diff --git a/bench/lib/eval/workspace.mjs b/bench/lib/eval/workspace.mjs
index 2c27c39..d814257 100644
--- a/bench/lib/eval/workspace.mjs
+++ b/bench/lib/eval/workspace.mjs
@@ -20,6 +20,19 @@ export function evalAssessmentPath(root, id, phase) {
   return path.join(evalWorkspaceDir(root, id, phase), 'assessment.json')
 }
 
+// Blinding: harness-written inputs must not name the run. The verification
+// record carries the real runId (and could embed it in check output or
+// workspace paths), so every string in it is rewritten to the blind handle
+// before it reaches the evaluator.
+function scrubRunIdentity(value, id, blindId) {
+  if (typeof value === 'string') return value.split(id).join(blindId)
+  if (Array.isArray(value)) return value.map((entry) => scrubRunIdentity(entry, id, blindId))
+  if (value && typeof value === 'object') {
+    return Object.fromEntries(Object.entries(value).map(([key, entry]) => [key.split(id).join(blindId), scrubRunIdentity(entry, id, blindId)]))
+  }
+  return value
+}
+
 function copyInput(root, relativePath, inputRoot) {
   const source = path.resolve(root, relativePath)
   if (!source.startsWith(`${path.resolve(root)}${path.sep}`)) throw new Error(`Input escapes repository: ${relativePath}`)
@@ -78,7 +91,7 @@ export function createEvalWorkspace(root, { id, phase, variantId, stageDir, veri
   writeJson(path.join(inputs, 'specs', 'nord-stage-4.variants.json'), projectVariants(variants, variantId))
   writeJson(path.join(inputs, 'rubric.json'), rubric)
   if (verificationPath && fs.existsSync(verificationPath)) {
-    fs.copyFileSync(verificationPath, path.join(inputs, 'verification.json'))
+    writeJson(path.join(inputs, 'verification.json'), scrubRunIdentity(readJson(verificationPath), id, blindId))
   }
   // Blind template: identified by the handle, never the model id.
   writeJson(path.join(workspace, 'assessment.json'), createAssessmentTemplate(rubric, blindId, phase))
diff --git a/tests/bench.test.mjs b/tests/bench.test.mjs
index f951adb..f4eb358 100644
--- a/tests/bench.test.mjs
+++ b/tests/bench.test.mjs
@@ -180,6 +180,24 @@ test('a run flows new → start → seal → score with sealed digests and a sco
     assert.ok(!evalWorkspace.workspace.includes(created.id), 'the model id is not printed in the evaluator path')
     assert.ok(fs.readFileSync(path.join(evalWorkspace.workspace, 'EVAL.md'), 'utf8').includes('blind'), 'EVAL.md states the evaluation is blind')
     assert.equal(readJson(evalWorkspace.assessment).runId, blindRunCode(created.id), 'the template is identified by the blind handle')
+    assert.equal(readJson(path.join(evalWorkspace.inputs, 'verification.json')).runId, evalWorkspace.blindId, 'the verification record is re-identified by the blind handle')
+    // No harness-written file in the evaluator workspace may name the real run.
+    // artifact/ is exempt: it is the candidate's own sealed bytes, pinned by the
+    // digest check, so the harness cannot rewrite it.
+    const harnessFiles = []
+    const collectHarnessFiles = (dir) => {
+      for (const entry of fs.readdirSync(dir, { withFileTypes: true })) {
+        const full = path.join(dir, entry.name)
+        if (entry.isDirectory()) {
+          if (full !== evalWorkspace.artifact) collectHarnessFiles(full)
+        } else harnessFiles.push(full)
+      }
+    }
+    collectHarnessFiles(evalWorkspace.workspace)
+    assert.ok(harnessFiles.some((file) => file.endsWith('verification.json')), 'the blinding sweep covers the verification record')
+    for (const file of harnessFiles) {
+      assert.ok(!fs.readFileSync(file).includes(created.id), `${path.relative(evalWorkspace.workspace, file)} leaks the run id`)
+    }
 
     const assessment = readJson(evalWorkspace.assessment)
     assessment.evaluator = 'Fixture evaluator'