CreatorBench v1.4: adaptive routing, sealed evidence, and review UI

.qa/memory/findings.jsonl

@@ -53,3 +53,7 @@
 {"id":"nodevideo-ui-topology-20260721","rootCause":"onboarding-operation-provider-proof-export-composed-as-one-card-dashboard","sev":"P0","symptom":"proof-dashboard-as-primary-product-instead-of-artifact-workspace","status":"fixed","area":"creator-product-topology","evidence":"harness/references/anti-references/nodevideo-proof-dashboard-as-product","fix":"guided-start+artifact-workspace+run-inspector+mobile-modes+topology-gate","ts":"2026-07-22T01:14:09.081Z","fp":"95a1b08805f3"}
 {"id":"NV-SR-001","sev":"P1","area":"creator responsive topology","symptom":"Artifact stage falls below majority width at tablet and xl boundary viewports","rootCause":"Three-column workspace stayed active below the width where both side rails could coexist with an artifact-dominant canvas","evidence":"GitHub Actions run 29885587329 and focused xl-boundary/tablet Playwright rerun","fix":"Switch tablet to mode-based surfaces at 1023px and compact rails from 1024px through 1350px","status":"fixed","ts":"2026-07-22T02:30:07.066Z","fp":"09e51987fd46"}
 {"id":"NV-ATLAS-001","sev":"P1","area":"tracking-atlas-responsive-topology","symptom":"The 834px tablet viewport retained the desktop navigation rail and overflowed horizontally","rootCause":"The Atlas single-column breakpoint was set to 820px, below the tested tablet width","evidence":".qa/runs/2026-07-21-tracking-atlas/pixels/tablet-dark.png and tablet-light.png before/after pixel rerun","fix":"Move the deliberate single-column navigation transition to 900px and assert document scrollWidth against innerWidth","status":"fixed","ts":"2026-07-22T03:24:00.000Z","fp":"nv-atlas-tablet-900"}
+{"id":"NV-CB-001","sev":"P1","area":"creatorbench-responsive-topology","symptom":"mobile-navigation-overflow","rootCause":"desktop-links-remained-unwrapped-at-mobile-breakpoint","evidence":"creatorbench-focused-desktop-mobile-playwright","fix":"compact-wrap-plus-overflow-assertions","status":"fixed","ts":"2026-07-22T04:51:57.719Z","fp":"7fa8eba8dce1"}
+{"id":"NV-CB-002","sev":"P0","area":"creatorbench-production-evidence-rendering","symptom":"real-null-cost-metric-crashed-public-dashboard","rootCause":"normalizer-preserved-null-as-numeric-value","evidence":"signed-in-vercel-preview-console-plus-12-pass-regression","fix":"normalize-all-optional-measured-numbers-through-finite-guard","status":"fixed","ts":"2026-07-22T04:59:24.159Z","fp":"c1fa451a2311"}
+{"id":"creatorbench-claim-honesty-v14","sev":"P0","area":"CreatorBench public claim honesty","symptom":"The v1.3 report described a 0.0% usable first-pass result even though zero private held-out instances had human editing-quality review.","rootCause":"derivePublicClaim emitted automatic and assisted usability language unconditionally from machine routing dispositions.","evidence":"v1.3 report output contrasted with v1.4 public report and src/lib/creatorbench-contracts.test.ts","fix":"Gate usability and silent-failure claims on complete human review coverage; otherwise report only review-required and safe-abstention routing.","status":"fixed","ts":"2026-07-22T10:05:11.681Z","fp":"e22af4c1239a"}
+{"id":"creatorbench-e2e-count-drift-v14","sev":"P1","area":"CreatorBench production UI evidence binding","symptom":"The focused UI proof expected stale pilot corpus counts after the benchmark expanded to the v1.4 corpus.","rootCause":"CreatorBench E2E assertions remained pinned to earlier 112-clip pilot values instead of frozen report values.","evidence":"tests/e2e/creatorbench.spec.ts; 40/40 local and 40/40 production five-viewport reruns","fix":"Bind assertions to the v1.4 evidence counts: 250 clips, 205 creator-disjoint sources, 6,128 instances, and 1,392 private held-out instances.","status":"fixed","ts":"2026-07-22T10:05:33.613Z","fp":"cbfe5598592d"}

.qa/memory/runs.jsonl

@@ -19,3 +19,6 @@
 {"evidenceDir":".qa/evidence/creator-topology","executor":"codex","gates":{"fullCheck":"pass","topology":"pass","creatorDesktopMobile":"12-pass-4-intentional-skip","build":"pass","typecheck":"pass"},"bar":{"durability":5,"functional":5,"responsive":5,"visual":4,"safety":5,"accessibility":5},"journeys":["guided-first-arrival","artifact-workspace","proposal-review","paid-executor-gate","durable-two-session","export-reopen","mobile-surface-modes"],"ts":"2026-07-22T01:13:56.489Z"}
 {"executor":"Codex","journeys":["Smart Reframe local journey"],"bar":{"functional":4,"coherence":4,"responsive":4,"accessibility":4,"proof":4},"gates":{"check":"pass","creatorProof":"pass"},"evidenceDir":".qa/evidence/smart-reframe","ts":"2026-07-22T02:19:26.333Z"}
 {"executor":"Codex","mode":"AUTHORIZED PRODUCTION","journeys":{"A0":"PASS","A1":"PASS(local catalog)","A2":"N/A(no external model; local guide)","A3":"PASS(hash-bound receipts)","A4":"PASS(downloadable 42.6s compilation)","A5":"PASS(desktop/tablet/mobile light+dark)","A6":"PASS(fail-closed license/hash validators)","A7":"N/A"},"bar":{"B1":2,"B2":2,"B3":2,"B4":2,"B5":2,"B6":2,"B7":1,"B8":2,"B9":2,"B10":2,"B11":2},"gates":{"atlasVerify":"8-pass","atlasE2E":"4-pass","typecheck":"pass","build":"pass","pixels":"6-pass-no-overflow"},"evidenceDir":".qa/runs/2026-07-21-tracking-atlas","notes":"Eight CC Attribution fixtures; six automatic YOLO11n routes and two explicit first-frame seed plus OpenCV routes. Fixture-bound proof only.","ts":"2026-07-22T03:24:00.000Z"}
+{"executor":"Codex","mode":"AUTHORIZED_PRODUCTION","journeys":{"A0":"PASS","A1":"PASS","A2":"PASS","A3":"PASS","A4":"PASS","A5":"PASS","A6":"PASS","A7":"PARTIAL_5_of_250"},"bar":{"B1":2,"B2":2,"B3":2,"B4":2,"B5":2,"B6":2,"B7":2,"B8":2,"B9":2,"B10":2,"B11":2},"gates":{"lint":"pass","typecheck":"pass","unit":"247-pass","build":"pass","creatorbenchE2E":"10-pass","acquisitionTarget":"fail-honest-5-of-250"},"evidenceDir":".qa/runs/2026-07-22-creatorbench-v1","notes":"dataset-and-human-review-incomplete-no-universal-claim","ts":"2026-07-22T04:51:57.778Z"}
+{"executor":"codex","journeys":{"A0":"PASS","A1":"PASS","A2":"SKIPPED","A3":"PASS","A4":"PASS","A5":"PASS","A6":"PASS","A7":"NA"},"bar":{"B1":2,"B2":1,"B3":2,"B4":2,"B5":2,"B6":1,"B7":1,"B8":2,"B9":2,"B10":1,"B11":2},"gates":["lint","typecheck","build","playwright14","proof","bloat"],"evidenceDir":".qa/runs/2026-07-21-creatorbench-render-pilot","ts":"2026-07-22T05:40:57.668Z"}
+{"executor":"Codex CreatorBench v1.4 production proof","journeys":{"A0":"PASS production shell and report route","A1":"N/A read-only benchmark surface","A2":"N/A no live AI execution on public report","A3":"PASS freeze receipt, provenance, and report binding","A4":"PASS JSON and CSV downloads plus export/reopen evidence","A5":"PASS five viewport projects, both themes, accessibility, and no overflow","A6":"PASS missing-report, infrastructure-only, null-metric, and blind-review states","A7":"N/A public evidence surface is not the durable agent execution route"},"bar":{"B1":2,"B2":2,"B3":2,"B4":2,"B5":2,"B6":2,"B7":1,"B8":2,"B9":2,"B10":2,"B11":2},"gates":{"repository":"PASS npm run check","proof":"PASS 16 CreatorBench proof gates","localE2E":"PASS 40/40 across five viewports","productionE2E":"PASS 40/40 across five viewports","localPixels":"PASS 6/6 theme and viewport captures","productionPixels":"PASS 6/6 theme and viewport captures","deploymentIdentity":"PASS Vercel build receipt, frozen report SHA, and source commit match"},"evidenceDir":"C:\\Users\\hshum\\.codex\\visualizations\\2026\\07\\20\\019f7d76-4ac0-7730-b4f0-9a8b64863b4e\\creatorbench-v1.4-qa","nextImprovement":"Complete authenticated durable human review labels and correction-time measurements before making usability or silent-failure claims.","ts":"2026-07-22T10:05:57.744Z"}

.ui/contract.json

@@ -485,6 +485,63 @@
           "forbidTexts": ["/ 100"]
         }
       ]
+    },
+    {
+      "id": "creatorbench",
+      "route": "/creatorbench",
+      "landmark": {
+        "testid": "creatorbench-overview"
+      },
+      "description": "Public measured-evidence surface for CreatorBench. It separates frozen private routing outcomes, workflow-admissible corpus coverage, public deterministic render proof, known gaps, route evidence, and bounded blinded review from the creator editing workspace.",
+      "controls": [
+        {
+          "id": "creatorbench-overview",
+          "role": "button",
+          "name": "Overview"
+        },
+        {
+          "id": "creatorbench-coverage",
+          "role": "button",
+          "name": "Coverage"
+        },
+        {
+          "id": "creatorbench-weaknesses",
+          "role": "button",
+          "name": "Weaknesses"
+        },
+        {
+          "id": "creatorbench-routes",
+          "role": "button",
+          "name": "Route evidence"
+        },
+        {
+          "id": "creatorbench-freeze",
+          "role": "button",
+          "name": "Freeze receipt"
+        },
+        {
+          "id": "creatorbench-review",
+          "role": "button",
+          "name": "Review lab",
+          "meaning": "Blinded reviewer surface. Machine findings and route identity remain hidden until judgment. Durable submission requires explicit review-data consent and a reachable Convex backend."
+        },
+        {
+          "id": "creatorbench-guide-input",
+          "role": "textbox",
+          "name": "Ask CreatorBench"
+        }
+      ],
+      "states": [
+        {
+          "id": "creatorbench-evaluated",
+          "meaning": "Exact counts and confidence intervals are loaded from the public report; zero human labels are disclosed rather than interpreted as editing quality.",
+          "requireTexts": [
+            "See what works, what needs help, and what fails.",
+            "Human editing quality and silent-failure incidence are unverified"
+          ]
+        }
+      ],
+      "agentGuidance": "Use this surface for measured evidence, not editing. Verify every rate against its numerator, denominator, interval, scope, and freeze receipt. Do not infer human usability from routing-only or decode-only results."
     }
   ]
 }

README.md

@@ -111,6 +111,36 @@ See [the generalized architecture](docs/architecture/GENERALIZED_CREATOR_PIPELIN
 integration contract](docs/architecture/EXECUTOR_INTEGRATION.md), [Higgsfield runbook](docs/operations/HIGGSFIELD_RUNBOOK.md),
 and [copy-ready end-to-end prompt suite](docs/testing/CREATOR_PIPELINE_TEST_PROMPTS.md).
 
+## CreatorBench
+
+`/creatorbench` is the evidence surface for measured generalization. It is separate from the creator
+workspace and the fixture-bound Artifact Atlas. Every benchmark request uses the same
+`nodevideo.creator-request/v1` contract, and every outcome is classified as automatic usable,
+assisted usable, review required, safe abstention, unsupported, technical failure, or silent
+failure. A rendered file is never counted as usable without a blinded human judgment.
+
+```powershell
+npm run creatorbench:acquire
+npm run creatorbench:manifests
+npm run creatorbench:dedupe
+npm run creatorbench:validate
+npm run creatorbench:evaluate
+npm run creatorbench:freeze
+$env:NODEVIDEO_CREATORBENCH_EVALUATOR_TOKEN='<post-freeze evaluator credential>'
+npm run creatorbench:evaluate:sealed
+npm run creatorbench:report
+npm run proof:creatorbench
+```
+
+Acquisition is fail-honest: the target remains 250 rights-cleared clips, 75 creator-disjoint sources,
+15 domains, 8 workflows, and 2,000 instances. If a source host or rights review prevents that target,
+the receipt publishes the achieved population and gap; it does not prefill the claim. See the
+[governance contract](docs/benchmarks/CREATORBENCH_V1_GOVERNANCE.md), [sealed evaluation
+boundary](docs/benchmarks/SEALED_EVALUATION.md), and [end-to-end benchmark prompts](docs/testing/CREATORBENCH_E2E_TEST_PROMPTS.md).
+The report command also emits the standalone machine-readable
+`benchmarks/creatorbench-v1/results/public-claim.json`; it is derived from the frozen held-out
+results and is scanned alongside the public report for private locators and hidden targets.
+
 ### Music rights modes
 
 | Music input | NodeVideo behavior |

apps/atlas/src/atlas.tsx

@@ -459,6 +459,9 @@ function AtlasApp() {
           <a href="/creator">
             <ArrowLeft /> Creator workspace
           </a>
+          <a href="/creatorbench">
+            <Gauge /> CreatorBench
+          </a>
           <p>Explore</p>
           {MODES.map(({ id, label, icon: Icon }) => (
             <button

apps/creatorbench/src/creatorbench-review-client.test.ts

@@ -0,0 +1,106 @@
+import { describe, expect, it, vi } from 'vitest';
+import {
+  CreatorBenchReviewClient,
+  type DurableReviewInput,
+  type ReviewBackend,
+} from './creatorbench-review-client';
+
+class MemoryStorage implements Storage {
+  private readonly values = new Map<string, string>();
+  get length() {
+    return this.values.size;
+  }
+  clear() {
+    this.values.clear();
+  }
+  getItem(key: string) {
+    return this.values.get(key) ?? null;
+  }
+  key(index: number) {
+    return [...this.values.keys()][index] ?? null;
+  }
+  removeItem(key: string) {
+    this.values.delete(key);
+  }
+  setItem(key: string, value: string) {
+    this.values.set(key, value);
+  }
+}
+
+const input: DurableReviewInput = {
+  benchmarkVersion: 'creatorbench-v1.4',
+  instanceId: 'instance:public:1',
+  resultId: 'result:public:1',
+  split: 'public-test',
+  blindedVariantIds: ['variant:a', 'variant:b'],
+  usability: 'usable_after_minor_correction',
+  correctionTimeSeconds: 18,
+  reasonCodes: ['wrong_subject'],
+  explicitOptIn: true,
+};
+
+function backend(): ReviewBackend {
+  const records: Array<Record<string, unknown>> = [];
+  return {
+    claimAssignment: vi.fn(async (args) => {
+      records.push({ ...args, status: 'assigned', blind: true });
+    }),
+    submitReview: vi.fn(async (args) => {
+      const record = records.find((item) => item.assignmentId === args.assignmentId);
+      if (record) Object.assign(record, args, { status: 'completed' });
+    }),
+    listReviewerHistory: vi.fn(async ({ reviewerRef }) =>
+      records.filter((item) => item.reviewerRef === reviewerRef),
+    ) as ReviewBackend['listReviewerHistory'],
+    deleteReviewerData: vi.fn(async ({ reviewerRef }) => {
+      records.splice(0, records.length);
+      return { deletedCount: 1, deletedAt: Date.now(), reviewerRef };
+    }),
+  };
+}
+
+describe('CreatorBench review client', () => {
+  it('requires explicit opt-in before creating an identity or contacting the backend', async () => {
+    const transport = backend();
+    const storage = new MemoryStorage();
+    const client = new CreatorBenchReviewClient(transport, storage);
+    await expect(client.submit({ ...input, explicitOptIn: false })).rejects.toThrow(/opt-in/u);
+    expect(transport.claimAssignment).not.toHaveBeenCalled();
+    expect(storage.length).toBe(0);
+  });
+
+  it('persists a pseudonymous blinded review and verifies the completed record', async () => {
+    const transport = backend();
+    const client = new CreatorBenchReviewClient(transport, new MemoryStorage());
+    const result = await client.submit(input);
+    expect(result.reviewerRef).toMatch(/^reviewer:[a-f\d]{32}$/u);
+    expect(result.confirmed).toMatchObject({ blind: true, status: 'completed' });
+    expect(transport.claimAssignment).toHaveBeenCalledWith(
+      expect.objectContaining({ consentVersion: 'creatorbench-review-consent/v1' }),
+    );
+  });
+
+  it('fails closed when the backend is unavailable', async () => {
+    const client = new CreatorBenchReviewClient(null, new MemoryStorage());
+    await expect(client.submit(input)).rejects.toThrow(/backend unavailable/u);
+  });
+
+  it('exports pseudonymous history and verifies deletion before clearing local identity', async () => {
+    const storage = new MemoryStorage();
+    const client = new CreatorBenchReviewClient(backend(), storage);
+    await client.submit(input);
+    const exported = JSON.parse(await client.exportHistory());
+    expect(exported.reviewerRef).toMatch(/^reviewer:/u);
+    expect(exported.records).toHaveLength(1);
+    const receipt = await client.deleteAll();
+    expect(receipt.deletedCount).toBe(1);
+    expect(storage.length).toBe(0);
+  });
+
+  it('rejects hidden target and private locator hints in blind variant identifiers', async () => {
+    const client = new CreatorBenchReviewClient(backend(), new MemoryStorage());
+    await expect(
+      client.submit({ ...input, blindedVariantIds: ['private-heldout-target-hint'] }),
+    ).rejects.toThrow(/prohibited/u);
+  });
+});