@mailwoman/corpus 9.4.0 → 10.0.0

package/out/no/recipes/street/led.d.ts

@@ -0,0 +1,24 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `no-street-led` — the Norwegian street-led slice, third orthography of the #901
+ *   leading-name-boundary family. The post-#920 NO row-read measured the residual as
+ *   ORDER-SENSITIVE decode: street-led forms ("Tangavegen 40, 5620 Tørvikbygd") carry a 30%
+ *   residual vs pc-first's 7% — the diacritic street head (…vegen/…veien with ø/å/æ) is the same
+ *   leading-name-before-comma boundary the FR bare-street and SI village forms exercise, in a
+ *   third orthography. All three real orders cycle per tuple (balanced polarity — the v1.9.9
+ *   lesson: a one-order gradient loses at convergence):
+ *
+ *   1. canonical  "«st» «n», «pc» «city»"   (the 30% residual class — the lesson)
+ *   2. city-first "«city», «pc», «st» «n»"
+ *   3. pc-first   "«pc» «city», «st» «n»"   (the 7% floor — the anchor the others converge to)
+ */
+import { type CorpusRecipe } from "#recipes/scaffold";
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export declare const noStreetLedRecipe: CorpusRecipe;
+//# sourceMappingURL=led.d.ts.map

package/out/no/recipes/street/led.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"led.d.ts","sourceRoot":"","sources":["../../../../lib/no/recipes/street/led.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAKH,OAAO,EAA4C,KAAK,YAAY,EAAiB,MAAM,mBAAmB,CAAA;AAE9G;;;GAGG;AACH,eAAO,MAAM,iBAAiB,EAAE,YA2G/B,CAAA"}

package/out/no/recipes/street/led.js

@@ -0,0 +1,114 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `no-street-led` — the Norwegian street-led slice, third orthography of the #901
+ *   leading-name-boundary family. The post-#920 NO row-read measured the residual as
+ *   ORDER-SENSITIVE decode: street-led forms ("Tangavegen 40, 5620 Tørvikbygd") carry a 30%
+ *   residual vs pc-first's 7% — the diacritic street head (…vegen/…veien with ø/å/æ) is the same
+ *   leading-name-before-comma boundary the FR bare-street and SI village forms exercise, in a
+ *   third orthography. All three real orders cycle per tuple (balanced polarity — the v1.9.9
+ *   lesson: a one-order gradient loses at convergence):
+ *
+ *   1. canonical  "«st» «n», «pc» «city»"   (the 30% residual class — the lesson)
+ *   2. city-first "«city», «pc», «st» «n»"
+ *   3. pc-first   "«pc» «city», «st» «n»"   (the 7% floor — the anchor the others converge to)
+ */
+import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils";
+import { TextSpliterator } from "spliterator";
+import { alignAndWrite, foldNOSurface, readTuples, sliceSourceID } from "#recipes/scaffold";
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export const noStreetLedRecipe = {
+    name: "no-street-led",
+    description: "NO street-led boundary form (#901 family): '«st» «n», «pc» «city»' — diacritic street heads",
+    mode: "tuples",
+    options: [
+        {
+            flag: "--exclude-surfaces <path>",
+            description: "REQUIRED. The NO digit board's reserved surface list " +
+                "(mailwoman/eval-harness/fixtures/no-digits.surfaces.txt); every listed street is skipped.",
+        },
+    ],
+    async run(opts, write) {
+        makeMulberry32(opts.seed);
+        // THE SPLIT (ported from fr-fragment, #727 T2). Without it this slice trains on all 10,697
+        // NO surfaces, 1,952 of which the digit board reserves — so a Norway retrain would grade
+        // memorization of `Hallingrudveien` while claiming to measure the boundary form. There is no
+        // safe default: source-disjoint by street SURFACE is the discipline, so the flag throws.
+        const excludePath = opts.excludeSurfaces;
+        if (!excludePath) {
+            throw new Error("no-street-led: --exclude-surfaces is REQUIRED. Pass the NO digit board's reserved list " +
+                "(mailwoman/eval-harness/fixtures/no-digits.surfaces.txt) or this slice trains on its own eval set. " +
+                "Source-disjoint by street SURFACE is the split discipline; there is no safe default.");
+        }
+        const excluded = new Set();
+        for await (const line of TextSpliterator.fromAsync(excludePath)) {
+            const trimmed = line.trim();
+            if (trimmed && !trimmed.startsWith("#")) {
+                excluded.add(trimmed);
+            }
+        }
+        if (!excluded.size)
+            throw new Error(`no-street-led: --exclude-surfaces "${excludePath}" listed no surfaces`);
+        let read = 0;
+        let emitted = 0;
+        let skipped = 0;
+        let contaminated = 0;
+        for await (const t of readTuples(opts.input)) {
+            read++;
+            const street = String(t.street ?? "").trim();
+            const city = String(t.locality ?? "").trim();
+            const number = String(t.number ?? "").trim();
+            const postcode = String(t.postcode ?? "").trim();
+            if (!street || !city || !number || !postcode) {
+                skipped++;
+                continue;
+            }
+            // A surface on the digit board never enters training.
+            if (excluded.has(foldNOSurface(street))) {
+                contaminated++;
+                continue;
+            }
+            const order = read % 3;
+            let raw;
+            const components = {
+                street,
+                house_number: number,
+                postcode,
+                locality: city,
+            };
+            if (order === 0) {
+                raw = `${street} ${number}, ${postcode} ${city}`;
+            }
+            else if (order === 1) {
+                raw = `${city}, ${postcode}, ${street} ${number}`;
+            }
+            else {
+                raw = `${postcode} ${city}, ${street} ${number}`;
+            }
+            const source_id = sliceSourceID("synth-no-street-led", { ...components, o: String(order), v: String(read) });
+            const canonical = {
+                raw,
+                components,
+                country: "NO",
+                locale: "nb-NO",
+                source: "synth-no-street-led",
+                source_id,
+                corpus_version: "0.10.0",
+                license: "Synthetic — no-street-led; (street, number, postcode, city) from OpenAddresses NO (per-source attribution in the model card)",
+            };
+            if (alignAndWrite(write, canonical, "no-street-led")) {
+                emitted++;
+            }
+            else {
+                skipped++;
+            }
+        }
+        return { read, emitted, skipped, contaminated };
+    },
+};
+//# sourceMappingURL=led.js.map

package/out/no/recipes/street/led.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"led.js","sourceRoot":"","sources":["../../../../lib/no/recipes/street/led.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AACpE,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAA;AAE7C,OAAO,EAAE,aAAa,EAAE,aAAa,EAAE,UAAU,EAAqB,aAAa,EAAE,MAAM,mBAAmB,CAAA;AAE9G;;;GAGG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAiB;IAC9C,IAAI,EAAE,eAAe;IACrB,WAAW,EAAE,6FAA6F;IAC1G,IAAI,EAAE,QAAQ;IACd,OAAO,EAAE;QACR;YACC,IAAI,EAAE,2BAA2B;YACjC,WAAW,EACV,uDAAuD;gBACvD,2FAA2F;SAC5F;KACD;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QAEzB,2FAA2F;QAC3F,yFAAyF;QACzF,6FAA6F;QAC7F,yFAAyF;QACzF,MAAM,WAAW,GAAG,IAAI,CAAC,eAAe,CAAA;QAExC,IAAI,CAAC,WAAW,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CACd,yFAAyF;gBACxF,qGAAqG;gBACrG,sFAAsF,CACvF,CAAA;QACF,CAAC;QAED,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAU,CAAA;QAElC,IAAI,KAAK,EAAE,MAAM,IAAI,IAAI,eAAe,CAAC,SAAS,CAAC,WAAW,CAAC,EAAE,CAAC;YACjE,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,CAAA;YAE3B,IAAI,OAAO,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;gBACzC,QAAQ,CAAC,GAAG,CAAC,OAAO,CAAC,CAAA;YACtB,CAAC;QACF,CAAC;QAED,IAAI,CAAC,QAAQ,CAAC,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,sCAAsC,WAAW,sBAAsB,CAAC,CAAA;QAE5G,IAAI,IAAI,GAAG,CAAC,CAAA;QACZ,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,YAAY,GAAG,CAAC,CAAA;QAEpB,IAAI,KAAK,EAAE,MAAM,CAAC,IAAI,UAAU,CAAC,IAAI,CAAC,KAAM,CAAC,EAAE,CAAC;YAC/C,IAAI,EAAE,CAAA;YACN,MAAM,MAAM,GAAG,MAAM,CAAC,CAAC,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAC5C,MAAM,IAAI,GAAG,MAAM,CAAC,CAAC,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAC5C,MAAM,MAAM,GAAG,MAAM,CAAC,CAAC,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAC5C,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAEhD,IAAI,CAAC,MAAM,IAAI,CAAC,IAAI,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAC9C,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,sDAAsD;YACtD,IAAI,QAAQ,CAAC,GAAG,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;gBACzC,YAAY,EAAE,CAAA;gBAEd,SAAQ;YACT,CAAC;YAED,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,CAAA;YACtB,IAAI,GAAW,CAAA;YAEf,MAAM,UAAU,GAA2B;gBAC1C,MAAM;gBACN,YAAY,EAAE,MAAM;gBACpB,QAAQ;gBACR,QAAQ,EAAE,IAAI;aACd,CAAA;YAED,IAAI,KAAK,KAAK,CAAC,EAAE,CAAC;gBACjB,GAAG,GAAG,GAAG,MAAM,IAAI,MAAM,KAAK,QAAQ,IAAI,IAAI,EAAE,CAAA;YACjD,CAAC;iBAAM,IAAI,KAAK,KAAK,CAAC,EAAE,CAAC;gBACxB,GAAG,GAAG,GAAG,IAAI,KAAK,QAAQ,KAAK,MAAM,IAAI,MAAM,EAAE,CAAA;YAClD,CAAC;iBAAM,CAAC;gBACP,GAAG,GAAG,GAAG,QAAQ,IAAI,IAAI,KAAK,MAAM,IAAI,MAAM,EAAE,CAAA;YACjD,CAAC;YAED,MAAM,SAAS,GAAG,aAAa,CAAC,qBAAqB,EAAE,EAAE,GAAG,UAAU,EAAE,CAAC,EAAE,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;YAE5G,MAAM,SAAS,GAAG;gBACjB,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,qBAAqB;gBAC7B,SAAS;gBACT,cAAc,EAAE,QAAQ;gBACxB,OAAO,EACN,8HAA8H;aAC/H,CAAA;YAED,IAAI,aAAa,CAAC,KAAK,EAAE,SAAS,EAAE,eAAe,CAAC,EAAE,CAAC;gBACtD,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,YAAY,EAAE,CAAA;IAChD,CAAC;CACD,CAAA"}

package/out/recipes/bare/country.d.ts

@@ -0,0 +1,25 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `bare-country` — the country name as the WHOLE query. Addresses almost never consist of one bare
+ *   country name, so the class is out of the model's training distribution and the tag is a coin
+ *   flip: measured on the 2026-08-13 panel, France/Germany/United States/New Zealand parsed
+ *   `country` while Japan/China/Nigeria/Australia/Deutschland parsed `locality`. The retrieval-side
+ *   bare-country race (#1651) covers the answer either way; this slice closes the PARSE half so the
+ *   tag itself is right.
+ *
+ *   Source is the codex country table — every ISO canonical English name plus the curated surface
+ *   forms (endonyms and long-form names). The 2–3 letter CODES are deliberately excluded: a bare
+ *   `JP` or `GER` is an abbreviation register with its own ambiguity surface (`IN`, `DE`, `TO` are
+ *   English words), and teaching it here would be an unmeasured claim. Case augmentation is the
+ *   loader's job, as everywhere.
+ */
+import { type CorpusRecipe } from "#recipes/scaffold";
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export declare const bareCountryRecipe: CorpusRecipe;
+//# sourceMappingURL=country.d.ts.map

package/out/recipes/bare/country.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"country.d.ts","sourceRoot":"","sources":["../../../lib/recipes/bare/country.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAKH,OAAO,EAAiB,KAAK,YAAY,EAAiB,MAAM,mBAAmB,CAAA;AA+BnF;;;GAGG;AACH,eAAO,MAAM,iBAAiB,EAAE,YAqC/B,CAAA"}

package/out/recipes/bare/country.js

@@ -0,0 +1,88 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `bare-country` — the country name as the WHOLE query. Addresses almost never consist of one bare
+ *   country name, so the class is out of the model's training distribution and the tag is a coin
+ *   flip: measured on the 2026-08-13 panel, France/Germany/United States/New Zealand parsed
+ *   `country` while Japan/China/Nigeria/Australia/Deutschland parsed `locality`. The retrieval-side
+ *   bare-country race (#1651) covers the answer either way; this slice closes the PARSE half so the
+ *   tag itself is right.
+ *
+ *   Source is the codex country table — every ISO canonical English name plus the curated surface
+ *   forms (endonyms and long-form names). The 2–3 letter CODES are deliberately excluded: a bare
+ *   `JP` or `GER` is an abbreviation register with its own ambiguity surface (`IN`, `DE`, `TO` are
+ *   English words), and teaching it here would be an unmeasured claim. Case augmentation is the
+ *   loader's job, as everywhere.
+ */
+import { COUNTRY_SURFACE_FORMS, CountryNames, matchCountry } from "@mailwoman/codex/country";
+import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils";
+import { alignAndWrite, sliceSourceID } from "#recipes/scaffold";
+/**
+ * Surfaces shorter than this are the code register (`JP`, `GER`), not a name — excluded (see the module doc).
+ */
+const MIN_NAME_LENGTH = 4;
+/**
+ * Every (surface, iso2) pair the slice emits: the ISO canonical names plus each country's curated surface forms,
+ * deduplicated on the surface string (a form shared across countries — none known — would keep its first bearer).
+ */
+function* bareCountrySurfaces() {
+    const seen = new Set();
+    for (const [iso2, forms] of Object.entries(COUNTRY_SURFACE_FORMS)) {
+        for (const surface of forms) {
+            if (surface.length < MIN_NAME_LENGTH || seen.has(surface))
+                continue;
+            seen.add(surface);
+            yield { surface, iso2 };
+        }
+    }
+    for (const name of CountryNames) {
+        if (name.length < MIN_NAME_LENGTH || seen.has(name))
+            continue;
+        const matched = matchCountry(name);
+        if (!matched)
+            continue;
+        seen.add(name);
+        yield { surface: name, iso2: matched.iso2 };
+    }
+}
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export const bareCountryRecipe = {
+    name: "bare-country",
+    description: "The country name as the whole query (#1651 parse half): ISO names + curated endonyms, no codes",
+    mode: "generate",
+    async run(opts, write) {
+        // Seeded for parity with the other recipes; the surfaces themselves drive the content.
+        makeMulberry32(opts.seed);
+        let read = 0;
+        let emitted = 0;
+        let skipped = 0;
+        for (const { surface, iso2 } of bareCountrySurfaces()) {
+            read++;
+            const components = { country: surface };
+            const source_id = sliceSourceID("synth-bare-country", { country: surface, cc: iso2, v: String(read) });
+            const canonical = {
+                raw: surface,
+                components,
+                country: iso2,
+                locale: "und",
+                source: "synth-bare-country",
+                source_id,
+                corpus_version: "0.11.0",
+                license: "Synthetic — bare-country; surfaces from the codex country table (ISO 3166 + curated forms)",
+            };
+            if (alignAndWrite(write, canonical, "bare-country")) {
+                emitted++;
+            }
+            else {
+                skipped++;
+            }
+        }
+        return { read, emitted, skipped };
+    },
+};
+//# sourceMappingURL=country.js.map

package/out/recipes/bare/country.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"country.js","sourceRoot":"","sources":["../../../lib/recipes/bare/country.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,EAAE,qBAAqB,EAAE,YAAY,EAAE,YAAY,EAAE,MAAM,0BAA0B,CAAA;AAC5F,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AAEpE,OAAO,EAAE,aAAa,EAAqB,aAAa,EAAE,MAAM,mBAAmB,CAAA;AACnF;;GAEG;AACH,MAAM,eAAe,GAAG,CAAC,CAAA;AAEzB;;;GAGG;AACH,QAAQ,CAAC,CAAC,mBAAmB;IAC5B,MAAM,IAAI,GAAG,IAAI,GAAG,EAAU,CAAA;IAE9B,KAAK,MAAM,CAAC,IAAI,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,qBAAqB,CAAC,EAAE,CAAC;QACnE,KAAK,MAAM,OAAO,IAAI,KAAK,EAAE,CAAC;YAC7B,IAAI,OAAO,CAAC,MAAM,GAAG,eAAe,IAAI,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC;gBAAE,SAAQ;YACnE,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,CAAA;YACjB,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,CAAA;QACxB,CAAC;IACF,CAAC;IAED,KAAK,MAAM,IAAI,IAAI,YAAY,EAAE,CAAC;QACjC,IAAI,IAAI,CAAC,MAAM,GAAG,eAAe,IAAI,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC;YAAE,SAAQ;QAC7D,MAAM,OAAO,GAAG,YAAY,CAAC,IAAI,CAAC,CAAA;QAElC,IAAI,CAAC,OAAO;YAAE,SAAQ;QACtB,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,CAAA;QACd,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,CAAA;IAC5C,CAAC;AACF,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAiB;IAC9C,IAAI,EAAE,cAAc;IACpB,WAAW,EAAE,gGAAgG;IAC7G,IAAI,EAAE,UAAU;IAChB,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,uFAAuF;QACvF,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACzB,IAAI,IAAI,GAAG,CAAC,CAAA;QACZ,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QAEf,KAAK,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,mBAAmB,EAAE,EAAE,CAAC;YACvD,IAAI,EAAE,CAAA;YAEN,MAAM,UAAU,GAA2B,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;YAC/D,MAAM,SAAS,GAAG,aAAa,CAAC,oBAAoB,EAAE,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;YAEtG,MAAM,SAAS,GAAG;gBACjB,GAAG,EAAE,OAAO;gBACZ,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,KAAK;gBACb,MAAM,EAAE,oBAAoB;gBAC5B,SAAS;gBACT,cAAc,EAAE,QAAQ;gBACxB,OAAO,EAAE,4FAA4F;aACrG,CAAA;YAED,IAAI,aAAa,CAAC,KAAK,EAAE,SAAS,EAAE,cAAc,CAAC,EAAE,CAAC;gBACrD,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAClC,CAAC;CACD,CAAA"}

package/out/recipes/bare/postcode/eval.d.ts

@@ -0,0 +1,29 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Reserved bare-postcode capability cases. These strings are excluded from every country in the
+ *   training recipe and are read only by evaluation. Reservation is by normalized input string,
+ *   because `NNN NN` is shared by CZ, SK, SE and GR; excluding a Czech row alone could still train
+ *   the model on the identical Swedish input.
+ */
+export type BarePostcodeEvalFamily = "nnn_nn" | "nnnn_ll";
+export interface BarePostcodeEvalCase {
+    input: string;
+    country: "CZ" | "NL" | "SK";
+    family: BarePostcodeEvalFamily;
+}
+/**
+ * The 56 postcode strings measured across v5.0.0 through v5.5.0 and reserved from later training.
+ */
+export declare const BARE_POSTCODE_EVAL_CASES: readonly BarePostcodeEvalCase[];
+/**
+ * Normalize the model-visible postcode value for reservation checks.
+ */
+export declare function normalizeBarePostcodeSurface(surface: string): string;
+/**
+ * Whether a postcode string is reserved exclusively for the capability evaluation.
+ */
+export declare function isReservedBarePostcode(surface: string): boolean;
+//# sourceMappingURL=eval.d.ts.map

package/out/recipes/bare/postcode/eval.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval.d.ts","sourceRoot":"","sources":["../../../../lib/recipes/bare/postcode/eval.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,MAAM,MAAM,sBAAsB,GAAG,QAAQ,GAAG,SAAS,CAAA;AAEzD,MAAM,WAAW,oBAAoB;IACpC,KAAK,EAAE,MAAM,CAAA;IACb,OAAO,EAAE,IAAI,GAAG,IAAI,GAAG,IAAI,CAAA;IAC3B,MAAM,EAAE,sBAAsB,CAAA;CAC9B;AA0DD;;GAEG;AACH,eAAO,MAAM,wBAAwB,EAAE,SAAS,oBAAoB,EAInE,CAAA;AAED;;GAEG;AACH,wBAAgB,4BAA4B,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAEpE;AAID;;GAEG;AACH,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAE/D"}

package/out/recipes/bare/postcode/eval.js

@@ -0,0 +1,85 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Reserved bare-postcode capability cases. These strings are excluded from every country in the
+ *   training recipe and are read only by evaluation. Reservation is by normalized input string,
+ *   because `NNN NN` is shared by CZ, SK, SE and GR; excluding a Czech row alone could still train
+ *   the model on the identical Swedish input.
+ */
+const SPACED_FIVE_CZ = [
+    "120 00",
+    "130 00",
+    "140 00",
+    "150 00",
+    "160 00",
+    "170 00",
+    "180 00",
+    "190 00",
+    "602 00",
+    "603 00",
+    "612 00",
+    "616 00",
+    "621 00",
+    "628 00",
+    "635 00",
+    "639 00",
+    "702 00",
+    "708 00",
+    "710 00",
+    "721 00",
+    "779 00",
+    "746 01",
+    "747 05",
+    "750 02",
+];
+const SPACED_FIVE_SK = ["811 02", "821 01", "831 02", "841 01", "851 01", "949 01", "974 01", "010 01"];
+const DUTCH = [
+    "1011 AB",
+    "1013 BG",
+    "1015 CJ",
+    "1017 DK",
+    "1018 EL",
+    "1052 GM",
+    "1053 HN",
+    "1054 JP",
+    "2011 KR",
+    "2012 LS",
+    "2511 MT",
+    "2513 NV",
+    "3011 PW",
+    "3012 RX",
+    "3511 SA",
+    "3512 TB",
+    "4811 VC",
+    "5611 WD",
+    "6211 XE",
+    "6511 ZF",
+    "7511 AG",
+    "8011 BH",
+    "9711 CJ",
+    "9712 DK",
+];
+/**
+ * The 56 postcode strings measured across v5.0.0 through v5.5.0 and reserved from later training.
+ */
+export const BARE_POSTCODE_EVAL_CASES = [
+    ...SPACED_FIVE_CZ.map((input) => ({ input, country: "CZ", family: "nnn_nn" })),
+    ...SPACED_FIVE_SK.map((input) => ({ input, country: "SK", family: "nnn_nn" })),
+    ...DUTCH.map((input) => ({ input, country: "NL", family: "nnnn_ll" })),
+];
+/**
+ * Normalize the model-visible postcode value for reservation checks.
+ */
+export function normalizeBarePostcodeSurface(surface) {
+    return surface.trim().toUpperCase().replaceAll(/\s+/gu, "");
+}
+const RESERVED_SURFACES = new Set(BARE_POSTCODE_EVAL_CASES.map(({ input }) => normalizeBarePostcodeSurface(input)));
+/**
+ * Whether a postcode string is reserved exclusively for the capability evaluation.
+ */
+export function isReservedBarePostcode(surface) {
+    return RESERVED_SURFACES.has(normalizeBarePostcodeSurface(surface));
+}
+//# sourceMappingURL=eval.js.map

package/out/recipes/bare/postcode/eval.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval.js","sourceRoot":"","sources":["../../../../lib/recipes/bare/postcode/eval.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAUH,MAAM,cAAc,GAAG;IACtB,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;IACR,QAAQ;CACC,CAAA;AAEV,MAAM,cAAc,GAAG,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAU,CAAA;AAEhH,MAAM,KAAK,GAAG;IACb,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;IACT,SAAS;CACA,CAAA;AAEV;;GAEG;AACH,MAAM,CAAC,MAAM,wBAAwB,GAAoC;IACxE,GAAG,cAAc,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,IAAa,EAAE,MAAM,EAAE,QAAiB,EAAE,CAAC,CAAC;IAChG,GAAG,cAAc,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,IAAa,EAAE,MAAM,EAAE,QAAiB,EAAE,CAAC,CAAC;IAChG,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,IAAa,EAAE,MAAM,EAAE,SAAkB,EAAE,CAAC,CAAC;CACxF,CAAA;AAED;;GAEG;AACH,MAAM,UAAU,4BAA4B,CAAC,OAAe;IAC3D,OAAO,OAAO,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAA;AAC5D,CAAC;AAED,MAAM,iBAAiB,GAAG,IAAI,GAAG,CAAC,wBAAwB,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC,4BAA4B,CAAC,KAAK,CAAC,CAAC,CAAC,CAAA;AAEnH;;GAEG;AACH,MAAM,UAAU,sBAAsB,CAAC,OAAe;IACrD,OAAO,iBAAiB,CAAC,GAAG,CAAC,4BAA4B,CAAC,OAAO,CAAC,CAAC,CAAA;AACpE,CAAC"}

package/out/recipes/bare/postcode/index.d.ts

@@ -0,0 +1,62 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `bare-postcode` — a postcode standing ALONE, in the form its country writes.
+ *
+ *   The shape is absent from the corpus. A census of the 685,187,151 rows the v0.29.0 mixture reads
+ *   found ZERO two-token rows of `NNN NN` and one of `NNNN LL`, so `100 00` and `1012 LG` are out of
+ *   distribution and every model's answer for them is generalization off longer lines. That
+ *   generalization is what moved: held-out CZ/SK codes read 31/32 as a postcode under v5.0.0, v5.2.0
+ *   AND v5.3.0 — a seed pair agreeing, so not training noise — and 0/32 under v5.4.0, whose added
+ *   OpenStreetMap PK/BD/VN rows raise the leading-digit-is-a-house-number evidence at that opening
+ *   from 393 to 444 samples per epoch.
+ *
+ *   Neither existing repair reaches it. `nl-postcode` and `cz-pcfirst-preposition` teach the same
+ *   postcodes IN CONTEXT — every row they emit carries a street, a number and a city — and the
+ *   12,759 in-context `NNN NN` rows do not transfer to the bare input. The deterministic half,
+ *   `buildEmissionPriors`, caps near 0.95 at the default `biasScale` against a measured 1.67-to-3.78
+ *   nat deficit.
+ *
+ *   So this recipe emits the postcode and NOTHING else, which is the one thing no sibling does.
+ *
+ *   VERIFIED AGAINST THE PRIOR, NOT MERELY MATCHED TO IT. Every surface is run through
+ *   `detectKnownFormats` and refused unless the detector calls it a postcode. The recipe's rendering
+ *   table and `known-formats.ts`'s patterns have to agree, and matched constants would not prove
+ *   they do — the trained surface and the prior that boosts it come from one check.
+ */
+import { type CorpusRecipe } from "#recipes/scaffold";
+/**
+ * Choose distinct postcodes without inheriting the publisher's row order.
+ *
+ * Sorting first makes the result independent of input-file order; the seeded sample then gives a reproducible spread
+ * across the complete set instead of taking the first municipality or numeric prefix that happens to fill the cap.
+ */
+export declare function selectPostcodes(codes: Iterable<string>, limit: number, seed: number): string[];
+/**
+ * Return every required input path that is absent, preserving declaration order for diagnostics.
+ */
+export declare function findMissingPostcodeSources(paths: readonly string[], exists?: (path: string) => Promise<boolean>): Promise<string[]>;
+/**
+ * Every surface a country writes for one postcode, spaced form first, or `[]` when this recipe carries no form for that
+ * country or the code does not fit the one it carries.
+ *
+ * Exported because it is the half of the recipe a test can reach: `run` reads a 500 MB archive from the data root, and
+ * the contract worth pinning — that every surface this renders is one {@linkcode detectedAsPostcode} accepts — needs
+ * neither.
+ */
+export declare function renderBarePostcode(country: string, postcode: string): string[];
+/**
+ * Whether `known-formats.ts` reads this surface as a postcode across its whole span.
+ *
+ * A surface the detector does not recognize would train the model on a string the query-shape prior cannot then
+ * support, which is the divergence this recipe exists to close rather than widen.
+ */
+export declare function detectedAsPostcode(surface: string): boolean;
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export declare const barePostcodeRecipe: CorpusRecipe;
+//# sourceMappingURL=index.d.ts.map

package/out/recipes/bare/postcode/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../lib/recipes/bare/postcode/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAUH,OAAO,EAAiB,KAAK,YAAY,EAAiC,MAAM,mBAAmB,CAAA;AA+FnG;;;;;GAKG;AACH,wBAAgB,eAAe,CAAC,KAAK,EAAE,QAAQ,CAAC,MAAM,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,CAM9F;AAED;;GAEG;AACH,wBAAsB,0BAA0B,CAC/C,KAAK,EAAE,SAAS,MAAM,EAAE,EACxB,MAAM,GAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC,OAAO,CAAc,GACrD,OAAO,CAAC,MAAM,EAAE,CAAC,CAInB;AAED;;;;;;;GAOG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,MAAM,EAAE,CAM9E;AAED;;;;;GAKG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAI3D;AAED;;;GAGG;AACH,eAAO,MAAM,kBAAkB,EAAE,YAwIhC,CAAA"}