audrey 0.23.1 → 1.0.1

package/docs/paper/output/submission-bundle/benchmarks/output/guardbench-raw.json

@@ -0,0 +1,1271 @@
+{
+  "suite": "GuardBench comparative",
+  "generatedAt": "2026-05-15T17:52:12.761Z",
+  "manifestVersion": "0.2.0",
+  "provenance": {
+    "generatedAt": "2026-05-15T17:52:12.761Z",
+    "gitSha": "82b0e9979680acf751b9e80f6f90f8c6ac74befb",
+    "gitDirty": false,
+    "node": "v24.15.0",
+    "v8": "13.6.233.17-node.48",
+    "platform": "linux",
+    "arch": "x64",
+    "osRelease": "6.17.0-1013-azure",
+    "cpuModel": "AMD EPYC 9V74 80-Core Processor",
+    "cpuCount": 4,
+    "totalMemoryGb": 15.61,
+    "embeddingProvider": "mock",
+    "embeddingDimensions": 64,
+    "llmProvider": "mock"
+  },
+  "cases": [
+    {
+      "id": "GB-01",
+      "name": "Repeated failed shell command",
+      "expectedDecision": "block",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.9,
+          "passed": true,
+          "latencyMs": 6.135,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRPC8HJAWZEY0085CJJHK54G",
+            "failure:Bash:2026-05-15T17:52:01.099Z"
+          ],
+          "recommendedActions": [
+            "Do not repeat the exact failed action until the prior error is understood or the command is changed.",
+            "Before re-running Bash, check what changed since the last failure."
+          ],
+          "summary": "Blocked: this exact Bash action failed before. Caution: 1 memory signal, 1 medium severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.028,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.185,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8HKK5GH11P78E2KB66KW"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.742,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8HM92Q8XXJDSGGCKV8XP"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.418,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-02",
+      "name": "Required preflight procedure missing",
+      "expectedDecision": "block",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 1.96,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8HPQ1DDFJ3F929DEEJEB"
+          ],
+          "recommendedActions": [
+            "Do not proceed until the high-severity memory warning is addressed.",
+            "Apply this must-follow rule before acting."
+          ],
+          "summary": "Blocked: 1 memory signal, 1 high severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.006,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 0.298,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8HS2VFCEMKZT27Y7J289"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 0.475,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8HSJ7N9KKFGH3EZGTFWP"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.353,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-03",
+      "name": "Same command in a different file scope",
+      "expectedDecision": "warn",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-03",
+          "name": "Same command in a different file scope",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.55,
+          "passed": true,
+          "latencyMs": 2.654,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "failure:Bash:2026-05-15T17:52:01.365Z"
+          ],
+          "recommendedActions": [
+            "Before re-running Bash, check what changed since the last failure."
+          ],
+          "summary": "Caution: 1 memory signal, 1 medium severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-03",
+          "name": "Same command in a different file scope",
+          "expectedDecision": "warn",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.007,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-03",
+          "name": "Same command in a different file scope",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.55,
+          "passed": true,
+          "latencyMs": 0.079,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8J3S4YPWSTE26SD6SVPY"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-03",
+          "name": "Same command in a different file scope",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.35,
+          "passed": true,
+          "latencyMs": 0.492,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8J4BXV5DHC2RY9YBGZZC"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-03",
+          "name": "Same command in a different file scope",
+          "expectedDecision": "warn",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.381,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-04",
+      "name": "Same tool with changed command",
+      "expectedDecision": "warn",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.55,
+          "passed": true,
+          "latencyMs": 2.465,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "failure:Bash:2026-05-15T17:52:01.798Z"
+          ],
+          "recommendedActions": [
+            "Before re-running Bash, check what changed since the last failure."
+          ],
+          "summary": "Caution: 1 memory signal, 1 medium severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.006,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.55,
+          "passed": true,
+          "latencyMs": 0.052,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8J98CDFSGA2AG5E56TN0"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.35,
+          "passed": true,
+          "latencyMs": 0.315,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8J9Q0CMEHTHR4TPX8SYY"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.348,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-05",
+      "name": "Prior failure plus successful fix",
+      "expectedDecision": "allow",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "allow",
+          "decisionCorrect": true,
+          "riskScore": 0.2,
+          "passed": true,
+          "latencyMs": 2.485,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRPC8JAPXFTFGGG94QP185MS",
+            "failure:Bash:2026-05-15T17:52:01.877Z"
+          ],
+          "recommendedActions": [
+            "This exact action has succeeded since its last failure; proceed with normal validation.",
+            "Before re-running Bash, check what changed since the last failure."
+          ],
+          "summary": "Allowed: this exact Bash action has succeeded since the prior failure. Caution: 1 memory signal, 1 medium severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "allow",
+          "decisionCorrect": true,
+          "riskScore": 0,
+          "passed": true,
+          "latencyMs": 0.004,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.055,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8JBSK81DRW3SP4PGA3M0"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.35,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8JC7Z8S82XZQM0MC2VED"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "allow",
+          "decisionCorrect": true,
+          "riskScore": 0,
+          "passed": true,
+          "latencyMs": 0.322,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        }
+      ]
+    },
+    {
+      "id": "GB-06",
+      "name": "Recall vector table missing",
+      "expectedDecision": "block",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 2.159,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "recall:episodic:recall.vector_counts"
+          ],
+          "recommendedActions": [
+            "Do not proceed until the high-severity memory warning is addressed.",
+            "Run npx audrey status and npx audrey reembed before depending on memory.",
+            "Run npx audrey status and repair the degraded recall path before relying on Guard."
+          ],
+          "summary": "Blocked: 2 memory signals, 2 high severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.005,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 0.128,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8JEJYKMEDCJKMDKX3Q7H"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.267,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [
+            "Inspect degraded recall before relying on baseline output."
+          ],
+          "summary": "Recall returned partial-failure metadata but this baseline has no fail-closed guard.",
+          "recallErrors": [
+            {
+              "type": "episodic",
+              "stage": "recall.vector_counts",
+              "message": "Missing vector table vec_episodes"
+            }
+          ],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.334,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-07",
+      "name": "FTS failure under hybrid recall",
+      "expectedDecision": "block",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-07",
+          "name": "FTS failure under hybrid recall",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 1.561,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "recall:fts:recall.fts_lookup",
+            "01KRPC8JKB36TE59QKA7Z4V2DM"
+          ],
+          "recommendedActions": [
+            "Do not proceed until the high-severity memory warning is addressed.",
+            "Run npx audrey status and repair the degraded recall path before relying on Guard.",
+            "Apply this must-follow rule before acting."
+          ],
+          "summary": "Blocked: 2 memory signals, 2 high severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-07",
+          "name": "FTS failure under hybrid recall",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.005,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-07",
+          "name": "FTS failure under hybrid recall",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.103,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8JNWTHH9J03GFSYTHN4K"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-07",
+          "name": "FTS failure under hybrid recall",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.314,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8JPC7P0SJDFCJXF222DE"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-07",
+          "name": "FTS failure under hybrid recall",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.129,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [
+            "Inspect degraded recall before relying on baseline output."
+          ],
+          "summary": "Recall returned partial-failure metadata but this baseline has no fail-closed guard.",
+          "recallErrors": [
+            {
+              "type": "fts",
+              "stage": "recall.fts_lookup",
+              "message": "no such table: fts_episodes"
+            }
+          ],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-08",
+      "name": "Long secret near truncation boundary",
+      "expectedDecision": "block",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-08",
+          "name": "Long secret near truncation boundary",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.9,
+          "passed": true,
+          "latencyMs": 2.339,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRPC8JQFVTGQBPCSTSKTRPY7",
+            "failure:Bash:2026-05-15T17:52:02.287Z"
+          ],
+          "recommendedActions": [
+            "Do not repeat the exact failed action until the prior error is understood or the command is changed.",
+            "Before re-running Bash, check what changed since the last failure."
+          ],
+          "summary": "Blocked: this exact Bash action failed before. Caution: 1 memory signal, 1 medium severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-08",
+          "name": "Long secret near truncation boundary",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.007,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-08",
+          "name": "Long secret near truncation boundary",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.049,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8JRKBJR9Y6CTD0D1ZX47"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-08",
+          "name": "Long secret near truncation boundary",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.331,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8JS661GJEJJV12PR5YD3"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-08",
+          "name": "Long secret near truncation boundary",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.306,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-09",
+      "name": "Conflicting project instructions",
+      "expectedDecision": "block",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 1.963,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRPC8K2N9C3SKKD835K921Z8",
+            "01KRPC8K2PVBNMYZ2RBA7B2Q9X"
+          ],
+          "recommendedActions": [
+            "Do not proceed until the high-severity memory warning is addressed.",
+            "Apply this must-follow rule before acting."
+          ],
+          "summary": "Blocked: 2 memory signals, 2 high severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.004,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 0.088,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRPC8K3V9JQY1TQFJQDWGHGM",
+            "01KRPC8K3TKQJ65V280YRNH91B"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 0.345,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRPC8K4CWWGSHGAYZ5JDF62G",
+            "01KRPC8K4D0KF2C6EW79KC869P"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.329,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "id": "GB-10",
+      "name": "High-volume irrelevant memory noise",
+      "expectedDecision": "block",
+      "results": [
+        {
+          "system": "Audrey Guard",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 30.791,
+          "evidenceCount": 13,
+          "evidenceIds": [
+            "01KRPC8PQ72DA5K79S9YZ7N381",
+            "01KRPC8PQ6YCVWK55HP85M0JKB",
+            "01KRPC8PMZ7SZFK6P2HCZQF23X",
+            "01KRPC8PHVXXXJ1HRFGXQ9SNZD",
+            "01KRPC8PE7CP3E77NRQKFWB01Z",
+            "01KRPC8PC7C083T4QRW0PB54W0",
+            "01KRPC8P76C1BBHBKMW79XHVPA",
+            "01KRPC8NSJ25DKGHN9RM5EKGSZ",
+            "01KRPC8NSFC7N7AHWGCBNHXP2P",
+            "01KRPC8MWXZ9DVQJ2QAFM2EJJC",
+            "01KRPC8MV37S2ZR305M1PCPCJA",
+            "01KRPC8KZNCXB2CYDMJ6QVV5CJ",
+            "01KRPC8K5SHHV6HE5MQ10DSKAT"
+          ],
+          "recommendedActions": [
+            "Do not proceed until the high-severity memory warning is addressed.",
+            "Apply this must-follow rule before acting.",
+            "Treat this as uncertain context and verify before relying on it."
+          ],
+          "summary": "Blocked: 13 memory signals, 1 high severity, 12 medium severity found before acting.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": true,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.009,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": true,
+          "latencyMs": 0.462,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRPC8V0CK77K0V6ZKRC1T15A"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 1.051,
+          "evidenceCount": 5,
+          "evidenceIds": [
+            "01KRPC8VCE8VNRWCGWMC1VYNA9",
+            "01KRPC8VCG3GQ7EPDPV9RQ23JA",
+            "01KRPC8VG14K20MGW0C8N1WDGH",
+            "01KRPC8VK28WY2BM5BB3AR9NPA",
+            "01KRPC8V709AFR44CVFQB5MAFW"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": true,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.545,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "hasEvidenceForDecision": false,
+          "lineageTextMatched": false,
+          "requiredEvidenceMatched": false
+        }
+      ]
+    }
+  ],
+  "artifactRedactionSweep": {
+    "checkedAt": "2026-05-15T17:52:12.780Z",
+    "filesChecked": [
+      "benchmarks/output/guardbench-manifest.json",
+      "benchmarks/output/guardbench-raw.json",
+      "benchmarks/output/guardbench-summary.json"
+    ],
+    "seededSecretsChecked": 1,
+    "leakCount": 0,
+    "leaks": [],
+    "passed": true
+  }
+}