@tangle-network/agent-bench 0.4.1 → 0.4.3

package/dist/benchmarks/cad-design.js

@@ -1,7 +1,513 @@
-import {
-  createCadDesignAdapter
-} from "../chunk-53UPUNBZ.js";
-export {
-  createCadDesignAdapter
-};
+import { mkdtemp, readFile, writeFile } from "node:fs/promises";
+import { join } from "node:path";
+import { execFile } from "node:child_process";
+import { tmpdir } from "node:os";
+import { promisify } from "node:util";
+//#region src/benchmarks/cad-design.ts
+/**
+* CAD-Design adapter. Worker artifact = OpenSCAD source (`.scad`). Judge = the
+* REAL OpenSCAD engine: compile + export STL, then measure the produced solid
+* geometry against the task's spec. Fully deterministic — the authoritative CAD
+* kernel is the gate, not an LLM and not the agent's own claim.
+*
+* This is verifiable-reward CAD: an agent is far better at *writing* exact
+* parametric code than at clicking a GUI, and the kernel either produces the
+* specified geometry or it doesn't. Every judged artifact also renders a PNG,
+* so a refine loop's attempts become a watchable "model getting better" reel
+* (run-capsule consumes the renders).
+*
+* Spec checks are geometric and ungameable:
+*   - compiles      : `openscad -o out.stl` exits 0 (hard gate)
+*   - volumes       : disconnected-solid count within [min,max]
+*   - bbox          : overall X/Y/Z extent within bounds (the thing is the right size)
+*   - detail        : triangle count ≥ floor (not a degenerate single cube)
+*   - pitchedRoof   : the top band's XY footprint NARROWS vs the base — a flat
+*                     box can't fake this; a real gabled/hipped roof tapers
+*   - hollow        : interior cavity present (walls, not a solid block) — the
+*                     bbox volume materially exceeds the printed solid volume
+*
+* Requires only `openscad` + `xvfb-run` on PATH (no venv, no Docker, no network).
+*/
+const execFileAsync = promisify(execFile);
+/** Run openscad with xvfb (it needs a GL context even headless). */
+async function openscad(args, cwd) {
+	const { stdout, stderr } = await execFileAsync("xvfb-run", [
+		"-a",
+		"openscad",
+		...args
+	], {
+		cwd,
+		maxBuffer: 1024 * 1024 * 64,
+		timeout: 12e4
+	});
+	return {
+		stdout,
+		stderr
+	};
+}
+/** Parse an ASCII STL into the geometric measures the spec checks. Pure. */
+function measureStl(stl, volumesFromStderr) {
+	const verts = [];
+	const tris = [];
+	const re = /vertex\s+(-?[\d.eE+]+)\s+(-?[\d.eE+]+)\s+(-?[\d.eE+]+)/g;
+	let m;
+	const flat = [];
+	while (m = re.exec(stl)) flat.push([
+		Number(m[1]),
+		Number(m[2]),
+		Number(m[3])
+	]);
+	for (let i = 0; i + 2 < flat.length; i += 3) tris.push([
+		flat[i],
+		flat[i + 1],
+		flat[i + 2]
+	]);
+	for (const v of flat) verts.push(v);
+	const xs = verts.map((v) => v[0]);
+	const ys = verts.map((v) => v[1]);
+	const zs = verts.map((v) => v[2]);
+	const min = (a) => Math.min(...a);
+	const max = (a) => Math.max(...a);
+	const minZ = min(zs);
+	const maxZ = max(zs);
+	const h = maxZ - minZ || 1;
+	const footprint = (lo, hi) => {
+		const band = verts.filter((v) => v[2] >= minZ + lo * h && v[2] <= minZ + hi * h);
+		if (band.length < 3) return 0;
+		const bx = band.map((v) => v[0]);
+		const by = band.map((v) => v[1]);
+		return (max(bx) - min(bx)) * (max(by) - min(by));
+	};
+	let vol6 = 0;
+	for (const [a, b, c] of tris) vol6 += a[0] * (b[1] * c[2] - b[2] * c[1]) - a[1] * (b[0] * c[2] - b[2] * c[0]) + a[2] * (b[0] * c[1] - b[1] * c[0]);
+	const bbox = {
+		x: max(xs) - min(xs),
+		y: max(ys) - min(ys),
+		z: maxZ - minZ
+	};
+	return {
+		triangles: tris.length,
+		volumes: volumesFromStderr,
+		bbox,
+		baseFootprint: footprint(0, .6),
+		topFootprint: footprint(.75, 1),
+		solidVolume: Math.abs(vol6) / 6,
+		bboxVolume: bbox.x * bbox.y * bbox.z || 1
+	};
+}
+function inRange(v, [lo, hi]) {
+	return v >= lo && v <= hi;
+}
+/** Score the geometry against the spec — each check is a named pass/fail. */
+function scoreGeometry(g, spec) {
+	const checks = {};
+	if (spec.volumes) checks.volumes = inRange(g.volumes, spec.volumes);
+	if (spec.minTriangles != null) checks.detail = g.triangles >= spec.minTriangles;
+	if (spec.bbox?.x) checks.bboxX = inRange(g.bbox.x, spec.bbox.x);
+	if (spec.bbox?.y) checks.bboxY = inRange(g.bbox.y, spec.bbox.y);
+	if (spec.bbox?.z) checks.bboxZ = inRange(g.bbox.z, spec.bbox.z);
+	if (spec.pitchedRoof != null) checks.pitchedRoof = g.baseFootprint > 0 && g.topFootprint < spec.pitchedRoof * g.baseFootprint;
+	if (spec.hollowBelow != null) checks.hollow = g.solidVolume < spec.hollowBelow * g.bboxVolume;
+	const vals = Object.values(checks);
+	return {
+		checks,
+		score: vals.length ? vals.filter(Boolean).length / vals.length : 0
+	};
+}
+/** A parametric multi-story gabled building — the gold generator for the
+*  building-family tasks. They share geometry (hollow stacked shells + a gabled
+*  roof that tapers to a ridge + cut openings) and differ only in dimensions, so
+*  one verified generator yields a spec-passing oracle for each. */
+function buildingGold(w, d, sh, ns, rh, t = 3) {
+	return `
+w=${w}; d=${d}; sh=${sh}; ns=${ns}; rh=${rh}; t=${t};
+module shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }
+module roof(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }
+union(){
+  for(s=[0:ns-1]) translate([0,0,s*sh]) shell(w,d,sh,t);
+  for(s=[0:ns]) translate([0,0,s*sh-0.5]) cube([w,d,1]);
+  translate([0,0,ns*sh]) roof(w,d,rh);
+  for(s=[0:ns-1]) for(i=[0:floor((w-28)/22)]) translate([14+i*22,-1,s*sh+10]) cube([10,t+2,12]);
+  translate([w/2-7,-1,2]) cube([14,t+2,20]);
+}
+`.trim();
+}
+/** Build a building-family task (prompt + spec + gold) from its dimensions.
+*  The spec's bbox is derived from the dims with a ±~18% tolerance so the agent
+*  must respect the brief's size; the gold (buildingGold) lands exactly on the
+*  nominal dims, so it passes by construction. Keep rh > ns*sh/3 (else the roof
+*  base sits inside the top-25% band and pitchedRoof fails). */
+function buildingTask(o) {
+	const { id, desc, w, d, sh, ns, rh } = o;
+	const totalWall = ns * sh;
+	const totalH = totalWall + rh;
+	const r = Math.round;
+	const spec = {
+		volumes: [1, ns * 4 + 6],
+		bbox: {
+			x: [r(w * .82), r(w * 1.18)],
+			y: [r(d * .82), r((d + 1) * 1.2)],
+			z: [r(totalH * .82), r((totalH + 1) * 1.2)]
+		},
+		minTriangles: 50,
+		pitchedRoof: .55,
+		hollowBelow: .72
+	};
+	return {
+		id,
+		prompt: [
+			`Write OpenSCAD source for ${desc}.`,
+			"Requirements:",
+			`- footprint roughly ${w} (X) by ${d} (Y) units`,
+			ns > 1 ? `- ${ns} stories, total wall height ~${totalWall} units` : `- one story, wall height ~${sh} units`,
+			"- a PITCHED / gabled roof on top that tapers toward a ridge (NOT flat)",
+			"- hollow shell (walls with an interior cavity), not a solid block",
+			"- a door opening and several windows",
+			"Output ONLY the .scad source, no prose, no code fences."
+		].join("\n"),
+		meta: {
+			spec,
+			gold: buildingGold(w, d, sh, ns, rh)
+		}
+	};
+}
+/** The building family — one parametric skill (hit the bbox + pitched roof +
+*  hollow shell), sampled across footprints, heights, story counts, roof rises. */
+const BUILDING_CONFIGS = [
+	{
+		id: "cottage",
+		desc: "a single-story cottage",
+		w: 50,
+		d: 40,
+		sh: 30,
+		ns: 1,
+		rh: 18
+	},
+	{
+		id: "cabin",
+		desc: "a small one-room cabin",
+		w: 44,
+		d: 36,
+		sh: 26,
+		ns: 1,
+		rh: 22
+	},
+	{
+		id: "bungalow",
+		desc: "a wide single-story bungalow",
+		w: 70,
+		d: 52,
+		sh: 28,
+		ns: 1,
+		rh: 20
+	},
+	{
+		id: "barn",
+		desc: "a long deep barn",
+		w: 60,
+		d: 96,
+		sh: 42,
+		ns: 1,
+		rh: 26
+	},
+	{
+		id: "warehouse",
+		desc: "a large warehouse",
+		w: 110,
+		d: 80,
+		sh: 40,
+		ns: 1,
+		rh: 26
+	},
+	{
+		id: "chapel",
+		desc: "a narrow tall chapel with a steep roof",
+		w: 40,
+		d: 72,
+		sh: 48,
+		ns: 1,
+		rh: 34
+	},
+	{
+		id: "farmhouse",
+		desc: "a two-story farmhouse",
+		w: 72,
+		d: 56,
+		sh: 30,
+		ns: 2,
+		rh: 26
+	},
+	{
+		id: "two-story-manor",
+		desc: "a wide two-story manor",
+		w: 96,
+		d: 64,
+		sh: 32,
+		ns: 2,
+		rh: 28
+	},
+	{
+		id: "townhouse",
+		desc: "a narrow three-story townhouse",
+		w: 42,
+		d: 52,
+		sh: 30,
+		ns: 3,
+		rh: 34
+	},
+	{
+		id: "watchtower",
+		desc: "a tall narrow three-story watchtower",
+		w: 34,
+		d: 34,
+		sh: 30,
+		ns: 3,
+		rh: 32
+	}
+];
+/** A hollow cylindrical water tower with a conical roof — a non-box shape that
+*  still exercises pitched-roof (the cone tapers to an apex) + hollow + bbox. */
+const WATER_TOWER_GOLD = `
+$fn=48; r=24; bh=72; t=3; ch=26;
+union(){
+  difference(){ cylinder(h=bh, r=r); translate([0,0,t]) cylinder(h=bh, r=r-t); }
+  translate([0,0,bh]) cylinder(h=ch, r1=r, r2=0);
+}
+`.trim();
+/** An A-frame: the steep roof IS the walls — a hollow triangular prism. The most
+*  extreme taper in the set; tests that the agent can build a non-rectangular shell. */
+const AFRAME_GOLD = `
+w=44; d=60; h=66; t=3;
+module tri(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }
+union(){
+  difference(){ tri(w,d,h); translate([t,0,0]) tri(w-2*t,d,h-2*t*h/d); }
+  translate([w/2-7,-1,0]) cube([14,t+2,24]);
+}
+`.trim();
+const TASKS = [
+	{
+		id: "two-story-house",
+		prompt: [
+			"Write OpenSCAD source for a two-story house.",
+			"Requirements:",
+			"- footprint roughly 80 (X) by 60 (Y) units",
+			"- two stories, total wall height ~68 units",
+			"- a PITCHED / gabled roof on top (NOT a flat roof) — the roof must taper toward a ridge",
+			"- hollow shell (walls with an interior cavity), not a solid block",
+			"- at least one door opening and several windows",
+			"Output ONLY the .scad source, no prose, no code fences."
+		].join("\n"),
+		meta: {
+			spec: {
+				volumes: [1, 12],
+				bbox: {
+					x: [70, 110],
+					y: [50, 80],
+					z: [80, 130]
+				},
+				minTriangles: 60,
+				pitchedRoof: .55,
+				hollowBelow: .7
+			},
+			gold: `
+w=80; d=60; sh=34; t=3; rh=28;
+module shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }
+module roof(w,d,h){ translate([0,0,0]) rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }
+union(){
+  for(s=[0,1]) translate([0,0,s*sh]) shell(w,d,sh,t);
+  for(s=[0,1,2]) translate([0,0,s*sh-1]) cube([w,d,1]);
+  translate([0,0,2*sh]) roof(w,d,rh);
+  for(s=[0,1]) for(i=[0:2]) translate([16+i*22,-1,s*sh+10]) cube([12,t+2,14]);
+  translate([w/2-8,-1,2]) cube([16,t+2,22]);
+}
+`.trim()
+		}
+	},
+	...BUILDING_CONFIGS.map(buildingTask),
+	{
+		id: "water-tower",
+		prompt: [
+			"Write OpenSCAD source for a water tower: a tall hollow cylindrical tank with a conical roof.",
+			"Requirements:",
+			"- a hollow cylindrical tank, ~48 units in diameter, ~72 units tall, with an interior cavity",
+			"- a CONICAL roof on top that tapers to a point (a pitched roof)",
+			"- wall thickness around 3 units",
+			"Output ONLY the .scad source, no prose, no code fences."
+		].join("\n"),
+		meta: {
+			spec: {
+				volumes: [1, 4],
+				bbox: {
+					x: [40, 56],
+					y: [40, 56],
+					z: [88, 110]
+				},
+				minTriangles: 60,
+				pitchedRoof: .55,
+				hollowBelow: .8
+			},
+			gold: WATER_TOWER_GOLD
+		}
+	},
+	{
+		id: "a-frame-cabin",
+		prompt: [
+			"Write OpenSCAD source for an A-frame cabin: a steep triangular shell where the roof forms the walls.",
+			"Requirements:",
+			"- footprint roughly 44 (X) by 60 (Y) units",
+			"- a STEEP triangular cross-section ~66 units tall that tapers to a ridge at the top",
+			"- hollow inside (an interior cavity), not a solid wedge",
+			"- a door opening at the front",
+			"Output ONLY the .scad source, no prose, no code fences."
+		].join("\n"),
+		meta: {
+			spec: {
+				volumes: [1, 4],
+				bbox: {
+					x: [36, 52],
+					y: [50, 74],
+					z: [56, 80]
+				},
+				minTriangles: 30,
+				pitchedRoof: .45,
+				hollowBelow: .8
+			},
+			gold: AFRAME_GOLD
+		}
+	},
+	{
+		id: "hex-planter",
+		prompt: [
+			"Write OpenSCAD source for a hexagonal planter pot.",
+			"Requirements:",
+			"- a hollow hexagonal prism (6-sided), outer width ~50 units across, height ~45",
+			"- open at the top, closed at the bottom (a cavity for soil)",
+			"- a wall thickness around 3 units",
+			"Output ONLY the .scad source, no prose, no code fences."
+		].join("\n"),
+		meta: {
+			spec: {
+				volumes: [1, 3],
+				bbox: {
+					x: [40, 60],
+					y: [40, 60],
+					z: [40, 55]
+				},
+				minTriangles: 30,
+				hollowBelow: .7
+			},
+			gold: `
+$fn=6; or=25; h=45; t=3;
+difference(){
+  cylinder(h=h, r=or);
+  translate([0,0,t]) cylinder(h=h, r=or-t);
+}
+`.trim()
+		}
+	}
+];
+function createCadDesignAdapter() {
+	return {
+		name: "cad-design",
+		async preflight() {
+			try {
+				await execFileAsync("xvfb-run", [
+					"-a",
+					"openscad",
+					"--version"
+				], { timeout: 3e4 });
+			} catch (err) {
+				const msg = err instanceof Error ? err.message : String(err);
+				throw new Error(`cad-design preflight failed: ${msg}\nFix: install OpenSCAD + Xvfb (Debian/Ubuntu: sudo apt-get install -y openscad xvfb). The judge runs \`xvfb-run -a openscad -o out.stl model.scad\` — both must be on PATH.`);
+			}
+		},
+		async loadTasks(opts = {}) {
+			let tasks = TASKS;
+			if (opts.ids) tasks = tasks.filter((t) => opts.ids.includes(t.id));
+			if (opts.limit != null) tasks = tasks.slice(0, opts.limit);
+			return tasks.map((t) => ({
+				id: t.id,
+				prompt: t.prompt,
+				metadata: t.meta
+			}));
+		},
+		async goldArtifact(task) {
+			return task.metadata?.gold;
+		},
+		async judge(task, artifact) {
+			const spec = task.metadata.spec;
+			const src = artifact.trim();
+			if (!src) return {
+				resolved: false,
+				score: 0,
+				detail: "empty artifact"
+			};
+			const dir = await mkdtemp(join(tmpdir(), "cad-"));
+			const scad = join(dir, "model.scad");
+			const stlPath = join(dir, "model.stl");
+			const pngPath = join(dir, "model.png");
+			await writeFile(scad, src);
+			let stderr = "";
+			try {
+				stderr = (await openscad([
+					"-o",
+					stlPath,
+					scad
+				], dir)).stderr;
+			} catch (err) {
+				return {
+					resolved: false,
+					score: 0,
+					detail: `compile failed: ${(err instanceof Error ? err.message : String(err)).slice(0, 300)}`
+				};
+			}
+			try {
+				await openscad([
+					"-o",
+					pngPath,
+					"--imgsize=1100,850",
+					"--camera=40,30,40,55,0,25,260",
+					"--colorscheme=Tomorrow",
+					scad
+				], dir);
+			} catch {}
+			let stl;
+			try {
+				stl = await readFile(stlPath, "utf8");
+			} catch {
+				return {
+					resolved: false,
+					score: 0,
+					detail: "compiled but produced no STL (empty geometry)"
+				};
+			}
+			const volumes = Number(/Volumes:\s*(\d+)/.exec(stderr)?.[1] ?? "1");
+			const geo = measureStl(stl, volumes);
+			const { checks, score } = scoreGeometry(geo, spec);
+			return {
+				resolved: score === 1,
+				score,
+				detail: JSON.stringify({
+					checks,
+					geo: {
+						triangles: geo.triangles,
+						volumes: geo.volumes,
+						bbox: {
+							x: +geo.bbox.x.toFixed(1),
+							y: +geo.bbox.y.toFixed(1),
+							z: +geo.bbox.z.toFixed(1)
+						},
+						hollowRatio: +(geo.solidVolume / geo.bboxVolume).toFixed(3),
+						topVsBaseFootprint: geo.baseFootprint ? +(geo.topFootprint / geo.baseFootprint).toFixed(3) : null
+					},
+					renderPath: pngPath,
+					stlPath
+				})
+			};
+		}
+	};
+}
+//#endregion
+export { createCadDesignAdapter };
+
 //# sourceMappingURL=cad-design.js.map

package/dist/benchmarks/cad-design.js.map

@@ -1 +1 @@
-{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}
+{"version":3,"file":"cad-design.js","names":[],"sources":["../../src/benchmarks/cad-design.ts"],"sourcesContent":["/**\n * CAD-Design adapter. Worker artifact = OpenSCAD source (`.scad`). Judge = the\n * REAL OpenSCAD engine: compile + export STL, then measure the produced solid\n * geometry against the task's spec. Fully deterministic — the authoritative CAD\n * kernel is the gate, not an LLM and not the agent's own claim.\n *\n * This is verifiable-reward CAD: an agent is far better at *writing* exact\n * parametric code than at clicking a GUI, and the kernel either produces the\n * specified geometry or it doesn't. Every judged artifact also renders a PNG,\n * so a refine loop's attempts become a watchable \"model getting better\" reel\n * (run-capsule consumes the renders).\n *\n * Spec checks are geometric and ungameable:\n *   - compiles      : `openscad -o out.stl` exits 0 (hard gate)\n *   - volumes       : disconnected-solid count within [min,max]\n *   - bbox          : overall X/Y/Z extent within bounds (the thing is the right size)\n *   - detail        : triangle count ≥ floor (not a degenerate single cube)\n *   - pitchedRoof   : the top band's XY footprint NARROWS vs the base — a flat\n *                     box can't fake this; a real gabled/hipped roof tapers\n *   - hollow        : interior cavity present (walls, not a solid block) — the\n *                     bbox volume materially exceeds the printed solid volume\n *\n * Requires only `openscad` + `xvfb-run` on PATH (no venv, no Docker, no network).\n */\n\nimport { execFile } from 'node:child_process'\nimport { mkdtemp, readFile, writeFile } from 'node:fs/promises'\nimport { tmpdir } from 'node:os'\nimport { join } from 'node:path'\nimport { promisify } from 'node:util'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst execFileAsync = promisify(execFile)\n\n/** Spec assertions a CAD task can require. All are deterministic + geometric. */\nexport interface CadSpec {\n  volumes?: [number, number]\n  bbox?: { x?: [number, number]; y?: [number, number]; z?: [number, number] }\n  minTriangles?: number\n  /** Top-band XY footprint must be < this fraction of the base footprint. */\n  pitchedRoof?: number\n  /** Printed solid volume must be < this fraction of the bbox volume (hollow). */\n  hollowBelow?: number\n}\n\ninterface CadTaskMeta {\n  spec: CadSpec\n  /** A known-good .scad that satisfies the spec — the oracle for verify-judge. */\n  gold: string\n}\n\n/** Run openscad with xvfb (it needs a GL context even headless). */\nasync function openscad(args: string[], cwd: string): Promise<{ stdout: string; stderr: string }> {\n  // -a => auto-pick a free display; openscad writes Facets/Volumes to stderr.\n  const { stdout, stderr } = await execFileAsync('xvfb-run', ['-a', 'openscad', ...args], {\n    cwd,\n    maxBuffer: 1024 * 1024 * 64,\n    timeout: 120_000,\n  })\n  return { stdout, stderr }\n}\n\ninterface Geometry {\n  triangles: number\n  volumes: number\n  bbox: { x: number; y: number; z: number }\n  /** XY footprint area (extent_x * extent_y) of the bottom 60% and top 25% z-bands. */\n  baseFootprint: number\n  topFootprint: number\n  /** Convex-ish solid volume estimate from the STL (sum of signed tetra volumes). */\n  solidVolume: number\n  bboxVolume: number\n}\n\n/** Parse an ASCII STL into the geometric measures the spec checks. Pure. */\nfunction measureStl(stl: string, volumesFromStderr: number): Geometry {\n  const verts: Array<[number, number, number]> = []\n  const tris: Array<[[number, number, number], [number, number, number], [number, number, number]]> = []\n  const re = /vertex\\s+(-?[\\d.eE+]+)\\s+(-?[\\d.eE+]+)\\s+(-?[\\d.eE+]+)/g\n  let m: RegExpExecArray | null\n  const flat: Array<[number, number, number]> = []\n  while ((m = re.exec(stl))) flat.push([Number(m[1]), Number(m[2]), Number(m[3])])\n  for (let i = 0; i + 2 < flat.length; i += 3) tris.push([flat[i]!, flat[i + 1]!, flat[i + 2]!])\n  for (const v of flat) verts.push(v)\n\n  const xs = verts.map((v) => v[0])\n  const ys = verts.map((v) => v[1])\n  const zs = verts.map((v) => v[2])\n  const min = (a: number[]) => Math.min(...a)\n  const max = (a: number[]) => Math.max(...a)\n  const minZ = min(zs)\n  const maxZ = max(zs)\n  const h = maxZ - minZ || 1\n\n  // Footprint of a z-band = XY extent of the vertices within it.\n  const footprint = (lo: number, hi: number): number => {\n    const band = verts.filter((v) => v[2] >= minZ + lo * h && v[2] <= minZ + hi * h)\n    if (band.length < 3) return 0\n    const bx = band.map((v) => v[0])\n    const by = band.map((v) => v[1])\n    return (max(bx) - min(bx)) * (max(by) - min(by))\n  }\n\n  // Signed volume via the divergence theorem over triangles (|Σ v0·(v1×v2)/6|).\n  let vol6 = 0\n  for (const [a, b, c] of tris) {\n    vol6 +=\n      a[0] * (b[1] * c[2] - b[2] * c[1]) -\n      a[1] * (b[0] * c[2] - b[2] * c[0]) +\n      a[2] * (b[0] * c[1] - b[1] * c[0])\n  }\n  const bbox = { x: max(xs) - min(xs), y: max(ys) - min(ys), z: maxZ - minZ }\n  return {\n    triangles: tris.length,\n    volumes: volumesFromStderr,\n    bbox,\n    baseFootprint: footprint(0, 0.6),\n    topFootprint: footprint(0.75, 1),\n    solidVolume: Math.abs(vol6) / 6,\n    bboxVolume: bbox.x * bbox.y * bbox.z || 1,\n  }\n}\n\nfunction inRange(v: number, [lo, hi]: [number, number]): boolean {\n  return v >= lo && v <= hi\n}\n\n/** Score the geometry against the spec — each check is a named pass/fail. */\nfunction scoreGeometry(g: Geometry, spec: CadSpec): { checks: Record<string, boolean>; score: number } {\n  const checks: Record<string, boolean> = {}\n  if (spec.volumes) checks.volumes = inRange(g.volumes, spec.volumes)\n  if (spec.minTriangles != null) checks.detail = g.triangles >= spec.minTriangles\n  if (spec.bbox?.x) checks.bboxX = inRange(g.bbox.x, spec.bbox.x)\n  if (spec.bbox?.y) checks.bboxY = inRange(g.bbox.y, spec.bbox.y)\n  if (spec.bbox?.z) checks.bboxZ = inRange(g.bbox.z, spec.bbox.z)\n  if (spec.pitchedRoof != null)\n    checks.pitchedRoof = g.baseFootprint > 0 && g.topFootprint < spec.pitchedRoof * g.baseFootprint\n  if (spec.hollowBelow != null) checks.hollow = g.solidVolume < spec.hollowBelow * g.bboxVolume\n  const vals = Object.values(checks)\n  const score = vals.length ? vals.filter(Boolean).length / vals.length : 0\n  return { checks, score }\n}\n\n/** A parametric multi-story gabled building — the gold generator for the\n *  building-family tasks. They share geometry (hollow stacked shells + a gabled\n *  roof that tapers to a ridge + cut openings) and differ only in dimensions, so\n *  one verified generator yields a spec-passing oracle for each. */\nfunction buildingGold(w: number, d: number, sh: number, ns: number, rh: number, t = 3): string {\n  return `\nw=${w}; d=${d}; sh=${sh}; ns=${ns}; rh=${rh}; t=${t};\nmodule shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }\nmodule roof(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }\nunion(){\n  for(s=[0:ns-1]) translate([0,0,s*sh]) shell(w,d,sh,t);\n  for(s=[0:ns]) translate([0,0,s*sh-0.5]) cube([w,d,1]);\n  translate([0,0,ns*sh]) roof(w,d,rh);\n  for(s=[0:ns-1]) for(i=[0:floor((w-28)/22)]) translate([14+i*22,-1,s*sh+10]) cube([10,t+2,12]);\n  translate([w/2-7,-1,2]) cube([14,t+2,20]);\n}\n`.trim()\n}\n\n/** Build a building-family task (prompt + spec + gold) from its dimensions.\n *  The spec's bbox is derived from the dims with a ±~18% tolerance so the agent\n *  must respect the brief's size; the gold (buildingGold) lands exactly on the\n *  nominal dims, so it passes by construction. Keep rh > ns*sh/3 (else the roof\n *  base sits inside the top-25% band and pitchedRoof fails). */\nfunction buildingTask(o: { id: string; desc: string; w: number; d: number; sh: number; ns: number; rh: number }): {\n  id: string\n  prompt: string\n  meta: CadTaskMeta\n} {\n  const { id, desc, w, d, sh, ns, rh } = o\n  const totalWall = ns * sh\n  const totalH = totalWall + rh\n  const r = Math.round\n  const spec: CadSpec = {\n    volumes: [1, ns * 4 + 6],\n    bbox: {\n      x: [r(w * 0.82), r(w * 1.18)],\n      y: [r(d * 0.82), r((d + 1) * 1.2)],\n      z: [r(totalH * 0.82), r((totalH + 1) * 1.2)],\n    },\n    minTriangles: 50,\n    pitchedRoof: 0.55,\n    hollowBelow: 0.72,\n  }\n  const prompt = [\n    `Write OpenSCAD source for ${desc}.`,\n    'Requirements:',\n    `- footprint roughly ${w} (X) by ${d} (Y) units`,\n    ns > 1 ? `- ${ns} stories, total wall height ~${totalWall} units` : `- one story, wall height ~${sh} units`,\n    '- a PITCHED / gabled roof on top that tapers toward a ridge (NOT flat)',\n    '- hollow shell (walls with an interior cavity), not a solid block',\n    '- a door opening and several windows',\n    'Output ONLY the .scad source, no prose, no code fences.',\n  ].join('\\n')\n  return { id, prompt, meta: { spec, gold: buildingGold(w, d, sh, ns, rh) } }\n}\n\n/** The building family — one parametric skill (hit the bbox + pitched roof +\n *  hollow shell), sampled across footprints, heights, story counts, roof rises. */\nconst BUILDING_CONFIGS: Array<{ id: string; desc: string; w: number; d: number; sh: number; ns: number; rh: number }> = [\n  { id: 'cottage', desc: 'a single-story cottage', w: 50, d: 40, sh: 30, ns: 1, rh: 18 },\n  { id: 'cabin', desc: 'a small one-room cabin', w: 44, d: 36, sh: 26, ns: 1, rh: 22 },\n  { id: 'bungalow', desc: 'a wide single-story bungalow', w: 70, d: 52, sh: 28, ns: 1, rh: 20 },\n  { id: 'barn', desc: 'a long deep barn', w: 60, d: 96, sh: 42, ns: 1, rh: 26 },\n  { id: 'warehouse', desc: 'a large warehouse', w: 110, d: 80, sh: 40, ns: 1, rh: 26 },\n  { id: 'chapel', desc: 'a narrow tall chapel with a steep roof', w: 40, d: 72, sh: 48, ns: 1, rh: 34 },\n  { id: 'farmhouse', desc: 'a two-story farmhouse', w: 72, d: 56, sh: 30, ns: 2, rh: 26 },\n  { id: 'two-story-manor', desc: 'a wide two-story manor', w: 96, d: 64, sh: 32, ns: 2, rh: 28 },\n  { id: 'townhouse', desc: 'a narrow three-story townhouse', w: 42, d: 52, sh: 30, ns: 3, rh: 34 },\n  { id: 'watchtower', desc: 'a tall narrow three-story watchtower', w: 34, d: 34, sh: 30, ns: 3, rh: 32 },\n]\n\n/** A hollow cylindrical water tower with a conical roof — a non-box shape that\n *  still exercises pitched-roof (the cone tapers to an apex) + hollow + bbox. */\nconst WATER_TOWER_GOLD = `\n$fn=48; r=24; bh=72; t=3; ch=26;\nunion(){\n  difference(){ cylinder(h=bh, r=r); translate([0,0,t]) cylinder(h=bh, r=r-t); }\n  translate([0,0,bh]) cylinder(h=ch, r1=r, r2=0);\n}\n`.trim()\n\n/** An A-frame: the steep roof IS the walls — a hollow triangular prism. The most\n *  extreme taper in the set; tests that the agent can build a non-rectangular shell. */\nconst AFRAME_GOLD = `\nw=44; d=60; h=66; t=3;\nmodule tri(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }\nunion(){\n  difference(){ tri(w,d,h); translate([t,0,0]) tri(w-2*t,d,h-2*t*h/d); }\n  translate([w/2-7,-1,0]) cube([14,t+2,24]);\n}\n`.trim()\n\nconst TASKS: Array<{ id: string; prompt: string; meta: CadTaskMeta }> = [\n  {\n    id: 'two-story-house',\n    prompt: [\n      'Write OpenSCAD source for a two-story house.',\n      'Requirements:',\n      '- footprint roughly 80 (X) by 60 (Y) units',\n      '- two stories, total wall height ~68 units',\n      '- a PITCHED / gabled roof on top (NOT a flat roof) — the roof must taper toward a ridge',\n      '- hollow shell (walls with an interior cavity), not a solid block',\n      '- at least one door opening and several windows',\n      'Output ONLY the .scad source, no prose, no code fences.',\n    ].join('\\n'),\n    meta: {\n      spec: {\n        volumes: [1, 12],\n        bbox: { x: [70, 110], y: [50, 80], z: [80, 130] },\n        minTriangles: 60,\n        pitchedRoof: 0.55,\n        hollowBelow: 0.7,\n      },\n      gold: `\nw=80; d=60; sh=34; t=3; rh=28;\nmodule shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }\nmodule roof(w,d,h){ translate([0,0,0]) rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }\nunion(){\n  for(s=[0,1]) translate([0,0,s*sh]) shell(w,d,sh,t);\n  for(s=[0,1,2]) translate([0,0,s*sh-1]) cube([w,d,1]);\n  translate([0,0,2*sh]) roof(w,d,rh);\n  for(s=[0,1]) for(i=[0:2]) translate([16+i*22,-1,s*sh+10]) cube([12,t+2,14]);\n  translate([w/2-8,-1,2]) cube([16,t+2,22]);\n}\n`.trim(),\n    },\n  },\n  ...BUILDING_CONFIGS.map(buildingTask),\n  {\n    id: 'water-tower',\n    prompt: [\n      'Write OpenSCAD source for a water tower: a tall hollow cylindrical tank with a conical roof.',\n      'Requirements:',\n      '- a hollow cylindrical tank, ~48 units in diameter, ~72 units tall, with an interior cavity',\n      '- a CONICAL roof on top that tapers to a point (a pitched roof)',\n      '- wall thickness around 3 units',\n      'Output ONLY the .scad source, no prose, no code fences.',\n    ].join('\\n'),\n    meta: {\n      spec: {\n        volumes: [1, 4],\n        bbox: { x: [40, 56], y: [40, 56], z: [88, 110] },\n        minTriangles: 60,\n        pitchedRoof: 0.55,\n        hollowBelow: 0.8,\n      },\n      gold: WATER_TOWER_GOLD,\n    },\n  },\n  {\n    id: 'a-frame-cabin',\n    prompt: [\n      'Write OpenSCAD source for an A-frame cabin: a steep triangular shell where the roof forms the walls.',\n      'Requirements:',\n      '- footprint roughly 44 (X) by 60 (Y) units',\n      '- a STEEP triangular cross-section ~66 units tall that tapers to a ridge at the top',\n      '- hollow inside (an interior cavity), not a solid wedge',\n      '- a door opening at the front',\n      'Output ONLY the .scad source, no prose, no code fences.',\n    ].join('\\n'),\n    meta: {\n      spec: {\n        volumes: [1, 4],\n        bbox: { x: [36, 52], y: [50, 74], z: [56, 80] },\n        minTriangles: 30,\n        pitchedRoof: 0.45,\n        hollowBelow: 0.8,\n      },\n      gold: AFRAME_GOLD,\n    },\n  },\n  {\n    id: 'hex-planter',\n    prompt: [\n      'Write OpenSCAD source for a hexagonal planter pot.',\n      'Requirements:',\n      '- a hollow hexagonal prism (6-sided), outer width ~50 units across, height ~45',\n      '- open at the top, closed at the bottom (a cavity for soil)',\n      '- a wall thickness around 3 units',\n      'Output ONLY the .scad source, no prose, no code fences.',\n    ].join('\\n'),\n    meta: {\n      spec: {\n        volumes: [1, 3],\n        bbox: { x: [40, 60], y: [40, 60], z: [40, 55] },\n        minTriangles: 30,\n        hollowBelow: 0.7,\n      },\n      gold: `\n$fn=6; or=25; h=45; t=3;\ndifference(){\n  cylinder(h=h, r=or);\n  translate([0,0,t]) cylinder(h=h, r=or-t);\n}\n`.trim(),\n    },\n  },\n]\n\nexport function createCadDesignAdapter(): BenchmarkAdapter {\n  return {\n    name: 'cad-design',\n\n    async preflight() {\n      try {\n        await execFileAsync('xvfb-run', ['-a', 'openscad', '--version'], { timeout: 30_000 })\n      } catch (err) {\n        const msg = err instanceof Error ? err.message : String(err)\n        throw new Error(\n          `cad-design preflight failed: ${msg}\\n` +\n            `Fix: install OpenSCAD + Xvfb (Debian/Ubuntu: sudo apt-get install -y openscad xvfb). ` +\n            `The judge runs \\`xvfb-run -a openscad -o out.stl model.scad\\` — both must be on PATH.`,\n        )\n      }\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      let tasks = TASKS\n      if (opts.ids) tasks = tasks.filter((t) => opts.ids!.includes(t.id))\n      if (opts.limit != null) tasks = tasks.slice(0, opts.limit)\n      return tasks.map((t): BenchTask => ({ id: t.id, prompt: t.prompt, metadata: t.meta as unknown as Record<string, unknown> }))\n    },\n\n    async goldArtifact(task: BenchTask) {\n      const meta = task.metadata as unknown as CadTaskMeta | undefined\n      return meta?.gold\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const spec = (task.metadata as unknown as CadTaskMeta).spec\n      const src = artifact.trim()\n      if (!src) return { resolved: false, score: 0, detail: 'empty artifact' }\n\n      const dir = await mkdtemp(join(tmpdir(), 'cad-'))\n      const scad = join(dir, 'model.scad')\n      const stlPath = join(dir, 'model.stl')\n      const pngPath = join(dir, 'model.png')\n      await writeFile(scad, src)\n\n      // GATE 1: compile + export STL. Nonzero exit / CGAL error => does not compile.\n      let stderr = ''\n      try {\n        const r = await openscad(['-o', stlPath, scad], dir)\n        stderr = r.stderr\n      } catch (err) {\n        const msg = err instanceof Error ? err.message : String(err)\n        return { resolved: false, score: 0, detail: `compile failed: ${msg.slice(0, 300)}` }\n      }\n\n      // Render a PNG alongside (the artifact run-capsule animates). Non-fatal.\n      try {\n        await openscad(\n          ['-o', pngPath, '--imgsize=1100,850', '--camera=40,30,40,55,0,25,260', '--colorscheme=Tomorrow', scad],\n          dir,\n        )\n      } catch {\n        /* render is for the reel; geometry gate already has the STL */\n      }\n\n      let stl: string\n      try {\n        stl = await readFile(stlPath, 'utf8')\n      } catch {\n        return { resolved: false, score: 0, detail: 'compiled but produced no STL (empty geometry)' }\n      }\n      const volumes = Number(/Volumes:\\s*(\\d+)/.exec(stderr)?.[1] ?? '1')\n      const geo = measureStl(stl, volumes)\n      const { checks, score } = scoreGeometry(geo, spec)\n      const resolved = score === 1\n      const detail = JSON.stringify({\n        checks,\n        geo: {\n          triangles: geo.triangles,\n          volumes: geo.volumes,\n          bbox: { x: +geo.bbox.x.toFixed(1), y: +geo.bbox.y.toFixed(1), z: +geo.bbox.z.toFixed(1) },\n          hollowRatio: +(geo.solidVolume / geo.bboxVolume).toFixed(3),\n          topVsBaseFootprint: geo.baseFootprint ? +(geo.topFootprint / geo.baseFootprint).toFixed(3) : null,\n        },\n        renderPath: pngPath,\n        stlPath,\n      })\n      return { resolved, score, detail }\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgCA,MAAM,gBAAgB,UAAU,QAAQ;;AAoBxC,eAAe,SAAS,MAAgB,KAA0D;CAEhG,MAAM,EAAE,QAAQ,WAAW,MAAM,cAAc,YAAY;EAAC;EAAM;EAAY,GAAG;CAAI,GAAG;EACtF;EACA,WAAW,OAAO,OAAO;EACzB,SAAS;CACX,CAAC;CACD,OAAO;EAAE;EAAQ;CAAO;AAC1B;;AAeA,SAAS,WAAW,KAAa,mBAAqC;CACpE,MAAM,QAAyC,CAAC;CAChD,MAAM,OAA8F,CAAC;CACrG,MAAM,KAAK;CACX,IAAI;CACJ,MAAM,OAAwC,CAAC;CAC/C,OAAQ,IAAI,GAAG,KAAK,GAAG,GAAI,KAAK,KAAK;EAAC,OAAO,EAAE,EAAE;EAAG,OAAO,EAAE,EAAE;EAAG,OAAO,EAAE,EAAE;CAAC,CAAC;CAC/E,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK,QAAQ,KAAK,GAAG,KAAK,KAAK;EAAC,KAAK;EAAK,KAAK,IAAI;EAAK,KAAK,IAAI;CAAG,CAAC;CAC7F,KAAK,MAAM,KAAK,MAAM,MAAM,KAAK,CAAC;CAElC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,EAAE;CAChC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,EAAE;CAChC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,EAAE;CAChC,MAAM,OAAO,MAAgB,KAAK,IAAI,GAAG,CAAC;CAC1C,MAAM,OAAO,MAAgB,KAAK,IAAI,GAAG,CAAC;CAC1C,MAAM,OAAO,IAAI,EAAE;CACnB,MAAM,OAAO,IAAI,EAAE;CACnB,MAAM,IAAI,OAAO,QAAQ;CAGzB,MAAM,aAAa,IAAY,OAAuB;EACpD,MAAM,OAAO,MAAM,QAAQ,MAAM,EAAE,MAAM,OAAO,KAAK,KAAK,EAAE,MAAM,OAAO,KAAK,CAAC;EAC/E,IAAI,KAAK,SAAS,GAAG,OAAO;EAC5B,MAAM,KAAK,KAAK,KAAK,MAAM,EAAE,EAAE;EAC/B,MAAM,KAAK,KAAK,KAAK,MAAM,EAAE,EAAE;EAC/B,QAAQ,IAAI,EAAE,IAAI,IAAI,EAAE,MAAM,IAAI,EAAE,IAAI,IAAI,EAAE;CAChD;CAGA,IAAI,OAAO;CACX,KAAK,MAAM,CAAC,GAAG,GAAG,MAAM,MACtB,QACE,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,MAC/B,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,MAC/B,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE;CAEnC,MAAM,OAAO;EAAE,GAAG,IAAI,EAAE,IAAI,IAAI,EAAE;EAAG,GAAG,IAAI,EAAE,IAAI,IAAI,EAAE;EAAG,GAAG,OAAO;CAAK;CAC1E,OAAO;EACL,WAAW,KAAK;EAChB,SAAS;EACT;EACA,eAAe,UAAU,GAAG,EAAG;EAC/B,cAAc,UAAU,KAAM,CAAC;EAC/B,aAAa,KAAK,IAAI,IAAI,IAAI;EAC9B,YAAY,KAAK,IAAI,KAAK,IAAI,KAAK,KAAK;CAC1C;AACF;AAEA,SAAS,QAAQ,GAAW,CAAC,IAAI,KAAgC;CAC/D,OAAO,KAAK,MAAM,KAAK;AACzB;;AAGA,SAAS,cAAc,GAAa,MAAmE;CACrG,MAAM,SAAkC,CAAC;CACzC,IAAI,KAAK,SAAS,OAAO,UAAU,QAAQ,EAAE,SAAS,KAAK,OAAO;CAClE,IAAI,KAAK,gBAAgB,MAAM,OAAO,SAAS,EAAE,aAAa,KAAK;CACnE,IAAI,KAAK,MAAM,GAAG,OAAO,QAAQ,QAAQ,EAAE,KAAK,GAAG,KAAK,KAAK,CAAC;CAC9D,IAAI,KAAK,MAAM,GAAG,OAAO,QAAQ,QAAQ,EAAE,KAAK,GAAG,KAAK,KAAK,CAAC;CAC9D,IAAI,KAAK,MAAM,GAAG,OAAO,QAAQ,QAAQ,EAAE,KAAK,GAAG,KAAK,KAAK,CAAC;CAC9D,IAAI,KAAK,eAAe,MACtB,OAAO,cAAc,EAAE,gBAAgB,KAAK,EAAE,eAAe,KAAK,cAAc,EAAE;CACpF,IAAI,KAAK,eAAe,MAAM,OAAO,SAAS,EAAE,cAAc,KAAK,cAAc,EAAE;CACnF,MAAM,OAAO,OAAO,OAAO,MAAM;CAEjC,OAAO;EAAE;EAAQ,OADH,KAAK,SAAS,KAAK,OAAO,OAAO,CAAC,CAAC,SAAS,KAAK,SAAS;CACjD;AACzB;;;;;AAMA,SAAS,aAAa,GAAW,GAAW,IAAY,IAAY,IAAY,IAAI,GAAW;CAC7F,OAAO;IACL,EAAE,MAAM,EAAE,OAAO,GAAG,OAAO,GAAG,OAAO,GAAG,MAAM,EAAE;;;;;;;;;;EAUlD,KAAK;AACP;;;;;;AAOA,SAAS,aAAa,GAIpB;CACA,MAAM,EAAE,IAAI,MAAM,GAAG,GAAG,IAAI,IAAI,OAAO;CACvC,MAAM,YAAY,KAAK;CACvB,MAAM,SAAS,YAAY;CAC3B,MAAM,IAAI,KAAK;CACf,MAAM,OAAgB;EACpB,SAAS,CAAC,GAAG,KAAK,IAAI,CAAC;EACvB,MAAM;GACJ,GAAG,CAAC,EAAE,IAAI,GAAI,GAAG,EAAE,IAAI,IAAI,CAAC;GAC5B,GAAG,CAAC,EAAE,IAAI,GAAI,GAAG,GAAG,IAAI,KAAK,GAAG,CAAC;GACjC,GAAG,CAAC,EAAE,SAAS,GAAI,GAAG,GAAG,SAAS,KAAK,GAAG,CAAC;EAC7C;EACA,cAAc;EACd,aAAa;EACb,aAAa;CACf;CAWA,OAAO;EAAE;EAAI,QAVE;GACb,6BAA6B,KAAK;GAClC;GACA,uBAAuB,EAAE,UAAU,EAAE;GACrC,KAAK,IAAI,KAAK,GAAG,+BAA+B,UAAU,UAAU,6BAA6B,GAAG;GACpG;GACA;GACA;GACA;EACF,CAAC,CAAC,KAAK,IACW;EAAG,MAAM;GAAE;GAAM,MAAM,aAAa,GAAG,GAAG,IAAI,IAAI,EAAE;EAAE;CAAE;AAC5E;;;AAIA,MAAM,mBAAkH;CACtH;EAAE,IAAI;EAAW,MAAM;EAA0B,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CACrF;EAAE,IAAI;EAAS,MAAM;EAA0B,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CACnF;EAAE,IAAI;EAAY,MAAM;EAAgC,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CAC5F;EAAE,IAAI;EAAQ,MAAM;EAAoB,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CAC5E;EAAE,IAAI;EAAa,MAAM;EAAqB,GAAG;EAAK,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CACnF;EAAE,IAAI;EAAU,MAAM;EAA0C,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CACpG;EAAE,IAAI;EAAa,MAAM;EAAyB,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CACtF;EAAE,IAAI;EAAmB,MAAM;EAA0B,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CAC7F;EAAE,IAAI;EAAa,MAAM;EAAkC,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;CAC/F;EAAE,IAAI;EAAc,MAAM;EAAwC,GAAG;EAAI,GAAG;EAAI,IAAI;EAAI,IAAI;EAAG,IAAI;CAAG;AACxG;;;AAIA,MAAM,mBAAmB;;;;;;EAMvB,KAAK;;;AAIP,MAAM,cAAc;;;;;;;EAOlB,KAAK;AAEP,MAAM,QAAkE;CACtE;EACE,IAAI;EACJ,QAAQ;GACN;GACA;GACA;GACA;GACA;GACA;GACA;GACA;EACF,CAAC,CAAC,KAAK,IAAI;EACX,MAAM;GACJ,MAAM;IACJ,SAAS,CAAC,GAAG,EAAE;IACf,MAAM;KAAE,GAAG,CAAC,IAAI,GAAG;KAAG,GAAG,CAAC,IAAI,EAAE;KAAG,GAAG,CAAC,IAAI,GAAG;IAAE;IAChD,cAAc;IACd,aAAa;IACb,aAAa;GACf;GACA,MAAM;;;;;;;;;;;EAWV,KAAK;EACH;CACF;CACA,GAAG,iBAAiB,IAAI,YAAY;CACpC;EACE,IAAI;EACJ,QAAQ;GACN;GACA;GACA;GACA;GACA;GACA;EACF,CAAC,CAAC,KAAK,IAAI;EACX,MAAM;GACJ,MAAM;IACJ,SAAS,CAAC,GAAG,CAAC;IACd,MAAM;KAAE,GAAG,CAAC,IAAI,EAAE;KAAG,GAAG,CAAC,IAAI,EAAE;KAAG,GAAG,CAAC,IAAI,GAAG;IAAE;IAC/C,cAAc;IACd,aAAa;IACb,aAAa;GACf;GACA,MAAM;EACR;CACF;CACA;EACE,IAAI;EACJ,QAAQ;GACN;GACA;GACA;GACA;GACA;GACA;GACA;EACF,CAAC,CAAC,KAAK,IAAI;EACX,MAAM;GACJ,MAAM;IACJ,SAAS,CAAC,GAAG,CAAC;IACd,MAAM;KAAE,GAAG,CAAC,IAAI,EAAE;KAAG,GAAG,CAAC,IAAI,EAAE;KAAG,GAAG,CAAC,IAAI,EAAE;IAAE;IAC9C,cAAc;IACd,aAAa;IACb,aAAa;GACf;GACA,MAAM;EACR;CACF;CACA;EACE,IAAI;EACJ,QAAQ;GACN;GACA;GACA;GACA;GACA;GACA;EACF,CAAC,CAAC,KAAK,IAAI;EACX,MAAM;GACJ,MAAM;IACJ,SAAS,CAAC,GAAG,CAAC;IACd,MAAM;KAAE,GAAG,CAAC,IAAI,EAAE;KAAG,GAAG,CAAC,IAAI,EAAE;KAAG,GAAG,CAAC,IAAI,EAAE;IAAE;IAC9C,cAAc;IACd,aAAa;GACf;GACA,MAAM;;;;;;EAMV,KAAK;EACH;CACF;AACF;AAEA,SAAgB,yBAA2C;CACzD,OAAO;EACL,MAAM;EAEN,MAAM,YAAY;GAChB,IAAI;IACF,MAAM,cAAc,YAAY;KAAC;KAAM;KAAY;IAAW,GAAG,EAAE,SAAS,IAAO,CAAC;GACtF,SAAS,KAAK;IACZ,MAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC3D,MAAM,IAAI,MACR,gCAAgC,IAAI,6KAGtC;GACF;EACF;EAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;GACtC,IAAI,QAAQ;GACZ,IAAI,KAAK,KAAK,QAAQ,MAAM,QAAQ,MAAM,KAAK,IAAK,SAAS,EAAE,EAAE,CAAC;GAClE,IAAI,KAAK,SAAS,MAAM,QAAQ,MAAM,MAAM,GAAG,KAAK,KAAK;GACzD,OAAO,MAAM,KAAK,OAAkB;IAAE,IAAI,EAAE;IAAI,QAAQ,EAAE;IAAQ,UAAU,EAAE;GAA2C,EAAE;EAC7H;EAEA,MAAM,aAAa,MAAiB;GAElC,OADa,KAAK,UACL;EACf;EAEA,MAAM,MAAM,MAAiB,UAAuC;GAClE,MAAM,OAAQ,KAAK,SAAoC;GACvD,MAAM,MAAM,SAAS,KAAK;GAC1B,IAAI,CAAC,KAAK,OAAO;IAAE,UAAU;IAAO,OAAO;IAAG,QAAQ;GAAiB;GAEvE,MAAM,MAAM,MAAM,QAAQ,KAAK,OAAO,GAAG,MAAM,CAAC;GAChD,MAAM,OAAO,KAAK,KAAK,YAAY;GACnC,MAAM,UAAU,KAAK,KAAK,WAAW;GACrC,MAAM,UAAU,KAAK,KAAK,WAAW;GACrC,MAAM,UAAU,MAAM,GAAG;GAGzB,IAAI,SAAS;GACb,IAAI;IAEF,UAAS,MADO,SAAS;KAAC;KAAM;KAAS;IAAI,GAAG,GAAG,EAAA,CACxC;GACb,SAAS,KAAK;IAEZ,OAAO;KAAE,UAAU;KAAO,OAAO;KAAG,QAAQ,oBADhC,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAAA,CACQ,MAAM,GAAG,GAAG;IAAI;GACrF;GAGA,IAAI;IACF,MAAM,SACJ;KAAC;KAAM;KAAS;KAAsB;KAAiC;KAA0B;IAAI,GACrG,GACF;GACF,QAAQ,CAER;GAEA,IAAI;GACJ,IAAI;IACF,MAAM,MAAM,SAAS,SAAS,MAAM;GACtC,QAAQ;IACN,OAAO;KAAE,UAAU;KAAO,OAAO;KAAG,QAAQ;IAAgD;GAC9F;GACA,MAAM,UAAU,OAAO,mBAAmB,KAAK,MAAM,CAAC,GAAG,MAAM,GAAG;GAClE,MAAM,MAAM,WAAW,KAAK,OAAO;GACnC,MAAM,EAAE,QAAQ,UAAU,cAAc,KAAK,IAAI;GAcjD,OAAO;IAAE,UAbQ,UAAU;IAaR;IAAO,QAZX,KAAK,UAAU;KAC5B;KACA,KAAK;MACH,WAAW,IAAI;MACf,SAAS,IAAI;MACb,MAAM;OAAE,GAAG,CAAC,IAAI,KAAK,EAAE,QAAQ,CAAC;OAAG,GAAG,CAAC,IAAI,KAAK,EAAE,QAAQ,CAAC;OAAG,GAAG,CAAC,IAAI,KAAK,EAAE,QAAQ,CAAC;MAAE;MACxF,aAAa,EAAE,IAAI,cAAc,IAAI,WAAA,CAAY,QAAQ,CAAC;MAC1D,oBAAoB,IAAI,gBAAgB,EAAE,IAAI,eAAe,IAAI,cAAA,CAAe,QAAQ,CAAC,IAAI;KAC/F;KACA,YAAY;KACZ;IACF,CAC+B;GAAE;EACnC;CACF;AACF"}

package/dist/benchmarks/cadbench.d.ts

@@ -1,19 +1,6 @@
-import { BenchmarkAdapter } from './types.js';
-import '@tangle-network/agent-runtime/loops';
-
-/**
- * CADBench / BlenderLLM adapter (FreedomIntelligence/CADBench, arXiv:2412.14203).
- * Task = NL instruction → a Blender `bpy` script. Score = the paper's criteria
- * eval: render the produced model to standardized views, then a vision judge
- * (GPT-4o-class) marks each per-task criterion bullet pass/fail against the
- * rendered images + the script text. score = fraction of criteria satisfied.
- *
- * Data: the published dataset's `criteria` flattened to a bullet list (700 tasks,
- * 500 Simulative + 200 Wild). Point CADBENCH_PATH at the cleaned JSONL
- * ({id,name,instruction,type,criteria:string[]} per line). Judge creds from
- * TANGLE_API_KEY / ROUTER_BASE / JUDGE_MODEL (default gpt-4o).
- */
-
+import { BenchmarkAdapter } from "./types.js";
+//#region src/benchmarks/cadbench.d.ts
 declare function createCadBenchAdapter(): BenchmarkAdapter;
-
+//#endregion
 export { createCadBenchAdapter };
+//# sourceMappingURL=cadbench.d.ts.map

package/dist/benchmarks/cadbench.js

@@ -1,8 +1,2 @@
-import {
-  createCadBenchAdapter
-} from "../chunk-PUIRNYI7.js";
-import "../chunk-JTHWEDEW.js";
-export {
-  createCadBenchAdapter
-};
-//# sourceMappingURL=cadbench.js.map
+import { t as createCadBenchAdapter } from "../cadbench-DpQWZHp4.js";
+export { createCadBenchAdapter };

package/dist/benchmarks/cadgenbench.d.ts

@@ -1,22 +1,6 @@
-import { BenchmarkAdapter } from './types.js';
-import '@tangle-network/agent-runtime/loops';
-
-/**
- * CADGenBench adapter (huggingface/cadgenbench, Apache-2.0). Task = a part
- * description → a STEP B-rep solid (output.step). Score = the benchmark's OWN
- * deterministic geometric metric (cad_score): validity gate → PCA/ICP align to
- * the ground truth → point-cloud F1 + volume IoU + edge F1 + topology match.
- * NOT an LLM judge, NOT self-defined checks — the published CAD kernel decides.
- *
- * The official task set (private GT, server-side graded) isn't released yet, so
- * tasks here are seeded from the repo's dimension-named geometry fixtures (real
- * GT STEPs scored by the real scorer). When CADGENBENCH_DATA_DIR is set, swap
- * loadTasks to read the published fixtures' description.yaml + ground_truth.step.
- *
- * Requires the CADGenBench venv (CADGENBENCH_VENV) + clone (CADGENBENCH_DIR) +
- * xvfb (the scorer's alignment renders need a display).
- */
-
+import { BenchmarkAdapter } from "./types.js";
+//#region src/benchmarks/cadgenbench.d.ts
 declare function createCadGenBenchAdapter(): BenchmarkAdapter;
-
+//#endregion
 export { createCadGenBenchAdapter };
+//# sourceMappingURL=cadgenbench.d.ts.map

package/dist/benchmarks/cadgenbench.js

@@ -1,8 +1,2 @@
-import {
-  createCadGenBenchAdapter
-} from "../chunk-C7T7WEK2.js";
-import "../chunk-JTHWEDEW.js";
-export {
-  createCadGenBenchAdapter
-};
-//# sourceMappingURL=cadgenbench.js.map
+import { t as createCadGenBenchAdapter } from "../cadgenbench-DRhczfsG.js";
+export { createCadGenBenchAdapter };