@mailwoman/corpus 9.2.0 → 9.4.0

package/out/recipes/cz-pcfirst-preposition.js

@@ -0,0 +1,101 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `cz-pcfirst-preposition` — the Czech pc-first prepositional-locality slice, fourth orthography
+ *   of the #901 leading-name-boundary family. The #897 close-out read all 8 residual CZ rows as one
+ *   class: a LEADING postcode mis-assigned as house_number while the multi-word "nad/pod/u X"
+ *   locality shatters ("51244 Rokytnice nad Jizerou, Dolní Rokytnice 111" → street
+ *   'RokytnicenadJizerou' + house_number '51244'). That leading-5-digit confusion is the
+ *   anchor-pollution class whose decode-time OVERRIDE was correctly killed in #723 — this slice is
+ *   the model-first fix as DATA: real prepositional localities in the order that breaks, so the
+ *   model learns that a leading postcode before a multi-word name is a postcode. pc-first leads the
+ *   cycle (the lesson); canonical and city-first keep the polarity balanced (the v1.9.9 lesson).
+ */
+import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils";
+import { alignAndWrite, readTuples, sliceSourceID } from "#recipes/scaffold";
+/**
+ * The order-cycle slot for the STREET-LESS form (`«city» «pc», Česko`) — the exact surface of the
+ * `cz-full-praha-100-00` board row, whose absence from the street-bearing orders was the v4.5.0 no-promote's measured
+ * gap.
+ */
+const STREETLESS_ORDER = 3;
+/**
+ * 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 czPcFirstPrepositionRecipe = {
+    name: "cz-pcfirst-preposition",
+    description: "CZ pc-first + prepositional locality (#901 family): '«pc» «city nad X», «st» «n»' — the #723 class as data",
+    mode: "tuples",
+    async run(opts, write) {
+        makeMulberry32(opts.seed);
+        let read = 0;
+        let emitted = 0;
+        let skipped = 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;
+            }
+            const order = read % 4;
+            // The OFFICIAL Czech rendering spaces the PSČ as `NNN NN` ('512 44'); OpenAddresses stores it
+            // unspaced ('51244'), and a model trained only on the source form reads the spaced surface as
+            // house_number + garbage (the 'Praha 100 00' mangle). Alternate the two renderings so both
+            // orthographies are attested — the LABEL is the postcode either way.
+            const spaced = read % 2 === 0 && /^\d{5}$/.test(postcode);
+            const postcodeSurface = spaced ? `${postcode.slice(0, 3)} ${postcode.slice(3)}` : postcode;
+            let raw;
+            const components = order === STREETLESS_ORDER
+                ? { locality: city, postcode: postcodeSurface, country: "Česko" }
+                : { street, house_number: number, postcode: postcodeSurface, locality: city };
+            if (order === 0) {
+                raw = `${postcodeSurface} ${city}, ${street} ${number}`;
+            }
+            else if (order === 1) {
+                raw = `${street} ${number}, ${postcodeSurface} ${city}`;
+            }
+            else if (order === 2) {
+                raw = `${city}, ${postcodeSurface}, ${street} ${number}`;
+            }
+            else {
+                // order === STREETLESS_ORDER: the street-less form — `«city» «pc», Česko` — the exact surface of the
+                // cz-full-praha-100-00 board row. The v4.5.0 no-promote receipt measured the gap: every
+                // prior order was street-bearing, so the model never saw a spaced PSČ beside a bare
+                // locality and mangled 'Praha 100 00, Czechia' into house_number spans. Street/number
+                // stay OUT of the components for this form (they are not in the surface).
+                raw = `${city} ${postcodeSurface}, Česko`;
+            }
+            const source_id = sliceSourceID("synth-cz-pcfirst-preposition", {
+                ...components,
+                o: String(order),
+                s: spaced ? "1" : "0",
+                v: String(read),
+            });
+            const canonical = {
+                raw,
+                components,
+                country: "CZ",
+                locale: "cs-CZ",
+                source: "synth-cz-pcfirst-preposition",
+                source_id,
+                corpus_version: "0.10.0",
+                license: "Synthetic — cz-pcfirst-preposition; (street, number, postcode, city) from OpenAddresses CZ (per-source attribution in the model card)",
+            };
+            if (alignAndWrite(write, canonical, "cz-pcfirst-preposition")) {
+                emitted++;
+            }
+            else {
+                skipped++;
+            }
+        }
+        return { read, emitted, skipped };
+    },
+};
+//# sourceMappingURL=cz-pcfirst-preposition.js.map

package/out/recipes/cz-pcfirst-preposition.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cz-pcfirst-preposition.js","sourceRoot":"","sources":["../../lib/recipes/cz-pcfirst-preposition.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AAEpE,OAAO,EAAE,aAAa,EAAE,UAAU,EAAqB,aAAa,EAAE,MAAM,mBAAmB,CAAA;AAC/F;;;;GAIG;AACH,MAAM,gBAAgB,GAAG,CAAC,CAAA;AAE1B;;;GAGG;AACH,MAAM,CAAC,MAAM,0BAA0B,GAAiB;IACvD,IAAI,EAAE,wBAAwB;IAC9B,WAAW,EACV,4GAA4G;IAC7G,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;gBAET,SAAQ;YACT,CAAC;YAED,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,CAAA;YACtB,8FAA8F;YAC9F,8FAA8F;YAC9F,2FAA2F;YAC3F,qEAAqE;YACrE,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,KAAK,CAAC,IAAI,SAAS,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAA;YACzD,MAAM,eAAe,GAAG,MAAM,CAAC,CAAC,CAAC,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAA;YAC1F,IAAI,GAAW,CAAA;YAEf,MAAM,UAAU,GACf,KAAK,KAAK,gBAAgB;gBACzB,CAAC,CAAC,EAAE,QAAQ,EAAE,IAAI,EAAE,QAAQ,EAAE,eAAe,EAAE,OAAO,EAAE,OAAO,EAAE;gBACjE,CAAC,CAAC,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,EAAE,QAAQ,EAAE,eAAe,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAA;YAE/E,IAAI,KAAK,KAAK,CAAC,EAAE,CAAC;gBACjB,GAAG,GAAG,GAAG,eAAe,IAAI,IAAI,KAAK,MAAM,IAAI,MAAM,EAAE,CAAA;YACxD,CAAC;iBAAM,IAAI,KAAK,KAAK,CAAC,EAAE,CAAC;gBACxB,GAAG,GAAG,GAAG,MAAM,IAAI,MAAM,KAAK,eAAe,IAAI,IAAI,EAAE,CAAA;YACxD,CAAC;iBAAM,IAAI,KAAK,KAAK,CAAC,EAAE,CAAC;gBACxB,GAAG,GAAG,GAAG,IAAI,KAAK,eAAe,KAAK,MAAM,IAAI,MAAM,EAAE,CAAA;YACzD,CAAC;iBAAM,CAAC;gBACP,qGAAqG;gBACrG,wFAAwF;gBACxF,oFAAoF;gBACpF,sFAAsF;gBACtF,0EAA0E;gBAC1E,GAAG,GAAG,GAAG,IAAI,IAAI,eAAe,SAAS,CAAA;YAC1C,CAAC;YAED,MAAM,SAAS,GAAG,aAAa,CAAC,8BAA8B,EAAE;gBAC/D,GAAG,UAAU;gBACb,CAAC,EAAE,MAAM,CAAC,KAAK,CAAC;gBAChB,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG;gBACrB,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC;aACf,CAAC,CAAA;YAEF,MAAM,SAAS,GAAG;gBACjB,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM,EAAE,8BAA8B;gBACtC,SAAS;gBACT,cAAc,EAAE,QAAQ;gBACxB,OAAO,EACN,uIAAuI;aACxI,CAAA;YAED,IAAI,aAAa,CAAC,KAAK,EAAE,SAAS,EAAE,wBAAwB,CAAC,EAAE,CAAC;gBAC/D,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/fr-admin-split.d.ts

@@ -0,0 +1,38 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-admin-split` slice recipe — the FR admin-split coverage slice (night 2026-06-19,
+ *   surpass-v1.5.0). Teaches the model to SPLIT the département out of the locality on
+ *   bare/space/comma-delimited French place rows — the admin-deciding failure class the pre-GPU
+ *   self-validation proved moves the resolved coordinate (collision communes −61%; see
+ *   docs/articles/evals/experiments/2026-06-19-fr-admin-split-prevalidation.md). Ported from
+ *   scripts/build-fr-admin-split-slice.mjs.
+ *
+ *   Failure shapes (the model currently mis-handles all of these):
+ *
+ *   - `Thauron, Creuse` → région dropped to null (the comma+full-name miss)
+ *   - `Montredon, Lozère` → région = "ère" (the diacritic subword split, #727)
+ *   - (AU analog) `CANBERRA ACT` → the space-delimited admin fuse
+ *
+ *   The département is the essential admin unit for FR postal geography and maps to the `region`
+ *   component tag in our schema. We derive it DETERMINISTICALLY from the real postcode via codex
+ *   `departementForCodePostal` (first two digits = département) — salvage-first, no re-derived
+ *   table.
+ *
+ *   Data (`--communes`, opts.communes): REAL BAN (Base Adresse Nationale) commune+postcode+coord
+ *   tuples, one per line, TAB-separated `commune <TAB> postcode <TAB> lon <TAB> lat`. Build the
+ *   input TSV once from the BAN staging CSV (see the legacy script header). Anchor-ON by
+ *   construction: rows carry a REAL postcode token in `raw` + a `postcode` component, so the
+ *   training loader paints the anchor feature onto that span automatically. The trailing-postcode
+ *   anchor REINFORCES the FR split (FR postcode is trailing, unlike German PLZ-leading — the v0.9.2
+ *   scar is positional, not universal).
+ */
+import type { 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 frAdminSplitRecipe: CorpusRecipe;
+//# sourceMappingURL=fr-admin-split.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"fr-admin-split.d.ts","sourceRoot":"","sources":["../../lib/recipes/fr-admin-split.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AASH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAA;AA8GrD;;;GAGG;AACH,eAAO,MAAM,kBAAkB,EAAE,YAyFhC,CAAA"}

package/out/recipes/fr-admin-split.js

@@ -0,0 +1,192 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-admin-split` slice recipe — the FR admin-split coverage slice (night 2026-06-19,
+ *   surpass-v1.5.0). Teaches the model to SPLIT the département out of the locality on
+ *   bare/space/comma-delimited French place rows — the admin-deciding failure class the pre-GPU
+ *   self-validation proved moves the resolved coordinate (collision communes −61%; see
+ *   docs/articles/evals/experiments/2026-06-19-fr-admin-split-prevalidation.md). Ported from
+ *   scripts/build-fr-admin-split-slice.mjs.
+ *
+ *   Failure shapes (the model currently mis-handles all of these):
+ *
+ *   - `Thauron, Creuse` → région dropped to null (the comma+full-name miss)
+ *   - `Montredon, Lozère` → région = "ère" (the diacritic subword split, #727)
+ *   - (AU analog) `CANBERRA ACT` → the space-delimited admin fuse
+ *
+ *   The département is the essential admin unit for FR postal geography and maps to the `region`
+ *   component tag in our schema. We derive it DETERMINISTICALLY from the real postcode via codex
+ *   `departementForCodePostal` (first two digits = département) — salvage-first, no re-derived
+ *   table.
+ *
+ *   Data (`--communes`, opts.communes): REAL BAN (Base Adresse Nationale) commune+postcode+coord
+ *   tuples, one per line, TAB-separated `commune <TAB> postcode <TAB> lon <TAB> lat`. Build the
+ *   input TSV once from the BAN staging CSV (see the legacy script header). Anchor-ON by
+ *   construction: rows carry a REAL postcode token in `raw` + a `postcode` component, so the
+ *   training loader paints the anchor feature onto that span automatically. The trailing-postcode
+ *   anchor REINFORCES the FR split (FR postcode is trailing, unlike German PLZ-leading — the v0.9.2
+ *   scar is positional, not universal).
+ */
+import { departementForCodePostal } from "@mailwoman/codex/fr";
+import { tempRootPath } from "@mailwoman/core/data-root";
+import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils";
+import { CSVSpliterator, Delimiters } from "spliterator";
+import { stableSourceID } from "#adapters/utils";
+import { alignRow } from "#utils";
+const DEFAULT_COMMUNES = tempRootPath("reg", "fr-communes.tsv");
+const LICENSE = "BAN (Base Adresse Nationale) commune+postcode tuples, rendered admin-split — see ingest SOURCE";
+/**
+ * Read the distinct commune TSV (commune, postcode, lon, lat); derive the département name.
+ */
+async function readCommunes(path) {
+    const rows = [];
+    // The TSV is headerless — every line is a commune tuple — so `header: false` keeps row 1 instead of
+    // spending it on column names. Source is repo-generated (LF); even under CRLF only the trailing `lat`
+    // column carries a CR, and it's consumed via `Number()` (whitespace-trimming), so it stays harmless.
+    for await (const [commune, postcode, lon, lat] of CSVSpliterator.fromAsync(path, {
+        columnDelimiter: Delimiters.Tab,
+        header: false,
+    })) {
+        if (!commune || !postcode)
+            continue;
+        const dep = departementForCodePostal(postcode);
+        if (!dep)
+            continue; // bad/unmappable postcode — skip (CEDEX, etc.)
+        // Substring invariant: a département whose name isn't a clean token (none are) or a commune
+        // containing the département name would confuse alignment — both are vanishingly rare here.
+        rows.push({ commune, postcode, departement: dep.name, lon, lat });
+    }
+    return rows;
+}
+// 10% of communes render upper-case.
+const UPPER_LOCALITY_SHARE = 0.1;
+// Layouts: 25% bare-comma, 25% bare-comma-pc, 20% space-pc, 15% canonical-pc-first, 15% commune-pc.
+const BARE_COMMA_CUTOFF = 0.25;
+const BARE_COMMA_PC_CUTOFF = 0.5;
+const SPACE_PC_CUTOFF = 0.7;
+const CANONICAL_PC_FIRST_CUTOFF = 0.85;
+// 20% of rows append an explicit "France" + a `country` component.
+const APPEND_COUNTRY_SHARE = 0.2;
+/**
+ * Render one admin-split variant. The CORE teaching signal: the département, even as a full word after a comma or a
+ * space, is `region` — never folded into `locality`. Variants 1-3 are the failure class; 4-5 are canonical-FR
+ * preservation so the model doesn't over-fire region on every trailing token (and the bare commune still resolves).
+ */
+function render(random, c) {
+    const r = random();
+    const loc = random() < UPPER_LOCALITY_SHARE ? c.commune.toUpperCase() : c.commune;
+    const dep = c.departement;
+    const pc = c.postcode;
+    let out;
+    if (r < BARE_COMMA_CUTOFF) {
+        // 1. bare comma, NO postcode — the Thauron/#727 shape (anchor off)
+        out = { raw: `${loc}, ${dep}`, components: { locality: loc, region: dep }, order: "bare-comma" };
+    }
+    else if (r < BARE_COMMA_PC_CUTOFF) {
+        // 2. bare comma + postcode — anchor ON
+        out = {
+            raw: `${loc}, ${dep} ${pc}`,
+            components: { locality: loc, region: dep, postcode: pc },
+            order: "bare-comma-pc",
+        };
+    }
+    else if (r < SPACE_PC_CUTOFF) {
+        // 3. space-delimited admin (the AU `CANBERRA ACT` fuse applied to FR) — anchor ON
+        out = { raw: `${loc} ${dep} ${pc}`, components: { locality: loc, region: dep, postcode: pc }, order: "space-pc" };
+    }
+    else if (r < CANONICAL_PC_FIRST_CUTOFF) {
+        // 4. canonical FR postcode-first (NO département) — preservation, anchor ON
+        out = { raw: `${pc} ${loc}`, components: { postcode: pc, locality: loc }, order: "canonical-pc-first" };
+    }
+    else {
+        // 5. commune + postcode (NO département) — preservation, anchor ON
+        out = { raw: `${loc} ${pc}`, components: { locality: loc, postcode: pc }, order: "commune-pc" };
+    }
+    // fr.country preservation (the v1.8.0 #728 finding): the v1.8.0 slice's bare rows carried NO country
+    // token, so the model under-emitted country on FR (fr.country −3.5pp). ~20% of rows now append an
+    // explicit "France" + a `country` component — the model relearns to emit country WHEN the token is
+    // present without over-firing it on the (still-majority) country-less rows. Substring invariant holds.
+    if (random() < APPEND_COUNTRY_SHARE) {
+        out = {
+            raw: `${out.raw}, France`,
+            components: { ...out.components, country: "France" },
+            order: `${out.order}+fr`,
+        };
+    }
+    return out;
+}
+/**
+ * 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 frAdminSplitRecipe = {
+    name: "fr-admin-split",
+    description: "FR admin-split rows: BAN communes → split département into `region` (+ canonical-FR preservation)",
+    mode: "generate",
+    options: [
+        { flag: "--communes <tsv>", description: "BAN commune+postcode+coord TSV. Default /tmp/reg/fr-communes.tsv" },
+    ],
+    async run(opts, write) {
+        // Legacy build-fr-admin-split-slice.mjs seeded `mulberry32(opts.seed)`.
+        const random = makeMulberry32(opts.seed);
+        const count = opts.count ?? 60_000;
+        const source = opts.sourceName ?? "synth-fr-admin-split";
+        const communesPath = opts.communes ?? DEFAULT_COMMUNES;
+        const pool = await readCommunes(communesPath);
+        console.error(`  ${communesPath}: ${pool.length} communes with derived département`);
+        if (!pool.length) {
+            throw new Error("No communes — build the TSV from BAN first (see the recipe header).");
+        }
+        let emitted = 0;
+        let skipped = 0;
+        let guard = 0;
+        const orderCounts = {};
+        const N = pool.length;
+        while (emitted < count && guard++ < count * 12) {
+            const base = pool[Math.floor(random() * N)];
+            const { raw, components, order } = render(random, base);
+            // Alignment precondition: every component surface appears verbatim in raw.
+            const values = Object.values(components).filter((v) => Boolean(v));
+            if (!values.every((v) => raw.includes(v))) {
+                skipped++;
+                continue;
+            }
+            if (opts.golden) {
+                // Held-out eval slice for the centroid check — carries the truth coordinate.
+                write(JSON.stringify({ raw, components, country: "FR", lat: Number(base.lat), lon: Number(base.lon) }) + "\n");
+                emitted++;
+                orderCounts[order] = (orderCounts[order] ?? 0) + 1;
+                continue;
+            }
+            const sourceID = stableSourceID(source, {
+                locality: components.locality,
+                region: components.region,
+                postcode: components.postcode,
+            });
+            const canonical = {
+                raw,
+                components,
+                country: "FR",
+                locale: "fr-FR",
+                source,
+                source_id: sourceID,
+                corpus_version: "0.5.0",
+                license: LICENSE,
+            };
+            const aligned = alignRow(canonical);
+            if (aligned.kind !== "labeled" || !aligned.row) {
+                skipped++;
+                continue;
+            }
+            write(JSON.stringify({ ...aligned.row, synth_method: "fr-admin-split", synth_order: order, synth_base_id: null }) +
+                "\n");
+            emitted++;
+            orderCounts[order] = (orderCounts[order] ?? 0) + 1;
+        }
+        console.error(`  emitted=${emitted} skipped=${skipped} order-mix=${JSON.stringify(orderCounts)}`);
+        return { emitted, skipped };
+    },
+};
+//# sourceMappingURL=fr-admin-split.js.map

package/out/recipes/fr-admin-split.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fr-admin-split.js","sourceRoot":"","sources":["../../lib/recipes/fr-admin-split.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,EAAE,wBAAwB,EAAE,MAAM,qBAAqB,CAAA;AAC9D,OAAO,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAA;AAExD,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AACpE,OAAO,EAAE,cAAc,EAAE,UAAU,EAAE,MAAM,aAAa,CAAA;AAExD,OAAO,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAA;AAGhD,OAAO,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAA;AAEjC,MAAM,gBAAgB,GAAG,YAAY,CAAC,KAAK,EAAE,iBAAiB,CAAC,CAAA;AAC/D,MAAM,OAAO,GAAG,gGAAgG,CAAA;AAsBhH;;GAEG;AACH,KAAK,UAAU,YAAY,CAAC,IAAY;IACvC,MAAM,IAAI,GAAiB,EAAE,CAAA;IAE7B,oGAAoG;IACpG,sGAAsG;IACtG,qGAAqG;IACrG,IAAI,KAAK,EAAE,MAAM,CAAC,OAAO,EAAE,QAAQ,EAAE,GAAG,EAAE,GAAG,CAAC,IAAI,cAAc,CAAC,SAAS,CAAC,IAAI,EAAE;QAChF,eAAe,EAAE,UAAU,CAAC,GAAG;QAC/B,MAAM,EAAE,KAAK;KACb,CAAC,EAAE,CAAC;QACJ,IAAI,CAAC,OAAO,IAAI,CAAC,QAAQ;YAAE,SAAQ;QACnC,MAAM,GAAG,GAAG,wBAAwB,CAAC,QAAQ,CAAC,CAAA;QAE9C,IAAI,CAAC,GAAG;YAAE,SAAQ,CAAC,+CAA+C;QAClE,4FAA4F;QAC5F,4FAA4F;QAC5F,IAAI,CAAC,IAAI,CAAC,EAAE,OAAO,EAAE,QAAQ,EAAE,WAAW,EAAE,GAAG,CAAC,IAAI,EAAE,GAAG,EAAE,GAAG,EAAE,CAAC,CAAA;IAClE,CAAC;IAED,OAAO,IAAI,CAAA;AACZ,CAAC;AAED,qCAAqC;AACrC,MAAM,oBAAoB,GAAG,GAAG,CAAA;AAChC,oGAAoG;AACpG,MAAM,iBAAiB,GAAG,IAAI,CAAA;AAC9B,MAAM,oBAAoB,GAAG,GAAG,CAAA;AAChC,MAAM,eAAe,GAAG,GAAG,CAAA;AAC3B,MAAM,yBAAyB,GAAG,IAAI,CAAA;AACtC,mEAAmE;AACnE,MAAM,oBAAoB,GAAG,GAAG,CAAA;AAEhC;;;;GAIG;AACH,SAAS,MAAM,CAAC,MAAoB,EAAE,CAAa;IAClD,MAAM,CAAC,GAAG,MAAM,EAAE,CAAA;IAClB,MAAM,GAAG,GAAG,MAAM,EAAE,GAAG,oBAAoB,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAA;IACjF,MAAM,GAAG,GAAG,CAAC,CAAC,WAAW,CAAA;IACzB,MAAM,EAAE,GAAG,CAAC,CAAC,QAAQ,CAAA;IACrB,IAAI,GAAsB,CAAA;IAE1B,IAAI,CAAC,GAAG,iBAAiB,EAAE,CAAC;QAC3B,mEAAmE;QACnE,GAAG,GAAG,EAAE,GAAG,EAAE,GAAG,GAAG,KAAK,GAAG,EAAE,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,YAAY,EAAE,CAAA;IACjG,CAAC;SAAM,IAAI,CAAC,GAAG,oBAAoB,EAAE,CAAC;QACrC,uCAAuC;QACvC,GAAG,GAAG;YACL,GAAG,EAAE,GAAG,GAAG,KAAK,GAAG,IAAI,EAAE,EAAE;YAC3B,UAAU,EAAE,EAAE,QAAQ,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE,EAAE;YACxD,KAAK,EAAE,eAAe;SACtB,CAAA;IACF,CAAC;SAAM,IAAI,CAAC,GAAG,eAAe,EAAE,CAAC;QAChC,kFAAkF;QAClF,GAAG,GAAG,EAAE,GAAG,EAAE,GAAG,GAAG,IAAI,GAAG,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAA;IAClH,CAAC;SAAM,IAAI,CAAC,GAAG,yBAAyB,EAAE,CAAC;QAC1C,4EAA4E;QAC5E,GAAG,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,IAAI,GAAG,EAAE,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAA;IACxG,CAAC;SAAM,CAAC;QACP,mEAAmE;QACnE,GAAG,GAAG,EAAE,GAAG,EAAE,GAAG,GAAG,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE,EAAE,EAAE,KAAK,EAAE,YAAY,EAAE,CAAA;IAChG,CAAC;IAED,qGAAqG;IACrG,kGAAkG;IAClG,mGAAmG;IACnG,uGAAuG;IACvG,IAAI,MAAM,EAAE,GAAG,oBAAoB,EAAE,CAAC;QACrC,GAAG,GAAG;YACL,GAAG,EAAE,GAAG,GAAG,CAAC,GAAG,UAAU;YACzB,UAAU,EAAE,EAAE,GAAG,GAAG,CAAC,UAAU,EAAE,OAAO,EAAE,QAAQ,EAAE;YACpD,KAAK,EAAE,GAAG,GAAG,CAAC,KAAK,KAAK;SACxB,CAAA;IACF,CAAC;IAED,OAAO,GAAG,CAAA;AACX,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAiB;IAC/C,IAAI,EAAE,gBAAgB;IACtB,WAAW,EAAE,mGAAmG;IAChH,IAAI,EAAE,UAAU;IAChB,OAAO,EAAE;QACR,EAAE,IAAI,EAAE,kBAAkB,EAAE,WAAW,EAAE,kEAAkE,EAAE;KAC7G;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,wEAAwE;QACxE,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,MAAM,CAAA;QAClC,MAAM,MAAM,GAAG,IAAI,CAAC,UAAU,IAAI,sBAAsB,CAAA;QACxD,MAAM,YAAY,GAAG,IAAI,CAAC,QAAQ,IAAI,gBAAgB,CAAA;QAEtD,MAAM,IAAI,GAAG,MAAM,YAAY,CAAC,YAAY,CAAC,CAAA;QAE7C,OAAO,CAAC,KAAK,CAAC,KAAK,YAAY,KAAK,IAAI,CAAC,MAAM,oCAAoC,CAAC,CAAA;QAEpF,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CAAC,qEAAqE,CAAC,CAAA;QACvF,CAAC;QAED,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,KAAK,GAAG,CAAC,CAAA;QACb,MAAM,WAAW,GAA2B,EAAE,CAAA;QAC9C,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAA;QAErB,OAAO,OAAO,GAAG,KAAK,IAAI,KAAK,EAAE,GAAG,KAAK,GAAG,EAAE,EAAE,CAAC;YAChD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAE,CAAA;YAC5C,MAAM,EAAE,GAAG,EAAE,UAAU,EAAE,KAAK,EAAE,GAAG,MAAM,CAAC,MAAM,EAAE,IAAI,CAAC,CAAA;YAEvD,2EAA2E;YAC3E,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAe,EAAE,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAA;YAE/E,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC3C,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;gBACjB,6EAA6E;gBAC7E,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,EAAE,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,EAAE,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,CAAA;gBAE9G,OAAO,EAAE,CAAA;gBACT,WAAW,CAAC,KAAK,CAAC,GAAG,CAAC,WAAW,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAA;gBAElD,SAAQ;YACT,CAAC;YAED,MAAM,QAAQ,GAAG,cAAc,CAAC,MAAM,EAAE;gBACvC,QAAQ,EAAE,UAAU,CAAC,QAAQ;gBAC7B,MAAM,EAAE,UAAU,CAAC,MAAM;gBACzB,QAAQ,EAAE,UAAU,CAAC,QAAQ;aAC7B,CAAC,CAAA;YAEF,MAAM,SAAS,GAAiB;gBAC/B,GAAG;gBACH,UAAU;gBACV,OAAO,EAAE,IAAI;gBACb,MAAM,EAAE,OAAO;gBACf,MAAM;gBACN,SAAS,EAAE,QAAQ;gBACnB,cAAc,EAAE,OAAO;gBACvB,OAAO,EAAE,OAAO;aAChB,CAAA;YAED,MAAM,OAAO,GAAG,QAAQ,CAAC,SAAS,CAAC,CAAA;YAEnC,IAAI,OAAO,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,GAAG,EAAE,CAAC;gBAChD,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,KAAK,CACJ,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,OAAO,CAAC,GAAG,EAAE,YAAY,EAAE,gBAAgB,EAAE,WAAW,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC;gBAC1G,IAAI,CACL,CAAA;YAED,OAAO,EAAE,CAAA;YACT,WAAW,CAAC,KAAK,CAAC,GAAG,CAAC,WAAW,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAA;QACnD,CAAC;QAED,OAAO,CAAC,KAAK,CAAC,aAAa,OAAO,YAAY,OAAO,cAAc,IAAI,CAAC,SAAS,CAAC,WAAW,CAAC,EAAE,CAAC,CAAA;QAEjG,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAC5B,CAAC;CACD,CAAA"}

package/out/recipes/fr-bare-street.d.ts

@@ -0,0 +1,28 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-bare-street` slice recipe (#251) — the postcode-anchoring-imbalance change. BAN (and every
+ *   other comprehensive FR source) is postcode-COMPLETE, so the model learned the French
+ *   street→locality boundary as "the token after the 5-digit postcode," never as "comma + city." Strip
+ *   the postcode and it leaks the street's proper-noun tokens into the following locality ("Rue René
+ *   Cassin, Paris" → street="Rue Ren", locality="Cassin"). This recipe mints the MISSING distribution:
+ *   the BARE comma form, NO postcode, real `(street, number, city)` tuples from BAN (Licence Ouverte —
+ *   permissive; the model stays clean of ODbL, unlike the opt-in OSM rooftop slices).
+ *
+ *   Each tuple → `<n> <Rue/Avenue/…> <proper-noun name>, <City>` with the FR prefix split
+ *   ({@link decomposeFrStreet}: "Rue" → street_prefix, the rest → street). Tuples whose street carries
+ *   no recognized FR type word are skipped — the failing class is precisely the prefix-led street.
+ *
+ *   ⚠ Convention loss-mask: 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 slice.
+ */
+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 frBareStreetRecipe: CorpusRecipe;
+//# sourceMappingURL=fr-bare-street.d.ts.map

package/out/recipes/fr-bare-street.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"fr-bare-street.d.ts","sourceRoot":"","sources":["../../lib/recipes/fr-bare-street.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAMH,OAAO,EAA6B,KAAK,YAAY,EAAiB,MAAM,mBAAmB,CAAA;AAgB/F;;;GAGG;AACH,eAAO,MAAM,kBAAkB,EAAE,YA4HhC,CAAA"}

package/out/recipes/fr-bare-street.js

@@ -0,0 +1,144 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-bare-street` slice recipe (#251) — the postcode-anchoring-imbalance change. BAN (and every
+ *   other comprehensive FR source) is postcode-COMPLETE, so the model learned the French
+ *   street→locality boundary as "the token after the 5-digit postcode," never as "comma + city." Strip
+ *   the postcode and it leaks the street's proper-noun tokens into the following locality ("Rue René
+ *   Cassin, Paris" → street="Rue Ren", locality="Cassin"). This recipe mints the MISSING distribution:
+ *   the BARE comma form, NO postcode, real `(street, number, city)` tuples from BAN (Licence Ouverte —
+ *   permissive; the model stays clean of ODbL, unlike the opt-in OSM rooftop slices).
+ *
+ *   Each tuple → `<n> <Rue/Avenue/…> <proper-noun name>, <City>` with the FR prefix split
+ *   ({@link decomposeFrStreet}: "Rue" → street_prefix, the rest → street). Tuples whose street carries
+ *   no recognized FR type word are skipped — the failing class is precisely the prefix-led street.
+ *
+ *   ⚠ Convention loss-mask: 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 slice.
+ */
+import { FR_VOIE_TYPES } from "@mailwoman/codex/fr";
+import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils";
+import { decomposeFrStreet } from "#adapters/ban/street-decompose";
+import { alignAndWrite, readTuples, sliceSourceID } from "#recipes/scaffold";
+/**
+ * Canonical voie type (lowercase, accent-kept) → its most common written abbreviation, from the codex table's first
+ * entry. Types with no attested abbreviation stay canonical in the abbreviated form.
+ */
+const FR_VOIE_ABBREV = Object.fromEntries(Object.entries(FR_VOIE_TYPES).flatMap(([canonical, abbrevs]) => (abbrevs[0] ? [[canonical, abbrevs[0]]] : [])));
+/**
+ * The order-cycle slot for the BARE-STREET-ONLY form (`«voie» «name»`, no number, no locality) — the absence
+ * counterweight to the locality-terminated comma-free forms (see the cycle comment).
+ */
+const BARE_STREET_ONLY_FORM = 3;
+/**
+ * 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 frBareStreetRecipe = {
+    name: "fr-bare-street",
+    description: "FR bare street+city, NO postcode (#251): comma / comma-free / abbreviated-voie surfaces over one label set",
+    mode: "tuples",
+    async run(opts, write) {
+        // Seeded for parity with the other recipes; unused beyond reproducibility (the tuples drive the content).
+        makeMulberry32(opts.seed);
+        let read = 0;
+        let emitted = 0;
+        let skipped = 0;
+        for await (const t of readTuples(opts.input)) {
+            read++;
+            const fullStreet = String(t.street ?? "").trim();
+            const number = String(t.number ?? "").trim();
+            const locality = String(t.locality ?? "").trim();
+            if (!fullStreet || !number || !locality) {
+                skipped++;
+                continue;
+            }
+            const { prefix, street } = decomposeFrStreet(fullStreet);
+            // A no-prefix nom_voie ("La Ville Mois") is not the prefix-led class the numbered forms
+            // exercise — but as a BARE surface it is exactly the non-voie-led counterweight the v4.5.1
+            // probe showed missing ('Savile Row'-shaped spans still fell to the trailing-locality
+            // prior; the voie-led bare form protected only voie-led spans). Alternate rows emit the
+            // whole span as a bare street; the rest skip as before.
+            if (!prefix || !street) {
+                if (read % 2 === 0 && fullStreet.split(" ").length >= 2) {
+                    const bare = {
+                        raw: fullStreet,
+                        components: { street: fullStreet },
+                        country: "FR",
+                        locale: "fr-FR",
+                        source: "synth-fr-bare-street",
+                        source_id: sliceSourceID("synth-fr-bare-street", {
+                            street: fullStreet,
+                            f: "bare-nonvoie",
+                            v: String(read),
+                        }),
+                        corpus_version: "0.9.4",
+                        license: "Synthetic — fr-bare-street; (street, number, city) from BAN (Base Adresse Nationale, Licence Ouverte)",
+                    };
+                    if (alignAndWrite(write, bare, "fr-bare-street")) {
+                        emitted++;
+                    }
+                    else {
+                        skipped++;
+                    }
+                    continue;
+                }
+                skipped++;
+                continue;
+            }
+            // Four surfaces over the same tuple, cycled deterministically. The comma form was the
+            // original change; the COMMA-FREE form is the colloquial register users actually type
+            // ('12 rue de Rome Paris' — the street↔locality boundary with NO delimiter, the fr-fr
+            // panel's named loss); the ABBREVIATED form is the typeahead register the geocoder-tester
+            // FR slice attests at scale; and the BARE-STREET-ONLY form is the absence counterweight —
+            // without it, every delimiter-free surface in the mix ENDS in a locality, the model learns
+            // "trailing span = locality" as categorical, and bare street names across locales flip to
+            // locality wholesale (the v4.5.0 no-promote's measured erosion: 'Calle de Alcalá',
+            // 'Madison Square West', and COMER's fork all fell to that prior). Tags are identical
+            // where present; each component VALUE is the span as written (BIO alignment binds value
+            // to surface); the bare form carries NO number and NO locality because the surface has
+            // neither.
+            const form = read % 4;
+            const prefixSurface = form >= 2 ? (FR_VOIE_ABBREV[prefix.toLowerCase()] ?? prefix) : prefix;
+            const components = form === BARE_STREET_ONLY_FORM
+                ? { street_prefix: prefix, street }
+                : { house_number: number, street_prefix: prefixSurface, street, locality };
+            // When the tuple carries a WOF-attested neighbourhood, the comma slot renders the
+            // THREE-slot middle surface — the dependent-locality counterweight (the v4.5.1 erosion's
+            // untouched half: the two-slot comma-free endings squeezed the middle tag out).
+            const hood = String(t.neighbourhood ?? "").trim();
+            if (form === 0 && hood) {
+                components.dependent_locality = hood;
+            }
+            const raw = form === 0
+                ? hood
+                    ? `${number} ${prefix} ${street}, ${hood}, ${locality}`
+                    : `${number} ${prefix} ${street}, ${locality}`
+                : form === BARE_STREET_ONLY_FORM
+                    ? `${prefix} ${street}`
+                    : `${number} ${prefixSurface} ${street} ${locality}`;
+            const source_id = sliceSourceID("synth-fr-bare-street", { ...components, f: String(form), v: String(read) });
+            const canonical = {
+                raw,
+                components,
+                country: "FR",
+                locale: "fr-FR",
+                source: "synth-fr-bare-street",
+                source_id,
+                corpus_version: "0.9.4",
+                license: "Synthetic — fr-bare-street; (street, number, city) from BAN (Base Adresse Nationale, Licence Ouverte)",
+            };
+            if (alignAndWrite(write, canonical, "fr-bare-street")) {
+                emitted++;
+            }
+            else {
+                skipped++;
+            }
+        }
+        return { read, emitted, skipped };
+    },
+};
+//# sourceMappingURL=fr-bare-street.js.map

package/out/recipes/fr-bare-street.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fr-bare-street.js","sourceRoot":"","sources":["../../lib/recipes/fr-bare-street.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAEH,OAAO,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAA;AACnD,OAAO,EAAE,UAAU,IAAI,cAAc,EAAE,MAAM,uBAAuB,CAAA;AAEpE,OAAO,EAAE,iBAAiB,EAAE,MAAM,gCAAgC,CAAA;AAClE,OAAO,EAAE,aAAa,EAAE,UAAU,EAAqB,aAAa,EAAE,MAAM,mBAAmB,CAAA;AAE/F;;;GAGG;AACH,MAAM,cAAc,GAA2B,MAAM,CAAC,WAAW,CAChE,MAAM,CAAC,OAAO,CAAC,aAAa,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,EAAE,OAAO,CAAC,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAC9G,CAAA;AAED;;;GAGG;AACH,MAAM,qBAAqB,GAAG,CAAC,CAAA;AAE/B;;;GAGG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAiB;IAC/C,IAAI,EAAE,gBAAgB;IACtB,WAAW,EACV,4GAA4G;IAC7G,IAAI,EAAE,QAAQ;IACd,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,0GAA0G;QAC1G,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,UAAU,GAAG,MAAM,CAAC,CAAC,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAChD,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,UAAU,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACzC,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,iBAAiB,CAAC,UAAU,CAAC,CAAA;YAExD,wFAAwF;YACxF,2FAA2F;YAC3F,sFAAsF;YACtF,wFAAwF;YACxF,wDAAwD;YACxD,IAAI,CAAC,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC;gBACxB,IAAI,IAAI,GAAG,CAAC,KAAK,CAAC,IAAI,UAAU,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,EAAE,CAAC;oBACzD,MAAM,IAAI,GAAG;wBACZ,GAAG,EAAE,UAAU;wBACf,UAAU,EAAE,EAAE,MAAM,EAAE,UAAU,EAA4B;wBAC5D,OAAO,EAAE,IAAI;wBACb,MAAM,EAAE,OAAO;wBACf,MAAM,EAAE,sBAAsB;wBAC9B,SAAS,EAAE,aAAa,CAAC,sBAAsB,EAAE;4BAChD,MAAM,EAAE,UAAU;4BAClB,CAAC,EAAE,cAAc;4BACjB,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC;yBACf,CAAC;wBACF,cAAc,EAAE,OAAO;wBACvB,OAAO,EACN,uGAAuG;qBACxG,CAAA;oBAED,IAAI,aAAa,CAAC,KAAK,EAAE,IAAI,EAAE,gBAAgB,CAAC,EAAE,CAAC;wBAClD,OAAO,EAAE,CAAA;oBACV,CAAC;yBAAM,CAAC;wBACP,OAAO,EAAE,CAAA;oBACV,CAAC;oBAED,SAAQ;gBACT,CAAC;gBAED,OAAO,EAAE,CAAA;gBAET,SAAQ;YACT,CAAC;YAED,sFAAsF;YACtF,sFAAsF;YACtF,sFAAsF;YACtF,0FAA0F;YAC1F,0FAA0F;YAC1F,2FAA2F;YAC3F,0FAA0F;YAC1F,mFAAmF;YACnF,sFAAsF;YACtF,wFAAwF;YACxF,uFAAuF;YACvF,WAAW;YACX,MAAM,IAAI,GAAG,IAAI,GAAG,CAAC,CAAA;YACrB,MAAM,aAAa,GAAG,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,cAAc,CAAC,MAAM,CAAC,WAAW,EAAE,CAAC,IAAI,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAA;YAE3F,MAAM,UAAU,GACf,IAAI,KAAK,qBAAqB;gBAC7B,CAAC,CAAC,EAAE,aAAa,EAAE,MAAM,EAAE,MAAM,EAAE;gBACnC,CAAC,CAAC,EAAE,YAAY,EAAE,MAAM,EAAE,aAAa,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAA;YAE5E,kFAAkF;YAClF,yFAAyF;YACzF,gFAAgF;YAChF,MAAM,IAAI,GAAG,MAAM,CAAC,CAAC,CAAC,aAAa,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;YAEjD,IAAI,IAAI,KAAK,CAAC,IAAI,IAAI,EAAE,CAAC;gBACxB,UAAU,CAAC,kBAAkB,GAAG,IAAI,CAAA;YACrC,CAAC;YAED,MAAM,GAAG,GACR,IAAI,KAAK,CAAC;gBACT,CAAC,CAAC,IAAI;oBACL,CAAC,CAAC,GAAG,MAAM,IAAI,MAAM,IAAI,MAAM,KAAK,IAAI,KAAK,QAAQ,EAAE;oBACvD,CAAC,CAAC,GAAG,MAAM,IAAI,MAAM,IAAI,MAAM,KAAK,QAAQ,EAAE;gBAC/C,CAAC,CAAC,IAAI,KAAK,qBAAqB;oBAC/B,CAAC,CAAC,GAAG,MAAM,IAAI,MAAM,EAAE;oBACvB,CAAC,CAAC,GAAG,MAAM,IAAI,aAAa,IAAI,MAAM,IAAI,QAAQ,EAAE,CAAA;YAEvD,MAAM,SAAS,GAAG,aAAa,CAAC,sBAAsB,EAAE,EAAE,GAAG,UAAU,EAAE,CAAC,EAAE,MAAM,CAAC,IAAI,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,sBAAsB;gBAC9B,SAAS;gBACT,cAAc,EAAE,OAAO;gBACvB,OAAO,EACN,uGAAuG;aACxG,CAAA;YAED,IAAI,aAAa,CAAC,KAAK,EAAE,SAAS,EAAE,gBAAgB,CAAC,EAAE,CAAC;gBACvD,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/fr-fragment.d.ts

@@ -0,0 +1,77 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-fragment` slice recipe (#727 T2) — the HOUSE-NUMBER-LICENCE change.
+ *
+ *   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 slice
+ *   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 slice.
+ *
+ *   MIX. The slice is ~145K rows off a 120K-tuple draw; the intended corpus mix is **5–10%**, set at
+ *   assembly time by slice weight rather than by row count. Keep the cap: a slice 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 slice 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
+ *   change is slice 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 slice.
+ */
+import { type CorpusRecipe } from "#recipes/scaffold";
+export declare function frTitleCase(value: string): string;
+/**
+ * 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 frFragmentRecipe: CorpusRecipe;
+//# sourceMappingURL=fr-fragment.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"fr-fragment.d.ts","sourceRoot":"","sources":["../../lib/recipes/fr-fragment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoEG;AAMH,OAAO,EAA6B,KAAK,YAAY,EAAiB,MAAM,mBAAmB,CAAA;AAqD/F,wBAAgB,WAAW,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAajD;AAgBD;;;GAGG;AACH,eAAO,MAAM,gBAAgB,EAAE,YA6K9B,CAAA"}