@tangle-network/agent-bench 0.7.1 → 0.8.1

package/src/benchmarks/cadbench.ts

@@ -14,6 +14,7 @@
 import { readFile } from 'node:fs/promises'
 import type { BenchScore, BenchTask, BenchmarkAdapter, LoadOptions } from './types'
 import { renderBpy } from '../worker-blender'
+import { runBenchRouterTurn } from '../router-turn'
 
 interface CadBenchMeta {
   name: string
@@ -48,14 +49,26 @@ async function judgeCriteria(
     `Return ONLY a JSON array of exactly ${criteria.length} booleans (true=satisfied, false=not), in order, no prose.\n\nCRITERIA:\n${numbered}\n\nSCRIPT:\n\`\`\`python\n${script.slice(0, 6000)}\n\`\`\``
   const content: unknown[] = [{ type: 'text', text }]
   for (const url of renders) content.push({ type: 'image_url', image_url: { url } })
-  const res = await fetch(`${base}/chat/completions`, {
-    method: 'POST',
-    headers: { 'content-type': 'application/json', authorization: `Bearer ${key}` },
-    body: JSON.stringify({ model, max_tokens: 1500, temperature: 0, messages: [{ role: 'user', content }] }),
-  })
-  if (!res.ok) throw new Error(`judge ${model} ${res.status}: ${(await res.text()).slice(0, 200)}`)
-  const data = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }
-  const raw = data.choices?.[0]?.message?.content ?? ''
+  const turn = await runBenchRouterTurn(
+    {
+      routerBaseUrl: base,
+      routerKey: key,
+      profile: {
+        name: 'cadbench-vision-judge',
+        harness: 'cli-base',
+        model: {
+          provider: 'tangle-router',
+          default: model,
+          metadata: {
+            temperature: 0,
+            maxTokens: Number(process.env.JUDGE_MAX_TOKENS ?? 1500),
+          },
+        },
+      },
+    },
+    { messages: [{ role: 'user', content }] },
+  )
+  const raw = turn.finalText
   const m = /\[\s*(?:true|false)[\s\S]*?\]/i.exec(raw)
   if (!m) return { passed: criteria.map(() => false), note: `judge returned no parseable verdict: ${raw.slice(0, 80)}` }
   let arr: unknown

package/src/benchmarks/finresearchbench.ts

@@ -9,6 +9,7 @@
 
 import { readFile, stat } from 'node:fs/promises'
 import { join } from 'node:path'
+import { runBenchRouterTurn } from '../router-turn'
 import { benchRoot } from './_harness'
 import type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'
 
@@ -180,22 +181,25 @@ function parseJudgeScore(content: string): { score: number; raw: unknown } {
 async function runOfficialJudge(meta: FinResearchMeta, response: string): Promise<BenchScore> {
   if (!meta.judgeSystemPrompt) throw new Error(`FinResearchBench task ${meta.id} missing judge_system_prompt`)
   const router = routerConfig()
-  const res = await fetch(`${router.baseUrl}/chat/completions`, {
-    method: 'POST',
-    headers: { 'content-type': 'application/json', authorization: `Bearer ${router.key}` },
-    body: JSON.stringify({
-      model: router.model,
-      temperature: 0,
-      messages: [
-        { role: 'system', content: meta.judgeSystemPrompt },
-        { role: 'user', content: fillTemplate(meta, response) },
-      ],
-    }),
-  })
-  if (!res.ok) throw new Error(`FinResearchBench judge HTTP ${res.status}: ${(await res.text()).slice(0, 300)}`)
-  const body = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }
-  const content = body.choices?.[0]?.message?.content
-  if (typeof content !== 'string') throw new Error(`FinResearchBench judge returned no message content: ${JSON.stringify(body).slice(0, 300)}`)
+  const turn = await runBenchRouterTurn(
+    {
+      routerBaseUrl: router.baseUrl,
+      routerKey: router.key,
+      profile: {
+        name: 'finresearchbench-judge',
+        harness: 'cli-base',
+        model: {
+          provider: 'tangle-router',
+          default: router.model,
+          metadata: { temperature: 0 },
+        },
+        prompt: { systemPrompt: meta.judgeSystemPrompt },
+      },
+    },
+    fillTemplate(meta, response),
+  )
+  const content = turn.finalText
+  if (!content) throw new Error('FinResearchBench judge returned no message content')
   const { score, raw } = parseJudgeScore(content)
   return {
     resolved: score >= Number(process.env.FINRESEARCHBENCH_PASS_THRESHOLD ?? 0.8),

package/src/benchmarks/finsearchcomp.ts

@@ -32,6 +32,7 @@
 import { readFile } from 'node:fs/promises'
 import { join } from 'node:path'
 import { fileURLToPath } from 'node:url'
+import { runBenchRouterTurn } from '../router-turn'
 import type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'
 
 const BENCH_ROOT = fileURLToPath(new URL('../..', import.meta.url))
@@ -206,26 +207,25 @@ function parseJudgeOutput(content: string): { resolved: boolean; score: number;
 
 /** Run the record's own judge via the router. Fail loud on transport/parse errors. */
 async function runRecordJudge(meta: FinSearchMeta, response: string, router: JudgeRouter): Promise<BenchScore> {
-  const res = await fetch(`${router.baseUrl}/chat/completions`, {
-    method: 'POST',
-    headers: { 'content-type': 'application/json', authorization: `Bearer ${router.key}` },
-    body: JSON.stringify({
-      model: router.model,
-      temperature: 0,
-      messages: [
-        { role: 'system', content: meta.judgeSystemPrompt },
-        { role: 'user', content: fillJudgePrompt(meta, response) },
-      ],
-    }),
-  })
-  if (!res.ok) {
-    throw new Error(`FinSearchComp judge HTTP ${res.status}: ${(await res.text()).slice(0, 300)}`)
-  }
-  const body = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }
-  const content = body.choices?.[0]?.message?.content
-  if (typeof content !== 'string') {
-    throw new Error(`FinSearchComp judge returned no message content: ${JSON.stringify(body).slice(0, 300)}`)
-  }
+  const turn = await runBenchRouterTurn(
+    {
+      routerBaseUrl: router.baseUrl,
+      routerKey: router.key,
+      profile: {
+        name: 'finsearchcomp-judge',
+        harness: 'cli-base',
+        model: {
+          provider: 'tangle-router',
+          default: router.model,
+          metadata: { temperature: 0 },
+        },
+        prompt: { systemPrompt: meta.judgeSystemPrompt },
+      },
+    },
+    fillJudgePrompt(meta, response),
+  )
+  const content = turn.finalText
+  if (!content) throw new Error('FinSearchComp judge returned no message content')
   const { resolved, score, raw } = parseJudgeOutput(content)
   return {
     resolved,

package/src/benchmarks/frames.ts

@@ -25,6 +25,7 @@ import { readFile } from 'node:fs/promises'
 import { join } from 'node:path'
 import { fileURLToPath } from 'node:url'
 import { promisify } from 'node:util'
+import { runBenchRouterTurn } from '../router-turn'
 import type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'
 
 const execFileAsync = promisify(execFile)
@@ -322,24 +323,24 @@ async function tier2Judge(
   candidate: string,
   router: JudgeRouter,
 ): Promise<boolean> {
-  const res = await fetch(`${router.baseUrl}/chat/completions`, {
-    method: 'POST',
-    headers: { 'content-type': 'application/json', authorization: `Bearer ${router.key}` },
-    body: JSON.stringify({
-      model: router.model,
-      temperature: 0,
-      seed: 0,
-      messages: [{ role: 'user', content: JUDGE_PROMPT(question, gold, candidate) }],
-    }),
-  })
-  if (!res.ok) {
-    throw new Error(`FRAMES Tier-2 judge HTTP ${res.status}: ${(await res.text()).slice(0, 300)}`)
-  }
-  const body = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }
-  const content = body.choices?.[0]?.message?.content
-  if (typeof content !== 'string') {
-    throw new Error(`FRAMES Tier-2 judge returned no message content: ${JSON.stringify(body).slice(0, 300)}`)
-  }
+  const turn = await runBenchRouterTurn(
+    {
+      routerBaseUrl: router.baseUrl,
+      routerKey: router.key,
+      profile: {
+        name: 'frames-equivalence-judge',
+        harness: 'cli-base',
+        model: {
+          provider: 'tangle-router',
+          default: router.model,
+          metadata: { temperature: 0, seed: 0 },
+        },
+      },
+    },
+    JUDGE_PROMPT(question, gold, candidate),
+  )
+  const content = turn.finalText
+  if (!content) throw new Error('FRAMES Tier-2 judge returned no message content')
   const fenced = content.match(/```(?:json)?\s*([\s\S]*?)```/)
   const raw = (fenced ? fenced[1] : content)?.trim() ?? ''
   let parsed: { verdict?: unknown }

package/src/benchmarks/mcad-bench.test.mts

@@ -0,0 +1,455 @@
+/**
+ * MCAD adapter tests, in two halves.
+ *
+ * PURE: hand-written ASCII STL fixtures exercise the geometry engine with no
+ * OpenSCAD anywhere — a unit cube, an open box, two disjoint cubes, and a cube
+ * with a square through-channel. The fixtures are deliberately lattice-aligned,
+ * because that is the case a naive axis-aligned ray-parity test gets wrong.
+ *
+ * LIVE (gated on `openscad` + `xvfb-run`): the real judge, in BOTH directions.
+ * Accept — every `calibrated: true` task scores 1.0 on its own gold. Reject — the
+ * task-01 gold with its four holes deleted must fail the bore probes, and the same
+ * gold scaled to 90 x 54 x 18 must fail all three bbox axes. A judge that has only
+ * been shown passing inputs is not calibrated, so the reject direction is not
+ * optional here.
+ *
+ *   npx vitest run src/benchmarks/mcad-bench.test.mts
+ */
+
+import { execFileSync } from 'node:child_process'
+import { describe, expect, it } from 'vitest'
+import {
+  createMcadBenchAdapter,
+  measureMesh,
+  parseAsciiStl,
+  pointInSolid,
+  scoreAgainstSpec,
+  stripCodeFence,
+} from './mcad-bench'
+import { MCAD_GOLDS } from './mcad-golds'
+import { MCAD_TASKS } from './mcad-tasks'
+
+// ---------------------------------------------------------------------------
+// ASCII STL fixture builders (outward-facing winding, 2-manifold by construction)
+// ---------------------------------------------------------------------------
+
+type P = readonly [number, number, number]
+type T = readonly [P, P, P]
+
+/** A quad a-b-c-d as two triangles sharing the a-c diagonal. */
+function quad(a: P, b: P, c: P, d: P): T[] {
+  return [
+    [a, b, c],
+    [a, c, d],
+  ]
+}
+
+function boxTris(min: P, max: P): T[] {
+  const [x0, y0, z0] = min
+  const [x1, y1, z1] = max
+  const p000: P = [x0, y0, z0]
+  const p100: P = [x1, y0, z0]
+  const p110: P = [x1, y1, z0]
+  const p010: P = [x0, y1, z0]
+  const p001: P = [x0, y0, z1]
+  const p101: P = [x1, y0, z1]
+  const p111: P = [x1, y1, z1]
+  const p011: P = [x0, y1, z1]
+  return [
+    ...quad(p000, p010, p110, p100), // bottom, -Z
+    ...quad(p001, p101, p111, p011), // top, +Z
+    ...quad(p000, p100, p101, p001), // front, -Y
+    ...quad(p010, p011, p111, p110), // back, +Y
+    ...quad(p000, p001, p011, p010), // left, -X
+    ...quad(p100, p110, p111, p101), // right, +X
+  ]
+}
+
+function toStl(tris: T[]): string {
+  const facets = tris
+    .map(([a, b, c]) =>
+      [
+        '  facet normal 0 0 0',
+        '    outer loop',
+        `      vertex ${a.join(' ')}`,
+        `      vertex ${b.join(' ')}`,
+        `      vertex ${c.join(' ')}`,
+        '    endloop',
+        '  endfacet',
+      ].join('\n'),
+    )
+    .join('\n')
+  return `solid fixture\n${facets}\nendsolid fixture\n`
+}
+
+const UNIT_CUBE = toStl(boxTris([0, 0, 0], [1, 1, 1]))
+/** The unit cube with its two +Z triangles deleted — four edges left unpaired. */
+const OPEN_BOX = toStl(boxTris([0, 0, 0], [1, 1, 1]).filter((_, i) => i !== 2 && i !== 3))
+const TWO_CUBES = toStl([...boxTris([0, 0, 0], [1, 1, 1]), ...boxTris([3, 0, 0], [4, 1, 1])])
+
+/**
+ * A 6 x 6 x 6 block with a 2 x 2 square channel running through it in Z. Both end
+ * faces are annuli split into four trapezoids whose corner diagonals are shared, so
+ * the mesh stays 2-manifold. Enclosed volume: 216 - 24 = 192.
+ */
+function channelCubeTris(): T[] {
+  const o = 6
+  const a = 2
+  const b = 4
+  const tris: T[] = []
+  // outer side walls
+  const outer: Array<[P, P]> = [
+    [[0, 0, 0], [o, 0, 0]],
+    [[o, 0, 0], [o, o, 0]],
+    [[o, o, 0], [0, o, 0]],
+    [[0, o, 0], [0, 0, 0]],
+  ]
+  for (const [s, e] of outer) {
+    tris.push(...quad(s, e, [e[0], e[1], o], [s[0], s[1], o]))
+  }
+  // end-face annuli: four trapezoids per face, outer edge -> inner edge
+  const strips: Array<[P, P, P, P]> = [
+    [[0, 0, 0], [o, 0, 0], [b, a, 0], [a, a, 0]],
+    [[o, 0, 0], [o, o, 0], [b, b, 0], [b, a, 0]],
+    [[o, o, 0], [0, o, 0], [a, b, 0], [b, b, 0]],
+    [[0, o, 0], [0, 0, 0], [a, a, 0], [a, b, 0]],
+  ]
+  for (const [p, q, r, s] of strips) {
+    const up = (v: P): P => [v[0], v[1], o]
+    tris.push(...quad(up(p), up(q), up(r), up(s))) // +Z face
+    tris.push(...quad(s, r, q, p)) // -Z face, reversed winding
+  }
+  // channel walls, outward normals point INTO the void
+  tris.push(...quad([a, a, 0], [a, b, 0], [a, b, o], [a, a, o])) // +X
+  tris.push(...quad([b, a, 0], [b, a, o], [b, b, o], [b, b, 0])) // -X
+  tris.push(...quad([a, a, 0], [a, a, o], [b, a, o], [b, a, 0])) // +Y
+  tris.push(...quad([a, b, 0], [b, b, 0], [b, b, o], [a, b, o])) // -Y
+  return tris
+}
+
+const CHANNEL_CUBE = toStl(channelCubeTris())
+
+// ---------------------------------------------------------------------------
+// Pure geometry tests
+// ---------------------------------------------------------------------------
+
+describe('mcad geometry engine (pure ASCII STL fixtures)', () => {
+  it('measures a unit cube: watertight, one solid, volume 1', () => {
+    const g = measureMesh(parseAsciiStl(UNIT_CUBE))
+    expect(g.triangles).toBe(12)
+    expect(g.watertight).toBe(true)
+    expect(g.openEdges).toBe(0)
+    expect(g.degenerateFaces).toBe(0)
+    expect(g.solids).toBe(1)
+    expect(g.volume).toBeCloseTo(1, 9)
+    expect(g.bbox.size).toEqual({ x: 1, y: 1, z: 1 })
+  })
+
+  it('rejects an open box: four unpaired edges, not watertight', () => {
+    const g = measureMesh(parseAsciiStl(OPEN_BOX))
+    expect(g.triangles).toBe(10)
+    expect(g.watertight).toBe(false)
+    expect(g.openEdges).toBe(4)
+  })
+
+  it('counts two disjoint cubes as two solids', () => {
+    const g = measureMesh(parseAsciiStl(TWO_CUBES))
+    expect(g.watertight).toBe(true)
+    expect(g.solids).toBe(2)
+    expect(g.volume).toBeCloseTo(2, 9)
+    expect(g.bbox.size.x).toBe(4)
+  })
+
+  it('measures the through-channel cube: one solid, volume 192', () => {
+    const g = measureMesh(parseAsciiStl(CHANNEL_CUBE))
+    expect(g.watertight).toBe(true)
+    expect(g.openEdges).toBe(0)
+    expect(g.solids).toBe(1)
+    expect(g.degenerateFaces).toBe(0)
+    expect(g.volume).toBeCloseTo(192, 6)
+  })
+
+  it('returns an empty measurement for a triangle-free STL', () => {
+    const g = measureMesh(parseAsciiStl('solid empty\nendsolid empty\n'))
+    expect(g.triangles).toBe(0)
+    expect(g.watertight).toBe(false)
+  })
+
+  it('places probes correctly in the through-channel cube', () => {
+    const tris = parseAsciiStl(CHANNEL_CUBE)
+    // channel centre is void; the surrounding material is solid
+    expect(pointInSolid(tris, { x: 3, y: 3, z: 3 })).toBe(false)
+    expect(pointInSolid(tris, { x: 1, y: 1, z: 3 })).toBe(true)
+    expect(pointInSolid(tris, { x: 5, y: 5, z: 3 })).toBe(true)
+    expect(pointInSolid(tris, { x: 1, y: 5, z: 3 })).toBe(true)
+    // outside the block entirely, including straight up the channel
+    expect(pointInSolid(tris, { x: 3, y: 3, z: 9 })).toBe(false)
+    expect(pointInSolid(tris, { x: -1, y: 3, z: 3 })).toBe(false)
+  })
+})
+
+describe('ray-parity robustness on lattice-aligned probes', () => {
+  const cube = parseAsciiStl(toStl(boxTris([0, 0, 0], [2, 2, 2])))
+  const channel = parseAsciiStl(CHANNEL_CUBE)
+
+  it('answers correctly for points sitting on the cube edge lines', () => {
+    // Every one of these shares two coordinates with a cube corner, so an
+    // axis-aligned ray from it runs exactly along an edge or inside a face plane.
+    expect(pointInSolid(cube, { x: 3, y: 0, z: 0 })).toBe(false)
+    expect(pointInSolid(cube, { x: 3, y: 2, z: 2 })).toBe(false)
+    expect(pointInSolid(cube, { x: -1, y: 0, z: 2 })).toBe(false)
+    expect(pointInSolid(cube, { x: 1, y: 3, z: 0 })).toBe(false)
+    expect(pointInSolid(cube, { x: 1, y: 1, z: 1 })).toBe(true)
+  })
+
+  it('answers correctly for interior points lying in a channel-wall plane', () => {
+    // y = 2 and x = 2 are the channel wall planes; these points are in material,
+    // and an axis-aligned ray from each would graze the channel wall's edge.
+    expect(pointInSolid(channel, { x: 1, y: 2, z: 3 })).toBe(true)
+    expect(pointInSolid(channel, { x: 2, y: 1, z: 3 })).toBe(true)
+    expect(pointInSolid(channel, { x: 5, y: 2, z: 3 })).toBe(true)
+    expect(pointInSolid(channel, { x: 4, y: 5, z: 3 })).toBe(true)
+    // and for outside points on the same degenerate lines
+    expect(pointInSolid(channel, { x: 7, y: 2, z: 3 })).toBe(false)
+    expect(pointInSolid(channel, { x: 3, y: 2, z: 7 })).toBe(false)
+  })
+
+  it('is deterministic: the same point and mesh always give the same answer', () => {
+    const pts = [
+      { x: 3, y: 3, z: 3 },
+      { x: 1, y: 2, z: 3 },
+      { x: 2, y: 1, z: 3 },
+    ]
+    for (const p of pts) {
+      const first = pointInSolid(channel, p)
+      for (let i = 0; i < 5; i++) expect(pointInSolid(channel, p)).toBe(first)
+    }
+  })
+})
+
+describe('spec scoring', () => {
+  const tris = parseAsciiStl(CHANNEL_CUBE)
+  const geo = measureMesh(tris)
+
+  it('scores 1.0 and resolves when every asserted check passes', () => {
+    const s = scoreAgainstSpec(geo, tris, {
+      bbox: { x: [5.9, 6.1], y: [5.9, 6.1], z: [5.9, 6.1] },
+      volume: [190, 194],
+      solids: [1, 1],
+      minTriangles: 20,
+      probesInsideSolid: [[1, 1, 3]],
+      probesOutsideSolid: [[3, 3, 3]],
+    })
+    expect(s.score).toBe(1)
+    expect(s.resolved).toBe(true)
+    expect(s.failed).toEqual([])
+  })
+
+  it('gives partial credit and names every failed check with measured vs expected', () => {
+    const s = scoreAgainstSpec(geo, tris, {
+      bbox: { x: [10, 12] },
+      volume: [190, 194],
+      solids: [1, 1],
+      probesOutsideSolid: [[1, 1, 3]],
+    })
+    expect(s.resolved).toBe(false)
+    expect(s.score).toBeCloseTo(2 / 4, 9)
+    const names = s.failed.map((c) => c.name)
+    expect(names).toContain('bboxX')
+    expect(names.some((n) => n.startsWith('probeOutside[0]'))).toBe(true)
+    const bbox = s.failed.find((c) => c.name === 'bboxX')
+    expect(bbox?.measured).toBe('6')
+    expect(bbox?.expected).toBe('[10, 12]')
+  })
+
+  it('counts each bbox axis, each band and each probe as one check', () => {
+    const s = scoreAgainstSpec(geo, tris, {
+      bbox: { x: [5, 7], y: [5, 7], z: [5, 7] },
+      volume: [190, 194],
+      solids: [1, 1],
+      minTriangles: 1,
+      probesInsideSolid: [
+        [1, 1, 3],
+        [5, 5, 3],
+      ],
+      probesOutsideSolid: [[3, 3, 3]],
+    })
+    expect(s.checks).toHaveLength(9)
+  })
+})
+
+describe('artifact handling', () => {
+  it('strips a markdown fence the model may have wrapped its answer in', () => {
+    expect(stripCodeFence('```openscad\ncube([1,1,1]);\n```')).toBe('cube([1,1,1]);')
+    expect(stripCodeFence('```\ncube([1,1,1]);\n```')).toBe('cube([1,1,1]);')
+    expect(stripCodeFence('  cube([1,1,1]);  ')).toBe('cube([1,1,1]);')
+  })
+
+  it('loads tasks with spec/source/calibrated metadata and no openscad', async () => {
+    const a = createMcadBenchAdapter()
+    const tasks = await a.loadTasks()
+    expect(tasks).toHaveLength(MCAD_TASKS.length)
+    for (const t of tasks) {
+      const md = t.metadata as { spec: unknown; source: string; calibrated: boolean }
+      expect(md.spec).toBeTruthy()
+      expect(typeof md.source).toBe('string')
+      expect(typeof md.calibrated).toBe('boolean')
+      expect(typeof (await a.goldArtifact(t))).toBe('string')
+    }
+  })
+
+  it('honours ids and limit', async () => {
+    const a = createMcadBenchAdapter()
+    expect((await a.loadTasks({ ids: ['l-bracket'] })).map((t) => t.id)).toEqual(['l-bracket'])
+    expect(await a.loadTasks({ limit: 3 })).toHaveLength(3)
+  })
+
+  it('task 10 asks for separate bodies, never "one fused solid"', () => {
+    const t10 = MCAD_TASKS.find((t) => t.id === 'planetary-gear-stage')
+    expect(t10?.prompt).not.toContain('one fused solid')
+    expect(t10?.prompt).toContain('separate')
+    // every other task keeps the single-solid boilerplate
+    for (const t of MCAD_TASKS) {
+      if (t.id === 'planetary-gear-stage') continue
+      expect(t.prompt).toContain('one fused solid')
+    }
+  })
+
+  it('scores an empty artifact 0 without touching openscad', async () => {
+    const a = createMcadBenchAdapter()
+    const [t] = await a.loadTasks({ ids: ['calibration-block'] })
+    const s = await a.judge(t!, '   ')
+    expect(s.resolved).toBe(false)
+    expect(s.score).toBe(0)
+    expect(s.detail).toBe('empty artifact')
+  })
+})
+
+// ---------------------------------------------------------------------------
+// Live judge (needs openscad + xvfb-run)
+// ---------------------------------------------------------------------------
+
+function openscadPresent(): boolean {
+  try {
+    execFileSync('xvfb-run', ['-a', 'openscad', '--version'], { stdio: 'ignore', timeout: 60_000 })
+    return true
+  } catch {
+    return false
+  }
+}
+
+const OPENSCAD = openscadPresent()
+const JUDGE_TIMEOUT = 300_000
+
+/** Task 01's gold with the four through-holes deleted — the geometry is otherwise identical. */
+const BLOCK_WITHOUT_HOLES = `
+$fn=96;
+module chamfered_block(l, w, h, c) {
+  hull() {
+    translate([-l/2, -w/2, 0]) cube([l, w, h - c]);
+    translate([-l/2 + c, -w/2 + c, h - c]) cube([l - 2*c, w - 2*c, c]);
+  }
+}
+chamfered_block(100, 60, 20, 2);
+`
+
+/** Task 01's gold at 0.9 scale: 90 x 54 x 18, every feature otherwise correct.
+ *  Wrapped in a module rather than a bare `scale(){...}` block because OpenSCAD
+ *  rejects a module definition inside a transform's child block but accepts one
+ *  inside a module body — this keeps the fixture derived from the gold itself. */
+const BLOCK_SCALED_90 = `module gold() {\n${MCAD_GOLDS['calibration-block']}\n}\nscale([0.9, 0.9, 0.9]) gold();`
+
+describe.skipIf(!OPENSCAD)('mcad judge, live openscad', () => {
+  it('preflight passes when the toolchain is present', async () => {
+    await expect(createMcadBenchAdapter().preflight()).resolves.toBeUndefined()
+  })
+
+  for (const task of MCAD_TASKS.filter((t) => t.calibrated)) {
+    it(
+      `gold for ${task.id} resolves at score 1.0`,
+      async () => {
+        const a = createMcadBenchAdapter()
+        const [t] = await a.loadTasks({ ids: [task.id] })
+        const gold = await a.goldArtifact(t!)
+        expect(typeof gold).toBe('string')
+        const s = await a.judge(t!, gold as string)
+        const failed = JSON.parse(s.detail ?? '{}').failed ?? []
+        expect(failed, `failed checks: ${JSON.stringify(failed)}`).toEqual([])
+        expect(s.score).toBe(1)
+        expect(s.resolved).toBe(true)
+      },
+      JUDGE_TIMEOUT,
+    )
+  }
+
+  it(
+    'MUST REJECT: the calibration block with its four holes removed fails the bore probes',
+    async () => {
+      const a = createMcadBenchAdapter()
+      const [t] = await a.loadTasks({ ids: ['calibration-block'] })
+      const s = await a.judge(t!, BLOCK_WITHOUT_HOLES)
+      expect(s.resolved).toBe(false)
+      expect(s.score).toBeLessThan(1)
+      const failed = (JSON.parse(s.detail as string).failed as Array<{ name: string; measured: string }>).map(
+        (f) => f.name,
+      )
+      // all five "must be outside" probes now sit in solid material
+      const outside = failed.filter((n) => n.startsWith('probeOutside['))
+      const spec = MCAD_TASKS.find((x) => x.id === 'calibration-block')?.spec
+      expect(outside).toHaveLength(spec?.probesOutsideSolid?.length ?? 0)
+      expect(failed).toContain('volume')
+      // the block itself is still the right size, so bbox must NOT be blamed
+      expect(failed).not.toContain('bboxX')
+    },
+    JUDGE_TIMEOUT,
+  )
+
+  it(
+    'MUST REJECT: the calibration block scaled to 90 x 54 x 18 fails every bbox axis',
+    async () => {
+      const a = createMcadBenchAdapter()
+      const [t] = await a.loadTasks({ ids: ['calibration-block'] })
+      const s = await a.judge(t!, BLOCK_SCALED_90)
+      expect(s.resolved).toBe(false)
+      const detail = JSON.parse(s.detail as string) as {
+        failed: Array<{ name: string; measured: string; expected: string }>
+        geo: { bbox: { x: number; y: number; z: number } }
+      }
+      expect(detail.geo.bbox).toEqual({ x: 90, y: 54, z: 18 })
+      const failed = detail.failed.map((f) => f.name)
+      expect(failed).toContain('bboxX')
+      expect(failed).toContain('bboxY')
+      expect(failed).toContain('bboxZ')
+      const x = detail.failed.find((f) => f.name === 'bboxX')
+      expect(x?.measured).toBe('90')
+      expect(x?.expected).toBe('[99, 101]')
+    },
+    JUDGE_TIMEOUT,
+  )
+
+  it(
+    'MUST REJECT: a non-watertight surface is scored 0 by the hard gate',
+    async () => {
+      const a = createMcadBenchAdapter()
+      const [t] = await a.loadTasks({ ids: ['calibration-block'] })
+      const s = await a.judge(t!, 'polyhedron(points=[[0,0,0],[10,0,0],[0,10,0]], faces=[[0,1,2]]);')
+      expect(s.resolved).toBe(false)
+      expect(s.score).toBe(0)
+      expect(JSON.parse(s.detail as string).gate).toBe('watertight')
+    },
+    JUDGE_TIMEOUT,
+  )
+
+  it(
+    'MUST REJECT: source that does not compile is scored 0',
+    async () => {
+      const a = createMcadBenchAdapter()
+      const [t] = await a.loadTasks({ ids: ['calibration-block'] })
+      const s = await a.judge(t!, 'this is not openscad {{{')
+      expect(s.resolved).toBe(false)
+      expect(s.score).toBe(0)
+      expect(s.detail).toMatch(/compile failed/)
+    },
+    JUDGE_TIMEOUT,
+  )
+})