@mailwoman/corpus 3.0.0 → 4.0.0

package/out/src/synthesize-house-venue.js

@@ -0,0 +1,147 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   House-number + venue + street co-occurrence synthesizer. The v0.6.3 corrective shard.
+ *
+ *   The v0.6.2 step-20K diagnostic showed that adding synth-no-street counter-distribution regressed
+ *   house_number recall by ~4-5pp. DeepSeek's turn-8 root-cause:
+ *
+ *   1. Direct: `5th Avenue Theatre`-style adversarial venues teach the model that tokens like "5th"
+ *        belong to venues, not house_numbers. (Fixed in `synthesize-no-street.ts` by removing
+ *        digit+ordinal venue patterns.)
+ *   2. Distributional dilution: synth-no-street adds 122K rows where house_number is absent. The model's
+ *        training distribution shifts toward "house_number is rare," and it under-emits the tag at
+ *        inference.
+ *
+ *   This synthesizer fixes #2 directly. Each emitted row has ALL of: house_number, street, venue,
+ *   locality, region, postcode — a counter-example to "house_number is rare." Used as a companion
+ *   shard to synth-no-street; the v0.6.3 config weights synth-no-street at 0.5 and
+ *   synth-house-venue at 1.0 to recover the lost house_number signal.
+ *
+ *   Real-world shape: business cards, mailing labels, store directories — `"123 Main St, Sunrise
+ *   Bakery, Springfield, IL 62701"` is a perfectly ordinary address form.
+ *
+ *   Venue pool: PLAIN_VENUES from `synthesize-no-street.ts` (re-exported here). Adversarial venues
+ *   are deliberately NOT used here — the point is to teach co-occurrence, not to re-introduce
+ *   decompose-mode pressure.
+ */
+// ---------------------------------------------------------------------------------------------
+// Venue pool — PLAIN, no street-typing tokens. The point of this shard is to teach
+// house_number + venue coexistence, NOT to re-introduce decompose-mode pressure.
+// Adversarial venue names live in `synthesize-no-street.ts`.
+// ---------------------------------------------------------------------------------------------
+const PLAIN_VENUES = [
+    "Bob's Pizza",
+    "Acme Corporation",
+    "Joe's Diner",
+    "Sunrise Bakery",
+    "Maple Leaf Cafe",
+    "Riverside Garden Center",
+    "Tech Solutions Inc",
+    "Pacific Industries",
+    "Atlantic Holdings",
+    "Stellar Consulting",
+    "Greenfield Partners",
+    "Mountain View Studio",
+    "The Daily Grind",
+    "Sunset Bistro",
+    "Harvest Moon Florist",
+    "Iron Forge Brewing",
+    "Crescent City Bookstore",
+    "Lighthouse Insurance Group",
+    "Pinecrest Veterinary",
+    "Westwood Realty",
+];
+// ---------------------------------------------------------------------------------------------
+// Fallback street pool for tuples that didn't carry a `street` field. Plain street names
+// without typing-token ambiguity.
+// ---------------------------------------------------------------------------------------------
+const FALLBACK_STREETS = [
+    "Main St",
+    "Oak Ave",
+    "Pine Rd",
+    "Elm Dr",
+    "Cedar Ln",
+    "Maple Blvd",
+    "Birch Ct",
+    "Walnut Pl",
+    "Cherry Way",
+    "Spruce St",
+    "Park Ave",
+    "Lake Dr",
+    "Hill Rd",
+    "River Ln",
+    "Forest Blvd",
+];
+// ---------------------------------------------------------------------------------------------
+// House-number generator
+// ---------------------------------------------------------------------------------------------
+function randomHouseNumber(random) {
+    // Generate a plain numeric house number 1-9999. No fractions/ranges — those land in
+    // `data/eval/falsehoods/numbers.jsonl` as known edge cases, not training material.
+    const digits = Math.floor(random() * 4) + 1;
+    const max = Math.pow(10, digits);
+    const n = Math.floor(random() * max) + 1;
+    return String(n);
+}
+function pick(arr, random) {
+    return arr[Math.floor(random() * arr.length)];
+}
+function countryToLocale(country) {
+    switch (country) {
+        case "US":
+            return "en-US";
+        case "CA":
+            return "en-CA";
+        case "GB":
+            return "en-GB";
+        case "AU":
+            return "en-AU";
+        case "FR":
+            return "fr-FR";
+        case "DE":
+            return "de-DE";
+        default:
+            return "en-US";
+    }
+}
+// ---------------------------------------------------------------------------------------------
+// Synthesis
+// ---------------------------------------------------------------------------------------------
+export function synthesizeHouseVenueRow(base, opts = {}) {
+    const random = opts.random ?? Math.random;
+    const locale = countryToLocale(base.country);
+    const template = opts.forceTemplate ?? (random() < 0.5 ? "venue-after-street" : "venue-before-street");
+    const venue = pick(PLAIN_VENUES, random);
+    const street = base.street ?? pick(FALLBACK_STREETS, random);
+    const houseNumber = base.houseNumber ?? randomHouseNumber(random);
+    const components = {
+        house_number: houseNumber,
+        street,
+        venue,
+        locality: base.locality,
+        region: base.region,
+        postcode: base.postcode,
+    };
+    let raw;
+    switch (template) {
+        case "venue-after-street":
+            raw = `${houseNumber} ${street}, ${venue}, ${base.locality}, ${base.region} ${base.postcode}`;
+            break;
+        case "venue-before-street":
+            raw = `${venue}, ${houseNumber} ${street}, ${base.locality}, ${base.region} ${base.postcode}`;
+            break;
+    }
+    return { raw, components, locale, template };
+}
+/**
+ * Contract: every synthesized row carries BOTH house_number AND venue (the co-occurrence signal
+ * that synth-no-street's distributional shift cost the model). Used by tests + downstream
+ * consumers.
+ */
+export function hasHouseNumberAndVenue(components) {
+    return components.house_number !== undefined && components.venue !== undefined;
+}
+//# sourceMappingURL=synthesize-house-venue.js.map

package/out/src/synthesize-house-venue.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"synthesize-house-venue.js","sourceRoot":"","sources":["../../src/synthesize-house-venue.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AA6BH,gGAAgG;AAChG,mFAAmF;AACnF,iFAAiF;AACjF,6DAA6D;AAC7D,gGAAgG;AAEhG,MAAM,YAAY,GAA0B;IAC3C,aAAa;IACb,kBAAkB;IAClB,aAAa;IACb,gBAAgB;IAChB,iBAAiB;IACjB,yBAAyB;IACzB,oBAAoB;IACpB,oBAAoB;IACpB,mBAAmB;IACnB,oBAAoB;IACpB,qBAAqB;IACrB,sBAAsB;IACtB,iBAAiB;IACjB,eAAe;IACf,sBAAsB;IACtB,oBAAoB;IACpB,yBAAyB;IACzB,4BAA4B;IAC5B,sBAAsB;IACtB,iBAAiB;CACjB,CAAA;AAED,gGAAgG;AAChG,yFAAyF;AACzF,kCAAkC;AAClC,gGAAgG;AAEhG,MAAM,gBAAgB,GAA0B;IAC/C,SAAS;IACT,SAAS;IACT,SAAS;IACT,QAAQ;IACR,UAAU;IACV,YAAY;IACZ,UAAU;IACV,WAAW;IACX,YAAY;IACZ,WAAW;IACX,UAAU;IACV,SAAS;IACT,SAAS;IACT,UAAU;IACV,aAAa;CACb,CAAA;AAED,gGAAgG;AAChG,yBAAyB;AACzB,gGAAgG;AAEhG,SAAS,iBAAiB,CAAC,MAAoB;IAC9C,oFAAoF;IACpF,mFAAmF;IACnF,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAA;IAC3C,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,EAAE,EAAE,MAAM,CAAC,CAAA;IAChC,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,GAAG,CAAC,CAAA;IACxC,OAAO,MAAM,CAAC,CAAC,CAAC,CAAA;AACjB,CAAC;AAED,SAAS,IAAI,CAAI,GAAqB,EAAE,MAAoB;IAC3D,OAAO,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,MAAM,CAAC,CAAE,CAAA;AAC/C,CAAC;AAED,SAAS,eAAe,CAAC,OAAe;IACvC,QAAQ,OAAO,EAAE,CAAC;QACjB,KAAK,IAAI;YACR,OAAO,OAAO,CAAA;QACf,KAAK,IAAI;YACR,OAAO,OAAO,CAAA;QACf,KAAK,IAAI;YACR,OAAO,OAAO,CAAA;QACf,KAAK,IAAI;YACR,OAAO,OAAO,CAAA;QACf,KAAK,IAAI;YACR,OAAO,OAAO,CAAA;QACf,KAAK,IAAI;YACR,OAAO,OAAO,CAAA;QACf;YACC,OAAO,OAAO,CAAA;IAChB,CAAC;AACF,CAAC;AAED,gGAAgG;AAChG,YAAY;AACZ,gGAAgG;AAEhG,MAAM,UAAU,uBAAuB,CACtC,IAAyB,EACzB,OAAgC,EAAE;IAElC,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,MAAM,CAAA;IACzC,MAAM,MAAM,GAAG,eAAe,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;IAC5C,MAAM,QAAQ,GAAG,IAAI,CAAC,aAAa,IAAI,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,oBAAoB,CAAC,CAAC,CAAC,qBAAqB,CAAC,CAAA;IAEtG,MAAM,KAAK,GAAG,IAAI,CAAC,YAAY,EAAE,MAAM,CAAC,CAAA;IACxC,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,gBAAgB,EAAE,MAAM,CAAC,CAAA;IAC5D,MAAM,WAAW,GAAG,IAAI,CAAC,WAAW,IAAI,iBAAiB,CAAC,MAAM,CAAC,CAAA;IAEjE,MAAM,UAAU,GAA+B;QAC9C,YAAY,EAAE,WAAW;QACzB,MAAM;QACN,KAAK;QACL,QAAQ,EAAE,IAAI,CAAC,QAAQ;QACvB,MAAM,EAAE,IAAI,CAAC,MAAM;QACnB,QAAQ,EAAE,IAAI,CAAC,QAAQ;KACvB,CAAA;IAED,IAAI,GAAW,CAAA;IACf,QAAQ,QAAQ,EAAE,CAAC;QAClB,KAAK,oBAAoB;YACxB,GAAG,GAAG,GAAG,WAAW,IAAI,MAAM,KAAK,KAAK,KAAK,IAAI,CAAC,QAAQ,KAAK,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAA;YAC7F,MAAK;QACN,KAAK,qBAAqB;YACzB,GAAG,GAAG,GAAG,KAAK,KAAK,WAAW,IAAI,MAAM,KAAK,IAAI,CAAC,QAAQ,KAAK,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAA;YAC7F,MAAK;IACP,CAAC;IAED,OAAO,EAAE,GAAG,EAAE,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAA;AAC7C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,sBAAsB,CAAC,UAAsC;IAC5E,OAAO,UAAU,CAAC,YAAY,KAAK,SAAS,IAAI,UAAU,CAAC,KAAK,KAAK,SAAS,CAAA;AAC/E,CAAC"}

package/out/src/synthesize-intersection.d.ts

@@ -0,0 +1,48 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Intersection synthesizer — v0.7 coverage fix (night-3, DeepSeek-decided).
+ *
+ *   The 2026-05-29 harness diagnostic found the neural model emits
+ *   `intersection_a`/`intersection_b` with ~0.0001 probability on canonical
+ *   intersections ("Broadway & W 42nd St") — it never learned the tags, because
+ *   the corpus has NO intersection training signal (no generator, and real-data
+ *   adapters don't emit intersection-formatted rows). Intersections are 65 of
+ *   the 376 harness assertions (17%), all 0% neural. This generator produces the
+ *   missing signal as a small targeted supplement shard (synthesis-as-supplement
+ *   discipline: weight < 0.25, one-and-done).
+ *
+ *   Output is a `CanonicalRow` ({raw, components}); the corpus aligner turns it
+ *   into BIO labels (B-/I-intersection_a, O on the connector, B-/I-intersection_b).
+ *   Surface forms of both streets MUST occur verbatim in `raw` so alignment lands.
+ *
+ *   US-idiomatic only (the harness intersection cases are US: "X & Y, City, ST ZIP").
+ */
+import type { CanonicalRow } from "./types.js";
+export interface IntersectionBaseTuple {
+    locality: string;
+    region: string;
+    /** ZIP — optional; ~30% of synthetic intersections omit it (idiomatic). */
+    postcode?: string;
+    country: string;
+}
+export interface SynthesizedIntersectionRow {
+    raw: string;
+    components: CanonicalRow["components"];
+    locale: string;
+}
+export interface IntersectionSynthesisOpts {
+    random?: () => number;
+}
+/**
+ * Synthesize one US intersection row. Returns null on the rare degenerate case where the two
+ * streets collide (so alignment never has two identical surface forms to disambiguate).
+ */
+export declare function synthesizeIntersectionRow(base: IntersectionBaseTuple, opts?: IntersectionSynthesisOpts): SynthesizedIntersectionRow | null;
+/** A small built-in US city/region/zip pool for standalone shard generation + tests. */
+export declare const DEFAULT_US_BASES: ReadonlyArray<IntersectionBaseTuple>;
+/** Generate `count` intersection rows over the provided bases (round-robin). */
+export declare function generateIntersectionRows(count: number, bases?: ReadonlyArray<IntersectionBaseTuple>, opts?: IntersectionSynthesisOpts): SynthesizedIntersectionRow[];
+//# sourceMappingURL=synthesize-intersection.d.ts.map

package/out/src/synthesize-intersection.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"synthesize-intersection.d.ts","sourceRoot":"","sources":["../../src/synthesize-intersection.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,YAAY,CAAA;AAyC9C,MAAM,WAAW,qBAAqB;IACrC,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,EAAE,MAAM,CAAA;IACd,2EAA2E;IAC3E,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,OAAO,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,0BAA0B;IAC1C,GAAG,EAAE,MAAM,CAAA;IACX,UAAU,EAAE,YAAY,CAAC,YAAY,CAAC,CAAA;IACtC,MAAM,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,yBAAyB;IACzC,MAAM,CAAC,EAAE,MAAM,MAAM,CAAA;CACrB;AAkBD;;;GAGG;AACH,wBAAgB,yBAAyB,CACxC,IAAI,EAAE,qBAAqB,EAC3B,IAAI,GAAE,yBAA8B,GAClC,0BAA0B,GAAG,IAAI,CAoCnC;AAED,wFAAwF;AACxF,eAAO,MAAM,gBAAgB,EAAE,aAAa,CAAC,qBAAqB,CAWjE,CAAA;AAED,gFAAgF;AAChF,wBAAgB,wBAAwB,CACvC,KAAK,EAAE,MAAM,EACb,KAAK,GAAE,aAAa,CAAC,qBAAqB,CAAoB,EAC9D,IAAI,GAAE,yBAA8B,GAClC,0BAA0B,EAAE,CAU9B"}

package/out/src/synthesize-intersection.js

@@ -0,0 +1,138 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Intersection synthesizer — v0.7 coverage fix (night-3, DeepSeek-decided).
+ *
+ *   The 2026-05-29 harness diagnostic found the neural model emits
+ *   `intersection_a`/`intersection_b` with ~0.0001 probability on canonical
+ *   intersections ("Broadway & W 42nd St") — it never learned the tags, because
+ *   the corpus has NO intersection training signal (no generator, and real-data
+ *   adapters don't emit intersection-formatted rows). Intersections are 65 of
+ *   the 376 harness assertions (17%), all 0% neural. This generator produces the
+ *   missing signal as a small targeted supplement shard (synthesis-as-supplement
+ *   discipline: weight < 0.25, one-and-done).
+ *
+ *   Output is a `CanonicalRow` ({raw, components}); the corpus aligner turns it
+ *   into BIO labels (B-/I-intersection_a, O on the connector, B-/I-intersection_b).
+ *   Surface forms of both streets MUST occur verbatim in `raw` so alignment lands.
+ *
+ *   US-idiomatic only (the harness intersection cases are US: "X & Y, City, ST ZIP").
+ */
+/** Street name cores (no suffix) — proper-noun streets that often appear bare. */
+const STREET_CORES = [
+    "Main",
+    "Oak",
+    "Elm",
+    "Maple",
+    "Pine",
+    "Cedar",
+    "Park",
+    "Lake",
+    "Hill",
+    "Washington",
+    "Lincoln",
+    "Jefferson",
+    "Madison",
+    "Franklin",
+    "Market",
+    "Broad",
+    "Church",
+    "Mill",
+    "Highland",
+    "Sunset",
+    "Union",
+    "Spring",
+];
+/** Bare proper-noun streets that idiomatically take NO suffix. */
+const BARE_NAMES = ["Broadway", "Wall", "Bourbon", "Esplanade", "Riverside", "Lakeshore"];
+const ORDINALS = ["1st", "2nd", "3rd", "4th", "5th", "6th", "7th", "8th", "9th", "10th", "42nd", "23rd", "34th"];
+const SUFFIXES = ["St", "Ave", "Blvd", "Rd", "Dr", "Ln", "Way", "Pl", "Ct", "Pkwy", "Ter", "Cir"];
+const DIRECTIONALS = ["N", "S", "E", "W", "NE", "NW", "SE", "SW"];
+/** Connectors between the two streets. Whitespace-padded forms keep tokens clean for alignment.
+ *  `@` added in v0.7.2 — the harness uses it ("Main St @ Second Ave") and v0.7.1 had never seen it. */
+const CONNECTORS = [" & ", " and ", " at ", " / ", " @ "];
+function pick(arr, random) {
+    return arr[Math.floor(random() * arr.length)];
+}
+/** Build a single street surface form, e.g. "W 42nd St", "Broadway", "Main St", "N Oak Ave". */
+function buildStreetName(random) {
+    // ~20% bare proper-noun street (no suffix), else directional? + core/ordinal + suffix.
+    if (random() < 0.2)
+        return pick(BARE_NAMES, random);
+    const parts = [];
+    if (random() < 0.35)
+        parts.push(pick(DIRECTIONALS, random));
+    parts.push(random() < 0.45 ? pick(ORDINALS, random) : pick(STREET_CORES, random));
+    parts.push(pick(SUFFIXES, random));
+    return parts.join(" ");
+}
+/**
+ * Synthesize one US intersection row. Returns null on the rare degenerate case where the two
+ * streets collide (so alignment never has two identical surface forms to disambiguate).
+ */
+export function synthesizeIntersectionRow(base, opts = {}) {
+    const random = opts.random ?? Math.random;
+    if (base.country !== "US")
+        return null;
+    const a = buildStreetName(random);
+    let b = buildStreetName(random);
+    // Ensure distinct surface forms (and not a substring of each other — alignment needs unambiguous spans).
+    let tries = 0;
+    while ((b === a || a.includes(b) || b.includes(a)) && tries++ < 8)
+        b = buildStreetName(random);
+    if (b === a || a.includes(b) || b.includes(a))
+        return null;
+    const connector = pick(CONNECTORS, random);
+    // "corner of" prefix variant (~20%) — still labels the two streets identically.
+    const cornerPrefix = random() < 0.2 ? "corner of " : "";
+    const components = { intersection_a: a, intersection_b: b };
+    // v0.7.2: ~60% BARE (no locality tail). v0.7.1 always appended ", City, ST", so the model learned
+    // to read post-intersection text as a locality and fumbled the harness's bare "X & Y" cases
+    // (mislabeling the second street as a locality). Match the eval distribution.
+    const bare = random() < 0.6;
+    let raw;
+    if (bare) {
+        raw = `${cornerPrefix}${a}${connector}${b}`;
+    }
+    else {
+        const includePostcode = base.postcode != null && random() < 0.7;
+        const tail = includePostcode
+            ? `, ${base.locality}, ${base.region} ${base.postcode}`
+            : `, ${base.locality}, ${base.region}`;
+        raw = `${cornerPrefix}${a}${connector}${b}${tail}`;
+        components.locality = base.locality;
+        components.region = base.region;
+        if (includePostcode)
+            components.postcode = base.postcode;
+    }
+    return { raw, components, locale: "en-US" };
+}
+/** A small built-in US city/region/zip pool for standalone shard generation + tests. */
+export const DEFAULT_US_BASES = [
+    { locality: "New York", region: "NY", postcode: "10036", country: "US" },
+    { locality: "Chicago", region: "IL", postcode: "60613", country: "US" },
+    { locality: "Los Angeles", region: "CA", postcode: "90012", country: "US" },
+    { locality: "Seattle", region: "WA", postcode: "98109", country: "US" },
+    { locality: "Austin", region: "TX", postcode: "78701", country: "US" },
+    { locality: "Portland", region: "OR", postcode: "97205", country: "US" },
+    { locality: "Denver", region: "CO", postcode: "80202", country: "US" },
+    { locality: "Boston", region: "MA", postcode: "02116", country: "US" },
+    { locality: "Miami", region: "FL", postcode: "33130", country: "US" },
+    { locality: "Atlanta", region: "GA", postcode: "30303", country: "US" },
+];
+/** Generate `count` intersection rows over the provided bases (round-robin). */
+export function generateIntersectionRows(count, bases = DEFAULT_US_BASES, opts = {}) {
+    const random = opts.random ?? Math.random;
+    const out = [];
+    let guard = 0;
+    while (out.length < count && guard++ < count * 4) {
+        const base = bases[out.length % bases.length];
+        const row = synthesizeIntersectionRow(base, { random });
+        if (row)
+            out.push(row);
+    }
+    return out;
+}
+//# sourceMappingURL=synthesize-intersection.js.map

package/out/src/synthesize-intersection.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"synthesize-intersection.js","sourceRoot":"","sources":["../../src/synthesize-intersection.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAIH,kFAAkF;AAClF,MAAM,YAAY,GAAG;IACpB,MAAM;IACN,KAAK;IACL,KAAK;IACL,OAAO;IACP,MAAM;IACN,OAAO;IACP,MAAM;IACN,MAAM;IACN,MAAM;IACN,YAAY;IACZ,SAAS;IACT,WAAW;IACX,SAAS;IACT,UAAU;IACV,QAAQ;IACR,OAAO;IACP,QAAQ;IACR,MAAM;IACN,UAAU;IACV,QAAQ;IACR,OAAO;IACP,QAAQ;CACC,CAAA;AAEV,kEAAkE;AAClE,MAAM,UAAU,GAAG,CAAC,UAAU,EAAE,MAAM,EAAE,SAAS,EAAE,WAAW,EAAE,WAAW,EAAE,WAAW,CAAU,CAAA;AAElG,MAAM,QAAQ,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAU,CAAA;AAEzH,MAAM,QAAQ,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,CAAU,CAAA;AAE1G,MAAM,YAAY,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAU,CAAA;AAE1E;uGACuG;AACvG,MAAM,UAAU,GAAG,CAAC,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,CAAU,CAAA;AAoBlE,SAAS,IAAI,CAAI,GAAqB,EAAE,MAAoB;IAC3D,OAAO,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,MAAM,CAAC,CAAE,CAAA;AAC/C,CAAC;AAED,gGAAgG;AAChG,SAAS,eAAe,CAAC,MAAoB;IAC5C,uFAAuF;IACvF,IAAI,MAAM,EAAE,GAAG,GAAG;QAAE,OAAO,IAAI,CAAC,UAAU,EAAE,MAAM,CAAC,CAAA;IAEnD,MAAM,KAAK,GAAa,EAAE,CAAA;IAC1B,IAAI,MAAM,EAAE,GAAG,IAAI;QAAE,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAA;IAC3D,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAA;IACjF,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAA;IAClC,OAAO,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;AACvB,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,yBAAyB,CACxC,IAA2B,EAC3B,OAAkC,EAAE;IAEpC,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,MAAM,CAAA;IACzC,IAAI,IAAI,CAAC,OAAO,KAAK,IAAI;QAAE,OAAO,IAAI,CAAA;IAEtC,MAAM,CAAC,GAAG,eAAe,CAAC,MAAM,CAAC,CAAA;IACjC,IAAI,CAAC,GAAG,eAAe,CAAC,MAAM,CAAC,CAAA;IAC/B,yGAAyG;IACzG,IAAI,KAAK,GAAG,CAAC,CAAA;IACb,OAAO,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,IAAI,KAAK,EAAE,GAAG,CAAC;QAAE,CAAC,GAAG,eAAe,CAAC,MAAM,CAAC,CAAA;IAC9F,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC;QAAE,OAAO,IAAI,CAAA;IAE1D,MAAM,SAAS,GAAG,IAAI,CAAC,UAAU,EAAE,MAAM,CAAC,CAAA;IAC1C,gFAAgF;IAChF,MAAM,YAAY,GAAG,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE,CAAA;IAEvD,MAAM,UAAU,GAA+B,EAAE,cAAc,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,EAAE,CAAA;IAEvF,kGAAkG;IAClG,4FAA4F;IAC5F,8EAA8E;IAC9E,MAAM,IAAI,GAAG,MAAM,EAAE,GAAG,GAAG,CAAA;IAC3B,IAAI,GAAW,CAAA;IACf,IAAI,IAAI,EAAE,CAAC;QACV,GAAG,GAAG,GAAG,YAAY,GAAG,CAAC,GAAG,SAAS,GAAG,CAAC,EAAE,CAAA;IAC5C,CAAC;SAAM,CAAC;QACP,MAAM,eAAe,GAAG,IAAI,CAAC,QAAQ,IAAI,IAAI,IAAI,MAAM,EAAE,GAAG,GAAG,CAAA;QAC/D,MAAM,IAAI,GAAG,eAAe;YAC3B,CAAC,CAAC,KAAK,IAAI,CAAC,QAAQ,KAAK,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,QAAQ,EAAE;YACvD,CAAC,CAAC,KAAK,IAAI,CAAC,QAAQ,KAAK,IAAI,CAAC,MAAM,EAAE,CAAA;QACvC,GAAG,GAAG,GAAG,YAAY,GAAG,CAAC,GAAG,SAAS,GAAG,CAAC,GAAG,IAAI,EAAE,CAAA;QAClD,UAAU,CAAC,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAA;QACnC,UAAU,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAA;QAC/B,IAAI,eAAe;YAAE,UAAU,CAAC,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAA;IACzD,CAAC;IAED,OAAO,EAAE,GAAG,EAAE,UAAU,EAAE,MAAM,EAAE,OAAO,EAAE,CAAA;AAC5C,CAAC;AAED,wFAAwF;AACxF,MAAM,CAAC,MAAM,gBAAgB,GAAyC;IACrE,EAAE,QAAQ,EAAE,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACxE,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACvE,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IAC3E,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACvE,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACtE,EAAE,QAAQ,EAAE,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACxE,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACtE,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACtE,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;IACrE,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE;CACvE,CAAA;AAED,gFAAgF;AAChF,MAAM,UAAU,wBAAwB,CACvC,KAAa,EACb,QAA8C,gBAAgB,EAC9D,OAAkC,EAAE;IAEpC,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,MAAM,CAAA;IACzC,MAAM,GAAG,GAAiC,EAAE,CAAA;IAC5C,IAAI,KAAK,GAAG,CAAC,CAAA;IACb,OAAO,GAAG,CAAC,MAAM,GAAG,KAAK,IAAI,KAAK,EAAE,GAAG,KAAK,GAAG,CAAC,EAAE,CAAC;QAClD,MAAM,IAAI,GAAG,KAAK,CAAC,GAAG,CAAC,MAAM,GAAG,KAAK,CAAC,MAAM,CAAE,CAAA;QAC9C,MAAM,GAAG,GAAG,yBAAyB,CAAC,IAAI,EAAE,EAAE,MAAM,EAAE,CAAC,CAAA;QACvD,IAAI,GAAG;YAAE,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACvB,CAAC;IACD,OAAO,GAAG,CAAA;AACX,CAAC"}

package/out/src/synthesize-no-street.d.ts

@@ -0,0 +1,70 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   No-street address synthesizer — the counter-distribution that v0.6.1's synth-street shard is
+ *   missing. Generates BIO-labelable rows where there is NO street, NO house_number, NO
+ *   street_prefix, NO street_suffix, NO intersection — only some subset of {venue, locality,
+ *   region, postcode, country}.
+ *
+ *   Rationale: the [2026-05-28 night-2
+ *   postmortem](../../docs/articles/evals/2026-05-28-night-2-postmortem.md) and the [layer-1
+ *   eval](../../docs/articles/evals/2026-05-28-layer-1-morphology-fst.md) showed that synth-street
+ *   pushed the model into a high-confidence "decompose mode" that leaked into `dependent_locality`.
+ *   Per DeepSeek's turn-2 recipe, the model needs explicit counter-examples: addresses where the
+ *   model should NOT emit street labels. This synthesizer is that source.
+ *
+ *   Six row templates, each producing a {raw, components} pair with no street-side tags:
+ *
+ *   1. **Plain venue + locality + region + postcode** `"Bob's Pizza, Boston, MA 02101"`
+ *   2. **Adversarial venue (containing street-typing words)** `"Wall Street Industries, NY 10005"`,
+ *        `"5th Avenue Theater, Seattle, WA"`, `"Highway 61 Diner, Memphis TN"`. These are the rows
+ *        that v0.6.1's decompose-mode would mis-tag as street_prefix/suffix; explicit negative
+ *        training kills that signal.
+ *   3. **Locality + region + postcode (minimal)** — `"Boston, MA 02101"`
+ *   4. **Locality + region** — `"Boston, MA"`
+ *   5. **Postcode-only** — `"02101"`
+ *   6. **Country-only** — `"United States"`, `"France"` (rare in real data, but the model has seen these
+ *        and should not hallucinate streets on them).
+ *
+ *   Output is a `CanonicalRow` with no street-side components. Alignment will produce BIO labels
+ *   where every token is one of {`B-venue`, `I-venue`, `B-locality`, `I-locality`, `B-region`,
+ *   `B-postcode`, `B-country`, `I-country`, `O`} — explicitly never any street tag. That IS the
+ *   counter-example signal the model is missing.
+ *
+ *   This complements (does not replace) the existing US-base-tuple source used by
+ *   `synthesize-po-box.ts`; the same `NoStreetBaseTuple` shape is consumed.
+ */
+import type { CanonicalRow } from "./types.js";
+export interface NoStreetBaseTuple {
+    locality: string;
+    region: string;
+    postcode: string;
+    country: string;
+}
+export type NoStreetTemplate = "venue-plain" | "venue-adversarial" | "locality-region-postcode" | "locality-region" | "postcode-only" | "country-only";
+export interface NoStreetSynthesisOpts {
+    random?: () => number;
+    /** Override the template selection entirely (used by tests for deterministic coverage). */
+    forceTemplate?: NoStreetTemplate;
+}
+export interface SynthesizedNoStreetRow {
+    raw: string;
+    components: CanonicalRow["components"];
+    locale: string;
+    template: NoStreetTemplate;
+}
+/**
+ * Generate one no-street counter-example row for a base (locality, region, postcode, country)
+ * tuple. Picks a template by weighted random; the venue templates are the load-bearing
+ * counter-distribution against synth-street's decompose-mode pressure.
+ */
+export declare function synthesizeNoStreetRow(base: NoStreetBaseTuple, opts?: NoStreetSynthesisOpts): SynthesizedNoStreetRow | null;
+/**
+ * Convenience: assert at type-level that a synthesized row carries no street-side components. Used
+ * by tests + downstream consumers who want to verify the contract at runtime.
+ */
+export declare const STREET_SIDE_TAGS: readonly ["street", "street_prefix", "street_prefix_particle", "street_suffix", "house_number", "intersection_a", "intersection_b", "unit"];
+export declare function hasAnyStreetSideTag(components: CanonicalRow["components"]): boolean;
+//# sourceMappingURL=synthesize-no-street.d.ts.map

package/out/src/synthesize-no-street.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"synthesize-no-street.d.ts","sourceRoot":"","sources":["../../src/synthesize-no-street.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,YAAY,CAAA;AAM9C,MAAM,WAAW,iBAAiB;IACjC,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,EAAE,MAAM,CAAA;IACd,QAAQ,EAAE,MAAM,CAAA;IAChB,OAAO,EAAE,MAAM,CAAA;CACf;AAED,MAAM,MAAM,gBAAgB,GACzB,aAAa,GACb,mBAAmB,GACnB,0BAA0B,GAC1B,iBAAiB,GACjB,eAAe,GACf,cAAc,CAAA;AAEjB,MAAM,WAAW,qBAAqB;IACrC,MAAM,CAAC,EAAE,MAAM,MAAM,CAAA;IACrB,2FAA2F;IAC3F,aAAa,CAAC,EAAE,gBAAgB,CAAA;CAChC;AAED,MAAM,WAAW,sBAAsB;IACtC,GAAG,EAAE,MAAM,CAAA;IACX,UAAU,EAAE,YAAY,CAAC,YAAY,CAAC,CAAA;IACtC,MAAM,EAAE,MAAM,CAAA;IACd,QAAQ,EAAE,gBAAgB,CAAA;CAC1B;AA4HD;;;;GAIG;AACH,wBAAgB,qBAAqB,CACpC,IAAI,EAAE,iBAAiB,EACvB,IAAI,GAAE,qBAA0B,GAC9B,sBAAsB,GAAG,IAAI,CAqF/B;AAgBD;;;GAGG;AACH,eAAO,MAAM,gBAAgB,6IASnB,CAAA;AAEV,wBAAgB,mBAAmB,CAAC,UAAU,EAAE,YAAY,CAAC,YAAY,CAAC,GAAG,OAAO,CAKnF"}