;IAED;;;;;;;;;;;OAWG;IACH,KAAK,CAAC,KAAK,CACV,QAAkB,EAClB,MAA8F;QAE9F,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,aAAa,EAAE,CAAA;QAC1C,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,CAAC,CAAA;QAC1D,MAAM,MAAM,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,CAAA;QAClD,MAAM,IAAI,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,WAAW,CAAC,CAAA;QAChD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjC,MAAM,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAA;YAChC,IAAI,CAAC,CAAC,CAAC,GAAG,EAAE,CAAA;QACb,CAAC;QAED,MAAM,KAAK,GAA+B;YACzC,SAAS,EAAE,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;YACjE,cAAc,EAAE,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;SACpE,CAAA;QAED,IAAI,MAAM,EAAE,CAAC;YACZ,MAAM,GAAG,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,MAAM,IAAI,CAAC,CAAA;YAC3C,MAAM,EAAE,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,WAAW,GAAG,GAAG,CAAC,CAAA;YACnD,MAAM,EAAE,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,WAAW,CAAC,CAAA;YAC7C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACjC,EAAE,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,CAAA;gBACjC,MAAM,GAAG,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAA;gBAC9B,IAAI,GAAG;oBAAE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE;wBAAE,EAAE,CAAC,CAAC,GAAG,GAAG,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,CAAA;YACrE,CAAC;YACD,KAAK,CAAC,eAAe,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,SAAS,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,EAAE,GAAG,CAAC,CAAC,CAAA;YACjF,KAAK,CAAC,iBAAiB,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,SAAS,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC,CAAA;QAC/E,CAAC;aAAM,IAAI,OAAO,CAAC,UAAU,CAAC,QAAQ,CAAC,iBAAiB,CAAC,EAAE,CAAC;YAC3D,6FAA6F;YAC7F,0FAA0F;YAC1F,0EAA0E;YAC1E,KAAK,CAAC,eAAe,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,SAAS,EAAE,IAAI,YAAY,CAAC,IAAI,CAAC,WAAW,GAAG,kBAAkB,CAAC,EAAE;gBAC1G,CAAC;gBACD,IAAI,CAAC,WAAW;gBAChB,kBAAkB;aAClB,CAAC,CAAA;YACF,KAAK,CAAC,iBAAiB,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,SAAS,EAAE,IAAI,YAAY,CAAC,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC,CAAA;QAC/G,CAAC;QAED,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,CAAA;QACvC,MAAM,YAAY,GAAG,MAAM,CAAC,MAAM,CAAA;QAClC,IAAI,CAAC,YAAY;YAAE,MAAM,IAAI,KAAK,CAAC,6CAA6C,CAAC,CAAA;QACjF,MAAM,IAAI,GAAG,YAAY,CAAC,IAAoB,CAAA;QAC9C,MAAM,CAAC,EAAE,AAAD,EAAG,SAAS,CAAC,GAAG,YAAY,CAAC,IAAyC,CAAA;QAE9E,MAAM,MAAM,GAAe,EAAE,CAAA;QAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjC,MAAM,GAAG,GAAa,IAAI,KAAK,CAAC,SAAS,CAAC,CAAA;YAC1C,MAAM,IAAI,GAAG,CAAC,GAAG,SAAS,CAAA;YAC1B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE;gBAAE,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,GAAG,CAAC,CAAE,CAAA;YAC5D,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;QACjB,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,CAAA;IAC7B,CAAC;CACD"}

package/out/postcode-anchor.d.ts

@@ -0,0 +1,117 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Postcode anchor — the first member of the "anchor-based parsing" family (Direction D, #240). See
+ *   `docs/articles/plan/2026-06-03-anchor-based-parsing.md`.
+ *
+ *   A postcode is the most information-dense token in an address: a hierarchical geo-encoding that
+ *   places a query on Earth far more cheaply than the rest of the parse. This module lifts the
+ *   postcode out of the BIO sequence-labelling problem and treats it as a structured anchor. It
+ *   runs the same per-country shape regexes the decoder repair pass uses ({@link collectMatches}),
+ *   resolves each shaped span against a postcode gazetteer, and returns a SOFT signal: a country
+ *   posterior plus a calibrated confidence. It never decides a postcode's identity on its own — it
+ *   reports "this string is (or is not) a real postcode, in these countries, near here", and leaves
+ *   the parser to weigh that against the surrounding tokens.
+ *
+ *   Two design rules carried from the DeepSeek consult
+ *   (`.agents/skills/deepseek-consult/ds-pc-turn{1,2}-postcode-anchor.txt`):
+ *
+ *   - The country posterior is UNIFORM over the countries a string actually exists in. We never weight
+ *       by per-country postcode volume, because that skews "75001" toward whichever country owns
+ *       more 5-digit codes — the exact bias the anchor exists to avoid. Disambiguation is the
+ *       parser's job, using script, city tokens, and user locale.
+ *   - Confidence combines gazetteer MEMBERSHIP with country AMBIGUITY. A string that matches a postcode
+ *       regex but exists in no gazetteer (a bare `27`, or a 5-digit house number that is not a real
+ *       code) gets confidence 0, so the parser treats it as a house number. A real-but-ambiguous
+ *       code (`75001` in FR and US) gets moderate confidence. A real, single-country code gets
+ *       1.0.
+ */
+/** A gazetteer hit for a postcode string. `lat`/`lon` of 0 means "known postcode, no centroid yet". */
+export interface PostcodePlace {
+    country: string;
+    lat: number;
+    lon: number;
+}
+/**
+ * The minimal surface the anchor needs from a gazetteer. Implementations: an in-memory fake (tests)
+ * or a SQLite-backed lookup over the `postalcode-*.db` shards (`@mailwoman/resolver-wof-sqlite`).
+ * Keeping the seam this narrow lets a future FST/WASM resolver drop in without touching the anchor
+ * logic.
+ */
+export interface PostcodeResolver {
+    /** Exact-match lookup of a normalized postcode string across every country shard. */
+    lookup(postcode: string): PostcodePlace[];
+}
+export interface PostcodeAnchor {
+    /** The shaped substring as it appeared in the raw text, with char offsets. */
+    span: {
+        text: string;
+        start: number;
+        end: number;
+    };
+    /** The normalized form actually queried (uppercased, `D-` prefix stripped, whitespace collapsed). */
+    normalized: string;
+    /** Coordinate-bearing gazetteer hits — best-effort centroid(s), one representative per country. */
+    candidates: PostcodePlace[];
+    /**
+     * Uniform distribution over the countries the postcode exists in (membership,
+     * coordinate-independent).
+     */
+    posterior: Record<string, number>;
+    /** `1 - normalizedEntropy(posterior)` when the postcode exists; `0` when it is in no gazetteer. */
+    confidence: number;
+    /**
+     * `exact` — the string is a real postcode; `outward` — a GB unit (`SO4 3RX`) resolved to its
+     * outward district (`SO4`), the granularity the GB gazetteer is aggregated at (no penalty — it is
+     * a real, confident GB match); `fuzzy` — only an edit-distance-1 variant exists (a likely typo /
+     * OCR slip), so the confidence carries a penalty; `none` — in no gazetteer.
+     */
+    matchType: "exact" | "outward" | "fuzzy" | "none";
+    /**
+     * Structural house-number prior in [0, 1]: `1` for a code that cannot be a house number, and
+     * below `1` for a digit-only code sharing its comma-delimited segment with a street word (so it
+     * reads as a house number rather than a postcode). Already folded into {@link confidence}; exposed
+     * so a consumer can rank competing spans, or see why one was down-weighted, without re-deriving
+     * it.
+     */
+    positionFactor: number;
+}
+export interface ExtractPostcodeAnchorsOpts {
+    /**
+     * When an exact lookup finds nothing, retry Damerau–Levenshtein ≤1 variants to absorb typos and
+     * OCR slips (`75OO8` → `75008`). Off by default so existing callers keep exact-match behaviour.
+     */
+    fuzzy?: boolean;
+}
+/**
+ * Class-aware edit-distance-1 variants of a postcode string: deletions, same-class substitutions
+ * (digit↔digit, letter↔letter), same-class insertions, and adjacent transpositions. Restricting
+ * substitutions/insertions to the character's class mirrors how humans mistype or OCR a postcode (a
+ * digit becomes another digit, not a letter) and keeps the candidate set small.
+ */
+export declare function editDistance1Variants(s: string): string[];
+/**
+ * Normalize a shaped span to the canonical gazetteer key: uppercase, collapse internal whitespace
+ * to a single space, and strip the German `D-` courtesy prefix (the shards store `68161`, not
+ * `D-68161`).
+ */
+export declare function normalizePostcode(raw: string): string;
+/**
+ * The GB outward code of a normalized unit postcode — the part before the space when the inward
+ * half is `\d[A-Z]{2}` (`SO4 3RX` → `SO4`). The GB gazetteer is aggregated to outward codes (2.7M
+ * units is too large + too fine for an anchor), so the extractor retries the outward code when a
+ * full GB unit misses. Returns `null` for any string that isn't a GB unit postcode (so it never
+ * fires elsewhere).
+ */
+export declare function gbOutwardCode(normalized: string): string | null;
+/**
+ * Extract postcode anchors from raw text. For each postcode-shaped span, resolve it against the
+ * gazetteer and emit a soft anchor (country posterior + confidence). Spans that match a shape but
+ * exist in no gazetteer are still returned, with an empty posterior and confidence 0 — an explicit
+ * "looks like a postcode, but isn't one" so the caller can see the extractor fired and chose not to
+ * anchor.
+ */
+export declare function extractPostcodeAnchors(text: string, resolver: PostcodeResolver, opts?: ExtractPostcodeAnchorsOpts): PostcodeAnchor[];
+//# sourceMappingURL=postcode-anchor.d.ts.map

package/out/postcode-anchor.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"postcode-anchor.d.ts","sourceRoot":"","sources":["../postcode-anchor.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AAQH,uGAAuG;AACvG,MAAM,WAAW,aAAa;IAC7B,OAAO,EAAE,MAAM,CAAA;IACf,GAAG,EAAE,MAAM,CAAA;IACX,GAAG,EAAE,MAAM,CAAA;CACX;AAED;;;;;GAKG;AACH,MAAM,WAAW,gBAAgB;IAChC,qFAAqF;IACrF,MAAM,CAAC,QAAQ,EAAE,MAAM,GAAG,aAAa,EAAE,CAAA;CACzC;AAED,MAAM,WAAW,cAAc;IAC9B,8EAA8E;IAC9E,IAAI,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAA;IAClD,qGAAqG;IACrG,UAAU,EAAE,MAAM,CAAA;IAClB,mGAAmG;IACnG,UAAU,EAAE,aAAa,EAAE,CAAA;IAC3B;;;OAGG;IACH,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IACjC,mGAAmG;IACnG,UAAU,EAAE,MAAM,CAAA;IAClB;;;;;OAKG;IACH,SAAS,EAAE,OAAO,GAAG,SAAS,GAAG,OAAO,GAAG,MAAM,CAAA;IACjD;;;;;;OAMG;IACH,cAAc,EAAE,MAAM,CAAA;CACtB;AAED,MAAM,WAAW,0BAA0B;IAC1C;;;OAGG;IACH,KAAK,CAAC,EAAE,OAAO,CAAA;CACf;AAWD;;;;;GAKG;AACH,wBAAgB,qBAAqB,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,EAAE,CAczD;AAED;;;;GAIG;AACH,wBAAgB,iBAAiB,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,CAKrD;AAED;;;;;;GAMG;AACH,wBAAgB,aAAa,CAAC,UAAU,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CAI/D;AAuGD;;;;;;GAMG;AACH,wBAAgB,sBAAsB,CACrC,IAAI,EAAE,MAAM,EACZ,QAAQ,EAAE,gBAAgB,EAC1B,IAAI,GAAE,0BAA+B,GACnC,cAAc,EAAE,CAmElB"}