@struktur/benchmarks 2.6.0

package/dist/datasets.js

@@ -0,0 +1,392 @@
+// src/datasets/source.ts
+import { mkdir, readFile, writeFile } from "fs/promises";
+import path from "path";
+
+// src/config.ts
+var defaultDatasetDir = () => process.env.STRUKTUR_BENCHMARKS_DIR ?? `${process.env.HOME ?? "."}/.struktur/benchmarks`;
+
+// src/datasets/source.ts
+var readCachedCases = async (file) => {
+  try {
+    return JSON.parse(await readFile(file, "utf8"));
+  } catch {
+    return void 0;
+  }
+};
+async function loadDataset(dataset, options = {}) {
+  const base = path.join(options.cacheDir ?? defaultDatasetDir(), dataset.id, dataset.version);
+  const casesFile = path.join(base, "cases.json");
+  const rawDir = path.join(base, "raw");
+  const limit = options.limit;
+  const offset = options.offset ?? 0;
+  let cases = options.force ? void 0 : await readCachedCases(casesFile);
+  if (!cases) {
+    if (dataset.source.kind === "manual") {
+      const hint = dataset.source.hint ?? `download the dataset into ${rawDir}`;
+      throw new Error(`Dataset "${dataset.id}" requires manual download. ${hint}`);
+    }
+    await mkdir(rawDir, { recursive: true });
+    await dataset.download(rawDir, options.fetchImpl);
+    cases = await dataset.convert(rawDir);
+    await mkdir(base, { recursive: true });
+    await writeFile(casesFile, JSON.stringify(cases, null, 2), "utf8");
+    await writeFile(
+      path.join(base, "manifest.json"),
+      JSON.stringify(
+        {
+          id: dataset.id,
+          version: dataset.version,
+          license: dataset.license,
+          attribution: dataset.attribution,
+          generatedAt: (/* @__PURE__ */ new Date()).toISOString(),
+          count: cases.length
+        },
+        null,
+        2
+      ),
+      "utf8"
+    );
+  }
+  return cases.slice(offset, limit !== void 0 ? offset + limit : void 0);
+}
+var DEFAULT_HUB_BASE = "https://datasets-server.huggingface.co";
+async function downloadHubRows(source, fetchImpl = fetch) {
+  const base = source.baseUrl ?? DEFAULT_HUB_BASE;
+  const repo = encodeURIComponent(source.repo);
+  const config = source.config ?? "default";
+  const split = source.split ?? "train";
+  const pageSize = 100;
+  const rows = [];
+  let offset = 0;
+  for (; ; ) {
+    const url = `${base}/rows?dataset=${repo}&config=${config}&split=${split}&offset=${offset}&length=${pageSize}`;
+    const res = await fetchImpl(url);
+    if (!res.ok) {
+      throw new Error(`Hugging Face datasets-server error ${res.status} for ${url}`);
+    }
+    const body = await res.json();
+    const page = body.rows ?? [];
+    if (page.length === 0) break;
+    rows.push(...page.map((r) => r.row));
+    if (page.length < pageSize) break;
+    offset += page.length;
+  }
+  return rows;
+}
+
+// src/datasets/synthetic.ts
+function textArtifact(text, id) {
+  return {
+    id: id ?? `artifact-${crypto.randomUUID()}`,
+    type: "text",
+    raw: async () => Buffer.from(text, "utf8"),
+    contents: [{ text }]
+  };
+}
+function syntheticCases(datasetId, schema, records, opts = {}) {
+  return records.map((r, i) => ({
+    id: r.id ?? `${datasetId}-${i}`,
+    schema,
+    gold: r.gold,
+    artifacts: [textArtifact(r.text)],
+    tracks: opts.tracks ?? ["text"],
+    metrics: opts.metrics,
+    source: { dataset: datasetId, license: "synthetic" }
+  }));
+}
+var mulberry32 = (seed) => {
+  let a = seed >>> 0;
+  return () => {
+    a |= 0;
+    a = a + 1831565813 | 0;
+    let t = Math.imul(a ^ a >>> 15, 1 | a);
+    t = t + Math.imul(t ^ t >>> 7, 61 | t) ^ t;
+    return ((t ^ t >>> 14) >>> 0) / 4294967296;
+  };
+};
+var DEFAULT_STRING_VALUES = ["Acme Corp", "Example Ltd", "Northwind", "Globex", "Initech"];
+var DEFAULT_NUMBER_VALUES = ["42", "100.5", "7", "1500", "0.75"];
+function generateKeyValueCases(datasetId, fields, count, seed = 1) {
+  const rnd = mulberry32(seed);
+  const pick = (arr) => arr[Math.floor(rnd() * arr.length)];
+  const schema = {
+    type: "object",
+    properties: Object.fromEntries(fields.map((f) => [f.name, { type: f.type ?? "string" }])),
+    required: fields.map((f) => f.name),
+    additionalProperties: false
+  };
+  const cases = [];
+  for (let i = 0; i < count; i++) {
+    const lines = [];
+    const gold = {};
+    for (const f of fields) {
+      const values = f.values ?? (f.type === "number" ? DEFAULT_NUMBER_VALUES : DEFAULT_STRING_VALUES);
+      const raw = pick(values);
+      const value = f.type === "number" ? Number.parseFloat(raw) : raw;
+      lines.push(`${f.name}: ${raw}`);
+      gold[f.name] = value;
+    }
+    const text = lines.join("\n");
+    cases.push({
+      id: `${datasetId}-${i}`,
+      schema,
+      gold,
+      artifacts: [textArtifact(text, `artifact-${datasetId}-${i}`)],
+      tracks: ["text"],
+      source: { dataset: datasetId, license: "synthetic" }
+    });
+  }
+  return cases;
+}
+
+// src/datasets/longdoc.ts
+var mulberry322 = (seed) => {
+  let a = seed >>> 0;
+  return () => {
+    a |= 0;
+    a = a + 1831565813 | 0;
+    let t = Math.imul(a ^ a >>> 15, 1 | a);
+    t = t + Math.imul(t ^ t >>> 7, 61 | t) ^ t;
+    return ((t ^ t >>> 14) >>> 0) / 4294967296;
+  };
+};
+var SCHEMA = {
+  type: "object",
+  properties: {
+    total_records: { type: "number" },
+    total_amount: { type: "number" },
+    records: {
+      type: "array",
+      items: {
+        type: "object",
+        properties: {
+          id: { type: "string" },
+          customer: { type: "string" },
+          amount: { type: "number" },
+          notes: { type: "string" }
+        },
+        required: ["id", "customer", "amount", "notes"],
+        additionalProperties: false
+      }
+    }
+  },
+  required: ["total_records", "total_amount", "records"],
+  additionalProperties: false
+};
+var CUSTOMERS = [
+  "Acme Corp",
+  "Globex Inc",
+  "Initech LLC",
+  "Umbrella Corp",
+  "Soylent Corp",
+  "Hooli",
+  "Pied Piper",
+  "Stark Industries",
+  "Wayne Enterprises",
+  "Aperture Science"
+];
+var NOTES = [
+  "priority order",
+  "standard delivery",
+  "expedited shipping",
+  "payment pending",
+  "payment received",
+  "on hold \u2014 awaiting review",
+  "partially shipped",
+  "fully delivered",
+  "return requested",
+  "refund issued",
+  "backordered",
+  "customer pickup"
+];
+var buildRecords = (count, seed) => {
+  const rnd = mulberry322(seed);
+  const records = [];
+  for (let i = 0; i < count; i++) {
+    records.push({
+      id: `REC-${String(i + 1).padStart(5, "0")}`,
+      customer: CUSTOMERS[Math.floor(rnd() * CUSTOMERS.length)],
+      amount: Math.round(rnd() * 1e5) / 100,
+      notes: NOTES[Math.floor(rnd() * NOTES.length)]
+    });
+  }
+  return records;
+};
+var buildText = (records) => records.map(
+  (r) => `Record ${r.id}
+Customer: ${r.customer}
+Amount: ${r.amount.toFixed(2)}
+Notes: ${r.notes}
+`
+).join("\n---\n");
+var longDocArtifact = (id, text) => textArtifact(text, id);
+function longDocCase(spec) {
+  const records = buildRecords(spec.recordCount, spec.seed ?? 42);
+  const text = buildText(records);
+  const total = records.reduce((a, r) => a + r.amount, 0);
+  return {
+    id: `longdoc-${spec.id}`,
+    schema: SCHEMA,
+    gold: {
+      total_records: records.length,
+      total_amount: Math.round(total * 100) / 100,
+      records
+    },
+    artifacts: [longDocArtifact(`longdoc-${spec.id}`, text)],
+    tracks: ["text"],
+    metrics: {
+      default: "normalized",
+      fields: { amount: "tolerance", total_amount: "tolerance" },
+      arrays: [{ path: "records", key: "id" }]
+    },
+    source: { dataset: "longdoc-synthetic" }
+  };
+}
+var longDocCases = () => [
+  longDocCase({ id: "5k", recordCount: 220, seed: 1 }),
+  longDocCase({ id: "10k", recordCount: 440, seed: 2 }),
+  longDocCase({ id: "20k", recordCount: 880, seed: 3 })
+];
+
+// src/datasets/sroie.ts
+var REPO = "darentang/sroie";
+var CONFIG = "sroie";
+var LABELS = [
+  "O",
+  "B-COMPANY",
+  "I-COMPANY",
+  "B-DATE",
+  "I-DATE",
+  "B-ADDRESS",
+  "I-ADDRESS",
+  "B-TOTAL",
+  "I-TOTAL"
+];
+var SCHEMA2 = {
+  type: "object",
+  properties: {
+    company: { type: ["string", "null"] },
+    date: { type: ["string", "null"] },
+    address: { type: ["string", "null"] },
+    total: { type: ["string", "null"] }
+  },
+  additionalProperties: false
+};
+function reconstructEntities(words, tags) {
+  const out = {
+    company: null,
+    date: null,
+    address: null,
+    total: null
+  };
+  let current = null;
+  let buffer = [];
+  const flush = () => {
+    if (current && buffer.length > 0) {
+      out[current.toLowerCase()] = buffer.join(" ");
+    }
+    current = null;
+    buffer = [];
+  };
+  for (let i = 0; i < words.length; i++) {
+    const tag = LABELS[tags[i] ?? 0] ?? "O";
+    if (tag.startsWith("B-")) {
+      flush();
+      current = tag.slice(2);
+      buffer = [words[i] ?? ""];
+    } else if (tag.startsWith("I-") && current === tag.slice(2)) {
+      buffer.push(words[i] ?? "");
+    } else {
+      flush();
+    }
+  }
+  flush();
+  return out;
+}
+var imageUrl = (imagePath) => `https://huggingface.co/datasets/${REPO}/resolve/main/${imagePath}`;
+var imageArtifact = (id, url) => ({
+  id: `artifact-${id}`,
+  type: "image",
+  raw: async () => Buffer.from(""),
+  contents: [{ media: [{ type: "image", url }] }]
+});
+function sroieToCases(rows) {
+  const cases = [];
+  for (const row of rows) {
+    const words = row.words ?? [];
+    const tags = row.ner_tags ?? [];
+    const imagePath = row.image_path;
+    if (!imagePath) continue;
+    const gold = reconstructEntities(words, tags);
+    const id = row.id ?? imagePath;
+    const url = imageUrl(imagePath);
+    const text = words.join(" ");
+    cases.push({
+      id: `sroie-${id}`,
+      schema: SCHEMA2,
+      gold,
+      artifacts: [imageArtifact(id, url)],
+      tracks: ["text", "text+embedded"],
+      artifactsByTrack: {
+        text: [
+          {
+            id: `artifact-${id}-text`,
+            type: "text",
+            raw: async () => Buffer.from(text, "utf8"),
+            contents: [{ text }]
+          }
+        ],
+        "text+embedded": [imageArtifact(id, url)]
+      },
+      metrics: { default: "normalized" },
+      source: {
+        dataset: "sroie",
+        license: "ICDAR 2019 SROIE (research use)",
+        url: "https://github.com/zzzDavid/ICDAR-2019-SROIE",
+        attribution: "ICDAR 2019 SROIE (Huang et al., 2019)"
+      }
+    });
+  }
+  return cases;
+}
+var download = async (rawDir, fetchImpl) => {
+  const { writeFile: writeFile2 } = await import("fs/promises");
+  const path2 = await import("path");
+  const all = [];
+  for (const split of ["train", "test"]) {
+    const rows = await downloadHubRows(
+      { kind: "hub", repo: REPO, config: CONFIG, split },
+      fetchImpl
+    );
+    all.push(...rows);
+  }
+  await writeFile2(path2.join(rawDir, "rows.json"), JSON.stringify(all), "utf8");
+};
+var convert = async (rawDir) => {
+  const { readFile: readFile2 } = await import("fs/promises");
+  const path2 = await import("path");
+  const rows = JSON.parse(await readFile2(path2.join(rawDir, "rows.json"), "utf8"));
+  return sroieToCases(rows);
+};
+var sroie = {
+  id: "sroie",
+  version: "1",
+  source: { kind: "hub", repo: REPO, config: CONFIG },
+  license: "ICDAR 2019 SROIE (research use)",
+  attribution: "ICDAR 2019 SROIE (Huang et al., 2019)",
+  download,
+  convert
+};
+export {
+  downloadHubRows,
+  generateKeyValueCases,
+  loadDataset,
+  longDocCase,
+  longDocCases,
+  reconstructEntities,
+  sroie,
+  sroieToCases,
+  syntheticCases,
+  textArtifact
+};
+//# sourceMappingURL=datasets.js.map

package/dist/datasets.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/datasets/source.ts","../src/config.ts","../src/datasets/synthetic.ts","../src/datasets/longdoc.ts","../src/datasets/sroie.ts"],"sourcesContent":["import { mkdir, readFile, writeFile } from \"node:fs/promises\";\nimport path from \"node:path\";\nimport type { BenchmarkCase } from \"../types\";\nimport { defaultDatasetDir } from \"../config\";\n\n/** Injectable fetch for tests. */\nexport type FetchLike = (url: string, init?: RequestInit) => Promise<Response>;\n\n/** Where a dataset's raw files come from. */\nexport type DatasetSource =\n  | { kind: \"hub\"; repo: string; config?: string; split?: string; baseUrl?: string }\n  | { kind: \"url\"; url: string }\n  | { kind: \"manual\"; hint?: string };\n\nexport type Dataset = {\n  id: string;\n  /** Bump to invalidate the local cache. */\n  version: string;\n  source: DatasetSource;\n  license?: string;\n  attribution?: string;\n  /** Download raw files into `rawDir`. */\n  download: (rawDir: string, fetchImpl?: FetchLike) => Promise<void>;\n  /** Convert raw files in `rawDir` into benchmark cases. */\n  convert: (rawDir: string) => Promise<BenchmarkCase[]>;\n};\n\nexport type LoadDatasetOptions = {\n  cacheDir?: string;\n  /** Return only this many cases (defaults to all). */\n  limit?: number;\n  offset?: number;\n  fetchImpl?: FetchLike;\n  /** Re-download + re-convert even if cached. */\n  force?: boolean;\n};\n\nconst readCachedCases = async (file: string): Promise<BenchmarkCase[] | undefined> => {\n  try {\n    return JSON.parse(await readFile(file, \"utf8\")) as BenchmarkCase[];\n  } catch {\n    return undefined;\n  }\n};\n\n/**\n * Load a dataset's cases, downloading and converting on first use and caching\n * the converted result under `~/.struktur/benchmarks/<id>/<version>/`.\n */\nexport async function loadDataset(\n  dataset: Dataset,\n  options: LoadDatasetOptions = {},\n): Promise<BenchmarkCase[]> {\n  const base = path.join(options.cacheDir ?? defaultDatasetDir(), dataset.id, dataset.version);\n  const casesFile = path.join(base, \"cases.json\");\n  const rawDir = path.join(base, \"raw\");\n  const limit = options.limit;\n  const offset = options.offset ?? 0;\n\n  let cases = options.force ? undefined : await readCachedCases(casesFile);\n\n  if (!cases) {\n    if (dataset.source.kind === \"manual\") {\n      const hint = dataset.source.hint ?? `download the dataset into ${rawDir}`;\n      throw new Error(`Dataset \"${dataset.id}\" requires manual download. ${hint}`);\n    }\n\n    await mkdir(rawDir, { recursive: true });\n    await dataset.download(rawDir, options.fetchImpl);\n    cases = await dataset.convert(rawDir);\n\n    await mkdir(base, { recursive: true });\n    await writeFile(casesFile, JSON.stringify(cases, null, 2), \"utf8\");\n    await writeFile(\n      path.join(base, \"manifest.json\"),\n      JSON.stringify(\n        {\n          id: dataset.id,\n          version: dataset.version,\n          license: dataset.license,\n          attribution: dataset.attribution,\n          generatedAt: new Date().toISOString(),\n          count: cases.length,\n        },\n        null,\n        2,\n      ),\n      \"utf8\",\n    );\n  }\n\n  return cases.slice(offset, limit !== undefined ? offset + limit : undefined);\n}\n\n// ---------------------------------------------------------------------------\n// Hugging Face datasets-server helpers\n// ---------------------------------------------------------------------------\n\nconst DEFAULT_HUB_BASE = \"https://datasets-server.huggingface.co\";\n\n/** Fetch every row of a hub dataset via the datasets-server `/rows` endpoint. */\nexport async function downloadHubRows(\n  source: Extract<DatasetSource, { kind: \"hub\" }>,\n  fetchImpl: FetchLike = fetch,\n): Promise<Record<string, unknown>[]> {\n  const base = source.baseUrl ?? DEFAULT_HUB_BASE;\n  const repo = encodeURIComponent(source.repo);\n  const config = source.config ?? \"default\";\n  const split = source.split ?? \"train\";\n  const pageSize = 100;\n\n  const rows: Record<string, unknown>[] = [];\n  let offset = 0;\n  for (;;) {\n    const url = `${base}/rows?dataset=${repo}&config=${config}&split=${split}&offset=${offset}&length=${pageSize}`;\n    const res = await fetchImpl(url);\n    if (!res.ok) {\n      throw new Error(`Hugging Face datasets-server error ${res.status} for ${url}`);\n    }\n    const body = (await res.json()) as { rows?: { row: Record<string, unknown> }[] };\n    const page = body.rows ?? [];\n    if (page.length === 0) break;\n    rows.push(...page.map((r) => r.row));\n    if (page.length < pageSize) break;\n    offset += page.length;\n  }\n  return rows;\n}\n","/**\n * The benchmark measures the harness, not the model — the model is a fixed baseline.\n *\n * deepseek-v4-flash-vision-exp: ~$0.60/M tokens, vision-capable, supports\n * structured output (json_schema) via OpenRouter. Cheap enough for frequent\n * benchmark runs while testing the full matrix of text + image variants.\n */\nexport const BENCHMARK_MODEL = \"openrouter/deepseek/deepseek-v4-flash-vision-exp\";\n\n/** Default cache root for downloaded datasets. */\nexport const defaultDatasetDir = (): string =>\n  process.env.STRUKTUR_BENCHMARKS_DIR ?? `${process.env.HOME ?? \".\"}/.struktur/benchmarks`;\n\n/** Default cache root for extraction result cells. */\nexport const defaultResultsDir = (): string => `${defaultDatasetDir()}/results`;\n","import type { Artifact } from \"@struktur/sdk\";\nimport type { BenchmarkCase, FieldMetricSpec, Track } from \"../types\";\n\n/** Build a plain text artifact (no media). */\nexport function textArtifact(text: string, id?: string): Artifact {\n  return {\n    id: id ?? `artifact-${crypto.randomUUID()}`,\n    type: \"text\",\n    raw: async () => Buffer.from(text, \"utf8\"),\n    contents: [{ text }],\n  };\n}\n\nexport type SyntheticRecord = {\n  text: string;\n  gold: unknown;\n  id?: string;\n};\n\n/** Wrap user-supplied `(text, gold)` records into benchmark cases. */\nexport function syntheticCases(\n  datasetId: string,\n  schema: unknown,\n  records: SyntheticRecord[],\n  opts: { metrics?: FieldMetricSpec; tracks?: Track[] } = {},\n): BenchmarkCase[] {\n  return records.map((r, i) => ({\n    id: r.id ?? `${datasetId}-${i}`,\n    schema,\n    gold: r.gold,\n    artifacts: [textArtifact(r.text)],\n    tracks: opts.tracks ?? [\"text\"],\n    metrics: opts.metrics,\n    source: { dataset: datasetId, license: \"synthetic\" },\n  }));\n}\n\n// ---------------------------------------------------------------------------\n// Deterministic key-value generator\n// ---------------------------------------------------------------------------\n\nexport type SimpleField = {\n  name: string;\n  type?: \"string\" | \"number\";\n  /** Candidate values; picked deterministically. */\n  values?: string[];\n};\n\n/** Mulberry32 seeded PRNG for reproducible generation. */\nconst mulberry32 = (seed: number) => {\n  let a = seed >>> 0;\n  return () => {\n    a |= 0;\n    a = (a + 0x6d2b79f5) | 0;\n    let t = Math.imul(a ^ (a >>> 15), 1 | a);\n    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t;\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;\n  };\n};\n\nconst DEFAULT_STRING_VALUES = [\"Acme Corp\", \"Example Ltd\", \"Northwind\", \"Globex\", \"Initech\"];\nconst DEFAULT_NUMBER_VALUES = [\"42\", \"100.5\", \"7\", \"1500\", \"0.75\"];\n\n/**\n * Generate N deterministic text→JSON cases for schema/prompt comparison.\n * Each case renders `<Field Name>: <value>` lines and a gold object, so it\n * exercises flat key-value extraction with no network or files.\n */\nexport function generateKeyValueCases(\n  datasetId: string,\n  fields: SimpleField[],\n  count: number,\n  seed = 1,\n): BenchmarkCase[] {\n  const rnd = mulberry32(seed);\n  const pick = <T>(arr: readonly T[]): T => arr[Math.floor(rnd() * arr.length)]!;\n\n  const schema = {\n    type: \"object\",\n    properties: Object.fromEntries(fields.map((f) => [f.name, { type: f.type ?? \"string\" }])),\n    required: fields.map((f) => f.name),\n    additionalProperties: false,\n  };\n\n  const cases: BenchmarkCase[] = [];\n  for (let i = 0; i < count; i++) {\n    const lines: string[] = [];\n    const gold: Record<string, unknown> = {};\n    for (const f of fields) {\n      const values =\n        f.values ?? (f.type === \"number\" ? DEFAULT_NUMBER_VALUES : DEFAULT_STRING_VALUES);\n      const raw = pick(values);\n      const value = f.type === \"number\" ? Number.parseFloat(raw) : raw;\n      lines.push(`${f.name}: ${raw}`);\n      gold[f.name] = value;\n    }\n    const text = lines.join(\"\\n\");\n    cases.push({\n      id: `${datasetId}-${i}`,\n      schema,\n      gold,\n      artifacts: [textArtifact(text, `artifact-${datasetId}-${i}`)],\n      tracks: [\"text\"],\n      source: { dataset: datasetId, license: \"synthetic\" },\n    });\n  }\n  return cases;\n}\n","import type { Artifact } from \"@struktur/sdk\";\nimport type { BenchmarkCase } from \"../types\";\nimport { textArtifact } from \"./synthetic\";\n\n/**\n * Synthetic long documents: `recordCount` self-contained records with a\n * repeating structure. Document length scales with record count, so total\n * length is controllable and the document actually splits into chunks.\n *\n * The schema requires collecting/counting across the whole document\n * (arrays + aggregates: total_records, total_amount), which exercises\n * chunking + merge quality — not just per-record extraction. Notes are short\n * meaningful phrases (no filler) so output tokens stay proportional to\n * content, keeping cost/latency measurements honest.\n */\n\nexport type LongDocSpec = {\n  id: string;\n  recordCount: number;\n  /** Random seed for deterministic generation. */\n  seed?: number;\n};\n\nconst mulberry32 = (seed: number) => {\n  let a = seed >>> 0;\n  return () => {\n    a |= 0;\n    a = (a + 0x6d2b79f5) | 0;\n    let t = Math.imul(a ^ (a >>> 15), 1 | a);\n    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t;\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;\n  };\n};\n\nconst SCHEMA = {\n  type: \"object\",\n  properties: {\n    total_records: { type: \"number\" },\n    total_amount: { type: \"number\" },\n    records: {\n      type: \"array\",\n      items: {\n        type: \"object\",\n        properties: {\n          id: { type: \"string\" },\n          customer: { type: \"string\" },\n          amount: { type: \"number\" },\n          notes: { type: \"string\" },\n        },\n        required: [\"id\", \"customer\", \"amount\", \"notes\"],\n        additionalProperties: false,\n      },\n    },\n  },\n  required: [\"total_records\", \"total_amount\", \"records\"],\n  additionalProperties: false,\n} as const;\n\nconst CUSTOMERS = [\n  \"Acme Corp\",\n  \"Globex Inc\",\n  \"Initech LLC\",\n  \"Umbrella Corp\",\n  \"Soylent Corp\",\n  \"Hooli\",\n  \"Pied Piper\",\n  \"Stark Industries\",\n  \"Wayne Enterprises\",\n  \"Aperture Science\",\n] as const;\n\nconst NOTES = [\n  \"priority order\",\n  \"standard delivery\",\n  \"expedited shipping\",\n  \"payment pending\",\n  \"payment received\",\n  \"on hold — awaiting review\",\n  \"partially shipped\",\n  \"fully delivered\",\n  \"return requested\",\n  \"refund issued\",\n  \"backordered\",\n  \"customer pickup\",\n] as const;\n\nexport type LongDocRecord = { id: string; customer: string; amount: number; notes: string };\n\nconst buildRecords = (count: number, seed: number): LongDocRecord[] => {\n  const rnd = mulberry32(seed);\n  const records: LongDocRecord[] = [];\n  for (let i = 0; i < count; i++) {\n    records.push({\n      id: `REC-${String(i + 1).padStart(5, \"0\")}`,\n      customer: CUSTOMERS[Math.floor(rnd() * CUSTOMERS.length)]!,\n      amount: Math.round(rnd() * 100000) / 100,\n      notes: NOTES[Math.floor(rnd() * NOTES.length)]!,\n    });\n  }\n  return records;\n};\n\nconst buildText = (records: LongDocRecord[]): string =>\n  records\n    .map(\n      (r) =>\n        `Record ${r.id}\\nCustomer: ${r.customer}\\nAmount: ${r.amount.toFixed(2)}\\nNotes: ${r.notes}\\n`,\n    )\n    .join(\"\\n---\\n\");\n\nexport const longDocArtifact = (id: string, text: string): Artifact => textArtifact(text, id);\n\n/** Create a single long-document benchmark case. */\nexport function longDocCase(spec: LongDocSpec): BenchmarkCase {\n  const records = buildRecords(spec.recordCount, spec.seed ?? 42);\n  const text = buildText(records);\n  const total = records.reduce((a, r) => a + r.amount, 0);\n\n  return {\n    id: `longdoc-${spec.id}`,\n    schema: SCHEMA,\n    gold: {\n      total_records: records.length,\n      total_amount: Math.round(total * 100) / 100,\n      records,\n    },\n    artifacts: [longDocArtifact(`longdoc-${spec.id}`, text)],\n    tracks: [\"text\"],\n    metrics: {\n      default: \"normalized\",\n      fields: { amount: \"tolerance\", total_amount: \"tolerance\" },\n      arrays: [{ path: \"records\", key: \"id\" }],\n    },\n    source: { dataset: \"longdoc-synthetic\" },\n  };\n}\n\n/**\n * Long-document cases at ~5k / 10k / 20k tokens (4 chars/token).\n * Record count = targetTokens * 4 / ~90 chars per record.\n */\nexport const longDocCases = (): BenchmarkCase[] => [\n  longDocCase({ id: \"5k\", recordCount: 220, seed: 1 }),\n  longDocCase({ id: \"10k\", recordCount: 440, seed: 2 }),\n  longDocCase({ id: \"20k\", recordCount: 880, seed: 3 }),\n];\n","import type { Artifact } from \"@struktur/sdk\";\nimport type { BenchmarkCase } from \"../types\";\nimport { downloadHubRows, type Dataset, type FetchLike } from \"./source\";\n\n/**\n * SROIE (ICDAR 2019, scanned receipts) — key information extraction:\n * company, date, address, total.\n *\n * Backed by the Hugging Face mirror `darentang/sroie`, which exposes per-word\n * BIO tags. Rows are downloaded via the datasets-server API; receipt images\n * are referenced by URL (fetched by the model at inference time), so no bulk\n * image download is needed.\n */\n\nconst REPO = \"darentang/sroie\";\nconst CONFIG = \"sroie\";\n\nconst LABELS = [\n  \"O\",\n  \"B-COMPANY\",\n  \"I-COMPANY\",\n  \"B-DATE\",\n  \"I-DATE\",\n  \"B-ADDRESS\",\n  \"I-ADDRESS\",\n  \"B-TOTAL\",\n  \"I-TOTAL\",\n] as const;\n\nconst SCHEMA = {\n  type: \"object\",\n  properties: {\n    company: { type: [\"string\", \"null\"] },\n    date: { type: [\"string\", \"null\"] },\n    address: { type: [\"string\", \"null\"] },\n    total: { type: [\"string\", \"null\"] },\n  },\n  additionalProperties: false,\n} as const;\n\nexport type SroieRow = {\n  id?: string;\n  words?: string[];\n  ner_tags?: number[];\n  image_path?: string;\n};\n\n/** Reconstruct `company/date/address/total` from BIO-tagged words. */\nexport function reconstructEntities(\n  words: string[],\n  tags: number[],\n): Record<string, string | null> {\n  const out: Record<string, string | null> = {\n    company: null,\n    date: null,\n    address: null,\n    total: null,\n  };\n  let current: string | null = null;\n  let buffer: string[] = [];\n  const flush = () => {\n    if (current && buffer.length > 0) {\n      out[current.toLowerCase()] = buffer.join(\" \");\n    }\n    current = null;\n    buffer = [];\n  };\n\n  for (let i = 0; i < words.length; i++) {\n    const tag = LABELS[tags[i] ?? 0] ?? \"O\";\n    if (tag.startsWith(\"B-\")) {\n      flush();\n      current = tag.slice(2);\n      buffer = [words[i] ?? \"\"];\n    } else if (tag.startsWith(\"I-\") && current === tag.slice(2)) {\n      buffer.push(words[i] ?? \"\");\n    } else {\n      flush();\n    }\n  }\n  flush();\n  return out;\n}\n\nconst imageUrl = (imagePath: string): string =>\n  `https://huggingface.co/datasets/${REPO}/resolve/main/${imagePath}`;\n\nconst imageArtifact = (id: string, url: string): Artifact => ({\n  id: `artifact-${id}`,\n  type: \"image\",\n  raw: async () => Buffer.from(\"\"),\n  contents: [{ media: [{ type: \"image\", url }] }],\n});\n\n/** Convert raw HF rows into benchmark cases. */\nexport function sroieToCases(rows: SroieRow[]): BenchmarkCase[] {\n  const cases: BenchmarkCase[] = [];\n  for (const row of rows) {\n    const words = row.words ?? [];\n    const tags = row.ner_tags ?? [];\n    const imagePath = row.image_path;\n    if (!imagePath) continue;\n\n    const gold = reconstructEntities(words, tags);\n    const id = row.id ?? imagePath;\n    const url = imageUrl(imagePath);\n\n    // OCR text variant: words joined — a fair text-only baseline.\n    const text = words.join(\" \");\n\n    cases.push({\n      id: `sroie-${id}`,\n      schema: SCHEMA,\n      gold,\n      artifacts: [imageArtifact(id, url)],\n      tracks: [\"text\", \"text+embedded\"],\n      artifactsByTrack: {\n        text: [\n          {\n            id: `artifact-${id}-text`,\n            type: \"text\",\n            raw: async () => Buffer.from(text, \"utf8\"),\n            contents: [{ text }],\n          },\n        ],\n        \"text+embedded\": [imageArtifact(id, url)],\n      },\n      metrics: { default: \"normalized\" },\n      source: {\n        dataset: \"sroie\",\n        license: \"ICDAR 2019 SROIE (research use)\",\n        url: \"https://github.com/zzzDavid/ICDAR-2019-SROIE\",\n        attribution: \"ICDAR 2019 SROIE (Huang et al., 2019)\",\n      },\n    });\n  }\n  return cases;\n}\n\nconst download = async (rawDir: string, fetchImpl?: FetchLike): Promise<void> => {\n  const { writeFile } = await import(\"node:fs/promises\");\n  const path = await import(\"node:path\");\n  const all: SroieRow[] = [];\n  for (const split of [\"train\", \"test\"]) {\n    const rows = await downloadHubRows(\n      { kind: \"hub\", repo: REPO, config: CONFIG, split },\n      fetchImpl,\n    );\n    all.push(...(rows as SroieRow[]));\n  }\n  await writeFile(path.join(rawDir, \"rows.json\"), JSON.stringify(all), \"utf8\");\n};\n\nconst convert = async (rawDir: string): Promise<BenchmarkCase[]> => {\n  const { readFile } = await import(\"node:fs/promises\");\n  const path = await import(\"node:path\");\n  const rows = JSON.parse(await readFile(path.join(rawDir, \"rows.json\"), \"utf8\")) as SroieRow[];\n  return sroieToCases(rows);\n};\n\nexport const sroie: Dataset = {\n  id: \"sroie\",\n  version: \"1\",\n  source: { kind: \"hub\", repo: REPO, config: CONFIG },\n  license: \"ICDAR 2019 SROIE (research use)\",\n  attribution: \"ICDAR 2019 SROIE (Huang et al., 2019)\",\n  download,\n  convert,\n};\n"],"mappings":";AAAA,SAAS,OAAO,UAAU,iBAAiB;AAC3C,OAAO,UAAU;;;ACSV,IAAM,oBAAoB,MAC/B,QAAQ,IAAI,2BAA2B,GAAG,QAAQ,IAAI,QAAQ,GAAG;;;AD0BnE,IAAM,kBAAkB,OAAO,SAAuD;AACpF,MAAI;AACF,WAAO,KAAK,MAAM,MAAM,SAAS,MAAM,MAAM,CAAC;AAAA,EAChD,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAMA,eAAsB,YACpB,SACA,UAA8B,CAAC,GACL;AAC1B,QAAM,OAAO,KAAK,KAAK,QAAQ,YAAY,kBAAkB,GAAG,QAAQ,IAAI,QAAQ,OAAO;AAC3F,QAAM,YAAY,KAAK,KAAK,MAAM,YAAY;AAC9C,QAAM,SAAS,KAAK,KAAK,MAAM,KAAK;AACpC,QAAM,QAAQ,QAAQ;AACtB,QAAM,SAAS,QAAQ,UAAU;AAEjC,MAAI,QAAQ,QAAQ,QAAQ,SAAY,MAAM,gBAAgB,SAAS;AAEvE,MAAI,CAAC,OAAO;AACV,QAAI,QAAQ,OAAO,SAAS,UAAU;AACpC,YAAM,OAAO,QAAQ,OAAO,QAAQ,6BAA6B,MAAM;AACvE,YAAM,IAAI,MAAM,YAAY,QAAQ,EAAE,+BAA+B,IAAI,EAAE;AAAA,IAC7E;AAEA,UAAM,MAAM,QAAQ,EAAE,WAAW,KAAK,CAAC;AACvC,UAAM,QAAQ,SAAS,QAAQ,QAAQ,SAAS;AAChD,YAAQ,MAAM,QAAQ,QAAQ,MAAM;AAEpC,UAAM,MAAM,MAAM,EAAE,WAAW,KAAK,CAAC;AACrC,UAAM,UAAU,WAAW,KAAK,UAAU,OAAO,MAAM,CAAC,GAAG,MAAM;AACjE,UAAM;AAAA,MACJ,KAAK,KAAK,MAAM,eAAe;AAAA,MAC/B,KAAK;AAAA,QACH;AAAA,UACE,IAAI,QAAQ;AAAA,UACZ,SAAS,QAAQ;AAAA,UACjB,SAAS,QAAQ;AAAA,UACjB,aAAa,QAAQ;AAAA,UACrB,cAAa,oBAAI,KAAK,GAAE,YAAY;AAAA,UACpC,OAAO,MAAM;AAAA,QACf;AAAA,QACA;AAAA,QACA;AAAA,MACF;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,SAAO,MAAM,MAAM,QAAQ,UAAU,SAAY,SAAS,QAAQ,MAAS;AAC7E;AAMA,IAAM,mBAAmB;AAGzB,eAAsB,gBACpB,QACA,YAAuB,OACa;AACpC,QAAM,OAAO,OAAO,WAAW;AAC/B,QAAM,OAAO,mBAAmB,OAAO,IAAI;AAC3C,QAAM,SAAS,OAAO,UAAU;AAChC,QAAM,QAAQ,OAAO,SAAS;AAC9B,QAAM,WAAW;AAEjB,QAAM,OAAkC,CAAC;AACzC,MAAI,SAAS;AACb,aAAS;AACP,UAAM,MAAM,GAAG,IAAI,iBAAiB,IAAI,WAAW,MAAM,UAAU,KAAK,WAAW,MAAM,WAAW,QAAQ;AAC5G,UAAM,MAAM,MAAM,UAAU,GAAG;AAC/B,QAAI,CAAC,IAAI,IAAI;AACX,YAAM,IAAI,MAAM,sCAAsC,IAAI,MAAM,QAAQ,GAAG,EAAE;AAAA,IAC/E;AACA,UAAM,OAAQ,MAAM,IAAI,KAAK;AAC7B,UAAM,OAAO,KAAK,QAAQ,CAAC;AAC3B,QAAI,KAAK,WAAW,EAAG;AACvB,SAAK,KAAK,GAAG,KAAK,IAAI,CAAC,MAAM,EAAE,GAAG,CAAC;AACnC,QAAI,KAAK,SAAS,SAAU;AAC5B,cAAU,KAAK;AAAA,EACjB;AACA,SAAO;AACT;;;AE3HO,SAAS,aAAa,MAAc,IAAuB;AAChE,SAAO;AAAA,IACL,IAAI,MAAM,YAAY,OAAO,WAAW,CAAC;AAAA,IACzC,MAAM;AAAA,IACN,KAAK,YAAY,OAAO,KAAK,MAAM,MAAM;AAAA,IACzC,UAAU,CAAC,EAAE,KAAK,CAAC;AAAA,EACrB;AACF;AASO,SAAS,eACd,WACA,QACA,SACA,OAAwD,CAAC,GACxC;AACjB,SAAO,QAAQ,IAAI,CAAC,GAAG,OAAO;AAAA,IAC5B,IAAI,EAAE,MAAM,GAAG,SAAS,IAAI,CAAC;AAAA,IAC7B;AAAA,IACA,MAAM,EAAE;AAAA,IACR,WAAW,CAAC,aAAa,EAAE,IAAI,CAAC;AAAA,IAChC,QAAQ,KAAK,UAAU,CAAC,MAAM;AAAA,IAC9B,SAAS,KAAK;AAAA,IACd,QAAQ,EAAE,SAAS,WAAW,SAAS,YAAY;AAAA,EACrD,EAAE;AACJ;AAcA,IAAM,aAAa,CAAC,SAAiB;AACnC,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,SAAK;AACL,QAAK,IAAI,aAAc;AACvB,QAAI,IAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACvC,QAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,KAAK,CAAC,IAAK;AAC7C,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAEA,IAAM,wBAAwB,CAAC,aAAa,eAAe,aAAa,UAAU,SAAS;AAC3F,IAAM,wBAAwB,CAAC,MAAM,SAAS,KAAK,QAAQ,MAAM;AAO1D,SAAS,sBACd,WACA,QACA,OACA,OAAO,GACU;AACjB,QAAM,MAAM,WAAW,IAAI;AAC3B,QAAM,OAAO,CAAI,QAAyB,IAAI,KAAK,MAAM,IAAI,IAAI,IAAI,MAAM,CAAC;AAE5E,QAAM,SAAS;AAAA,IACb,MAAM;AAAA,IACN,YAAY,OAAO,YAAY,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,SAAS,CAAC,CAAC,CAAC;AAAA,IACxF,UAAU,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI;AAAA,IAClC,sBAAsB;AAAA,EACxB;AAEA,QAAM,QAAyB,CAAC;AAChC,WAAS,IAAI,GAAG,IAAI,OAAO,KAAK;AAC9B,UAAM,QAAkB,CAAC;AACzB,UAAM,OAAgC,CAAC;AACvC,eAAW,KAAK,QAAQ;AACtB,YAAM,SACJ,EAAE,WAAW,EAAE,SAAS,WAAW,wBAAwB;AAC7D,YAAM,MAAM,KAAK,MAAM;AACvB,YAAM,QAAQ,EAAE,SAAS,WAAW,OAAO,WAAW,GAAG,IAAI;AAC7D,YAAM,KAAK,GAAG,EAAE,IAAI,KAAK,GAAG,EAAE;AAC9B,WAAK,EAAE,IAAI,IAAI;AAAA,IACjB;AACA,UAAM,OAAO,MAAM,KAAK,IAAI;AAC5B,UAAM,KAAK;AAAA,MACT,IAAI,GAAG,SAAS,IAAI,CAAC;AAAA,MACrB;AAAA,MACA;AAAA,MACA,WAAW,CAAC,aAAa,MAAM,YAAY,SAAS,IAAI,CAAC,EAAE,CAAC;AAAA,MAC5D,QAAQ,CAAC,MAAM;AAAA,MACf,QAAQ,EAAE,SAAS,WAAW,SAAS,YAAY;AAAA,IACrD,CAAC;AAAA,EACH;AACA,SAAO;AACT;;;ACpFA,IAAMA,cAAa,CAAC,SAAiB;AACnC,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,SAAK;AACL,QAAK,IAAI,aAAc;AACvB,QAAI,IAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACvC,QAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,KAAK,CAAC,IAAK;AAC7C,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAEA,IAAM,SAAS;AAAA,EACb,MAAM;AAAA,EACN,YAAY;AAAA,IACV,eAAe,EAAE,MAAM,SAAS;AAAA,IAChC,cAAc,EAAE,MAAM,SAAS;AAAA,IAC/B,SAAS;AAAA,MACP,MAAM;AAAA,MACN,OAAO;AAAA,QACL,MAAM;AAAA,QACN,YAAY;AAAA,UACV,IAAI,EAAE,MAAM,SAAS;AAAA,UACrB,UAAU,EAAE,MAAM,SAAS;AAAA,UAC3B,QAAQ,EAAE,MAAM,SAAS;AAAA,UACzB,OAAO,EAAE,MAAM,SAAS;AAAA,QAC1B;AAAA,QACA,UAAU,CAAC,MAAM,YAAY,UAAU,OAAO;AAAA,QAC9C,sBAAsB;AAAA,MACxB;AAAA,IACF;AAAA,EACF;AAAA,EACA,UAAU,CAAC,iBAAiB,gBAAgB,SAAS;AAAA,EACrD,sBAAsB;AACxB;AAEA,IAAM,YAAY;AAAA,EAChB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAM,QAAQ;AAAA,EACZ;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAIA,IAAM,eAAe,CAAC,OAAe,SAAkC;AACrE,QAAM,MAAMA,YAAW,IAAI;AAC3B,QAAM,UAA2B,CAAC;AAClC,WAAS,IAAI,GAAG,IAAI,OAAO,KAAK;AAC9B,YAAQ,KAAK;AAAA,MACX,IAAI,OAAO,OAAO,IAAI,CAAC,EAAE,SAAS,GAAG,GAAG,CAAC;AAAA,MACzC,UAAU,UAAU,KAAK,MAAM,IAAI,IAAI,UAAU,MAAM,CAAC;AAAA,MACxD,QAAQ,KAAK,MAAM,IAAI,IAAI,GAAM,IAAI;AAAA,MACrC,OAAO,MAAM,KAAK,MAAM,IAAI,IAAI,MAAM,MAAM,CAAC;AAAA,IAC/C,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,IAAM,YAAY,CAAC,YACjB,QACG;AAAA,EACC,CAAC,MACC,UAAU,EAAE,EAAE;AAAA,YAAe,EAAE,QAAQ;AAAA,UAAa,EAAE,OAAO,QAAQ,CAAC,CAAC;AAAA,SAAY,EAAE,KAAK;AAAA;AAC9F,EACC,KAAK,SAAS;AAEZ,IAAM,kBAAkB,CAAC,IAAY,SAA2B,aAAa,MAAM,EAAE;AAGrF,SAAS,YAAY,MAAkC;AAC5D,QAAM,UAAU,aAAa,KAAK,aAAa,KAAK,QAAQ,EAAE;AAC9D,QAAM,OAAO,UAAU,OAAO;AAC9B,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,CAAC;AAEtD,SAAO;AAAA,IACL,IAAI,WAAW,KAAK,EAAE;AAAA,IACtB,QAAQ;AAAA,IACR,MAAM;AAAA,MACJ,eAAe,QAAQ;AAAA,MACvB,cAAc,KAAK,MAAM,QAAQ,GAAG,IAAI;AAAA,MACxC;AAAA,IACF;AAAA,IACA,WAAW,CAAC,gBAAgB,WAAW,KAAK,EAAE,IAAI,IAAI,CAAC;AAAA,IACvD,QAAQ,CAAC,MAAM;AAAA,IACf,SAAS;AAAA,MACP,SAAS;AAAA,MACT,QAAQ,EAAE,QAAQ,aAAa,cAAc,YAAY;AAAA,MACzD,QAAQ,CAAC,EAAE,MAAM,WAAW,KAAK,KAAK,CAAC;AAAA,IACzC;AAAA,IACA,QAAQ,EAAE,SAAS,oBAAoB;AAAA,EACzC;AACF;AAMO,IAAM,eAAe,MAAuB;AAAA,EACjD,YAAY,EAAE,IAAI,MAAM,aAAa,KAAK,MAAM,EAAE,CAAC;AAAA,EACnD,YAAY,EAAE,IAAI,OAAO,aAAa,KAAK,MAAM,EAAE,CAAC;AAAA,EACpD,YAAY,EAAE,IAAI,OAAO,aAAa,KAAK,MAAM,EAAE,CAAC;AACtD;;;ACnIA,IAAM,OAAO;AACb,IAAM,SAAS;AAEf,IAAM,SAAS;AAAA,EACb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAMC,UAAS;AAAA,EACb,MAAM;AAAA,EACN,YAAY;AAAA,IACV,SAAS,EAAE,MAAM,CAAC,UAAU,MAAM,EAAE;AAAA,IACpC,MAAM,EAAE,MAAM,CAAC,UAAU,MAAM,EAAE;AAAA,IACjC,SAAS,EAAE,MAAM,CAAC,UAAU,MAAM,EAAE;AAAA,IACpC,OAAO,EAAE,MAAM,CAAC,UAAU,MAAM,EAAE;AAAA,EACpC;AAAA,EACA,sBAAsB;AACxB;AAUO,SAAS,oBACd,OACA,MAC+B;AAC/B,QAAM,MAAqC;AAAA,IACzC,SAAS;AAAA,IACT,MAAM;AAAA,IACN,SAAS;AAAA,IACT,OAAO;AAAA,EACT;AACA,MAAI,UAAyB;AAC7B,MAAI,SAAmB,CAAC;AACxB,QAAM,QAAQ,MAAM;AAClB,QAAI,WAAW,OAAO,SAAS,GAAG;AAChC,UAAI,QAAQ,YAAY,CAAC,IAAI,OAAO,KAAK,GAAG;AAAA,IAC9C;AACA,cAAU;AACV,aAAS,CAAC;AAAA,EACZ;AAEA,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAM,MAAM,OAAO,KAAK,CAAC,KAAK,CAAC,KAAK;AACpC,QAAI,IAAI,WAAW,IAAI,GAAG;AACxB,YAAM;AACN,gBAAU,IAAI,MAAM,CAAC;AACrB,eAAS,CAAC,MAAM,CAAC,KAAK,EAAE;AAAA,IAC1B,WAAW,IAAI,WAAW,IAAI,KAAK,YAAY,IAAI,MAAM,CAAC,GAAG;AAC3D,aAAO,KAAK,MAAM,CAAC,KAAK,EAAE;AAAA,IAC5B,OAAO;AACL,YAAM;AAAA,IACR;AAAA,EACF;AACA,QAAM;AACN,SAAO;AACT;AAEA,IAAM,WAAW,CAAC,cAChB,mCAAmC,IAAI,iBAAiB,SAAS;AAEnE,IAAM,gBAAgB,CAAC,IAAY,SAA2B;AAAA,EAC5D,IAAI,YAAY,EAAE;AAAA,EAClB,MAAM;AAAA,EACN,KAAK,YAAY,OAAO,KAAK,EAAE;AAAA,EAC/B,UAAU,CAAC,EAAE,OAAO,CAAC,EAAE,MAAM,SAAS,IAAI,CAAC,EAAE,CAAC;AAChD;AAGO,SAAS,aAAa,MAAmC;AAC9D,QAAM,QAAyB,CAAC;AAChC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,SAAS,CAAC;AAC5B,UAAM,OAAO,IAAI,YAAY,CAAC;AAC9B,UAAM,YAAY,IAAI;AACtB,QAAI,CAAC,UAAW;AAEhB,UAAM,OAAO,oBAAoB,OAAO,IAAI;AAC5C,UAAM,KAAK,IAAI,MAAM;AACrB,UAAM,MAAM,SAAS,SAAS;AAG9B,UAAM,OAAO,MAAM,KAAK,GAAG;AAE3B,UAAM,KAAK;AAAA,MACT,IAAI,SAAS,EAAE;AAAA,MACf,QAAQA;AAAA,MACR;AAAA,MACA,WAAW,CAAC,cAAc,IAAI,GAAG,CAAC;AAAA,MAClC,QAAQ,CAAC,QAAQ,eAAe;AAAA,MAChC,kBAAkB;AAAA,QAChB,MAAM;AAAA,UACJ;AAAA,YACE,IAAI,YAAY,EAAE;AAAA,YAClB,MAAM;AAAA,YACN,KAAK,YAAY,OAAO,KAAK,MAAM,MAAM;AAAA,YACzC,UAAU,CAAC,EAAE,KAAK,CAAC;AAAA,UACrB;AAAA,QACF;AAAA,QACA,iBAAiB,CAAC,cAAc,IAAI,GAAG,CAAC;AAAA,MAC1C;AAAA,MACA,SAAS,EAAE,SAAS,aAAa;AAAA,MACjC,QAAQ;AAAA,QACN,SAAS;AAAA,QACT,SAAS;AAAA,QACT,KAAK;AAAA,QACL,aAAa;AAAA,MACf;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,IAAM,WAAW,OAAO,QAAgB,cAAyC;AAC/E,QAAM,EAAE,WAAAC,WAAU,IAAI,MAAM,OAAO,aAAkB;AACrD,QAAMC,QAAO,MAAM,OAAO,MAAW;AACrC,QAAM,MAAkB,CAAC;AACzB,aAAW,SAAS,CAAC,SAAS,MAAM,GAAG;AACrC,UAAM,OAAO,MAAM;AAAA,MACjB,EAAE,MAAM,OAAO,MAAM,MAAM,QAAQ,QAAQ,MAAM;AAAA,MACjD;AAAA,IACF;AACA,QAAI,KAAK,GAAI,IAAmB;AAAA,EAClC;AACA,QAAMD,WAAUC,MAAK,KAAK,QAAQ,WAAW,GAAG,KAAK,UAAU,GAAG,GAAG,MAAM;AAC7E;AAEA,IAAM,UAAU,OAAO,WAA6C;AAClE,QAAM,EAAE,UAAAC,UAAS,IAAI,MAAM,OAAO,aAAkB;AACpD,QAAMD,QAAO,MAAM,OAAO,MAAW;AACrC,QAAM,OAAO,KAAK,MAAM,MAAMC,UAASD,MAAK,KAAK,QAAQ,WAAW,GAAG,MAAM,CAAC;AAC9E,SAAO,aAAa,IAAI;AAC1B;AAEO,IAAM,QAAiB;AAAA,EAC5B,IAAI;AAAA,EACJ,SAAS;AAAA,EACT,QAAQ,EAAE,MAAM,OAAO,MAAM,MAAM,QAAQ,OAAO;AAAA,EAClD,SAAS;AAAA,EACT,aAAa;AAAA,EACb;AAAA,EACA;AACF;","names":["mulberry32","SCHEMA","writeFile","path","readFile"]}

package/dist/expose/gold-score.d.ts

@@ -0,0 +1,15 @@
+import { type ExposeGold } from "./schema";
+/**
+ * Remove `images` / `floorplans` fields from a gold (or prediction) so a
+ * text-only run isn't penalised for a field it cannot legitimately populate.
+ *
+ * For gold: wraps in `real_estate_property` first (matching the schema), then
+ * strips image fields. For predictions: strips image fields as-is.
+ */
+export declare const stripImages: (gold: ExposeGold | unknown, isGold?: boolean) => unknown;
+/**
+ * `stripImages` + feature level-merge, applied to both gold and prediction
+ * before scoring. Call this instead of `stripImages` in scoring paths.
+ */
+export declare const normalizeForScoring: (gold: ExposeGold | unknown, isGold?: boolean) => unknown;
+//# sourceMappingURL=gold-score.d.ts.map

package/dist/expose/gold-score.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"gold-score.d.ts","sourceRoot":"","sources":["../../src/expose/gold-score.ts"],"names":[],"mappings":"AAAA,OAAO,EAAY,KAAK,UAAU,EAAE,MAAM,UAAU,CAAC;AAmDrD;;;;;;GAMG;AACH,eAAO,MAAM,WAAW,GAAI,MAAM,UAAU,GAAG,OAAO,EAAE,SAAS,OAAO,KAAG,OACe,CAAC;AAE3F;;;GAGG;AACH,eAAO,MAAM,mBAAmB,GAAI,MAAM,UAAU,GAAG,OAAO,EAAE,SAAS,OAAO,KAAG,OAChC,CAAC"}

package/dist/expose/gold.d.ts

@@ -0,0 +1,8 @@
+import type { ExposeGold } from "./schema";
+export declare const sandstrasse: ExposeGold;
+export declare const primePulse: ExposeGold;
+export declare const flottenstrasse: ExposeGold;
+export declare const dock100: ExposeGold;
+export declare const holzhauser: ExposeGold;
+export declare const elsenstrasse: ExposeGold;
+//# sourceMappingURL=gold.d.ts.map

package/dist/expose/gold.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"gold.d.ts","sourceRoot":"","sources":["../../src/expose/gold.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,UAAU,CAAC;AAO3C,eAAO,MAAM,WAAW,EAAE,UAuDzB,CAAC;AAKF,eAAO,MAAM,UAAU,EAAE,UA4ExB,CAAC;AAKF,eAAO,MAAM,cAAc,EAAE,UA8H5B,CAAC;AAKF,eAAO,MAAM,OAAO,EAAE,UAoJrB,CAAC;AAKF,eAAO,MAAM,UAAU,EAAE,UAqGxB,CAAC;AAKF,eAAO,MAAM,YAAY,EAAE,UAmJ1B,CAAC"}

package/dist/expose/instructions.d.ts

@@ -0,0 +1,2 @@
+export declare function buildEstateInstructions(): string;
+//# sourceMappingURL=instructions.d.ts.map

package/dist/expose/instructions.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"instructions.d.ts","sourceRoot":"","sources":["../../src/expose/instructions.ts"],"names":[],"mappings":"AAMA,wBAAgB,uBAAuB,IAAI,MAAM,CA6IhD"}

package/dist/expose/metric.d.ts

@@ -0,0 +1,7 @@
+import type { FieldMetricSpec } from "../types";
+/**
+ * Scoring rules for the estate exposes. usages + features are order-insensitive
+ * sets; numeric fields use tolerance; prose fields use content-word coverage.
+ */
+export declare const exposeMetric: FieldMetricSpec;
+//# sourceMappingURL=metric.d.ts.map

package/dist/expose/metric.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"metric.d.ts","sourceRoot":"","sources":["../../src/expose/metric.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,UAAU,CAAC;AAEhD;;;GAGG;AACH,eAAO,MAAM,YAAY,EAAE,eAqB1B,CAAC"}