@mailwoman/corpus 6.3.0 → 6.5.0

package/out/src/shard-recipes/fr-fragment.d.ts

@@ -0,0 +1,73 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-fragment` shard recipe (#727 T2) — the HOUSE-NUMBER-LICENCE lever.
+ *
+ *   The measured problem (T1c, `2026-07-16-t1c-fragment-board-verdict.md`): the shipped model scores
+ *   **0.925** on `<n> Rue X` and **0.215** on `Rue X`. Same streets, same model; the only difference
+ *   is a leading number. It has learned that a house number LICENSES a street reading at all — strip
+ *   it and a designator-led phrase parses as a locality, designator included:
+ *
+ *   ```
+ *   "Rue Montmartre"        -> locality="Rue Montmartre"
+ *   "Allee Poque"           -> locality="Allee Poque"
+ *   ```
+ *
+ *   `Rue` can only mean street in French. The model is not mislabelling an ambiguous toponym; it is
+ *   mislabelling `Rue`.
+ *
+ *   WHY THE EXISTING RECIPE DOESN'T COVER THIS. {@link frBareStreetRecipe} (#251) mints
+ *   `<n> Rue <name>, <City>` — the bare COMMA form, no postcode. It targets postcode-anchoring
+ *   imbalance, and every row it emits still carries a house number AND a locality. It cannot teach
+ *   the class above, because it never shows the model a street standing alone.
+ *
+ *   WHAT THIS MINTS. Five street forms — the first three carry NO house number and NO locality, which
+ *   is the whole point — plus the counter-distribution:
+ *
+ *   1. `bare-street`         "Rue Montmartre"
+ *   2. `street-particle`     "Rue de la Paix"
+ *   3. `date-name`           "Allee du 11 Novembre 1918"
+ *   4. `street-housenumber`  "12 Rue Montmartre"     — the anchor, so the licence isn't UNLEARNED
+ *   5. `alnum-housenumber`   "12 bis Rue Montmartre"
+ *   6. `bare-locality`       "Mery-sur-Oise"         — NEGATIVE: a bare toponym that IS a locality
+ *
+ *   The **admin/street homonym** class (`Rue de Rome`) is deliberately NOT a separate form. Homonym
+ *   streets are already in the register and fall into 1–3 naturally; there is nothing different to
+ *   teach about them, because the lesson is identical — the designator makes it a street. The
+ *   fragment board scores them separately (they need measuring, not minting), and the recipe would
+ *   need a second pass over the commune set to label them, buying a `synth_method` string and no
+ *   training signal.
+ *
+ *   FORM 7 IS NOT OPTIONAL. T1c's standing prediction: the board's `bare-locality` cell reads 0.980
+ *   for the WRONG REASON — the model calls everything without a house number a locality, and on bare
+ *   localities that is accidentally right. Teach bare streets alone and the model has every incentive
+ *   to flip that default rather than learn the distinction, trading a 0.215 for a 0.980. The shard
+ *   must show BOTH bare forms so the discriminating evidence is the designator, which is the only
+ *   thing that actually distinguishes them. This is the same counter-distribution principle
+ *   {@link noStreetRecipe} established after synth-street pushed the model into "decompose mode".
+ *
+ *   SPLIT. `--exclude-surfaces` takes the fragment board's reserved surface list
+ *   (`mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt`). Every listed surface is
+ *   skipped — source-disjoint by normalized street SURFACE, never by record row. Row-disjoint leaks
+ *   the surface across the boundary and measures memorization of `Rue de Rivoli` while claiming
+ *   generalization to unseen streets. The recipe REFUSES to run without the list rather than
+ *   silently minting a contaminated shard.
+ *
+ *   MIX. The shard is ~145K rows off a 120K-tuple draw; the intended corpus mix is **5–10%**, set at
+ *   assembly time by shard weight rather than by row count. Keep the cap: a shard that fixes fragments
+ *   by degrading full addresses has moved the failure, not fixed it — which is what
+ *   `street-housenumber` / `alnum-housenumber` on the fragment board and the global parity floor are
+ *   there to catch. `date-name` is ~0.5% of the shard because BAN only holds ~1,418 date-name streets
+ *   after filtering and the tuple extractor already takes every one; if that class needs more, the
+ *   lever is shard weight, not invented data.
+ *
+ *   ⚠ Convention loss-mask: like {@link frBareStreetRecipe}, this recipe TEACHES FR `street_prefix`.
+ *   The conventions loss-mask forbids it for FR and will `-inf` these gold labels (the v1.6.0 ~7M-loss
+ *   blow-up). Disable that mask for any run including this shard.
+ */
+import { type ShardRecipe } from "./scaffold.ts";
+export declare function frTitleCase(value: string): string;
+export declare const frFragmentRecipe: ShardRecipe;
+//# sourceMappingURL=fr-fragment.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"fr-fragment.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/fr-fragment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoEG;AAKH,OAAO,EAA6C,KAAK,WAAW,EAAiB,MAAM,eAAe,CAAA;AA4C1G,wBAAgB,WAAW,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAajD;AAQD,eAAO,MAAM,gBAAgB,EAAE,WAmK9B,CAAA"}

package/out/src/shard-recipes/fr-fragment.js

@@ -0,0 +1,255 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-fragment` shard recipe (#727 T2) — the HOUSE-NUMBER-LICENCE lever.
+ *
+ *   The measured problem (T1c, `2026-07-16-t1c-fragment-board-verdict.md`): the shipped model scores
+ *   **0.925** on `<n> Rue X` and **0.215** on `Rue X`. Same streets, same model; the only difference
+ *   is a leading number. It has learned that a house number LICENSES a street reading at all — strip
+ *   it and a designator-led phrase parses as a locality, designator included:
+ *
+ *   ```
+ *   "Rue Montmartre"        -> locality="Rue Montmartre"
+ *   "Allee Poque"           -> locality="Allee Poque"
+ *   ```
+ *
+ *   `Rue` can only mean street in French. The model is not mislabelling an ambiguous toponym; it is
+ *   mislabelling `Rue`.
+ *
+ *   WHY THE EXISTING RECIPE DOESN'T COVER THIS. {@link frBareStreetRecipe} (#251) mints
+ *   `<n> Rue <name>, <City>` — the bare COMMA form, no postcode. It targets postcode-anchoring
+ *   imbalance, and every row it emits still carries a house number AND a locality. It cannot teach
+ *   the class above, because it never shows the model a street standing alone.
+ *
+ *   WHAT THIS MINTS. Five street forms — the first three carry NO house number and NO locality, which
+ *   is the whole point — plus the counter-distribution:
+ *
+ *   1. `bare-street`         "Rue Montmartre"
+ *   2. `street-particle`     "Rue de la Paix"
+ *   3. `date-name`           "Allee du 11 Novembre 1918"
+ *   4. `street-housenumber`  "12 Rue Montmartre"     — the anchor, so the licence isn't UNLEARNED
+ *   5. `alnum-housenumber`   "12 bis Rue Montmartre"
+ *   6. `bare-locality`       "Mery-sur-Oise"         — NEGATIVE: a bare toponym that IS a locality
+ *
+ *   The **admin/street homonym** class (`Rue de Rome`) is deliberately NOT a separate form. Homonym
+ *   streets are already in the register and fall into 1–3 naturally; there is nothing different to
+ *   teach about them, because the lesson is identical — the designator makes it a street. The
+ *   fragment board scores them separately (they need measuring, not minting), and the recipe would
+ *   need a second pass over the commune set to label them, buying a `synth_method` string and no
+ *   training signal.
+ *
+ *   FORM 7 IS NOT OPTIONAL. T1c's standing prediction: the board's `bare-locality` cell reads 0.980
+ *   for the WRONG REASON — the model calls everything without a house number a locality, and on bare
+ *   localities that is accidentally right. Teach bare streets alone and the model has every incentive
+ *   to flip that default rather than learn the distinction, trading a 0.215 for a 0.980. The shard
+ *   must show BOTH bare forms so the discriminating evidence is the designator, which is the only
+ *   thing that actually distinguishes them. This is the same counter-distribution principle
+ *   {@link noStreetRecipe} established after synth-street pushed the model into "decompose mode".
+ *
+ *   SPLIT. `--exclude-surfaces` takes the fragment board's reserved surface list
+ *   (`mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt`). Every listed surface is
+ *   skipped — source-disjoint by normalized street SURFACE, never by record row. Row-disjoint leaks
+ *   the surface across the boundary and measures memorization of `Rue de Rivoli` while claiming
+ *   generalization to unseen streets. The recipe REFUSES to run without the list rather than
+ *   silently minting a contaminated shard.
+ *
+ *   MIX. The shard is ~145K rows off a 120K-tuple draw; the intended corpus mix is **5–10%**, set at
+ *   assembly time by shard weight rather than by row count. Keep the cap: a shard that fixes fragments
+ *   by degrading full addresses has moved the failure, not fixed it — which is what
+ *   `street-housenumber` / `alnum-housenumber` on the fragment board and the global parity floor are
+ *   there to catch. `date-name` is ~0.5% of the shard because BAN only holds ~1,418 date-name streets
+ *   after filtering and the tuple extractor already takes every one; if that class needs more, the
+ *   lever is shard weight, not invented data.
+ *
+ *   ⚠ Convention loss-mask: like {@link frBareStreetRecipe}, this recipe TEACHES FR `street_prefix`.
+ *   The conventions loss-mask forbids it for FR and will `-inf` these gold labels (the v1.6.0 ~7M-loss
+ *   blow-up). Disable that mask for any run including this shard.
+ */
+import { readFileSync } from "node:fs";
+import { decomposeFrStreet } from "../adapters/ban/street-decompose.js";
+import { alignAndWrite, makeMulberry32, readTuples, shardSourceID } from "./scaffold.js";
+/** House numbers, weighted toward the small values that dominate real BAN rows. */
+const HOUSE_NUMBERS = [
+    1, 1, 2, 2, 3, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 16, 18, 20, 21, 24, 27, 30, 33, 42, 57, 68, 84, 102, 115, 140,
+];
+/** FR alphanumeric house-number forms. `bis`/`ter` are separated; a bare letter is suffixed. */
+const ALNUM_SUFFIXES = ["bis", "ter", "A", "B"];
+const norm = (value) => value
+    .normalize("NFD")
+    .replace(/[̀-ͯ]/g, "")
+    .toLowerCase()
+    .replace(/\s+/g, " ")
+    .trim();
+/**
+ * French commune convention: capitalize each element, leave the joining particles lowercase. `saint-jean-de-luz` →
+ * `Saint-Jean-de-Luz`, not `Saint-Jean-De-Luz`.
+ *
+ * Needed because BAN's sharded DBs keep only `locality_base` — normalized, lowercase, accent-stripped. Emitting that
+ * verbatim would teach the counter-distribution that a lowercase accent-stripped string is a locality, which is not a
+ * fact about French and would not match the fragment board (which reconstructs the same casing). The accents are gone
+ * from the source and cannot be recovered here; the casing can.
+ */
+const FR_LOWER = new Set([
+    "le",
+    "la",
+    "les",
+    "de",
+    "du",
+    "des",
+    "d",
+    "l",
+    "sur",
+    "sous",
+    "en",
+    "aux",
+    "au",
+    "et",
+    "lez",
+]);
+export function frTitleCase(value) {
+    const cap = (token, first) => !first && FR_LOWER.has(token) ? token : token.charAt(0).toUpperCase() + token.slice(1);
+    return value
+        .split(" ")
+        .map((word, wordIndex) => word
+        .split("-")
+        .map((bit, bitIndex) => cap(bit, wordIndex === 0 && bitIndex === 0))
+        .join("-"))
+        .join(" ");
+}
+/** Does the street name carry a particle? Decides the particle vs bare classification. */
+const PARTICLE = /\b(de la|de l'|du|des|de|d'|le|la|les)\b/i;
+/** Does the street name carry date material (a year, or a day + French month)? */
+const DATEISH = /\b(1[0-9]|20)\d{2}\b|\b\d{1,2}\s+(janvier|f[ée]vrier|mars|avril|mai|juin|juillet|ao[ûu]t|septembre|octobre|novembre|d[ée]cembre)\b/i;
+export const frFragmentRecipe = {
+    name: "fr-fragment",
+    description: "FR street fragments with NO house number (#727 T2): the house-number-licence lever — bare/particle/date-name/homonym + the bare-locality counter",
+    mode: "tuples",
+    options: [
+        {
+            flag: "--exclude-surfaces <path>",
+            description: "REQUIRED. The fragment board's reserved surface list; every listed street is skipped.",
+        },
+        {
+            flag: "--hn-prob <n>",
+            description: "Share of rows carrying a house number (default 0.35 — the anchor, not the point)",
+        },
+        {
+            flag: "--bare-prob <n>",
+            description: "Share of NO-house-number rows that are bare LOCALITIES (default 0.25 — the counter)",
+        },
+    ],
+    async run(opts, write) {
+        const random = makeMulberry32(opts.seed);
+        const excludePath = opts.excludeSurfaces;
+        if (!excludePath) {
+            throw new Error("fr-fragment: --exclude-surfaces is REQUIRED. Pass the fragment board's reserved list " +
+                "(mailwoman/eval-harness/fixtures/ban-fragments-fr.surfaces.txt) or this shard trains on its own eval set. " +
+                "Source-disjoint by street SURFACE is the split discipline; there is no safe default.");
+        }
+        const excluded = new Set(readFileSync(excludePath, "utf8")
+            .split("\n")
+            .map((line) => line.trim())
+            .filter((line) => line && !line.startsWith("#")));
+        if (!excluded.size)
+            throw new Error(`fr-fragment: --exclude-surfaces "${excludePath}" listed no surfaces`);
+        const hnProb = opts.hnProb ?? 0.35;
+        const bareLocalityProb = opts.bareProb ?? 0.25;
+        // The locality pool is harvested from the tuples themselves — every BAN row carries its
+        // commune, so the counter-distribution needs no second source.
+        const localities = new Set();
+        let read = 0;
+        let emitted = 0;
+        let skipped = 0;
+        let contaminated = 0;
+        for await (const tuple of readTuples(opts.input)) {
+            read++;
+            const fullStreet = String(tuple.street ?? "").trim();
+            const locality = String(tuple.locality ?? "").trim();
+            if (locality)
+                localities.add(locality);
+            if (!fullStreet) {
+                skipped++;
+                continue;
+            }
+            // THE SPLIT. A surface on the fragment board never enters training.
+            if (excluded.has(norm(fullStreet))) {
+                contaminated++;
+                continue;
+            }
+            const { prefix, street } = decomposeFrStreet(fullStreet);
+            // The failing class is the designator-led street. A no-prefix nom_voie ("La Ville Mois")
+            // is a different problem and would muddy the signal.
+            if (!prefix || !street) {
+                skipped++;
+                continue;
+            }
+            const carriesNumber = random() < hnProb;
+            const components = { street_prefix: prefix, street };
+            let raw = `${prefix} ${street}`;
+            let klass = DATEISH.test(street) ? "date-name" : PARTICLE.test(street) ? "street-particle" : "bare-street";
+            if (carriesNumber) {
+                const number = HOUSE_NUMBERS[Math.floor(random() * HOUSE_NUMBERS.length)];
+                const alnum = random() < 0.25;
+                const suffix = ALNUM_SUFFIXES[Math.floor(random() * ALNUM_SUFFIXES.length)];
+                const houseNumber = alnum
+                    ? suffix === "bis" || suffix === "ter"
+                        ? `${number} ${suffix}`
+                        : `${number}${suffix}`
+                    : String(number);
+                components.house_number = houseNumber;
+                raw = `${houseNumber} ${prefix} ${street}`;
+                klass = alnum ? "alnum-housenumber" : "street-housenumber";
+            }
+            const sourceID = shardSourceID("synth-fr-fragment", { ...components, v: String(read) });
+            if (alignAndWrite(write, {
+                raw,
+                components,
+                country: "FR",
+                locale: "fr-FR",
+                source: "synth-fr-fragment",
+                source_id: sourceID,
+                corpus_version: "0.9.4",
+                license: "Synthetic — fr-fragment; (street, commune) from BAN (Base Adresse Nationale, Licence Ouverte)",
+            }, `fr-fragment:${klass}`)) {
+                emitted++;
+            }
+            else {
+                skipped++;
+            }
+        }
+        // ---- the counter-distribution: bare localities, NO street anywhere in the row -------------
+        // Minted last so the locality pool is complete. Without these the model can satisfy every row
+        // above by flipping its default from "bare => locality" to "bare => street", which trades one
+        // broken prior for another and would show up as bare-locality collapsing on the board.
+        const pool = [...localities].sort();
+        const wanted = Math.round((emitted / Math.max(1, 1 - bareLocalityProb)) * bareLocalityProb);
+        for (let i = 0; i < wanted && pool.length; i++) {
+            // BAN gives `locality_base` normalized; restore the casing the fragment board also
+            // reconstructs, so train and eval show the model the same shape of French.
+            const name = frTitleCase(pool[Math.floor(random() * pool.length)]);
+            const sourceID = shardSourceID("synth-fr-fragment", { locality: name, v: `neg-${i}` });
+            if (alignAndWrite(write, {
+                raw: name,
+                components: { locality: name },
+                country: "FR",
+                locale: "fr-FR",
+                source: "synth-fr-fragment",
+                source_id: sourceID,
+                corpus_version: "0.9.4",
+                license: "Synthetic — fr-fragment counter-distribution; commune from BAN (Licence Ouverte)",
+            }, "fr-fragment:bare-locality")) {
+                emitted++;
+            }
+            else {
+                skipped++;
+            }
+        }
+        if (contaminated) {
+            console.error(`fr-fragment: skipped ${contaminated} rows whose street surface is reserved by the fragment board`);
+        }
+        return { read, emitted, skipped };
+    },
+};
+//# sourceMappingURL=fr-fragment.js.map

package/out/src/shard-recipes/fr-fragment.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fr-fragment.js","sourceRoot":"","sources":["../../../src/shard-recipes/fr-fragment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoEG;AAEH,OAAO,EAAE,YAAY,EAAE,MAAM,SAAS,CAAA;AAEtC,OAAO,EAAE,iBAAiB,EAAE,MAAM,qCAAqC,CAAA;AACvE,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,UAAU,EAAoB,aAAa,EAAE,MAAM,eAAe,CAAA;AAE1G,mFAAmF;AACnF,MAAM,aAAa,GAAG;IACrB,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG;CACrH,CAAA;AACD,gGAAgG;AAChG,MAAM,cAAc,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,GAAG,CAAC,CAAA;AAE/C,MAAM,IAAI,GAAG,CAAC,KAAa,EAAU,EAAE,CACtC,KAAK;KACH,SAAS,CAAC,KAAK,CAAC;KAChB,OAAO,CAAC,QAAQ,EAAE,EAAE,CAAC;KACrB,WAAW,EAAE;KACb,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC;KACpB,IAAI,EAAE,CAAA;AAET;;;;;;;;GAQG;AACH,MAAM,QAAQ,GAAG,IAAI,GAAG,CAAC;IACxB,IAAI;IACJ,IAAI;IACJ,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,KAAK;IACL,GAAG;IACH,GAAG;IACH,KAAK;IACL,MAAM;IACN,IAAI;IACJ,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,KAAK;CACL,CAAC,CAAA;AAEF,MAAM,UAAU,WAAW,CAAC,KAAa;IACxC,MAAM,GAAG,GAAG,CAAC,KAAa,EAAE,KAAc,EAAU,EAAE,CACrD,CAAC,KAAK,IAAI,QAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAA;IAEvF,OAAO,KAAK;SACV,KAAK,CAAC,GAAG,CAAC;SACV,GAAG,CAAC,CAAC,IAAI,EAAE,SAAS,EAAE,EAAE,CACxB,IAAI;SACF,KAAK,CAAC,GAAG,CAAC;SACV,GAAG,CAAC,CAAC,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,EAAE,SAAS,KAAK,CAAC,IAAI,QAAQ,KAAK,CAAC,CAAC,CAAC;SACnE,IAAI,CAAC,GAAG,CAAC,CACX;SACA,IAAI,CAAC,GAAG,CAAC,CAAA;AACZ,CAAC;AAED,0FAA0F;AAC1F,MAAM,QAAQ,GAAG,2CAA2C,CAAA;AAC5D,kFAAkF;AAClF,MAAM,OAAO,GACZ,qIAAqI,CAAA;AAEtI,MAAM,CAAC,MAAM,gBAAgB,GAAgB;IAC5C,IAAI,EAAE,aAAa;IACnB,WAAW,EACV,kJAAkJ;IACnJ,IAAI,EAAE,QAAQ;IACd,OAAO,EAAE;QACR;YACC,IAAI,EAAE,2BAA2B;YACjC,WAAW,EAAE,uFAAuF;SACpG;QACD;YACC,IAAI,EAAE,eAAe;YACrB,WAAW,EAAE,kFAAkF;SAC/F;QACD;YACC,IAAI,EAAE,iBAAiB;YACvB,WAAW,EAAE,qFAAqF;SAClG;KACD;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,WAAW,GAAG,IAAI,CAAC,eAAe,CAAA;QAExC,IAAI,CAAC,WAAW,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CACd,uFAAuF;gBACtF,4GAA4G;gBAC5G,sFAAsF,CACvF,CAAA;QACF,CAAC;QAED,MAAM,QAAQ,GAAG,IAAI,GAAG,CACvB,YAAY,CAAC,WAAW,EAAE,MAAM,CAAC;aAC/B,KAAK,CAAC,IAAI,CAAC;aACX,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;aAC1B,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,CACjD,CAAA;QAED,IAAI,CAAC,QAAQ,CAAC,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,oCAAoC,WAAW,sBAAsB,CAAC,CAAA;QAE1G,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAA;QAClC,MAAM,gBAAgB,GAAG,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAA;QAE9C,wFAAwF;QACxF,+DAA+D;QAC/D,MAAM,UAAU,GAAG,IAAI,GAAG,EAAU,CAAA;QAEpC,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,KAAK,IAAI,UAAU,CAAC,IAAI,CAAC,KAAM,CAAC,EAAE,CAAC;YACnD,IAAI,EAAE,CAAA;YACN,MAAM,UAAU,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YACpD,MAAM,QAAQ,GAAG,MAAM,CAAC,KAAK,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAEpD,IAAI,QAAQ;gBAAE,UAAU,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;YAEtC,IAAI,CAAC,UAAU,EAAE,CAAC;gBACjB,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YAED,oEAAoE;YACpE,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC;gBACpC,YAAY,EAAE,CAAA;gBACd,SAAQ;YACT,CAAC;YAED,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,iBAAiB,CAAC,UAAU,CAAC,CAAA;YAExD,yFAAyF;YACzF,qDAAqD;YACrD,IAAI,CAAC,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC;gBACxB,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YAED,MAAM,aAAa,GAAG,MAAM,EAAE,GAAG,MAAM,CAAA;YACvC,MAAM,UAAU,GAA2B,EAAE,aAAa,EAAE,MAAM,EAAE,MAAM,EAAE,CAAA;YAC5E,IAAI,GAAG,GAAG,GAAG,MAAM,IAAI,MAAM,EAAE,CAAA;YAC/B,IAAI,KAAK,GAAG,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,aAAa,CAAA;YAE1G,IAAI,aAAa,EAAE,CAAC;gBACnB,MAAM,MAAM,GAAG,aAAa,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,aAAa,CAAC,MAAM,CAAC,CAAE,CAAA;gBAC1E,MAAM,KAAK,GAAG,MAAM,EAAE,GAAG,IAAI,CAAA;gBAC7B,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,cAAc,CAAC,MAAM,CAAC,CAAE,CAAA;gBAC5E,MAAM,WAAW,GAAG,KAAK;oBACxB,CAAC,CAAC,MAAM,KAAK,KAAK,IAAI,MAAM,KAAK,KAAK;wBACrC,CAAC,CAAC,GAAG,MAAM,IAAI,MAAM,EAAE;wBACvB,CAAC,CAAC,GAAG,MAAM,GAAG,MAAM,EAAE;oBACvB,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAA;gBACjB,UAAU,CAAC,YAAY,GAAG,WAAW,CAAA;gBACrC,GAAG,GAAG,GAAG,WAAW,IAAI,MAAM,IAAI,MAAM,EAAE,CAAA;gBAC1C,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,mBAAmB,CAAC,CAAC,CAAC,oBAAoB,CAAA;YAC3D,CAAC;YAED,MAAM,QAAQ,GAAG,aAAa,CAAC,mBAAmB,EAAE,EAAE,GAAG,UAAU,EAAE,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;YAEvF,IACC,aAAa,CACZ,KAAK,EACL;gBACC,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,mBAAmB;gBAC3B,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,+FAA+F;aACxG,EACD,eAAe,KAAK,EAAE,CACtB,EACA,CAAC;gBACF,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,8FAA8F;QAC9F,8FAA8F;QAC9F,8FAA8F;QAC9F,uFAAuF;QACvF,MAAM,IAAI,GAAG,CAAC,GAAG,UAAU,CAAC,CAAC,IAAI,EAAE,CAAA;QACnC,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,gBAAgB,CAAC,CAAC,GAAG,gBAAgB,CAAC,CAAA;QAE3F,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAChD,mFAAmF;YACnF,2EAA2E;YAC3E,MAAM,IAAI,GAAG,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,CAAE,CAAC,CAAA;YACnE,MAAM,QAAQ,GAAG,aAAa,CAAC,mBAAmB,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,EAAE,CAAC,CAAA;YAEtF,IACC,aAAa,CACZ,KAAK,EACL;gBACC,GAAG,EAAE,IAAI;gBACT,UAAU,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE;gBAC9B,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,mBAAmB;gBAC3B,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,kFAAkF;aAC3F,EACD,2BAA2B,CAC3B,EACA,CAAC;gBACF,OAAO,EAAE,CAAA;YACV,CAAC;iBAAM,CAAC;gBACP,OAAO,EAAE,CAAA;YACV,CAAC;QACF,CAAC;QAED,IAAI,YAAY,EAAE,CAAC;YAClB,OAAO,CAAC,KAAK,CAAC,wBAAwB,YAAY,8DAA8D,CAAC,CAAA;QAClH,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAClC,CAAC;CACD,CAAA"}

package/out/src/shard-recipes/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAkBH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAA;AAOhD,cAAc,eAAe,CAAA;AA2B7B,4BAA4B;AAC5B,eAAO,MAAM,aAAa,EAAE,WAAW,CAAC,MAAM,EAAE,WAAW,CAA4C,CAAA;AAEvG,+CAA+C;AAC/C,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,SAAS,CAEpE;AAED,yCAAyC;AACzC,wBAAgB,gBAAgB,IAAI,SAAS,WAAW,EAAE,CAEzD"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAoBH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAA;AAOhD,cAAc,eAAe,CAAA;AA6B7B,4BAA4B;AAC5B,eAAO,MAAM,aAAa,EAAE,WAAW,CAAC,MAAM,EAAE,WAAW,CAA4C,CAAA;AAEvG,+CAA+C;AAC/C,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,SAAS,CAEpE;AAED,yCAAyC;AACzC,wBAAgB,gBAAgB,IAAI,SAAS,WAAW,EAAE,CAEzD"}

package/out/src/shard-recipes/index.js

@@ -13,12 +13,14 @@ import { countryBalancedRecipe } from "./country-balanced.js";
 import { czPcFirstPrepositionRecipe } from "./cz-pcfirst-preposition.js";
 import { frAdminSplitRecipe } from "./fr-admin-split.js";
 import { frBareStreetRecipe } from "./fr-bare-street.js";
+import { frFragmentRecipe } from "./fr-fragment.js";
 import { frOrderRecipe } from "./fr-order.js";
 import { germanRecipe } from "./german.js";
 import { houseVenueRecipe } from "./house-venue.js";
 import { intersectionRecipe } from "./intersection.js";
 import { localeRecipe } from "./locale.js";
 import { nlPostcodeRecipe } from "./nl-postcode.js";
+import { noFragmentRecipe } from "./no-fragment.js";
 import { noStreetLedRecipe } from "./no-street-led.js";
 import { noStreetRecipe } from "./no-street.js";
 import { poBoxCedexRecipe } from "./po-box-cedex.js";
@@ -45,9 +47,11 @@ const RECIPES = [
     frOrderRecipe,
     frAdminSplitRecipe,
     frBareStreetRecipe,
+    frFragmentRecipe,
     czPcFirstPrepositionRecipe,
     nlPostcodeRecipe,
     noStreetLedRecipe,
+    noFragmentRecipe,
     siBareVillageRecipe,
     countryBalancedRecipe,
     boundaryStressRecipe,

package/out/src/shard-recipes/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/shard-recipes/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,EAAE,sBAAsB,EAAE,MAAM,wBAAwB,CAAA;AAC/D,OAAO,EAAE,oBAAoB,EAAE,MAAM,sBAAsB,CAAA;AAC3D,OAAO,EAAE,qBAAqB,EAAE,MAAM,uBAAuB,CAAA;AAC7D,OAAO,EAAE,0BAA0B,EAAE,MAAM,6BAA6B,CAAA;AACxE,OAAO,EAAE,kBAAkB,EAAE,MAAM,qBAAqB,CAAA;AACxD,OAAO,EAAE,kBAAkB,EAAE,MAAM,qBAAqB,CAAA;AACxD,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAA;AAC7C,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAA;AAC1C,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAA;AACtD,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAA;AAC1C,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,iBAAiB,EAAE,MAAM,oBAAoB,CAAA;AACtD,OAAO,EAAE,cAAc,EAAE,MAAM,gBAAgB,CAAA;AAC/C,OAAO,EAAE,gBAAgB,EAAE,MAAM,mBAAmB,CAAA;AACpD,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAA;AAEzC,OAAO,EAAE,mBAAmB,EAAE,MAAM,sBAAsB,CAAA;AAC1D,OAAO,EAAE,iBAAiB,EAAE,MAAM,mBAAmB,CAAA;AACrD,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAA;AAC1C,OAAO,EAAE,UAAU,EAAE,MAAM,WAAW,CAAA;AAEtC,cAAc,eAAe,CAAA;AAE7B,iDAAiD;AACjD,MAAM,OAAO,GAA2B;IACvC,YAAY;IACZ,gBAAgB;IAChB,iBAAiB;IACjB,cAAc;IACd,gBAAgB;IAChB,WAAW;IACX,gBAAgB;IAChB,UAAU;IACV,kBAAkB;IAClB,YAAY;IACZ,YAAY;IACZ,aAAa;IACb,kBAAkB;IAClB,kBAAkB;IAClB,0BAA0B;IAC1B,gBAAgB;IAChB,iBAAiB;IACjB,mBAAmB;IACnB,qBAAqB;IACrB,oBAAoB;IACpB,sBAAsB;CACtB,CAAA;AAED,4BAA4B;AAC5B,MAAM,CAAC,MAAM,aAAa,GAAqC,IAAI,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,CAAA;AAEvG,+CAA+C;AAC/C,MAAM,UAAU,cAAc,CAAC,IAAY;IAC1C,OAAO,aAAa,CAAC,GAAG,CAAC,IAAI,CAAC,CAAA;AAC/B,CAAC;AAED,yCAAyC;AACzC,MAAM,UAAU,gBAAgB;IAC/B,OAAO,OAAO,CAAA;AACf,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/shard-recipes/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,EAAE,sBAAsB,EAAE,MAAM,wBAAwB,CAAA;AAC/D,OAAO,EAAE,oBAAoB,EAAE,MAAM,sBAAsB,CAAA;AAC3D,OAAO,EAAE,qBAAqB,EAAE,MAAM,uBAAuB,CAAA;AAC7D,OAAO,EAAE,0BAA0B,EAAE,MAAM,6BAA6B,CAAA;AACxE,OAAO,EAAE,kBAAkB,EAAE,MAAM,qBAAqB,CAAA;AACxD,OAAO,EAAE,kBAAkB,EAAE,MAAM,qBAAqB,CAAA;AACxD,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAA;AAC7C,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAA;AAC1C,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAA;AACtD,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAA;AAC1C,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,iBAAiB,EAAE,MAAM,oBAAoB,CAAA;AACtD,OAAO,EAAE,cAAc,EAAE,MAAM,gBAAgB,CAAA;AAC/C,OAAO,EAAE,gBAAgB,EAAE,MAAM,mBAAmB,CAAA;AACpD,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAA;AAEzC,OAAO,EAAE,mBAAmB,EAAE,MAAM,sBAAsB,CAAA;AAC1D,OAAO,EAAE,iBAAiB,EAAE,MAAM,mBAAmB,CAAA;AACrD,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAA;AAC1C,OAAO,EAAE,UAAU,EAAE,MAAM,WAAW,CAAA;AAEtC,cAAc,eAAe,CAAA;AAE7B,iDAAiD;AACjD,MAAM,OAAO,GAA2B;IACvC,YAAY;IACZ,gBAAgB;IAChB,iBAAiB;IACjB,cAAc;IACd,gBAAgB;IAChB,WAAW;IACX,gBAAgB;IAChB,UAAU;IACV,kBAAkB;IAClB,YAAY;IACZ,YAAY;IACZ,aAAa;IACb,kBAAkB;IAClB,kBAAkB;IAClB,gBAAgB;IAChB,0BAA0B;IAC1B,gBAAgB;IAChB,iBAAiB;IACjB,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,oBAAoB;IACpB,sBAAsB;CACtB,CAAA;AAED,4BAA4B;AAC5B,MAAM,CAAC,MAAM,aAAa,GAAqC,IAAI,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,CAAA;AAEvG,+CAA+C;AAC/C,MAAM,UAAU,cAAc,CAAC,IAAY;IAC1C,OAAO,aAAa,CAAC,GAAG,CAAC,IAAI,CAAC,CAAA;AAC/B,CAAC;AAED,yCAAyC;AACzC,MAAM,UAAU,gBAAgB;IAC/B,OAAO,OAAO,CAAA;AACf,CAAC"}

package/out/src/shard-recipes/no-fragment.d.ts

@@ -0,0 +1,39 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `no-fragment` — the Norwegian house-number-licence lever (Track B, 2026-07-16). The mirror of
+ *   `fr-fragment`, which earned +50pp on the same defect shape in French.
+ *
+ *   WHY THIS EXISTS AND `no-street-led` DOES NOT SUFFICE. Board 3 (the NO digit board) measured that
+ *   `synth-no-street-led`'s three forms — all carrying postcode+city — are ALREADY at 0.940-0.968 on
+ *   a model with zero Norwegian rows. The headroom is in the forms that shard never emits:
+ *
+ *     bare-street-hn   "Hallingrudveien 32"     0.693   — no postcode competing, still fails 31%
+ *     slash-hn         "Øvrabø 124/1"           0.650   — cadastral gnr/bnr, ONE component
+ *
+ *   THE MECHANISM the board exposed: `Hallingrudveien 32` -> locality + postcode, while
+ *   `Hallingrudveien 32, 3370 Vikersund` parses perfectly. The street loses its street reading and
+ *   the digit loses its anchor TOGETHER. That is Track A's bare-street LICENCE in Norwegian — the
+ *   model will not read a street without its postcode/locality partner — not a digit-ownership prior.
+ *   fr-fragment fixed exactly this in French by teaching the street WITHOUT its partners.
+ *
+ *   THE COUNTER-DISTRIBUTION IS THE POINT (fr-fragment's lesson, and board 2's bare-locality guard).
+ *   Teaching bare `{street} {number}` alone lets the model satisfy every row by flipping its default
+ *   from "bare toponym -> locality" to "bare toponym -> street", trading one broken prior for
+ *   another. Two counter-classes hold the line:
+ *     - bare LOCALITIES (no street) so "bare -> street" is not free.
+ *     - bare POSTCODES so the model does not learn to stop emitting postcode to win the digit — the
+ *       board 3 bare-pc negative class (1.000) must HOLD.
+ *
+ *   SLASH HAZARD, pinned deliberately: NO `124/1` is ONE house_number (cadastral gnr/bnr). AU
+ *   `12/345` is TWO (unit + house_number). This shard teaches the Norwegian reading; a future AU
+ *   intra-word-split shard (B5) must not generalize over it. The two are locale-gated by design.
+ *
+ *   SPLIT: `--exclude-surfaces` REQUIRED (throws otherwise) — the digit board's reserved surface
+ *   list. Diacritic-KEEPING normalizer, matching the board (see the norm docstring).
+ */
+import { type ShardRecipe } from "./scaffold.ts";
+export declare const noFragmentRecipe: ShardRecipe;
+//# sourceMappingURL=no-fragment.d.ts.map

package/out/src/shard-recipes/no-fragment.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"no-fragment.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/no-fragment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmCG;AAIH,OAAO,EAA6C,KAAK,WAAW,EAAiB,MAAM,eAAe,CAAA;AAgB1G,eAAO,MAAM,gBAAgB,EAAE,WA8I9B,CAAA"}

package/out/src/shard-recipes/no-fragment.js

@@ -0,0 +1,175 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `no-fragment` — the Norwegian house-number-licence lever (Track B, 2026-07-16). The mirror of
+ *   `fr-fragment`, which earned +50pp on the same defect shape in French.
+ *
+ *   WHY THIS EXISTS AND `no-street-led` DOES NOT SUFFICE. Board 3 (the NO digit board) measured that
+ *   `synth-no-street-led`'s three forms — all carrying postcode+city — are ALREADY at 0.940-0.968 on
+ *   a model with zero Norwegian rows. The headroom is in the forms that shard never emits:
+ *
+ *     bare-street-hn   "Hallingrudveien 32"     0.693   — no postcode competing, still fails 31%
+ *     slash-hn         "Øvrabø 124/1"           0.650   — cadastral gnr/bnr, ONE component
+ *
+ *   THE MECHANISM the board exposed: `Hallingrudveien 32` -> locality + postcode, while
+ *   `Hallingrudveien 32, 3370 Vikersund` parses perfectly. The street loses its street reading and
+ *   the digit loses its anchor TOGETHER. That is Track A's bare-street LICENCE in Norwegian — the
+ *   model will not read a street without its postcode/locality partner — not a digit-ownership prior.
+ *   fr-fragment fixed exactly this in French by teaching the street WITHOUT its partners.
+ *
+ *   THE COUNTER-DISTRIBUTION IS THE POINT (fr-fragment's lesson, and board 2's bare-locality guard).
+ *   Teaching bare `{street} {number}` alone lets the model satisfy every row by flipping its default
+ *   from "bare toponym -> locality" to "bare toponym -> street", trading one broken prior for
+ *   another. Two counter-classes hold the line:
+ *     - bare LOCALITIES (no street) so "bare -> street" is not free.
+ *     - bare POSTCODES so the model does not learn to stop emitting postcode to win the digit — the
+ *       board 3 bare-pc negative class (1.000) must HOLD.
+ *
+ *   SLASH HAZARD, pinned deliberately: NO `124/1` is ONE house_number (cadastral gnr/bnr). AU
+ *   `12/345` is TWO (unit + house_number). This shard teaches the Norwegian reading; a future AU
+ *   intra-word-split shard (B5) must not generalize over it. The two are locale-gated by design.
+ *
+ *   SPLIT: `--exclude-surfaces` REQUIRED (throws otherwise) — the digit board's reserved surface
+ *   list. Diacritic-KEEPING normalizer, matching the board (see the norm docstring).
+ */
+import { readFileSync } from "node:fs";
+import { alignAndWrite, makeMulberry32, readTuples, shardSourceID } from "./scaffold.js";
+/**
+ * The surface key — MUST match the NO digit board's `norm_surface` and `no-street-led`'s `norm`: NFC, lowercase,
+ * collapse whitespace, KEEP diacritics. Stripping them (fr-fragment's norm) would fold `Tømmerlien` -> `tommerlien`,
+ * miss the board's reserved `tømmerlien`, and leak the surface.
+ */
+const norm = (value) => value.normalize("NFC").toLowerCase().replace(/\s+/g, " ").trim();
+/** Title-case a Kartverket ALL-CAPS locality (HELLVIK -> Hellvik); #690, all-caps is OOD. */
+const titleNO = (value) => value
+    .split(/\s+/)
+    .map((w) => (w ? w[0].toUpperCase() + w.slice(1).toLowerCase() : w))
+    .join(" ");
+export const noFragmentRecipe = {
+    name: "no-fragment",
+    description: "NO street fragments — the house-number-licence lever (Track B): '«st» «n»' / bare «st» with NO postcode partner, + bare-locality & bare-postcode counters",
+    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).",
+        },
+        {
+            flag: "--bare-street-prob <n>",
+            description: "Share of street rows emitted as a BARE street, no number (default 0.30 — the pure licence signal)",
+        },
+        {
+            flag: "--counter-prob <n>",
+            description: "Share of ALL rows that are counter-distribution (bare locality OR bare postcode) (default 0.30)",
+        },
+        {
+            flag: "--long-number-boost <n>",
+            description: "knob 3: emit N copies of each street+number row whose number has >= --long-number-min-digits digits — oversample the failing long-number class the model calls a postcode (default 1 = no boost)",
+        },
+        {
+            flag: "--long-number-min-digits <n>",
+            description: "knob 3: minimum digit count for a number to be 'long' and boosted (default 3)",
+        },
+    ],
+    async run(opts, write) {
+        const random = makeMulberry32(opts.seed);
+        const excludePath = opts.excludeSurfaces;
+        if (!excludePath) {
+            throw new Error("no-fragment: --exclude-surfaces is REQUIRED. Pass the NO digit board's reserved list " +
+                "(mailwoman/eval-harness/fixtures/no-digits.surfaces.txt) or this shard trains on its own eval set. " +
+                "Source-disjoint by street SURFACE is the split discipline; there is no safe default.");
+        }
+        const excluded = new Set(readFileSync(excludePath, "utf8")
+            .split("\n")
+            .map((line) => line.trim())
+            .filter((line) => line && !line.startsWith("#")));
+        if (!excluded.size)
+            throw new Error(`no-fragment: --exclude-surfaces "${excludePath}" listed no surfaces`);
+        const bareStreetProb = opts.bareProb ?? 0.3;
+        const counterProb = opts.counterProb ?? 0.3;
+        const longNumberBoost = Math.max(1, Math.floor(opts.longNumberBoost ?? 1));
+        const longNumberMinDigits = opts.longNumberMinDigits ?? 3;
+        // Harvested from the tuples — every NO row carries its locality and postcode, so the two
+        // counter-classes need no second source.
+        const localities = new Set();
+        const postcodes = new Set();
+        let read = 0;
+        let emitted = 0;
+        let skipped = 0;
+        let contaminated = 0;
+        let emitSeq = 0;
+        const emit = (raw, components, klass) => {
+            // emitSeq keeps every emit distinct — knob 3 emits N copies of one long-number row, and
+            // (components, read) alone would collide their source_id and let downstream dedup drop the boost.
+            const source_id = shardSourceID("synth-no-fragment", { ...components, k: klass, v: `${read}:${emitSeq++}` });
+            const canonical = {
+                raw,
+                components,
+                country: "NO",
+                locale: "nb-NO",
+                source: "synth-no-fragment",
+                source_id,
+                corpus_version: "0.11.0",
+                license: "Synthetic — no-fragment; (street, number, postcode, city) from OpenAddresses NO / Kartverket",
+            };
+            if (alignAndWrite(write, canonical, "no-fragment"))
+                emitted++;
+            else
+                skipped++;
+        };
+        for await (const tuple of readTuples(opts.input)) {
+            read++;
+            const street = String(tuple.street ?? "").trim();
+            const locality = titleNO(String(tuple.locality ?? "").trim());
+            const number = String(tuple.number ?? "").trim();
+            const postcode = String(tuple.postcode ?? "").trim();
+            if (locality)
+                localities.add(locality);
+            if (postcode)
+                postcodes.add(postcode);
+            if (!street) {
+                skipped++;
+                continue;
+            }
+            // THE SPLIT. A surface on the digit board never enters training.
+            if (excluded.has(norm(street))) {
+                contaminated++;
+                continue;
+            }
+            // COUNTER-DISTRIBUTION — drawn from the harvested pools, not this row's street. Half bare
+            // localities (so "bare -> street" is not free), half bare postcodes (so the model does not
+            // stop emitting postcode to win the digit — board 3's bare-pc must hold).
+            if (random() < counterProb) {
+                if (random() < 0.5 && localities.size) {
+                    const loc = [...localities][Math.floor(random() * localities.size)];
+                    emit(loc, { locality: loc }, "counter-bare-locality");
+                }
+                else if (postcodes.size) {
+                    const pc = [...postcodes][Math.floor(random() * postcodes.size)];
+                    emit(pc, { postcode: pc }, "counter-bare-postcode");
+                }
+                continue;
+            }
+            // THE SIGNAL. A street with NO postcode/locality partner. Either bare, or street+number —
+            // both are the forms board 3 measured as the headroom (bare-street-hn 0.693, slash-hn 0.650).
+            if (!number || random() < bareStreetProb) {
+                emit(street, { street }, "bare-street");
+            }
+            else {
+                const klass = number.includes("/") ? "slash-hn" : "street-hn";
+                // knob 3: the failing class is street + LONG number (Leppdalsvegen 1285 -> postcode). The
+                // digit count, not the slash, is what tips the length prior toward postcode. Oversample
+                // those rows to fight the prior with volume and teach the street/number boundary directly.
+                const digits = (number.match(/\d/g) ?? []).length;
+                const copies = digits >= longNumberMinDigits ? longNumberBoost : 1;
+                for (let c = 0; c < copies; c++) {
+                    emit(`${street} ${number}`, { street, house_number: number }, klass);
+                }
+            }
+        }
+        return { read, emitted, skipped, contaminated };
+    },
+};
+//# sourceMappingURL=no-fragment.js.map

package/out/src/shard-recipes/no-fragment.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"no-fragment.js","sourceRoot":"","sources":["../../../src/shard-recipes/no-fragment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmCG;AAEH,OAAO,EAAE,YAAY,EAAE,MAAM,SAAS,CAAA;AAEtC,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,UAAU,EAAoB,aAAa,EAAE,MAAM,eAAe,CAAA;AAE1G;;;;GAIG;AACH,MAAM,IAAI,GAAG,CAAC,KAAa,EAAU,EAAE,CAAC,KAAK,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAA;AAExG,6FAA6F;AAC7F,MAAM,OAAO,GAAG,CAAC,KAAa,EAAU,EAAE,CACzC,KAAK;KACH,KAAK,CAAC,KAAK,CAAC;KACZ,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,WAAW,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;KACpE,IAAI,CAAC,GAAG,CAAC,CAAA;AAEZ,MAAM,CAAC,MAAM,gBAAgB,GAAgB;IAC5C,IAAI,EAAE,aAAa;IACnB,WAAW,EACV,2JAA2J;IAC5J,IAAI,EAAE,QAAQ;IACd,OAAO,EAAE;QACR;YACC,IAAI,EAAE,2BAA2B;YACjC,WAAW,EACV,gHAAgH;SACjH;QACD;YACC,IAAI,EAAE,wBAAwB;YAC9B,WAAW,EAAE,mGAAmG;SAChH;QACD;YACC,IAAI,EAAE,oBAAoB;YAC1B,WAAW,EAAE,iGAAiG;SAC9G;QACD;YACC,IAAI,EAAE,yBAAyB;YAC/B,WAAW,EACV,kMAAkM;SACnM;QACD;YACC,IAAI,EAAE,8BAA8B;YACpC,WAAW,EAAE,+EAA+E;SAC5F;KACD;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,WAAW,GAAG,IAAI,CAAC,eAAe,CAAA;QAExC,IAAI,CAAC,WAAW,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CACd,uFAAuF;gBACtF,qGAAqG;gBACrG,sFAAsF,CACvF,CAAA;QACF,CAAC;QAED,MAAM,QAAQ,GAAG,IAAI,GAAG,CACvB,YAAY,CAAC,WAAW,EAAE,MAAM,CAAC;aAC/B,KAAK,CAAC,IAAI,CAAC;aACX,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;aAC1B,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,CACjD,CAAA;QAED,IAAI,CAAC,QAAQ,CAAC,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,oCAAoC,WAAW,sBAAsB,CAAC,CAAA;QAE1G,MAAM,cAAc,GAAG,IAAI,CAAC,QAAQ,IAAI,GAAG,CAAA;QAC3C,MAAM,WAAW,GAAG,IAAI,CAAC,WAAW,IAAI,GAAG,CAAA;QAC3C,MAAM,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,eAAe,IAAI,CAAC,CAAC,CAAC,CAAA;QAC1E,MAAM,mBAAmB,GAAG,IAAI,CAAC,mBAAmB,IAAI,CAAC,CAAA;QAEzD,yFAAyF;QACzF,yCAAyC;QACzC,MAAM,UAAU,GAAG,IAAI,GAAG,EAAU,CAAA;QACpC,MAAM,SAAS,GAAG,IAAI,GAAG,EAAU,CAAA;QAEnC,IAAI,IAAI,GAAG,CAAC,CAAA;QACZ,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,YAAY,GAAG,CAAC,CAAA;QACpB,IAAI,OAAO,GAAG,CAAC,CAAA;QAEf,MAAM,IAAI,GAAG,CAAC,GAAW,EAAE,UAAkC,EAAE,KAAa,EAAQ,EAAE;YACrF,wFAAwF;YACxF,kGAAkG;YAClG,MAAM,SAAS,GAAG,aAAa,CAAC,mBAAmB,EAAE,EAAE,GAAG,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,IAAI,IAAI,OAAO,EAAE,EAAE,EAAE,CAAC,CAAA;YAC5G,MAAM,SAAS,GAAG;gBACjB,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,mBAAmB;gBAC3B,SAAS;gBACT,cAAc,EAAE,QAAQ;gBACxB,OAAO,EAAE,8FAA8F;aACvG,CAAA;YAED,IAAI,aAAa,CAAC,KAAK,EAAE,SAAS,EAAE,aAAa,CAAC;gBAAE,OAAO,EAAE,CAAA;;gBACxD,OAAO,EAAE,CAAA;QACf,CAAC,CAAA;QAED,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,UAAU,CAAC,IAAI,CAAC,KAAM,CAAC,EAAE,CAAC;YACnD,IAAI,EAAE,CAAA;YACN,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAChD,MAAM,QAAQ,GAAG,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAA;YAC7D,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAChD,MAAM,QAAQ,GAAG,MAAM,CAAC,KAAK,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAEpD,IAAI,QAAQ;gBAAE,UAAU,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;YACtC,IAAI,QAAQ;gBAAE,SAAS,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;YAErC,IAAI,CAAC,MAAM,EAAE,CAAC;gBACb,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YAED,iEAAiE;YACjE,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;gBAChC,YAAY,EAAE,CAAA;gBACd,SAAQ;YACT,CAAC;YAED,0FAA0F;YAC1F,2FAA2F;YAC3F,0EAA0E;YAC1E,IAAI,MAAM,EAAE,GAAG,WAAW,EAAE,CAAC;gBAC5B,IAAI,MAAM,EAAE,GAAG,GAAG,IAAI,UAAU,CAAC,IAAI,EAAE,CAAC;oBACvC,MAAM,GAAG,GAAG,CAAC,GAAG,UAAU,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,UAAU,CAAC,IAAI,CAAC,CAAE,CAAA;oBAEpE,IAAI,CAAC,GAAG,EAAE,EAAE,QAAQ,EAAE,GAAG,EAAE,EAAE,uBAAuB,CAAC,CAAA;gBACtD,CAAC;qBAAM,IAAI,SAAS,CAAC,IAAI,EAAE,CAAC;oBAC3B,MAAM,EAAE,GAAG,CAAC,GAAG,SAAS,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,SAAS,CAAC,IAAI,CAAC,CAAE,CAAA;oBAEjE,IAAI,CAAC,EAAE,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,EAAE,uBAAuB,CAAC,CAAA;gBACpD,CAAC;gBACD,SAAQ;YACT,CAAC;YAED,0FAA0F;YAC1F,8FAA8F;YAC9F,IAAI,CAAC,MAAM,IAAI,MAAM,EAAE,GAAG,cAAc,EAAE,CAAC;gBAC1C,IAAI,CAAC,MAAM,EAAE,EAAE,MAAM,EAAE,EAAE,aAAa,CAAC,CAAA;YACxC,CAAC;iBAAM,CAAC;gBACP,MAAM,KAAK,GAAG,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,WAAW,CAAA;gBAC7D,0FAA0F;gBAC1F,wFAAwF;gBACxF,2FAA2F;gBAC3F,MAAM,MAAM,GAAG,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAA;gBACjD,MAAM,MAAM,GAAG,MAAM,IAAI,mBAAmB,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,CAAA;gBAElE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;oBACjC,IAAI,CAAC,GAAG,MAAM,IAAI,MAAM,EAAE,EAAE,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,EAAE,EAAE,KAAK,CAAC,CAAA;gBACrE,CAAC;YACF,CAAC;QACF,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,YAAY,EAAE,CAAA;IAChD,CAAC;CACD,CAAA"}

package/out/src/shard-recipes/no-street-led.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"no-street-led.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/no-street-led.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAA6C,KAAK,WAAW,EAAiB,MAAM,eAAe,CAAA;AAE1G,eAAO,MAAM,iBAAiB,EAAE,WA2D/B,CAAA"}
+{"version":3,"file":"no-street-led.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/no-street-led.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAIH,OAAO,EAA6C,KAAK,WAAW,EAAiB,MAAM,eAAe,CAAA;AAW1G,eAAO,MAAM,iBAAiB,EAAE,WAkG/B,CAAA"}

package/out/src/shard-recipes/no-street-led.js

@@ -15,16 +15,49 @@
  *   2. city-first "«city», «pc», «st» «n»"
  *   3. pc-first   "«pc» «city», «st» «n»"   (the 7% floor — the anchor the others converge to)
  */
+import { readFileSync } from "node:fs";
 import { alignAndWrite, makeMulberry32, readTuples, shardSourceID } from "./scaffold.js";
+/**
+ * The surface key. MUST match the NO digit board's `norm_surface` (scratchpad/build-no-board.py): NFC, lowercase,
+ * collapse whitespace — and critically, KEEP diacritics. fr-fragment's `norm` strips them (NFD + combining-mark
+ * removal), which is right for French but would collapse `Tømmerlien` → `tommerlien` here, so the shard's exclusion
+ * check would never match the board's reserved `tømmerlien` and the split would leak silently. Diacritic street heads
+ * (…vegen/…veien with ø/å/æ) are the whole point of this shard's boundary; folding them away is not an option.
+ */
+const norm = (value) => value.normalize("NFC").toLowerCase().replace(/\s+/g, " ").trim();
 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 shard 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 shard trains on its own eval set. " +
+                "Source-disjoint by street SURFACE is the split discipline; there is no safe default.");
+        }
+        const excluded = new Set(readFileSync(excludePath, "utf8")
+            .split("\n")
+            .map((line) => line.trim())
+            .filter((line) => line && !line.startsWith("#")));
+        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();
@@ -35,6 +68,11 @@ export const noStreetLedRecipe = {
                 skipped++;
                 continue;
             }
+            // A surface on the digit board never enters training.
+            if (excluded.has(norm(street))) {
+                contaminated++;
+                continue;
+            }
             const order = read % 3;
             let raw;
             const components = {
@@ -70,7 +108,7 @@ export const noStreetLedRecipe = {
                 skipped++;
             }
         }
-        return { read, emitted, skipped };
+        return { read, emitted, skipped, contaminated };
     },
 };
 //# sourceMappingURL=no-street-led.js.map

package/out/src/shard-recipes/no-street-led.js.map

@@ -1 +1 @@
-{"version":3,"file":"no-street-led.js","sourceRoot":"","sources":["../../../src/shard-recipes/no-street-led.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,UAAU,EAAoB,aAAa,EAAE,MAAM,eAAe,CAAA;AAE1G,MAAM,CAAC,MAAM,iBAAiB,GAAgB;IAC7C,IAAI,EAAE,eAAe;IACrB,WAAW,EAAE,6FAA6F;IAC1G,IAAI,EAAE,QAAQ;IACd,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,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,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;gBACT,SAAQ;YACT,CAAC;YACD,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,CAAA;YACtB,IAAI,GAAW,CAAA;YACf,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;YACD,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;YAC5G,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,CAAA;IAClC,CAAC;CACD,CAAA"}
+{"version":3,"file":"no-street-led.js","sourceRoot":"","sources":["../../../src/shard-recipes/no-street-led.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAE,YAAY,EAAE,MAAM,SAAS,CAAA;AAEtC,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,UAAU,EAAoB,aAAa,EAAE,MAAM,eAAe,CAAA;AAE1G;;;;;;GAMG;AACH,MAAM,IAAI,GAAG,CAAC,KAAa,EAAU,EAAE,CAAC,KAAK,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAA;AAExG,MAAM,CAAC,MAAM,iBAAiB,GAAgB;IAC7C,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,CACvB,YAAY,CAAC,WAAW,EAAE,MAAM,CAAC;aAC/B,KAAK,CAAC,IAAI,CAAC;aACX,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;aAC1B,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,CACjD,CAAA;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;gBACT,SAAQ;YACT,CAAC;YAED,sDAAsD;YACtD,IAAI,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;gBAChC,YAAY,EAAE,CAAA;gBACd,SAAQ;YACT,CAAC;YACD,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,CAAA;YACtB,IAAI,GAAW,CAAA;YACf,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;YACD,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;YAC5G,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/src/shard-recipes/scaffold.d.ts

@@ -73,12 +73,32 @@ export interface ShardRecipeOpts {
     hnProb?: number;
     communes?: string;
     multilocaleCount?: number;
+    /**
+     * `fr-fragment` / `no-fragment` / `no-street-led`: the eval board's reserved street-surface list. REQUIRED for those
+     * recipes — a shard that trains on its own eval set measures memorization. See their docstrings.
+     */
+    excludeSurfaces?: string;
+    /** `no-fragment`: share of rows that are counter-distribution (bare locality OR bare postcode). */
+    counterProb?: number;
+    /**
+     * `no-fragment` knob 3: emit N copies of each street+number row whose number has >= longNumberMinDigits digits
+     * (oversample the failing long-number class). Default 1 = no boost.
+     */
+    longNumberBoost?: number;
+    /** `no-fragment` knob 3: minimum digit count for a number to count as "long" and be boosted. Default 3. */
+    longNumberMinDigits?: number;
 }
 /** Tally a recipe returns. */
 export interface ShardStats {
     read?: number;
     emitted: number;
     skipped: number;
+    /**
+     * Rows dropped because their street SURFACE is reserved by an eval board (`--exclude-surfaces`). Separate from
+     * `skipped` on purpose: a nonzero value is the audit trail that the train/eval split actually fired. Zero when a
+     * recipe has no board split.
+     */
+    contaminated?: number;
 }
 /** A single declared recipe-specific option flag (for the command's --help). */
 export interface ShardRecipeOption {

package/out/src/shard-recipes/scaffold.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"scaffold.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/scaffold.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAKH;;;;GAIG;AACH,wBAAgB,aAAa,CAAC,SAAS,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,CAAC,GAAG,MAAM,CAElG;AAED,+FAA+F;AAC/F,MAAM,WAAW,UAAU;IAC1B,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,CAAC,CAAC,EAAE,MAAM,GAAG,OAAO,CAAA;CACpB;AAED;;;GAGG;AACH,wBAAgB,OAAO,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,MAAM,CAQlD;AAED,6CAA6C;AAC7C,eAAO,MAAM,UAAU,gBAAU,CAAA;AAEjC;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,MAAM,CAUzD;AAED,mGAAmG;AACnG,wBAAuB,UAAU,CAAC,KAAK,EAAE,MAAM,GAAG,cAAc,CAAC,UAAU,CAAC,CAgB3E;AAED,kGAAkG;AAClG,MAAM,WAAW,iBAAiB;IACjC,GAAG,EAAE,MAAM,CAAA;IACX,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAClC,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,MAAM,CAAA;IACd,SAAS,EAAE,MAAM,CAAA;IACjB,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,OAAO,CAAC,EAAE,MAAM,CAAA;CAChB;AAED;;;GAGG;AACH,wBAAgB,aAAa,CAC5B,KAAK,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,EAC7B,SAAS,EAAE,iBAAiB,EAC5B,WAAW,EAAE,MAAM,EACnB,WAAW,GAAE,MAAM,GAAG,IAAW,GAC/B,OAAO,CAOT;AAED,oGAAoG;AACpG,MAAM,WAAW,eAAe;IAC/B,MAAM,EAAE,MAAM,CAAA;IACd,IAAI,EAAE,MAAM,CAAA;IACZ,QAAQ,EAAE,MAAM,CAAA;IAChB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB,UAAU,CAAC,EAAE,MAAM,CAAA;IAEnB,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,gBAAgB,CAAC,EAAE,MAAM,CAAA;CACzB;AAED,8BAA8B;AAC9B,MAAM,WAAW,UAAU;IAC1B,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,OAAO,EAAE,MAAM,CAAA;IACf,OAAO,EAAE,MAAM,CAAA;CACf;AAED,gFAAgF;AAChF,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;CACnB;AAED,yEAAyE;AACzE,MAAM,WAAW,WAAW;IAC3B,sEAAsE;IACtE,IAAI,EAAE,MAAM,CAAA;IACZ,gDAAgD;IAChD,WAAW,EAAE,MAAM,CAAA;IACnB,gFAAgF;IAChF,IAAI,EAAE,QAAQ,GAAG,UAAU,CAAA;IAC3B,qEAAqE;IACrE,OAAO,CAAC,EAAE,iBAAiB,EAAE,CAAA;IAC7B;;;OAGG;IACH,GAAG,CAAC,IAAI,EAAE,eAAe,EAAE,KAAK,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,UAAU,CAAC,CAAA;CAC9E"}
+{"version":3,"file":"scaffold.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/scaffold.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAKH;;;;GAIG;AACH,wBAAgB,aAAa,CAAC,SAAS,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,CAAC,GAAG,MAAM,CAElG;AAED,+FAA+F;AAC/F,MAAM,WAAW,UAAU;IAC1B,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,CAAC,CAAC,EAAE,MAAM,GAAG,OAAO,CAAA;CACpB;AAED;;;GAGG;AACH,wBAAgB,OAAO,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,MAAM,CAQlD;AAED,6CAA6C;AAC7C,eAAO,MAAM,UAAU,gBAAU,CAAA;AAEjC;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,MAAM,CAUzD;AAED,mGAAmG;AACnG,wBAAuB,UAAU,CAAC,KAAK,EAAE,MAAM,GAAG,cAAc,CAAC,UAAU,CAAC,CAgB3E;AAED,kGAAkG;AAClG,MAAM,WAAW,iBAAiB;IACjC,GAAG,EAAE,MAAM,CAAA;IACX,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAClC,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,MAAM,CAAA;IACd,SAAS,EAAE,MAAM,CAAA;IACjB,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,OAAO,CAAC,EAAE,MAAM,CAAA;CAChB;AAED;;;GAGG;AACH,wBAAgB,aAAa,CAC5B,KAAK,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,EAC7B,SAAS,EAAE,iBAAiB,EAC5B,WAAW,EAAE,MAAM,EACnB,WAAW,GAAE,MAAM,GAAG,IAAW,GAC/B,OAAO,CAOT;AAED,oGAAoG;AACpG,MAAM,WAAW,eAAe;IAC/B,MAAM,EAAE,MAAM,CAAA;IACd,IAAI,EAAE,MAAM,CAAA;IACZ,QAAQ,EAAE,MAAM,CAAA;IAChB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB,UAAU,CAAC,EAAE,MAAM,CAAA;IAEnB,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB;;;OAGG;IACH,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,mGAAmG;IACnG,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB;;;OAGG;IACH,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,2GAA2G;IAC3G,mBAAmB,CAAC,EAAE,MAAM,CAAA;CAC5B;AAED,8BAA8B;AAC9B,MAAM,WAAW,UAAU;IAC1B,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,OAAO,EAAE,MAAM,CAAA;IACf,OAAO,EAAE,MAAM,CAAA;IACf;;;;OAIG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;CACrB;AAED,gFAAgF;AAChF,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;CACnB;AAED,yEAAyE;AACzE,MAAM,WAAW,WAAW;IAC3B,sEAAsE;IACtE,IAAI,EAAE,MAAM,CAAA;IACZ,gDAAgD;IAChD,WAAW,EAAE,MAAM,CAAA;IACnB,gFAAgF;IAChF,IAAI,EAAE,QAAQ,GAAG,UAAU,CAAA;IAC3B,qEAAqE;IACrE,OAAO,CAAC,EAAE,iBAAiB,EAAE,CAAA;IAC7B;;;OAGG;IACH,GAAG,CAAC,IAAI,EAAE,eAAe,EAAE,KAAK,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,UAAU,CAAC,CAAA;CAC9E"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mailwoman/corpus",
-  "version": "6.3.0",
+  "version": "6.5.0",
   "description": "Mailwoman corpus pipeline: BIO-labeled dataset builder for the neural classifier.",
   "license": "AGPL-3.0-only OR LicenseRef-Commercial",
   "repository": {
@@ -82,14 +82,14 @@
   },
   "dependencies": {
     "@dsnp/parquetjs": "1.8.8",
-    "@mailwoman/codex": "6.3.0",
-    "@mailwoman/core": "6.3.0",
-    "@mailwoman/formatter": "6.3.0",
+    "@mailwoman/codex": "6.5.0",
+    "@mailwoman/core": "6.5.0",
+    "@mailwoman/formatter": "6.5.0",
     "csv-parse": "^7.0.1",
     "fast-glob": "^3.3.3",
     "fastest-levenshtein": "^1.0.16",
     "lru-cache": "^11.5.2",
-    "spliterator": "^3.2.0"
+    "spliterator": "^4.0.0"
   },
   "devDependencies": {
     "@types/node": ">=26.1.1"