@mailwoman/corpus 4.16.2 → 5.1.0

package/out/src/shard-recipes/boundary-stress.js

@@ -0,0 +1,77 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `boundary-stress` shard recipe (#375) — the boundary-instability augmentation. Self-generates
+ *   `--count` rows from {@link synthesizeBoundaryStressRow}'s weighted template mix (the v1.7.0,
+ *   DeepSeek-tuned 2026-06-18 composition), aligns each to BIO, and emits a labeled JSONL. The
+ *   lever for the taxonomy's #1 parser family (the boundary-wobble class). Ported from
+ *   scripts/build-boundary-stress-shard.mjs.
+ *
+ *   `synthesizeBoundaryStressRow` is NOT re-exported from the corpus index — imported directly here.
+ */
+import { alignRow } from "../align.js";
+import { synthesizeBoundaryStressRow } from "../synthesize-boundary-stress.js";
+import { makeMulberry32, shardSourceID } from "./scaffold.js";
+// Revised composition (v1.7.0, DeepSeek-tuned 2026-06-18): `bare-locality` ~11% (recover the 84% locality
+// drop on bare "City, STATE" rows WITHOUT becoming a locality-first majority), and
+// house-number-before:after = 7:3 (FR's dominant order is number-BEFORE; 30% after breaks the order-bias
+// shortcut without risking FR hn-before accuracy). The three original non-number shapes keep the bulk.
+// Weights sum to 1.0. Key order is load-bearing — it drives the cumulative thresholds below.
+const WEIGHTS = {
+    "street-eats-affix": 0.22,
+    "comma-less-city-state": 0.22,
+    "fr-prefix": 0.18,
+    "bare-locality": 0.11,
+    "house-number-before-street": 0.189,
+    "house-number-after-street": 0.081,
+};
+const CUM = (() => {
+    let acc = 0;
+    return Object.entries(WEIGHTS).map(([t, w]) => [t, (acc += w)]);
+})();
+function pickTemplate(r) {
+    const x = r();
+    for (const [t, c] of CUM)
+        if (x <= c)
+            return t;
+    return CUM[CUM.length - 1][0];
+}
+export const boundaryStressRecipe = {
+    name: "boundary-stress",
+    description: "Boundary-instability rows (#375): weighted template mix → synthesizeBoundaryStressRow → aligned BIO",
+    mode: "generate",
+    async run(opts, write) {
+        // Emit PRNG: the legacy build-boundary-stress-shard.mjs seeded mulberry32(opts.seed).
+        const random = makeMulberry32(opts.seed);
+        const count = opts.count ?? 20000;
+        let emitted = 0;
+        let skipped = 0;
+        for (let i = 0; i < count; i++) {
+            const row = synthesizeBoundaryStressRow(undefined, { random, forceTemplate: pickTemplate(random) });
+            const country = row.locale.split("-")[1] ?? "US";
+            const source_id = shardSourceID("synth-boundary-stress", { ...row.components, v: String(i) });
+            const canonical = {
+                raw: row.raw,
+                components: row.components,
+                country,
+                locale: row.locale,
+                source: "synth-boundary-stress",
+                source_id,
+                corpus_version: "0.6.0",
+                license: "Synthetic — boundary-stress; derived from public-domain locality/region tuples",
+            };
+            const aligned = alignRow(canonical);
+            if (aligned.kind !== "labeled") {
+                skipped++;
+                continue;
+            }
+            // Match the base corpus parquet schema: flat synth_method / synth_base_id, not a nested `synth`.
+            write(JSON.stringify({ ...aligned.row, synth_method: `boundary-stress:${row.template}`, synth_base_id: null }) + "\n");
+            emitted++;
+        }
+        return { emitted, skipped };
+    },
+};
+//# sourceMappingURL=boundary-stress.js.map

package/out/src/shard-recipes/boundary-stress.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"boundary-stress.js","sourceRoot":"","sources":["../../../src/shard-recipes/boundary-stress.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,OAAO,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAA;AACtC,OAAO,EAA+B,2BAA2B,EAAE,MAAM,kCAAkC,CAAA;AAC3G,OAAO,EAAE,cAAc,EAAoB,aAAa,EAAE,MAAM,eAAe,CAAA;AAE/E,0GAA0G;AAC1G,mFAAmF;AACnF,yGAAyG;AACzG,uGAAuG;AACvG,6FAA6F;AAC7F,MAAM,OAAO,GAA2C;IACvD,mBAAmB,EAAE,IAAI;IACzB,uBAAuB,EAAE,IAAI;IAC7B,WAAW,EAAE,IAAI;IACjB,eAAe,EAAE,IAAI;IACrB,4BAA4B,EAAE,KAAK;IACnC,2BAA2B,EAAE,KAAK;CAClC,CAAA;AACD,MAAM,GAAG,GAA4C,CAAC,GAAG,EAAE;IAC1D,IAAI,GAAG,GAAG,CAAC,CAAA;IAEX,OAAQ,MAAM,CAAC,OAAO,CAAC,OAAO,CAA6C,CAAC,GAAG,CAC9E,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,CAAC,CAAqC,CAC/D,CAAA;AACF,CAAC,CAAC,EAAE,CAAA;AACJ,SAAS,YAAY,CAAC,CAAe;IACpC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAA;IAEb,KAAK,MAAM,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,GAAG;QAAE,IAAI,CAAC,IAAI,CAAC;YAAE,OAAO,CAAC,CAAA;IAE9C,OAAO,GAAG,CAAC,GAAG,CAAC,MAAM,GAAG,CAAC,CAAE,CAAC,CAAC,CAAC,CAAA;AAC/B,CAAC;AAED,MAAM,CAAC,MAAM,oBAAoB,GAAgB;IAChD,IAAI,EAAE,iBAAiB;IACvB,WAAW,EAAE,qGAAqG;IAClH,IAAI,EAAE,UAAU;IAChB,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,sFAAsF;QACtF,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,KAAK,CAAA;QACjC,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QAEf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,EAAE,EAAE,CAAC;YAChC,MAAM,GAAG,GAAG,2BAA2B,CAAC,SAAS,EAAE,EAAE,MAAM,EAAE,aAAa,EAAE,YAAY,CAAC,MAAM,CAAC,EAAE,CAAC,CAAA;YACnG,MAAM,OAAO,GAAG,GAAG,CAAC,MAAM,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,IAAI,CAAA;YAChD,MAAM,SAAS,GAAG,aAAa,CAAC,uBAAuB,EAAE,EAAE,GAAG,GAAG,CAAC,UAAU,EAAE,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;YAC7F,MAAM,SAAS,GAAG;gBACjB,GAAG,EAAE,GAAG,CAAC,GAAG;gBACZ,UAAU,EAAE,GAAG,CAAC,UAAU;gBAC1B,OAAO;gBACP,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,MAAM,EAAE,uBAAuB;gBAC/B,SAAS;gBACT,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,gFAAgF;aACzF,CAAA;YACD,MAAM,OAAO,GAAG,QAAQ,CAAC,SAA2C,CAAC,CAAA;YAErE,IAAI,OAAO,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;gBAChC,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YACD,iGAAiG;YACjG,KAAK,CACJ,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,OAAO,CAAC,GAAG,EAAE,YAAY,EAAE,mBAAmB,GAAG,CAAC,QAAQ,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,GAAG,IAAI,CAC/G,CAAA;YACD,OAAO,EAAE,CAAA;QACV,CAAC;QAED,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAC5B,CAAC;CACD,CAAA"}

package/out/src/shard-recipes/country-balanced.d.ts

@@ -0,0 +1,28 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `country-balanced` shard recipe — the BALANCED, MODEL-FIRST country-coverage shard (#464). The
+ *   shipped model is STARVED on `country` (P=R=F1=0 on the homograph eval), so this fills the void
+ *   the way the `unit` shard did, but built to AVOID over-firing "trailing token ⇒ country". Three
+ *   ingredients, ported faithfully from scripts/build-country-shard-balanced.mjs:
+ *
+ *   1. Breadth/recall — real OA skeletons (US/DE/FR/IT/NL) with a country token in a varied surface form
+ *        from `@mailwoman/codex/country` (canonical / endonym / ISO code), + ~30% country-ABSENT
+ *        negatives (teach O-emission, the precision floor).
+ *   2. Homograph CONTRAST pairs — each true country-name homograph (Georgia, Jordan, Lebanon, Mexico,
+ *        Peru, Turkey) rendered BOTH ways: as `country` (foreign-city context) AND as the US
+ *        `region`/`locality` (US-ZIP context). Teaches that the label is CONTEXTUAL, not
+ *        positional.
+ *   3. Code-as-region negatives — 2-letter codes that are both a US state abbrev and an ISO country code
+ *        (CA/GA/IN/MA/PA/AL) in US-ZIP context → must read as `region`, never `country`.
+ *
+ *   `--golden` emits a held-out synthetic val over the VT (US) + Berlin (DE) holdouts. This is a
+ *   `generate`-mode recipe that still reads REAL tuples off disk — `--count` bounds the OUTPUT, not
+ *   the input. The passed `random` (the framework LCG) is consumed in the exact call order the
+ *   legacy script used.
+ */
+import { type ShardRecipe } from "./scaffold.js";
+export declare const countryBalancedRecipe: ShardRecipe;
+//# sourceMappingURL=country-balanced.d.ts.map

package/out/src/shard-recipes/country-balanced.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"country-balanced.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/country-balanced.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAUH,OAAO,EAAkB,KAAK,WAAW,EAAE,MAAM,eAAe,CAAA;AA+UhE,eAAO,MAAM,qBAAqB,EAAE,WAwFnC,CAAA"}

package/out/src/shard-recipes/country-balanced.js

@@ -0,0 +1,370 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `country-balanced` shard recipe — the BALANCED, MODEL-FIRST country-coverage shard (#464). The
+ *   shipped model is STARVED on `country` (P=R=F1=0 on the homograph eval), so this fills the void
+ *   the way the `unit` shard did, but built to AVOID over-firing "trailing token ⇒ country". Three
+ *   ingredients, ported faithfully from scripts/build-country-shard-balanced.mjs:
+ *
+ *   1. Breadth/recall — real OA skeletons (US/DE/FR/IT/NL) with a country token in a varied surface form
+ *        from `@mailwoman/codex/country` (canonical / endonym / ISO code), + ~30% country-ABSENT
+ *        negatives (teach O-emission, the precision floor).
+ *   2. Homograph CONTRAST pairs — each true country-name homograph (Georgia, Jordan, Lebanon, Mexico,
+ *        Peru, Turkey) rendered BOTH ways: as `country` (foreign-city context) AND as the US
+ *        `region`/`locality` (US-ZIP context). Teaches that the label is CONTEXTUAL, not
+ *        positional.
+ *   3. Code-as-region negatives — 2-letter codes that are both a US state abbrev and an ISO country code
+ *        (CA/GA/IN/MA/PA/AL) in US-ZIP context → must read as `region`, never `country`.
+ *
+ *   `--golden` emits a held-out synthetic val over the VT (US) + Berlin (DE) holdouts. This is a
+ *   `generate`-mode recipe that still reads REAL tuples off disk — `--count` bounds the OUTPUT, not
+ *   the input. The passed `random` (the framework LCG) is consumed in the exact call order the
+ *   legacy script used.
+ */
+import { spawnSync } from "node:child_process";
+import { COUNTRY_SURFACE_FORMS, CountryNames } from "@mailwoman/codex/country";
+import { stableSourceID } from "../adapter.js";
+import { alignRow } from "../align.js";
+import { makeMulberry32 } from "./scaffold.js";
+// v2: the country TOKEN is decoupled from the skeleton's locale and drawn from a BROAD pool — every
+// ISO canonical name + every curated surface form (endonyms/abbrevs). Surface forms are over-weighted
+// so endonyms/abbrevs ("Deutschland","USA","NL") get strong signal.
+const COUNTRY_FORM_POOL = (() => {
+    const surface = Object.values(COUNTRY_SURFACE_FORMS).flat(); // endonyms + abbrevs + canonical (curated)
+    const names = [...CountryNames];
+    // all ~249 ISO canonical English names (breadth)
+    return { surface, names };
+})();
+const COUNTRY_ABSENT_PROB = 0.3; // negatives: rows with NO country token → teach golden precision
+// Multi-locale OA sources. region = implied admin where the extract is single-region (US states, DE
+// Saxony); countrywide extracts (FR/IT/NL) read region from the CSV when present.
+const SOURCES = [
+    { zip: "/tmp/oa-cache/us__ia__statewide.zip", csv: "us/ia/statewide.csv", iso2: "US", region: "IA", order: "us" },
+    { zip: "/tmp/oa-cache/us__il__cook.zip", csv: "us/il/cook.csv", iso2: "US", region: "IL", order: "us" },
+    { zip: "/tmp/oa-cache/us__mt__statewide.zip", csv: "us/mt/statewide.csv", iso2: "US", region: "MT", order: "us" },
+    { zip: "/tmp/oa-cache/us__sd__statewide.zip", csv: "us/sd/statewide.csv", iso2: "US", region: "SD", order: "us" },
+    { zip: "/tmp/oa-cache/de__sn__statewide.zip", csv: "de/sn/statewide.csv", iso2: "DE", region: "", order: "eu" },
+    { zip: "/tmp/oa-cache/fr__countrywide.zip", csv: "fr/countrywide.csv", iso2: "FR", region: "", order: "fr" },
+    // ES uses the Spanish IGN schema, not the OA standard columns — skipped here (codex still recognizes
+    // "España"/"Spain"). A dedicated IGN adapter is a follow-up.
+    { zip: "/tmp/oa-cache/it__countrywide.zip", csv: "it/countrywide.csv", iso2: "IT", region: "", order: "eu" },
+    { zip: "/tmp/oa-cache/nl__countrywide.zip", csv: "nl/countrywide.csv", iso2: "NL", region: "", order: "eu" },
+];
+// Held-out for --golden: Vermont (US holdout) + Berlin (DE holdout) — geographic split, never trained.
+const EVAL_SOURCES = [
+    { zip: "/tmp/oa-cache/us__vt__statewide.zip", csv: "us/vt/statewide.csv", iso2: "US", region: "VT", order: "us" },
+    { zip: "/tmp/oa-cache/de__berlin.zip", csv: "de/berlin.csv", iso2: "DE", region: "", order: "eu" },
+];
+function splitCSV(line) {
+    const out = [];
+    let cur = "", inQ = false;
+    for (let i = 0; i < line.length; i++) {
+        const c = line[i];
+        if (inQ) {
+            if (c === '"') {
+                if (line[i + 1] === '"') {
+                    cur += '"';
+                    i++;
+                }
+                else
+                    inQ = false;
+            }
+            else
+                cur += c;
+        }
+        else if (c === '"')
+            inQ = true;
+        else if (c === ",") {
+            out.push(cur);
+            cur = "";
+        }
+        else
+            cur += c;
+    }
+    out.push(cur);
+    return out;
+}
+function readTuples(source, limit) {
+    // countrywide extracts (FR/IT/NL) are GB-scale — cap the bytes with `head` (read ~8 lines per wanted
+    // tuple to survive dedup/skips) so the toString stays under V8's string limit.
+    const maxLines = Math.max(limit * 8, 20000) + 1;
+    const r = spawnSync("bash", ["-c", `unzip -p "${source.zip}" "${source.csv}" | head -n ${maxLines}`], {
+        maxBuffer: 1024 * 1024 * 1024,
+        encoding: "buffer",
+    });
+    if (r.status !== 0) {
+        console.error(`  WARN: unzip failed for ${source.zip} (status ${r.status})`);
+        return [];
+    }
+    const lines = r.stdout.toString("utf8").split(/\r?\n/);
+    if (lines.length < 2)
+        return [];
+    const header = splitCSV(lines[0]).map((h) => h.trim().toLowerCase());
+    const idx = (n) => header.indexOf(n);
+    const iNum = idx("number"), iStreet = idx("street"), iCity = idx("city"), iRegion = idx("region"), iPost = idx("postcode");
+    const get = (cells, i) => (i >= 0 && i < cells.length ? (cells[i] ?? "").trim() : "");
+    const tuples = [];
+    const seen = new Set();
+    for (let li = 1; li < lines.length && tuples.length < limit; li++) {
+        if (!lines[li])
+            continue;
+        const cells = splitCSV(lines[li]);
+        const street = get(cells, iStreet), locality = get(cells, iCity), house_number = get(cells, iNum);
+        if (!street || !locality || !house_number)
+            continue;
+        const key = `${house_number}|${street}|${locality}`.toLowerCase();
+        if (seen.has(key))
+            continue;
+        seen.add(key);
+        tuples.push({
+            house_number,
+            street,
+            locality,
+            region: get(cells, iRegion) || source.region,
+            postcode: get(cells, iPost),
+            iso2: source.iso2,
+            order: source.order,
+        });
+    }
+    return tuples;
+}
+/** Pick a country token from the BROAD pool, or null (a country-absent negative). v2. */
+function pickCountry(random) {
+    if (random() < COUNTRY_ABSENT_PROB)
+        return null; // negative — teaches "trailing token != always country"
+    // 60% curated surface forms (endonym/abbrev variety), 40% broad ISO canonical names (coverage).
+    const pool = random() < 0.6 ? COUNTRY_FORM_POOL.surface : COUNTRY_FORM_POOL.names;
+    return pool[Math.floor(random() * pool.length)];
+}
+/** Render the address body in native-ish order. `country` null → a country-ABSENT negative row. */
+function renderCountry(random, t, country) {
+    const { house_number: hn, street, locality: loc, region: reg, postcode: pc, order } = t;
+    const components = { house_number: hn, street, locality: loc };
+    if (reg)
+        components.region = reg;
+    if (pc)
+        components.postcode = pc;
+    let body;
+    if (order === "us") {
+        const regPc = [reg, pc].filter(Boolean).join(" ");
+        body = `${hn} ${street}, ${loc}${regPc ? ", " + regPc : ""}`;
+    }
+    else if (order === "fr") {
+        // French is NUMBER-street, postcode-city: "84 Route de la Fontaine, 75008 Paris".
+        body = `${hn} ${street}, ${[pc, loc].filter(Boolean).join(" ")}`;
+    }
+    else {
+        // DE/IT/NL: street-number, postcode-city: "Pariser Platz 1, 10117 Berlin".
+        const pcCity = [pc, loc].filter(Boolean).join(" ");
+        body = `${street} ${hn}, ${pcCity}`;
+    }
+    if (!country) {
+        // Negative: a normal address, NO country token/component. Teaches that a trailing region/city/
+        // postcode is NOT a country (counters the v1 golden over-firing).
+        return { fmt: "negative", raw: body, components };
+    }
+    const withC = { ...components, country };
+    const r = random();
+    if (r < 0.8)
+        return { fmt: "full", raw: `${body}, ${country}`, components: withC };
+    if (r < 0.92)
+        return { fmt: "full-nl", raw: `${body}\n${country}`, components: withC };
+    const bareBody = order === "us" || order === "fr" ? `${hn} ${street}, ${loc}` : `${street} ${hn}, ${loc}`;
+    return {
+        fmt: "bare",
+        raw: `${bareBody}, ${country}`,
+        components: { house_number: hn, street, locality: loc, country },
+    };
+}
+const HOMOGRAPHS = [
+    {
+        surface: "Georgia",
+        iso2: "GE",
+        cities: ["Tbilisi", "Batumi", "Kutaisi", "Rustavi"],
+        us: { role: "region", locality: "Atlanta", region: "Georgia", postcodes: ["30309", "31401", "30601", "31201"] },
+    },
+    {
+        surface: "Jordan",
+        iso2: "JO",
+        cities: ["Amman", "Irbid", "Zarqa", "Aqaba"],
+        us: { role: "locality", locality: "Jordan", region: "MN", postcodes: ["55352"] },
+    },
+    {
+        surface: "Lebanon",
+        iso2: "LB",
+        cities: ["Beirut", "Tripoli", "Sidon", "Byblos"],
+        us: { role: "locality", locality: "Lebanon", region: "TN", postcodes: ["37087", "03766", "17042", "45036"] },
+    },
+    {
+        surface: "Mexico",
+        iso2: "MX",
+        cities: ["Guadalajara", "Monterrey", "Puebla", "Oaxaca"],
+        us: { role: "locality", locality: "Mexico", region: "MO", postcodes: ["65265"] },
+    },
+    {
+        surface: "Peru",
+        iso2: "PE",
+        cities: ["Cusco", "Arequipa", "Trujillo", "Iquitos"],
+        us: { role: "locality", locality: "Peru", region: "IL", postcodes: ["61354", "46970"] },
+    },
+    {
+        surface: "Turkey",
+        iso2: "TR",
+        cities: ["Ankara", "Izmir", "Bursa", "Antalya"],
+        us: { role: "locality", locality: "Turkey", region: "TX", postcodes: ["79261", "28393"] },
+    },
+];
+const ABBREV_REGIONS = [
+    { code: "CA", localities: ["Los Angeles", "Sacramento", "San Diego"], postcodes: ["90012", "95814", "92101"] }, // California / Canada
+    { code: "GA", localities: ["Atlanta", "Savannah", "Macon"], postcodes: ["30309", "31401", "31201"] }, // Georgia(US) / Georgia
+    { code: "IN", localities: ["Indianapolis", "Fort Wayne"], postcodes: ["46204", "46802"] }, // Indiana / India
+    { code: "MA", localities: ["Boston", "Worcester"], postcodes: ["02108", "01608"] }, // Massachusetts / Morocco
+    { code: "PA", localities: ["Philadelphia", "Pittsburgh"], postcodes: ["19103", "15222"] }, // Pennsylvania / Panama
+    { code: "AL", localities: ["Birmingham", "Montgomery"], postcodes: ["35203", "36104"] }, // Alabama / Albania
+];
+const STREET_POOL = [
+    "Main Street",
+    "Oak Avenue",
+    "Park Road",
+    "Elm Street",
+    "Hill Road",
+    "Market Street",
+    "Church Street",
+    "King Street",
+    "2nd Avenue",
+    "Maple Drive",
+];
+const pick = (random, arr) => arr[Math.floor(random() * arr.length)];
+const houseNo = (random) => String(1 + Math.floor(random() * 998));
+/**
+ * A homograph CONTRAST row: ~half render the surface as `country` (foreign city), half as the US `region`/`locality`
+ * (US ZIP, NO country). Returns iso2 for provenance.
+ */
+function renderHomograph(random) {
+    const h = pick(random, HOMOGRAPHS);
+    const hn = houseNo(random), street = pick(random, STREET_POOL);
+    if (random() < 0.5) {
+        const city = pick(random, h.cities);
+        const withStreet = random() < 0.6;
+        const raw = withStreet ? `${hn} ${street}, ${city}, ${h.surface}` : `${city}, ${h.surface}`;
+        const components = withStreet
+            ? { house_number: hn, street, locality: city, country: h.surface }
+            : { locality: city, country: h.surface };
+        return { fmt: "homograph-country", raw, components, iso2: h.iso2 };
+    }
+    const pc = pick(random, h.us.postcodes);
+    if (h.us.role === "region") {
+        // surface is the US STATE: "123 Oak Ave, Atlanta, Georgia 30309" → region, no country
+        return {
+            fmt: "homograph-us-region",
+            raw: `${hn} ${street}, ${h.us.locality}, ${h.surface} ${pc}`,
+            components: { house_number: hn, street, locality: h.us.locality, region: h.surface, postcode: pc },
+            iso2: "US",
+        };
+    }
+    // surface is the US CITY: "123 Oak Ave, Lebanon, TN 37087" → locality, no country
+    return {
+        fmt: "homograph-us-locality",
+        raw: `${hn} ${street}, ${h.surface}, ${h.us.region} ${pc}`,
+        components: { house_number: hn, street, locality: h.surface, region: h.us.region, postcode: pc },
+        iso2: "US",
+    };
+}
+/** An abbrev-as-region negative: "123 Main St, Los Angeles, CA 90012" → region CA, NO country. */
+function renderAbbrevRegion(random) {
+    const a = pick(random, ABBREV_REGIONS);
+    const hn = houseNo(random), street = pick(random, STREET_POOL), locality = pick(random, a.localities), postcode = pick(random, a.postcodes);
+    return {
+        fmt: "abbrev-region",
+        raw: `${hn} ${street}, ${locality}, ${a.code} ${postcode}`,
+        components: { house_number: hn, street, locality, region: a.code, postcode },
+        iso2: "US",
+    };
+}
+const HOMOGRAPH_FRAC = 0.22; // share of rows that are homograph contrast pairs
+const ABBREV_FRAC = 0.08; // share that are code-as-region negatives (cumulative with HOMOGRAPH_FRAC)
+export const countryBalancedRecipe = {
+    name: "country-balanced",
+    description: "Balanced model-first country rows (#464): OA skeletons + ISO surface forms + homograph contrast pairs",
+    mode: "generate",
+    options: [{ flag: "--golden", description: "Emit the held-out VT+Berlin eval slice" }],
+    async run(opts, write) {
+        if (opts.count == null)
+            throw new Error("country-balanced recipe requires --count <N>");
+        const count = opts.count;
+        // Legacy build-country-shard-balanced.mjs seeded mulberry32 with the raw seed: `const random = mulberry32(opts.seed)`.
+        const random = makeMulberry32(opts.seed);
+        const source = opts.sourceName ?? "synth-country";
+        const sources = opts.golden ? EVAL_SOURCES : SOURCES;
+        const perSource = Math.ceil((count * 3) / sources.length); // over-read; balance locales
+        const pool = [];
+        for (const s of sources) {
+            const t = readTuples(s, perSource);
+            console.error(`  ${s.csv} (${s.iso2}): ${t.length} tuples`);
+            for (const x of t)
+                pool.push(x);
+        }
+        if (pool.length === 0) {
+            throw new Error("No tuples — are the cached OA zips present in /tmp/oa-cache?");
+        }
+        let emitted = 0;
+        let skipped = 0;
+        let guard = 0;
+        const N = pool.length;
+        while (emitted < count && guard++ < count * 8) {
+            // Mix three row types: homograph contrast (the distinction), code-as-region negatives, and the
+            // breadth/recall main path (random ISO form on an OA skeleton, ~30% country-absent).
+            const roll = random();
+            let rendered;
+            let rowISO2;
+            if (roll < HOMOGRAPH_FRAC) {
+                const h = renderHomograph(random);
+                rendered = h;
+                rowISO2 = h.iso2;
+            }
+            else if (roll < HOMOGRAPH_FRAC + ABBREV_FRAC) {
+                const a = renderAbbrevRegion(random);
+                rendered = a;
+                rowISO2 = a.iso2;
+            }
+            else {
+                const t = pool[Math.floor(random() * N)];
+                const country = pickCountry(random); // may be null → a country-absent negative row
+                rendered = renderCountry(random, t, country);
+                rowISO2 = t.iso2;
+                if (country && !rendered.raw.includes(country)) {
+                    skipped++;
+                    continue;
+                }
+            }
+            const { raw, components } = rendered;
+            const localeTag = rowISO2 === "US" ? "en-US" : `${rowISO2.toLowerCase()}-${rowISO2}`;
+            if (opts.golden) {
+                write(JSON.stringify({ raw, components, country: rowISO2 }) + "\n");
+                emitted++;
+                continue;
+            }
+            const canonical = {
+                raw,
+                components,
+                country: rowISO2,
+                locale: localeTag,
+                source,
+                source_id: stableSourceID(source, components),
+                corpus_version: "0.4.0",
+                license: "OpenAddresses multi-locale skeletons + injected ISO-3166 country surface forms (codex)",
+            };
+            const aligned = alignRow(canonical);
+            if (aligned.kind !== "labeled" || !aligned.row) {
+                skipped++;
+                continue;
+            }
+            write(JSON.stringify({ ...aligned.row, synth_method: "country", synth_base_id: null }) + "\n");
+            emitted++;
+        }
+        return { emitted, skipped };
+    },
+};
+//# sourceMappingURL=country-balanced.js.map

package/out/src/shard-recipes/country-balanced.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"country-balanced.js","sourceRoot":"","sources":["../../../src/shard-recipes/country-balanced.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,OAAO,EAAE,SAAS,EAAE,MAAM,oBAAoB,CAAA;AAE9C,OAAO,EAAE,qBAAqB,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAA;AAG9E,OAAO,EAAE,cAAc,EAAE,MAAM,eAAe,CAAA;AAC9C,OAAO,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAA;AAEtC,OAAO,EAAE,cAAc,EAAoB,MAAM,eAAe,CAAA;AAEhE,oGAAoG;AACpG,sGAAsG;AACtG,oEAAoE;AACpE,MAAM,iBAAiB,GAAG,CAAC,GAAG,EAAE;IAC/B,MAAM,OAAO,GAAG,MAAM,CAAC,MAAM,CAAC,qBAAqB,CAAC,CAAC,IAAI,EAAE,CAAA,CAAC,2CAA2C;IACvG,MAAM,KAAK,GAAG,CAAC,GAAG,YAAY,CAAC,CAAA;IAE/B,iDAAiD;IACjD,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,CAAA;AAC1B,CAAC,CAAC,EAAE,CAAA;AACJ,MAAM,mBAAmB,GAAG,GAAG,CAAA,CAAC,iEAAiE;AAWjG,oGAAoG;AACpG,kFAAkF;AAClF,MAAM,OAAO,GAA6B;IACzC,EAAE,GAAG,EAAE,qCAAqC,EAAE,GAAG,EAAE,qBAAqB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE;IACjH,EAAE,GAAG,EAAE,gCAAgC,EAAE,GAAG,EAAE,gBAAgB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE;IACvG,EAAE,GAAG,EAAE,qCAAqC,EAAE,GAAG,EAAE,qBAAqB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE;IACjH,EAAE,GAAG,EAAE,qCAAqC,EAAE,GAAG,EAAE,qBAAqB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE;IACjH,EAAE,GAAG,EAAE,qCAAqC,EAAE,GAAG,EAAE,qBAAqB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE;IAC/G,EAAE,GAAG,EAAE,mCAAmC,EAAE,GAAG,EAAE,oBAAoB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE;IAC5G,qGAAqG;IACrG,6DAA6D;IAC7D,EAAE,GAAG,EAAE,mCAAmC,EAAE,GAAG,EAAE,oBAAoB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE;IAC5G,EAAE,GAAG,EAAE,mCAAmC,EAAE,GAAG,EAAE,oBAAoB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE;CAC5G,CAAA;AACD,uGAAuG;AACvG,MAAM,YAAY,GAA6B;IAC9C,EAAE,GAAG,EAAE,qCAAqC,EAAE,GAAG,EAAE,qBAAqB,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE;IACjH,EAAE,GAAG,EAAE,8BAA8B,EAAE,GAAG,EAAE,eAAe,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE;CAClG,CAAA;AAaD,SAAS,QAAQ,CAAC,IAAY;IAC7B,MAAM,GAAG,GAAa,EAAE,CAAA;IACxB,IAAI,GAAG,GAAG,EAAE,EACX,GAAG,GAAG,KAAK,CAAA;IAEZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACtC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAA;QAEjB,IAAI,GAAG,EAAE,CAAC;YACT,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;gBACf,IAAI,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,GAAG,EAAE,CAAC;oBACzB,GAAG,IAAI,GAAG,CAAA;oBACV,CAAC,EAAE,CAAA;gBACJ,CAAC;;oBAAM,GAAG,GAAG,KAAK,CAAA;YACnB,CAAC;;gBAAM,GAAG,IAAI,CAAC,CAAA;QAChB,CAAC;aAAM,IAAI,CAAC,KAAK,GAAG;YAAE,GAAG,GAAG,IAAI,CAAA;aAC3B,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;YACpB,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;YACb,GAAG,GAAG,EAAE,CAAA;QACT,CAAC;;YAAM,GAAG,IAAI,CAAC,CAAA;IAChB,CAAC;IACD,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IAEb,OAAO,GAAG,CAAA;AACX,CAAC;AAED,SAAS,UAAU,CAAC,MAAqB,EAAE,KAAa;IACvD,qGAAqG;IACrG,+EAA+E;IAC/E,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,EAAE,KAAK,CAAC,GAAG,CAAC,CAAA;IAC/C,MAAM,CAAC,GAAG,SAAS,CAAC,MAAM,EAAE,CAAC,IAAI,EAAE,aAAa,MAAM,CAAC,GAAG,MAAM,MAAM,CAAC,GAAG,eAAe,QAAQ,EAAE,CAAC,EAAE;QACrG,SAAS,EAAE,IAAI,GAAG,IAAI,GAAG,IAAI;QAC7B,QAAQ,EAAE,QAAQ;KAClB,CAAC,CAAA;IAEF,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpB,OAAO,CAAC,KAAK,CAAC,4BAA4B,MAAM,CAAC,GAAG,YAAY,CAAC,CAAC,MAAM,GAAG,CAAC,CAAA;QAE5E,OAAO,EAAE,CAAA;IACV,CAAC;IACD,MAAM,KAAK,GAAG,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAA;IAEtD,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,EAAE,CAAA;IAC/B,MAAM,MAAM,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,CAAA;IACrE,MAAM,GAAG,GAAG,CAAC,CAAS,EAAU,EAAE,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAA;IACpD,MAAM,IAAI,GAAG,GAAG,CAAC,QAAQ,CAAC,EACzB,OAAO,GAAG,GAAG,CAAC,QAAQ,CAAC,EACvB,KAAK,GAAG,GAAG,CAAC,MAAM,CAAC,EACnB,OAAO,GAAG,GAAG,CAAC,QAAQ,CAAC,EACvB,KAAK,GAAG,GAAG,CAAC,UAAU,CAAC,CAAA;IACxB,MAAM,GAAG,GAAG,CAAC,KAAe,EAAE,CAAS,EAAU,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;IAC/G,MAAM,MAAM,GAAmB,EAAE,CAAA;IACjC,MAAM,IAAI,GAAG,IAAI,GAAG,EAAU,CAAA;IAE9B,KAAK,IAAI,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,KAAK,CAAC,MAAM,IAAI,MAAM,CAAC,MAAM,GAAG,KAAK,EAAE,EAAE,EAAE,EAAE,CAAC;QACnE,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YAAE,SAAQ;QACxB,MAAM,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAE,CAAC,CAAA;QAClC,MAAM,MAAM,GAAG,GAAG,CAAC,KAAK,EAAE,OAAO,CAAC,EACjC,QAAQ,GAAG,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC,EAC5B,YAAY,GAAG,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,CAAA;QAEhC,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,IAAI,CAAC,YAAY;YAAE,SAAQ;QACnD,MAAM,GAAG,GAAG,GAAG,YAAY,IAAI,MAAM,IAAI,QAAQ,EAAE,CAAC,WAAW,EAAE,CAAA;QAEjE,IAAI,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC;YAAE,SAAQ;QAC3B,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,CAAA;QACb,MAAM,CAAC,IAAI,CAAC;YACX,YAAY;YACZ,MAAM;YACN,QAAQ;YACR,MAAM,EAAE,GAAG,CAAC,KAAK,EAAE,OAAO,CAAC,IAAI,MAAM,CAAC,MAAM;YAC5C,QAAQ,EAAE,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC;YAC3B,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,KAAK,EAAE,MAAM,CAAC,KAAK;SACnB,CAAC,CAAA;IACH,CAAC;IAED,OAAO,MAAM,CAAA;AACd,CAAC;AAED,yFAAyF;AACzF,SAAS,WAAW,CAAC,MAAoB;IACxC,IAAI,MAAM,EAAE,GAAG,mBAAmB;QAAE,OAAO,IAAI,CAAA,CAAC,wDAAwD;IACxG,gGAAgG;IAChG,MAAM,IAAI,GAAG,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,iBAAiB,CAAC,OAAO,CAAC,CAAC,CAAC,iBAAiB,CAAC,KAAK,CAAA;IAEjF,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,CAAE,CAAA;AACjD,CAAC;AAED,mGAAmG;AACnG,SAAS,aAAa,CACrB,MAAoB,EACpB,CAAe,EACf,OAAsB;IAEtB,MAAM,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE,EAAE,KAAK,EAAE,GAAG,CAAC,CAAA;IACvF,MAAM,UAAU,GAA0C,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,EAAE,CAAA;IAErG,IAAI,GAAG;QAAE,UAAU,CAAC,MAAM,GAAG,GAAG,CAAA;IAEhC,IAAI,EAAE;QAAE,UAAU,CAAC,QAAQ,GAAG,EAAE,CAAA;IAChC,IAAI,IAAY,CAAA;IAEhB,IAAI,KAAK,KAAK,IAAI,EAAE,CAAC;QACpB,MAAM,KAAK,GAAG,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;QACjD,IAAI,GAAG,GAAG,EAAE,IAAI,MAAM,KAAK,GAAG,GAAG,KAAK,CAAC,CAAC,CAAC,IAAI,GAAG,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAA;IAC7D,CAAC;SAAM,IAAI,KAAK,KAAK,IAAI,EAAE,CAAC;QAC3B,kFAAkF;QAClF,IAAI,GAAG,GAAG,EAAE,IAAI,MAAM,KAAK,CAAC,EAAE,EAAE,GAAG,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAA;IACjE,CAAC;SAAM,CAAC;QACP,2EAA2E;QAC3E,MAAM,MAAM,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;QAClD,IAAI,GAAG,GAAG,MAAM,IAAI,EAAE,KAAK,MAAM,EAAE,CAAA;IACpC,CAAC;IAED,IAAI,CAAC,OAAO,EAAE,CAAC;QACd,+FAA+F;QAC/F,kEAAkE;QAClE,OAAO,EAAE,GAAG,EAAE,UAAU,EAAE,GAAG,EAAE,IAAI,EAAE,UAAU,EAAE,CAAA;IAClD,CAAC;IACD,MAAM,KAAK,GAA0C,EAAE,GAAG,UAAU,EAAE,OAAO,EAAE,CAAA;IAC/E,MAAM,CAAC,GAAG,MAAM,EAAE,CAAA;IAElB,IAAI,CAAC,GAAG,GAAG;QAAE,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE,GAAG,IAAI,KAAK,OAAO,EAAE,EAAE,UAAU,EAAE,KAAK,EAAE,CAAA;IAElF,IAAI,CAAC,GAAG,IAAI;QAAE,OAAO,EAAE,GAAG,EAAE,SAAS,EAAE,GAAG,EAAE,GAAG,IAAI,KAAK,OAAO,EAAE,EAAE,UAAU,EAAE,KAAK,EAAE,CAAA;IACtF,MAAM,QAAQ,GAAG,KAAK,KAAK,IAAI,IAAI,KAAK,KAAK,IAAI,CAAC,CAAC,CAAC,GAAG,EAAE,IAAI,MAAM,KAAK,GAAG,EAAE,CAAC,CAAC,CAAC,GAAG,MAAM,IAAI,EAAE,KAAK,GAAG,EAAE,CAAA;IAEzG,OAAO;QACN,GAAG,EAAE,MAAM;QACX,GAAG,EAAE,GAAG,QAAQ,KAAK,OAAO,EAAE;QAC9B,UAAU,EAAE,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,EAAE,OAAO,EAAE;KAChE,CAAA;AACF,CAAC;AAYD,MAAM,UAAU,GAAyB;IACxC;QACC,OAAO,EAAE,SAAS;QAClB,IAAI,EAAE,IAAI;QACV,MAAM,EAAE,CAAC,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,CAAC;QACnD,EAAE,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,EAAE,SAAS,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,OAAO,CAAC,EAAE;KAC/G;IACD;QACC,OAAO,EAAE,QAAQ;QACjB,IAAI,EAAE,IAAI;QACV,MAAM,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,OAAO,CAAC;QAC5C,EAAE,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,OAAO,CAAC,EAAE;KAChF;IACD;QACC,OAAO,EAAE,SAAS;QAClB,IAAI,EAAE,IAAI;QACV,MAAM,EAAE,CAAC,QAAQ,EAAE,SAAS,EAAE,OAAO,EAAE,QAAQ,CAAC;QAChD,EAAE,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,OAAO,CAAC,EAAE;KAC5G;IACD;QACC,OAAO,EAAE,QAAQ;QACjB,IAAI,EAAE,IAAI;QACV,MAAM,EAAE,CAAC,aAAa,EAAE,WAAW,EAAE,QAAQ,EAAE,QAAQ,CAAC;QACxD,EAAE,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,OAAO,CAAC,EAAE;KAChF;IACD;QACC,OAAO,EAAE,MAAM;QACf,IAAI,EAAE,IAAI;QACV,MAAM,EAAE,CAAC,OAAO,EAAE,UAAU,EAAE,UAAU,EAAE,SAAS,CAAC;QACpD,EAAE,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE;KACvF;IACD;QACC,OAAO,EAAE,QAAQ;QACjB,IAAI,EAAE,IAAI;QACV,MAAM,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,OAAO,EAAE,SAAS,CAAC;QAC/C,EAAE,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE;KACzF;CACD,CAAA;AAOD,MAAM,cAAc,GAA4B;IAC/C,EAAE,IAAI,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,aAAa,EAAE,YAAY,EAAE,WAAW,CAAC,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,CAAC,EAAE,EAAE,sBAAsB;IACtI,EAAE,IAAI,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,SAAS,EAAE,UAAU,EAAE,OAAO,CAAC,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,CAAC,EAAE,EAAE,wBAAwB;IAC9H,EAAE,IAAI,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,cAAc,EAAE,YAAY,CAAC,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE,EAAE,kBAAkB;IAC7G,EAAE,IAAI,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,QAAQ,EAAE,WAAW,CAAC,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE,EAAE,0BAA0B;IAC9G,EAAE,IAAI,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,cAAc,EAAE,YAAY,CAAC,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE,EAAE,wBAAwB;IACnH,EAAE,IAAI,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,YAAY,EAAE,YAAY,CAAC,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE,EAAE,oBAAoB;CAC7G,CAAA;AACD,MAAM,WAAW,GAAsB;IACtC,aAAa;IACb,YAAY;IACZ,WAAW;IACX,YAAY;IACZ,WAAW;IACX,eAAe;IACf,eAAe;IACf,aAAa;IACb,YAAY;IACZ,aAAa;CACb,CAAA;AACD,MAAM,IAAI,GAAG,CAAI,MAAoB,EAAE,GAAiB,EAAK,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,MAAM,CAAC,CAAE,CAAA;AACvG,MAAM,OAAO,GAAG,CAAC,MAAoB,EAAU,EAAE,CAAC,MAAM,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC,CAAA;AAExF;;;GAGG;AACH,SAAS,eAAe,CAAC,MAAoB;IAM5C,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,UAAU,CAAC,CAAA;IAClC,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,CAAC,EACzB,MAAM,GAAG,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC,CAAA;IAEnC,IAAI,MAAM,EAAE,GAAG,GAAG,EAAE,CAAC;QACpB,MAAM,IAAI,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAA;QACnC,MAAM,UAAU,GAAG,MAAM,EAAE,GAAG,GAAG,CAAA;QACjC,MAAM,GAAG,GAAG,UAAU,CAAC,CAAC,CAAC,GAAG,EAAE,IAAI,MAAM,KAAK,IAAI,KAAK,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,GAAG,IAAI,KAAK,CAAC,CAAC,OAAO,EAAE,CAAA;QAC3F,MAAM,UAAU,GAA0C,UAAU;YACnE,CAAC,CAAC,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE;YAClE,CAAC,CAAC,EAAE,QAAQ,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAA;QAEzC,OAAO,EAAE,GAAG,EAAE,mBAAmB,EAAE,GAAG,EAAE,UAAU,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;IACnE,CAAC;IACD,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,EAAE,CAAC,SAAS,CAAC,CAAA;IAEvC,IAAI,CAAC,CAAC,EAAE,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAC5B,sFAAsF;QACtF,OAAO;YACN,GAAG,EAAE,qBAAqB;YAC1B,GAAG,EAAE,GAAG,EAAE,IAAI,MAAM,KAAK,CAAC,CAAC,EAAE,CAAC,QAAQ,KAAK,CAAC,CAAC,OAAO,IAAI,EAAE,EAAE;YAC5D,UAAU,EAAE,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,EAAE,CAAC,CAAC,OAAO,EAAE,QAAQ,EAAE,EAAE,EAAE;YAClG,IAAI,EAAE,IAAI;SACV,CAAA;IACF,CAAC;IAED,kFAAkF;IAClF,OAAO;QACN,GAAG,EAAE,uBAAuB;QAC5B,GAAG,EAAE,GAAG,EAAE,IAAI,MAAM,KAAK,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,EAAE,CAAC,MAAM,IAAI,EAAE,EAAE;QAC1D,UAAU,EAAE,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC,EAAE,CAAC,MAAM,EAAE,QAAQ,EAAE,EAAE,EAAE;QAChG,IAAI,EAAE,IAAI;KACV,CAAA;AACF,CAAC;AAED,kGAAkG;AAClG,SAAS,kBAAkB,CAAC,MAAoB;IAM/C,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,cAAc,CAAC,CAAA;IACtC,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,CAAC,EACzB,MAAM,GAAG,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC,EAClC,QAAQ,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,UAAU,CAAC,EACrC,QAAQ,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,SAAS,CAAC,CAAA;IAErC,OAAO;QACN,GAAG,EAAE,eAAe;QACpB,GAAG,EAAE,GAAG,EAAE,IAAI,MAAM,KAAK,QAAQ,KAAK,CAAC,CAAC,IAAI,IAAI,QAAQ,EAAE;QAC1D,UAAU,EAAE,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC,CAAC,IAAI,EAAE,QAAQ,EAAE;QAC5E,IAAI,EAAE,IAAI;KACV,CAAA;AACF,CAAC;AAED,MAAM,cAAc,GAAG,IAAI,CAAA,CAAC,kDAAkD;AAC9E,MAAM,WAAW,GAAG,IAAI,CAAA,CAAC,2EAA2E;AAEpG,MAAM,CAAC,MAAM,qBAAqB,GAAgB;IACjD,IAAI,EAAE,kBAAkB;IACxB,WAAW,EAAE,uGAAuG;IACpH,IAAI,EAAE,UAAU;IAChB,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,wCAAwC,EAAE,CAAC;IACtF,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,IAAI,IAAI,CAAC,KAAK,IAAI,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,8CAA8C,CAAC,CAAA;QACvF,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAA;QACxB,uHAAuH;QACvH,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,MAAM,GAAG,IAAI,CAAC,UAAU,IAAI,eAAe,CAAA;QACjD,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,OAAO,CAAA;QACpD,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,CAAA,CAAC,6BAA6B;QAEvF,MAAM,IAAI,GAAmB,EAAE,CAAA;QAE/B,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;YACzB,MAAM,CAAC,GAAG,UAAU,CAAC,CAAC,EAAE,SAAS,CAAC,CAAA;YAClC,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,IAAI,MAAM,CAAC,CAAC,MAAM,SAAS,CAAC,CAAA;YAE3D,KAAK,MAAM,CAAC,IAAI,CAAC;gBAAE,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;QAChC,CAAC;QAED,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACvB,MAAM,IAAI,KAAK,CAAC,8DAA8D,CAAC,CAAA;QAChF,CAAC;QAED,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,KAAK,GAAG,CAAC,CAAA;QACb,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAA;QAErB,OAAO,OAAO,GAAG,KAAK,IAAI,KAAK,EAAE,GAAG,KAAK,GAAG,CAAC,EAAE,CAAC;YAC/C,+FAA+F;YAC/F,qFAAqF;YACrF,MAAM,IAAI,GAAG,MAAM,EAAE,CAAA;YACrB,IAAI,QAAyF,CAAA;YAC7F,IAAI,OAAe,CAAA;YAEnB,IAAI,IAAI,GAAG,cAAc,EAAE,CAAC;gBAC3B,MAAM,CAAC,GAAG,eAAe,CAAC,MAAM,CAAC,CAAA;gBACjC,QAAQ,GAAG,CAAC,CAAA;gBACZ,OAAO,GAAG,CAAC,CAAC,IAAI,CAAA;YACjB,CAAC;iBAAM,IAAI,IAAI,GAAG,cAAc,GAAG,WAAW,EAAE,CAAC;gBAChD,MAAM,CAAC,GAAG,kBAAkB,CAAC,MAAM,CAAC,CAAA;gBACpC,QAAQ,GAAG,CAAC,CAAA;gBACZ,OAAO,GAAG,CAAC,CAAC,IAAI,CAAA;YACjB,CAAC;iBAAM,CAAC;gBACP,MAAM,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAE,CAAA;gBACzC,MAAM,OAAO,GAAG,WAAW,CAAC,MAAM,CAAC,CAAA,CAAC,8CAA8C;gBAClF,QAAQ,GAAG,aAAa,CAAC,MAAM,EAAE,CAAC,EAAE,OAAO,CAAC,CAAA;gBAC5C,OAAO,GAAG,CAAC,CAAC,IAAI,CAAA;gBAEhB,IAAI,OAAO,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE,CAAC;oBAChD,OAAO,EAAE,CAAA;oBACT,SAAQ;gBACT,CAAC;YACF,CAAC;YACD,MAAM,EAAE,GAAG,EAAE,UAAU,EAAE,GAAG,QAAQ,CAAA;YACpC,MAAM,SAAS,GAAG,OAAO,KAAK,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,WAAW,EAAE,IAAI,OAAO,EAAE,CAAA;YAEpF,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;gBACjB,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,EAAE,UAAU,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC,GAAG,IAAI,CAAC,CAAA;gBACnE,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YACD,MAAM,SAAS,GAAiB;gBAC/B,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,OAAO;gBAChB,MAAM,EAAE,SAAS;gBACjB,MAAM;gBACN,SAAS,EAAE,cAAc,CAAC,MAAM,EAAE,UAAU,CAAC;gBAC7C,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,wFAAwF;aACjG,CAAA;YACD,MAAM,OAAO,GAAG,QAAQ,CAAC,SAAS,CAAC,CAAA;YAEnC,IAAI,OAAO,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,GAAG,EAAE,CAAC;gBAChD,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YACD,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,OAAO,CAAC,GAAG,EAAE,YAAY,EAAE,SAAS,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,GAAG,IAAI,CAAC,CAAA;YAC9F,OAAO,EAAE,CAAA;QACV,CAAC;QAED,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAC5B,CAAC;CACD,CAAA"}

package/out/src/shard-recipes/fr-admin-split.d.ts

@@ -0,0 +1,34 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-admin-split` shard recipe — the FR admin-split coverage shard (night 2026-06-19,
+ *   surpass-v1.5.0). Teaches the model to SPLIT the département out of the locality on
+ *   bare/space/comma-delimited French place rows — the admin-deciding failure class the pre-GPU
+ *   self-validation proved moves the resolved coordinate (collision communes −61%; see
+ *   docs/articles/evals/2026-06-19-fr-admin-split-prevalidation.md). Ported from
+ *   scripts/build-fr-admin-split-shard.mjs.
+ *
+ *   Failure shapes (the model currently mis-handles all of these):
+ *
+ *   - `Thauron, Creuse` → région dropped to null (the comma+full-name miss)
+ *   - `Montredon, Lozère` → région = "ère" (the diacritic subword split, #727)
+ *   - (AU analog) `CANBERRA ACT` → the space-delimited admin fuse
+ *
+ *   The département is the essential admin unit for FR postal geography and maps to the `region`
+ *   component tag in our schema. We derive it DETERMINISTICALLY from the real postcode via codex
+ *   `departementForCodePostal` (first two digits = département) — salvage-first, no re-derived
+ *   table.
+ *
+ *   Data (`--communes`, opts.communes): REAL BAN (Base Adresse Nationale) commune+postcode+coord
+ *   tuples, one per line, TAB-separated `commune <TAB> postcode <TAB> lon <TAB> lat`. Build the
+ *   input TSV once from the BAN staging CSV (see the legacy script header). Anchor-ON by
+ *   construction: rows carry a REAL postcode token in `raw` + a `postcode` component, so the
+ *   training loader paints the anchor feature onto that span automatically. The trailing-postcode
+ *   anchor REINFORCES the FR split (FR postcode is trailing, unlike German PLZ-leading — the v0.9.2
+ *   scar is positional, not universal).
+ */
+import { type ShardRecipe } from "./scaffold.js";
+export declare const frAdminSplitRecipe: ShardRecipe;
+//# sourceMappingURL=fr-admin-split.d.ts.map

package/out/src/shard-recipes/fr-admin-split.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"fr-admin-split.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/fr-admin-split.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAWH,OAAO,EAAkB,KAAK,WAAW,EAAE,MAAM,eAAe,CAAA;AA0FhE,eAAO,MAAM,kBAAkB,EAAE,WAgFhC,CAAA"}