@mailwoman/corpus 4.16.2 → 5.1.0

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

@@ -0,0 +1,175 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-admin-split` shard recipe — the FR admin-split coverage shard (night 2026-06-19,
+ *   surpass-v1.5.0). Teaches the model to SPLIT the département out of the locality on
+ *   bare/space/comma-delimited French place rows — the admin-deciding failure class the pre-GPU
+ *   self-validation proved moves the resolved coordinate (collision communes −61%; see
+ *   docs/articles/evals/2026-06-19-fr-admin-split-prevalidation.md). Ported from
+ *   scripts/build-fr-admin-split-shard.mjs.
+ *
+ *   Failure shapes (the model currently mis-handles all of these):
+ *
+ *   - `Thauron, Creuse` → région dropped to null (the comma+full-name miss)
+ *   - `Montredon, Lozère` → région = "ère" (the diacritic subword split, #727)
+ *   - (AU analog) `CANBERRA ACT` → the space-delimited admin fuse
+ *
+ *   The département is the essential admin unit for FR postal geography and maps to the `region`
+ *   component tag in our schema. We derive it DETERMINISTICALLY from the real postcode via codex
+ *   `departementForCodePostal` (first two digits = département) — salvage-first, no re-derived
+ *   table.
+ *
+ *   Data (`--communes`, opts.communes): REAL BAN (Base Adresse Nationale) commune+postcode+coord
+ *   tuples, one per line, TAB-separated `commune <TAB> postcode <TAB> lon <TAB> lat`. Build the
+ *   input TSV once from the BAN staging CSV (see the legacy script header). Anchor-ON by
+ *   construction: rows carry a REAL postcode token in `raw` + a `postcode` component, so the
+ *   training loader paints the anchor feature onto that span automatically. The trailing-postcode
+ *   anchor REINFORCES the FR split (FR postcode is trailing, unlike German PLZ-leading — the v0.9.2
+ *   scar is positional, not universal).
+ */
+import { createReadStream } from "node:fs";
+import { createInterface } from "node:readline";
+import { departementForCodePostal } from "@mailwoman/codex/fr";
+import { stableSourceID } from "../adapter.js";
+import { alignRow } from "../align.js";
+import { makeMulberry32 } from "./scaffold.js";
+const DEFAULT_COMMUNES = "/tmp/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 = [];
+    const rl = createInterface({ input: createReadStream(path, { encoding: "utf8" }), crlfDelay: Infinity });
+    for await (const line of rl) {
+        if (!line)
+            continue;
+        const [commune, postcode, lon, lat] = line.split("\t");
+        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;
+}
+/**
+ * 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() < 0.1 ? c.commune.toUpperCase() : c.commune;
+    const dep = c.departement;
+    const pc = c.postcode;
+    let out;
+    if (r < 0.25) {
+        // 1. bare comma, NO postcode — the Thauron/#727 shape (anchor off)
+        out = { raw: `${loc}, ${dep}`, components: { locality: loc, region: dep }, order: "bare-comma" };
+    }
+    else if (r < 0.5) {
+        // 2. bare comma + postcode — anchor ON
+        out = {
+            raw: `${loc}, ${dep} ${pc}`,
+            components: { locality: loc, region: dep, postcode: pc },
+            order: "bare-comma-pc",
+        };
+    }
+    else if (r < 0.7) {
+        // 3. space-delimited admin (the AU `CANBERRA ACT` fuse applied to FR) — anchor ON
+        out = { raw: `${loc} ${dep} ${pc}`, components: { locality: loc, region: dep, postcode: pc }, order: "space-pc" };
+    }
+    else if (r < 0.85) {
+        // 4. canonical FR postcode-first (NO département) — preservation, anchor ON
+        out = { raw: `${pc} ${loc}`, components: { postcode: pc, locality: loc }, order: "canonical-pc-first" };
+    }
+    else {
+        // 5. commune + postcode (NO département) — preservation, anchor ON
+        out = { raw: `${loc} ${pc}`, components: { locality: loc, postcode: pc }, order: "commune-pc" };
+    }
+    // fr.country preservation (the v1.8.0 #728 finding): the v1.8.0 shard's bare rows carried NO country
+    // token, so the model under-emitted country on FR (fr.country −3.5pp). ~20% of rows now append an
+    // explicit "France" + a `country` component — the model relearns to emit country WHEN the token is
+    // present without over-firing it on the (still-majority) country-less rows. Substring invariant holds.
+    if (random() < 0.2) {
+        out = {
+            raw: `${out.raw}, France`,
+            components: { ...out.components, country: "France" },
+            order: `${out.order}+fr`,
+        };
+    }
+    return out;
+}
+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-shard.mjs seeded `mulberry32(opts.seed)`.
+        const random = makeMulberry32(opts.seed);
+        const count = opts.count ?? 60000;
+        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 === 0) {
+            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 gate — 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/src/shard-recipes/fr-admin-split.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fr-admin-split.js","sourceRoot":"","sources":["../../../src/shard-recipes/fr-admin-split.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,EAAE,gBAAgB,EAAE,MAAM,SAAS,CAAA;AAC1C,OAAO,EAAE,eAAe,EAAE,MAAM,eAAe,CAAA;AAE/C,OAAO,EAAE,wBAAwB,EAAE,MAAM,qBAAqB,CAAA;AAG9D,OAAO,EAAE,cAAc,EAAE,MAAM,eAAe,CAAA;AAC9C,OAAO,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAA;AAEtC,OAAO,EAAE,cAAc,EAAoB,MAAM,eAAe,CAAA;AAEhE,MAAM,gBAAgB,GAAG,0BAA0B,CAAA;AACnD,MAAM,OAAO,GAAG,gGAAgG,CAAA;AAkBhH,gGAAgG;AAChG,KAAK,UAAU,YAAY,CAAC,IAAY;IACvC,MAAM,IAAI,GAAiB,EAAE,CAAA;IAC7B,MAAM,EAAE,GAAG,eAAe,CAAC,EAAE,KAAK,EAAE,gBAAgB,CAAC,IAAI,EAAE,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,QAAQ,EAAE,CAAC,CAAA;IAExG,IAAI,KAAK,EAAE,MAAM,IAAI,IAAI,EAAE,EAAE,CAAC;QAC7B,IAAI,CAAC,IAAI;YAAE,SAAQ;QACnB,MAAM,CAAC,OAAO,EAAE,QAAQ,EAAE,GAAG,EAAE,GAAG,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAA;QAEtD,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;;;;GAIG;AACH,SAAS,MAAM,CAAC,MAAoB,EAAE,CAAa;IAClD,MAAM,CAAC,GAAG,MAAM,EAAE,CAAA;IAClB,MAAM,GAAG,GAAG,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAA;IAChE,MAAM,GAAG,GAAG,CAAC,CAAC,WAAW,CAAA;IACzB,MAAM,EAAE,GAAG,CAAC,CAAC,QAAQ,CAAA;IACrB,IAAI,GAAsB,CAAA;IAE1B,IAAI,CAAC,GAAG,IAAI,EAAE,CAAC;QACd,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,GAAG,EAAE,CAAC;QACpB,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,GAAG,EAAE,CAAC;QACpB,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,IAAI,EAAE,CAAC;QACrB,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,GAAG,EAAE,CAAC;QACpB,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,MAAM,CAAC,MAAM,kBAAkB,GAAgB;IAC9C,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,KAAK,CAAA;QACjC,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;QAC7C,OAAO,CAAC,KAAK,CAAC,KAAK,YAAY,KAAK,IAAI,CAAC,MAAM,oCAAoC,CAAC,CAAA;QAEpF,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACvB,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;gBACT,SAAQ;YACT,CAAC;YAED,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;gBACjB,4EAA4E;gBAC5E,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;gBAC9G,OAAO,EAAE,CAAA;gBACT,WAAW,CAAC,KAAK,CAAC,GAAG,CAAC,WAAW,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAA;gBAClD,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;YACF,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;YACD,MAAM,OAAO,GAAG,QAAQ,CAAC,SAAS,CAAC,CAAA;YAEnC,IAAI,OAAO,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,GAAG,EAAE,CAAC;gBAChD,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YACD,KAAK,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;YACD,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/src/shard-recipes/fr-bare-street.d.ts

@@ -0,0 +1,24 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-bare-street` shard recipe (#251) — the postcode-anchoring-imbalance lever. 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 shards).
+ *
+ *   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 shard.
+ */
+import { type ShardRecipe } from "./scaffold.js";
+export declare const frBareStreetRecipe: ShardRecipe;
+//# sourceMappingURL=fr-bare-street.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"fr-bare-street.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/fr-bare-street.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,OAAO,EAA6C,KAAK,WAAW,EAAiB,MAAM,eAAe,CAAA;AAE1G,eAAO,MAAM,kBAAkB,EAAE,WAwDhC,CAAA"}

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

@@ -0,0 +1,76 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-bare-street` shard recipe (#251) — the postcode-anchoring-imbalance lever. 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 shards).
+ *
+ *   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 shard.
+ */
+import { decomposeFrStreet } from "../adapters/ban/street-decompose.js";
+import { alignAndWrite, makeMulberry32, readTuples, shardSourceID } from "./scaffold.js";
+export const frBareStreetRecipe = {
+    name: "fr-bare-street",
+    description: "FR bare comma-form street+city, NO postcode (#251): '<n> Rue <name>, <City>' — the postcode-anchoring lever",
+    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);
+            // The failing class is the prefix-led FR street; a no-prefix nom_voie ("La Ville Mois") isn't it.
+            if (!prefix || !street) {
+                skipped++;
+                continue;
+            }
+            const components = {
+                house_number: number,
+                street_prefix: prefix,
+                street,
+                locality,
+            };
+            // BARE comma form, number-before (FR's dominant order), NO postcode — the whole point.
+            const raw = `${number} ${prefix} ${street}, ${locality}`;
+            const source_id = shardSourceID("synth-fr-bare-street", { ...components, 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/src/shard-recipes/fr-bare-street.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fr-bare-street.js","sourceRoot":"","sources":["../../../src/shard-recipes/fr-bare-street.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAEH,OAAO,EAAE,iBAAiB,EAAE,MAAM,qCAAqC,CAAA;AACvE,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,UAAU,EAAoB,aAAa,EAAE,MAAM,eAAe,CAAA;AAE1G,MAAM,CAAC,MAAM,kBAAkB,GAAgB;IAC9C,IAAI,EAAE,gBAAgB;IACtB,WAAW,EACV,6GAA6G;IAC9G,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;gBACT,SAAQ;YACT,CAAC;YACD,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,iBAAiB,CAAC,UAAU,CAAC,CAAA;YAExD,kGAAkG;YAClG,IAAI,CAAC,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC;gBACxB,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YACD,MAAM,UAAU,GAA2B;gBAC1C,YAAY,EAAE,MAAM;gBACpB,aAAa,EAAE,MAAM;gBACrB,MAAM;gBACN,QAAQ;aACR,CAAA;YACD,uFAAuF;YACvF,MAAM,GAAG,GAAG,GAAG,MAAM,IAAI,MAAM,IAAI,MAAM,KAAK,QAAQ,EAAE,CAAA;YACxD,MAAM,SAAS,GAAG,aAAa,CAAC,sBAAsB,EAAE,EAAE,GAAG,UAAU,EAAE,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;YAC3F,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;gBAAE,OAAO,EAAE,CAAA;;gBAC3D,OAAO,EAAE,CAAA;QACf,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAClC,CAAC;CACD,CAAA"}

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

@@ -0,0 +1,34 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-order` shard recipe — French reversed-order coverage (#560). Reads REAL OpenAddresses FR
+ *   tuples (`fr/countrywide.csv` from the cached zip), then for each row picks — by
+ *   `--reversed-fraction` (default 0.5) — whether to render in CANONICAL French order
+ *   (number-street, postcode-city) or one of four REVERSED / postcode-first variants the
+ *   v4.4.0→v0.5.0 regression exposed (the model misses house_number in every reversed one):
+ *
+ *   - A: "47110 Sainte-Livrade-sur-Lot, 69 Allée du Bugatel" (postcode city, HN street)
+ *   - B: "Sainte-Livrade-sur-Lot, 47110, 619 Impasse de la Rose" (city, postcode, HN street)
+ *   - C: "Sainte-Livrade-sur-Lot 59 bis Rue des Ecuries 47110" (city HN street postcode — NO commas)
+ *   - D: "47110, 6 rue de la république, Sainte-Livrade-sur-Lot" (postcode, HN street, city)
+ *
+ *   Sub-modes ride alongside order: `bis`/`ter`/`quater` ordinal suffixes in house_number, and
+ *   ALL-CAPS locality. `--golden` emits a held-out reversed-order eval slice with a different
+ *   seed.
+ *
+ *   The inline synthesis (the OA-CSV reader, the ordinal/all-caps tables, the canonical + reversed
+ *   renderers) is ported faithfully from scripts/build-fr-order-shard.mjs. This is a
+ *   `generate`-mode recipe that still reads REAL tuples off disk — `--count` bounds the OUTPUT, not
+ *   the input. The passed `random` (the framework LCG) is consumed in the exact call order the
+ *   legacy script used.
+ *
+ *   NOT ported (diagnostic-only, no effect on emitted bytes): the post-run `runSpanCheck` self-check
+ *   (it reads the finished file back with a separate PRNG and prints to stderr; the recipe's output
+ *   stream is still open during `run`), and the dead `renderReversed` helper (the legacy `main`
+ *   inlined the variant logic and never called it).
+ */
+import { type ShardRecipe } from "./scaffold.js";
+export declare const frOrderRecipe: ShardRecipe;
+//# sourceMappingURL=fr-order.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"fr-order.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/fr-order.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AASH,OAAO,EAAkB,KAAK,WAAW,EAAE,MAAM,eAAe,CAAA;AAwHhE,eAAO,MAAM,aAAa,EAAE,WAoH3B,CAAA"}

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

@@ -0,0 +1,238 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-order` shard recipe — French reversed-order coverage (#560). Reads REAL OpenAddresses FR
+ *   tuples (`fr/countrywide.csv` from the cached zip), then for each row picks — by
+ *   `--reversed-fraction` (default 0.5) — whether to render in CANONICAL French order
+ *   (number-street, postcode-city) or one of four REVERSED / postcode-first variants the
+ *   v4.4.0→v0.5.0 regression exposed (the model misses house_number in every reversed one):
+ *
+ *   - A: "47110 Sainte-Livrade-sur-Lot, 69 Allée du Bugatel" (postcode city, HN street)
+ *   - B: "Sainte-Livrade-sur-Lot, 47110, 619 Impasse de la Rose" (city, postcode, HN street)
+ *   - C: "Sainte-Livrade-sur-Lot 59 bis Rue des Ecuries 47110" (city HN street postcode — NO commas)
+ *   - D: "47110, 6 rue de la république, Sainte-Livrade-sur-Lot" (postcode, HN street, city)
+ *
+ *   Sub-modes ride alongside order: `bis`/`ter`/`quater` ordinal suffixes in house_number, and
+ *   ALL-CAPS locality. `--golden` emits a held-out reversed-order eval slice with a different
+ *   seed.
+ *
+ *   The inline synthesis (the OA-CSV reader, the ordinal/all-caps tables, the canonical + reversed
+ *   renderers) is ported faithfully from scripts/build-fr-order-shard.mjs. This is a
+ *   `generate`-mode recipe that still reads REAL tuples off disk — `--count` bounds the OUTPUT, not
+ *   the input. The passed `random` (the framework LCG) is consumed in the exact call order the
+ *   legacy script used.
+ *
+ *   NOT ported (diagnostic-only, no effect on emitted bytes): the post-run `runSpanCheck` self-check
+ *   (it reads the finished file back with a separate PRNG and prints to stderr; the recipe's output
+ *   stream is still open during `run`), and the dead `renderReversed` helper (the legacy `main`
+ *   inlined the variant logic and never called it).
+ */
+import { spawnSync } from "node:child_process";
+import { stableSourceID } from "../adapter.js";
+import { alignRow } from "../align.js";
+import { makeMulberry32 } from "./scaffold.js";
+const SOURCE = { zip: "/tmp/oa-cache/fr__countrywide.zip", csv: "fr/countrywide.csv" };
+// Ordinal suffixes used in French house numbers (BAN corpus), to cover the "8 bis" sub-mode.
+const ORDINAL_SUFFIXES = ["bis", "ter", "quater"];
+// Probability that a row gets an ordinal suffix injected (matches the golden's ~10-15% rate).
+const ORDINAL_PROB = 0.12;
+// Probability that a locality renders ALL-CAPS (another sub-mode: "SAINTE-LIVRADE-SUR-LOT").
+const ALLCAPS_PROB = 0.1;
+/** Minimal RFC-4180-ish splitter (handles quoted fields with doubled-quote escaping). */
+function splitCSV(line) {
+    const out = [];
+    let cur = "";
+    let inQ = false;
+    for (let i = 0; i < line.length; i++) {
+        const c = line[i];
+        if (inQ) {
+            if (c === '"') {
+                if (line[i + 1] === '"') {
+                    cur += '"';
+                    i++;
+                }
+                else
+                    inQ = false;
+            }
+            else
+                cur += c;
+        }
+        else if (c === '"')
+            inQ = true;
+        else if (c === ",") {
+            out.push(cur);
+            cur = "";
+        }
+        else
+            cur += c;
+    }
+    out.push(cur);
+    return out;
+}
+/**
+ * Stream FR tuples out of the cached OA zip. The countrywide extract is GB-scale; cap with `head` to stay under V8's
+ * string limit. Only keeps rows with a house_number (the shard's core signal) and a postcode (required for
+ * reversed-order rendering to be meaningful).
+ */
+function readTuples(limit) {
+    const maxLines = Math.max(limit * 8, 40000) + 1;
+    const r = spawnSync("bash", ["-c", `unzip -p "${SOURCE.zip}" "${SOURCE.csv}" | head -n ${maxLines}`], {
+        maxBuffer: 1024 * 1024 * 1024,
+        encoding: "buffer",
+    });
+    if (r.status !== 0) {
+        console.error(`  WARN: unzip failed for ${SOURCE.zip} (status ${r.status})`);
+        return [];
+    }
+    const lines = r.stdout.toString("utf8").split(/\r?\n/);
+    if (lines.length < 2)
+        return [];
+    const header = splitCSV(lines[0]).map((h) => h.trim().toLowerCase());
+    const idx = (name) => header.indexOf(name);
+    const iNum = idx("number"), iStreet = idx("street"), iCity = idx("city"), iPost = idx("postcode");
+    const get = (cells, i) => (i >= 0 && i < cells.length ? (cells[i] ?? "").trim() : "");
+    const tuples = [];
+    const seen = new Set();
+    for (let li = 1; li < lines.length && tuples.length < limit; li++) {
+        if (!lines[li])
+            continue;
+        const cells = splitCSV(lines[li]);
+        const street = get(cells, iStreet);
+        const locality = get(cells, iCity);
+        const house_number = get(cells, iNum);
+        const postcode = get(cells, iPost);
+        // Require all four fields: HN is the signal; postcode drives reversed-order variants.
+        if (!street || !locality || !house_number || !postcode)
+            continue;
+        const key = `${house_number}|${street}|${locality}|${postcode}`.toLowerCase();
+        if (seen.has(key))
+            continue;
+        seen.add(key);
+        tuples.push({ house_number, street, locality, postcode });
+    }
+    return tuples;
+}
+/**
+ * Optionally augment a house_number with a French ordinal suffix ("59 bis", "4 ter"). Appended with a space so it forms
+ * one multi-token house_number string that alignRow can still locate verbatim.
+ */
+function maybeAddOrdinal(random, house_number) {
+    if (random() >= ORDINAL_PROB)
+        return house_number;
+    const suffix = ORDINAL_SUFFIXES[Math.floor(random() * ORDINAL_SUFFIXES.length)];
+    // Vary suffix case: "bis" (lower) vs "BIS" (upper) — a real-world split in the golden.
+    return `${house_number} ${random() < 0.5 ? suffix : suffix.toUpperCase()}`;
+}
+/** Render a tuple in CANONICAL French order: "9 Rue de la Promenade, 01200 Villes". */
+function renderCanonical(hn, street, postcode, locality) {
+    const raw = `${hn} ${street}, ${postcode} ${locality}`;
+    return { raw, components: { house_number: hn, street, postcode, locality } };
+}
+export const frOrderRecipe = {
+    name: "fr-order",
+    description: "French reversed-order rows (#560): real OA FR tuples rendered canonical + 4 postcode-first variants",
+    mode: "generate",
+    options: [
+        { flag: "--reversed-fraction <p>", description: "Fraction rendered reversed-order. Default 0.5" },
+        { flag: "--golden", description: "Emit the held-out reversed-order eval slice" },
+    ],
+    async run(opts, write) {
+        if (opts.count == null)
+            throw new Error("fr-order recipe requires --count <N>");
+        const count = opts.count;
+        // Legacy build-fr-order-shard.mjs seeded mulberry32 with the raw seed: `const random = mulberry32(opts.seed)`.
+        // (The omitted diagnostic runSpanCheck used a separate mulberry32(opts.seed + 1) — not part of generation.)
+        const random = makeMulberry32(opts.seed);
+        const source = opts.sourceName ?? "synth-fr-order";
+        const reversedFraction = opts.reversedFraction ?? 0.5;
+        // Over-read from the CSV so the dedup + filter pass can fill `count` rows.
+        const poolLimit = Math.max(count * 8, 40000);
+        const pool = readTuples(poolLimit);
+        console.error(`  ${SOURCE.csv}: ${pool.length} unique tuples (capped read)`);
+        if (pool.length === 0) {
+            throw new Error("No FR tuples found — is /tmp/oa-cache/fr__countrywide.zip present?");
+        }
+        let emitted = 0;
+        let skipped = 0;
+        let guard = 0;
+        const N = pool.length;
+        while (emitted < count && guard++ < count * 10) {
+            const base = pool[Math.floor(random() * N)];
+            const { street, postcode } = base;
+            const locality = random() < ALLCAPS_PROB ? base.locality.toUpperCase() : base.locality;
+            const house_number = maybeAddOrdinal(random, base.house_number);
+            // Pick canonical vs reversed by --reversed-fraction.
+            const isReversed = random() < reversedFraction;
+            let rendered;
+            if (isReversed) {
+                const variantRoll = random();
+                let raw;
+                if (variantRoll < 0.25) {
+                    // Variant A: postcode+city as a unit, then HN+street
+                    raw = `${postcode} ${locality}, ${house_number} ${street}`;
+                }
+                else if (variantRoll < 0.5) {
+                    // Variant B: city, then postcode, then HN+street (comma-separated, postcode isolated)
+                    raw = `${locality}, ${postcode}, ${house_number} ${street}`;
+                }
+                else if (variantRoll < 0.75) {
+                    // Variant C: no commas — locality HN street postcode (the "run-together" format)
+                    raw = `${locality} ${house_number} ${street} ${postcode}`;
+                }
+                else {
+                    // Variant D: postcode, HN+street, city (reversed top-to-bottom)
+                    raw = `${postcode}, ${house_number} ${street}, ${locality}`;
+                }
+                rendered = { raw, components: { house_number, street, postcode, locality } };
+            }
+            else {
+                rendered = renderCanonical(house_number, street, postcode, locality);
+            }
+            const { raw, components } = rendered;
+            // Safety check: every component must appear verbatim in raw (alignment precondition).
+            const componentValues = Object.values(components).filter(Boolean);
+            if (!componentValues.every((v) => raw.includes(v))) {
+                skipped++;
+                continue;
+            }
+            // --golden: emit per-locale-f1 eval rows ({raw, components, country:"FR"}).
+            if (opts.golden) {
+                write(JSON.stringify({ raw, components, country: "FR" }) + "\n");
+                emitted++;
+                continue;
+            }
+            const sourceID = stableSourceID(source, {
+                street: components.street,
+                house_number: components.house_number,
+                locality: components.locality,
+                postcode: components.postcode,
+            });
+            const canonical = {
+                raw,
+                components,
+                country: "FR",
+                locale: "fr-FR",
+                source,
+                source_id: sourceID,
+                corpus_version: "0.5.0",
+                license: "OpenAddresses FR countrywide tuples, rendered canonical + reversed-order — see ingest SOURCE",
+            };
+            const aligned = alignRow(canonical);
+            if (aligned.kind !== "labeled" || !aligned.row) {
+                skipped++;
+                continue;
+            }
+            write(JSON.stringify({
+                ...aligned.row,
+                synth_method: "fr-order",
+                synth_order: isReversed ? "reversed" : "canonical",
+                synth_base_id: null,
+            }) + "\n");
+            emitted++;
+        }
+        return { emitted, skipped };
+    },
+};
+//# sourceMappingURL=fr-order.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"fr-order.js","sourceRoot":"","sources":["../../../src/shard-recipes/fr-order.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,EAAE,SAAS,EAAE,MAAM,oBAAoB,CAAA;AAI9C,OAAO,EAAE,cAAc,EAAE,MAAM,eAAe,CAAA;AAC9C,OAAO,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAA;AAEtC,OAAO,EAAE,cAAc,EAAoB,MAAM,eAAe,CAAA;AAEhE,MAAM,MAAM,GAAG,EAAE,GAAG,EAAE,mCAAmC,EAAE,GAAG,EAAE,oBAAoB,EAAE,CAAA;AAEtF,6FAA6F;AAC7F,MAAM,gBAAgB,GAAsB,CAAC,KAAK,EAAE,KAAK,EAAE,QAAQ,CAAC,CAAA;AACpE,8FAA8F;AAC9F,MAAM,YAAY,GAAG,IAAI,CAAA;AACzB,6FAA6F;AAC7F,MAAM,YAAY,GAAG,GAAG,CAAA;AAUxB,yFAAyF;AACzF,SAAS,QAAQ,CAAC,IAAY;IAC7B,MAAM,GAAG,GAAa,EAAE,CAAA;IACxB,IAAI,GAAG,GAAG,EAAE,CAAA;IACZ,IAAI,GAAG,GAAG,KAAK,CAAA;IAEf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACtC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAA;QAEjB,IAAI,GAAG,EAAE,CAAC;YACT,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;gBACf,IAAI,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,GAAG,EAAE,CAAC;oBACzB,GAAG,IAAI,GAAG,CAAA;oBACV,CAAC,EAAE,CAAA;gBACJ,CAAC;;oBAAM,GAAG,GAAG,KAAK,CAAA;YACnB,CAAC;;gBAAM,GAAG,IAAI,CAAC,CAAA;QAChB,CAAC;aAAM,IAAI,CAAC,KAAK,GAAG;YAAE,GAAG,GAAG,IAAI,CAAA;aAC3B,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;YACpB,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;YACb,GAAG,GAAG,EAAE,CAAA;QACT,CAAC;;YAAM,GAAG,IAAI,CAAC,CAAA;IAChB,CAAC;IACD,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IAEb,OAAO,GAAG,CAAA;AACX,CAAC;AAED;;;;GAIG;AACH,SAAS,UAAU,CAAC,KAAa;IAChC,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,EAAE,KAAK,CAAC,GAAG,CAAC,CAAA;IAC/C,MAAM,CAAC,GAAG,SAAS,CAAC,MAAM,EAAE,CAAC,IAAI,EAAE,aAAa,MAAM,CAAC,GAAG,MAAM,MAAM,CAAC,GAAG,eAAe,QAAQ,EAAE,CAAC,EAAE;QACrG,SAAS,EAAE,IAAI,GAAG,IAAI,GAAG,IAAI;QAC7B,QAAQ,EAAE,QAAQ;KAClB,CAAC,CAAA;IAEF,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpB,OAAO,CAAC,KAAK,CAAC,4BAA4B,MAAM,CAAC,GAAG,YAAY,CAAC,CAAC,MAAM,GAAG,CAAC,CAAA;QAE5E,OAAO,EAAE,CAAA;IACV,CAAC;IACD,MAAM,KAAK,GAAG,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAA;IAEtD,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,EAAE,CAAA;IAC/B,MAAM,MAAM,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,CAAA;IACrE,MAAM,GAAG,GAAG,CAAC,IAAY,EAAU,EAAE,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,CAAA;IAC1D,MAAM,IAAI,GAAG,GAAG,CAAC,QAAQ,CAAC,EACzB,OAAO,GAAG,GAAG,CAAC,QAAQ,CAAC,EACvB,KAAK,GAAG,GAAG,CAAC,MAAM,CAAC,EACnB,KAAK,GAAG,GAAG,CAAC,UAAU,CAAC,CAAA;IACxB,MAAM,GAAG,GAAG,CAAC,KAAe,EAAE,CAAS,EAAU,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;IAC/G,MAAM,MAAM,GAAc,EAAE,CAAA;IAC5B,MAAM,IAAI,GAAG,IAAI,GAAG,EAAU,CAAA;IAE9B,KAAK,IAAI,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,KAAK,CAAC,MAAM,IAAI,MAAM,CAAC,MAAM,GAAG,KAAK,EAAE,EAAE,EAAE,EAAE,CAAC;QACnE,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YAAE,SAAQ;QACxB,MAAM,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAE,CAAC,CAAA;QAClC,MAAM,MAAM,GAAG,GAAG,CAAC,KAAK,EAAE,OAAO,CAAC,CAAA;QAClC,MAAM,QAAQ,GAAG,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC,CAAA;QAClC,MAAM,YAAY,GAAG,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,CAAA;QACrC,MAAM,QAAQ,GAAG,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC,CAAA;QAElC,sFAAsF;QACtF,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,IAAI,CAAC,YAAY,IAAI,CAAC,QAAQ;YAAE,SAAQ;QAChE,MAAM,GAAG,GAAG,GAAG,YAAY,IAAI,MAAM,IAAI,QAAQ,IAAI,QAAQ,EAAE,CAAC,WAAW,EAAE,CAAA;QAE7E,IAAI,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC;YAAE,SAAQ;QAC3B,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,CAAA;QACb,MAAM,CAAC,IAAI,CAAC,EAAE,YAAY,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,EAAE,CAAC,CAAA;IAC1D,CAAC;IAED,OAAO,MAAM,CAAA;AACd,CAAC;AAED;;;GAGG;AACH,SAAS,eAAe,CAAC,MAAoB,EAAE,YAAoB;IAClE,IAAI,MAAM,EAAE,IAAI,YAAY;QAAE,OAAO,YAAY,CAAA;IACjD,MAAM,MAAM,GAAG,gBAAgB,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,gBAAgB,CAAC,MAAM,CAAC,CAAE,CAAA;IAEhF,uFAAuF;IACvF,OAAO,GAAG,YAAY,IAAI,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,WAAW,EAAE,EAAE,CAAA;AAC3E,CAAC;AAED,uFAAuF;AACvF,SAAS,eAAe,CACvB,EAAU,EACV,MAAc,EACd,QAAgB,EAChB,QAAgB;IAEhB,MAAM,GAAG,GAAG,GAAG,EAAE,IAAI,MAAM,KAAK,QAAQ,IAAI,QAAQ,EAAE,CAAA;IAEtD,OAAO,EAAE,GAAG,EAAE,UAAU,EAAE,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,EAAE,EAAE,CAAA;AAC7E,CAAC;AAED,MAAM,CAAC,MAAM,aAAa,GAAgB;IACzC,IAAI,EAAE,UAAU;IAChB,WAAW,EAAE,qGAAqG;IAClH,IAAI,EAAE,UAAU;IAChB,OAAO,EAAE;QACR,EAAE,IAAI,EAAE,yBAAyB,EAAE,WAAW,EAAE,+CAA+C,EAAE;QACjG,EAAE,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,6CAA6C,EAAE;KAChF;IACD,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK;QACpB,IAAI,IAAI,CAAC,KAAK,IAAI,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,sCAAsC,CAAC,CAAA;QAC/E,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAA;QACxB,+GAA+G;QAC/G,4GAA4G;QAC5G,MAAM,MAAM,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QACxC,MAAM,MAAM,GAAG,IAAI,CAAC,UAAU,IAAI,gBAAgB,CAAA;QAClD,MAAM,gBAAgB,GAAG,IAAI,CAAC,gBAAgB,IAAI,GAAG,CAAA;QAErD,2EAA2E;QAC3E,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,EAAE,KAAK,CAAC,CAAA;QAC5C,MAAM,IAAI,GAAG,UAAU,CAAC,SAAS,CAAC,CAAA;QAClC,OAAO,CAAC,KAAK,CAAC,KAAK,MAAM,CAAC,GAAG,KAAK,IAAI,CAAC,MAAM,8BAA8B,CAAC,CAAA;QAE5E,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACvB,MAAM,IAAI,KAAK,CAAC,oEAAoE,CAAC,CAAA;QACtF,CAAC;QAED,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,OAAO,GAAG,CAAC,CAAA;QACf,IAAI,KAAK,GAAG,CAAC,CAAA;QACb,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAA;QAErB,OAAO,OAAO,GAAG,KAAK,IAAI,KAAK,EAAE,GAAG,KAAK,GAAG,EAAE,EAAE,CAAC;YAChD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAE,CAAA;YAC5C,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,IAAI,CAAA;YACjC,MAAM,QAAQ,GAAG,MAAM,EAAE,GAAG,YAAY,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAA;YACtF,MAAM,YAAY,GAAG,eAAe,CAAC,MAAM,EAAE,IAAI,CAAC,YAAY,CAAC,CAAA;YAE/D,qDAAqD;YACrD,MAAM,UAAU,GAAG,MAAM,EAAE,GAAG,gBAAgB,CAAA;YAE9C,IAAI,QAA4E,CAAA;YAEhF,IAAI,UAAU,EAAE,CAAC;gBAChB,MAAM,WAAW,GAAG,MAAM,EAAE,CAAA;gBAC5B,IAAI,GAAW,CAAA;gBAEf,IAAI,WAAW,GAAG,IAAI,EAAE,CAAC;oBACxB,qDAAqD;oBACrD,GAAG,GAAG,GAAG,QAAQ,IAAI,QAAQ,KAAK,YAAY,IAAI,MAAM,EAAE,CAAA;gBAC3D,CAAC;qBAAM,IAAI,WAAW,GAAG,GAAG,EAAE,CAAC;oBAC9B,sFAAsF;oBACtF,GAAG,GAAG,GAAG,QAAQ,KAAK,QAAQ,KAAK,YAAY,IAAI,MAAM,EAAE,CAAA;gBAC5D,CAAC;qBAAM,IAAI,WAAW,GAAG,IAAI,EAAE,CAAC;oBAC/B,iFAAiF;oBACjF,GAAG,GAAG,GAAG,QAAQ,IAAI,YAAY,IAAI,MAAM,IAAI,QAAQ,EAAE,CAAA;gBAC1D,CAAC;qBAAM,CAAC;oBACP,gEAAgE;oBAChE,GAAG,GAAG,GAAG,QAAQ,KAAK,YAAY,IAAI,MAAM,KAAK,QAAQ,EAAE,CAAA;gBAC5D,CAAC;gBACD,QAAQ,GAAG,EAAE,GAAG,EAAE,UAAU,EAAE,EAAE,YAAY,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,EAAE,EAAE,CAAA;YAC7E,CAAC;iBAAM,CAAC;gBACP,QAAQ,GAAG,eAAe,CAAC,YAAY,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,CAAC,CAAA;YACrE,CAAC;YAED,MAAM,EAAE,GAAG,EAAE,UAAU,EAAE,GAAG,QAAQ,CAAA;YAEpC,sFAAsF;YACtF,MAAM,eAAe,GAAG,MAAM,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,MAAM,CAAC,OAAO,CAAa,CAAA;YAE7E,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;gBACpD,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YAED,4EAA4E;YAC5E,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;gBACjB,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,EAAE,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,GAAG,IAAI,CAAC,CAAA;gBAChE,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YAED,MAAM,QAAQ,GAAG,cAAc,CAAC,MAAM,EAAE;gBACvC,MAAM,EAAE,UAAU,CAAC,MAAM;gBACzB,YAAY,EAAE,UAAU,CAAC,YAAY;gBACrC,QAAQ,EAAE,UAAU,CAAC,QAAQ;gBAC7B,QAAQ,EAAE,UAAU,CAAC,QAAQ;aAC7B,CAAC,CAAA;YACF,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,8FAA8F;aACvG,CAAA;YACD,MAAM,OAAO,GAAG,QAAQ,CAAC,SAAS,CAAC,CAAA;YAEnC,IAAI,OAAO,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,GAAG,EAAE,CAAC;gBAChD,OAAO,EAAE,CAAA;gBACT,SAAQ;YACT,CAAC;YACD,KAAK,CACJ,IAAI,CAAC,SAAS,CAAC;gBACd,GAAG,OAAO,CAAC,GAAG;gBACd,YAAY,EAAE,UAAU;gBACxB,WAAW,EAAE,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,WAAW;gBAClD,aAAa,EAAE,IAAI;aACnB,CAAC,GAAG,IAAI,CACT,CAAA;YACD,OAAO,EAAE,CAAA;QACV,CAAC;QAED,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,CAAA;IAC5B,CAAC;CACD,CAAA"}

package/out/src/shard-recipes/german.d.ts

@@ -0,0 +1,19 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `german` shard recipe — German coverage rows from REAL OpenAddresses tuples (Berlin + Saxony,
+ *   cached zips). Each sampled tuple is rendered via {@link synthesizeGermanRow} in BOTH orders —
+ *   `--intl-fraction` (default 0.4) in international order (house-first / postcode-after-city), the
+ *   rest in idiomatic German order — then aligned to BIO. Generate-mode: it builds a tuple pool
+ *   from the cached zips, then draws `--count` rows from it with the passed `random` (so the emit
+ *   stream matches the legacy reservoir-sample loop). Ported from scripts/build-german-shard.mjs.
+ *
+ *   ORDER ROBUSTNESS (2026-06-06): mixing the two renderings stops a native-only shard from teaching
+ *   German order so well it reads the US/feed-order eval as a "collapse". See
+ *   docs/articles/evals/2026-06-06-anchor-pilot.md (the order-artifact correction).
+ */
+import { type ShardRecipe } from "./scaffold.js";
+export declare const germanRecipe: ShardRecipe;
+//# sourceMappingURL=german.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"german.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/german.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAOH,OAAO,EAAkB,KAAK,WAAW,EAAE,MAAM,eAAe,CAAA;AA4FhE,eAAO,MAAM,YAAY,EAAE,WAkF1B,CAAA"}