audrey 0.23.1 → 1.0.0

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

@@ -0,0 +1,1171 @@
+{
+  "suite": "GuardBench comparative",
+  "generatedAt": "2026-05-13T23:33:51.220Z",
+  "manifestVersion": "0.2.0",
+  "provenance": {
+    "generatedAt": "2026-05-13T23:33:51.221Z",
+    "gitSha": "970752172441967c3ede79562eca69b08efb1f12",
+    "gitDirty": false,
+    "node": "v24.14.1",
+    "v8": "13.6.233.17-node.44",
+    "platform": "linux",
+    "arch": "x64",
+    "osRelease": "6.17.0-1010-azure",
+    "cpuModel": "AMD EPYC 7763 64-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": 8.899,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRHV0X9M33M5ZXDZ9YXYVBJN",
+            "failure:Bash:2026-05-13T23:33:47.445Z"
+          ],
+          "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": [],
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.038,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 0.27,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XAWMDCXTZB3AAWH585M"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 1.119,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XBBNDQN9FZB3T6YSG9Z"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-01",
+          "name": "Repeated failed shell command",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.573,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 2.566,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XCAFQE8M3MM50FDRPEP"
+          ],
+          "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": [],
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.005,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": false,
+          "latencyMs": 0.421,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XD9H0KTNY150TDNZ45F"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": false,
+          "latencyMs": 0.646,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XDQJJ3RJ24MEREZTWBB"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-02",
+          "name": "Required preflight procedure missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.509,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 3.463,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "failure:Bash:2026-05-13T23:33:47.603Z"
+          ],
+          "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": [],
+          "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": [],
+          "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": false,
+          "latencyMs": 0.082,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XFM7T2FZHATHJKY7D8S"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-03",
+          "name": "Same command in a different file scope",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.533,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XG1FT5R9QVJPK0GWCRZ"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "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": 0.524,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 3.097,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "failure:Bash:2026-05-13T23:33:47.677Z"
+          ],
+          "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": [],
+          "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": 0.01,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.069,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XHWDZY3FFYNYY79JHR0"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "warn",
+          "decisionCorrect": true,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.457,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XJAZH9TEAX0HXHS8CYD"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-04",
+          "name": "Same tool with changed command",
+          "expectedDecision": "warn",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.471,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 3.393,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRHV0XKCBJJXKWKSQW5NGZGJ",
+            "failure:Bash:2026-05-13T23:33:47.754Z"
+          ],
+          "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": [],
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "allow",
+          "decisionCorrect": true,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.008,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.098,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XMBZEH7TP2QT34DTB45"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.532,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XMSE6G3K0N1AD76K374"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-05",
+          "name": "Prior failure plus successful fix",
+          "expectedDecision": "allow",
+          "decision": "allow",
+          "decisionCorrect": true,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.514,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        }
+      ]
+    },
+    {
+      "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": 2.909,
+          "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": [],
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.008,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": false,
+          "latencyMs": 0.195,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XPQPW9GYR5M3TKX3BMS"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.347,
+          "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": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-06",
+          "name": "Recall vector table missing",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.427,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 2.024,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "recall:fts:recall.fts_lookup",
+            "01KRHV0XR0C1AF66E8C87BRDRK"
+          ],
+          "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": [],
+          "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": 0.007,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 0.126,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XRX3N2Q68CQSDFY9BA7"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-07",
+          "name": "FTS failure under hybrid recall",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.467,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XSBVXJJ3ZT6GYZQZ2BY"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-07",
+          "name": "FTS failure under hybrid recall",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.55,
+          "passed": false,
+          "latencyMs": 0.256,
+          "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": [],
+          "requiredEvidenceMatched": true
+        }
+      ]
+    },
+    {
+      "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": 2.879,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRHV0XT8ZD1431EA3P176C84",
+            "failure:Bash:2026-05-13T23:33:47.976Z"
+          ],
+          "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": [],
+          "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": [],
+          "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": 0.072,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XV6HNP678SEPD3N30BB"
+          ],
+          "recommendedActions": [
+            "Check the recent failed event before repeating a similar action."
+          ],
+          "summary": "Recent-window baseline found a failed Bash event.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-08",
+          "name": "Long secret near truncation boundary",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 0.451,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0XVM0E0SG23G9MZPGSED"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-08",
+          "name": "Long secret near truncation boundary",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.431,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 2.629,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRHV0XWHVTFB5Q7NVEJ6PYSA",
+            "01KRHV0XWFX8H82KAABRSMBEZY"
+          ],
+          "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": [],
+          "requiredEvidenceMatched": true
+        },
+        {
+          "system": "No Memory",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "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": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": false,
+          "latencyMs": 0.125,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRHV0XXJ3Z1K7DMJEWP3HWY7",
+            "01KRHV0XXGRTMH80WYV6D42VHX"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": false,
+          "latencyMs": 0.442,
+          "evidenceCount": 2,
+          "evidenceIds": [
+            "01KRHV0XXXCP5GB55SPAMC2B09",
+            "01KRHV0XXZSQCH0W7ZSFY54BDA"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-09",
+          "name": "Conflicting project instructions",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.458,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "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": 29.711,
+          "evidenceCount": 13,
+          "evidenceIds": [
+            "01KRHV0YFDERDNPYS8JG019DXV",
+            "01KRHV0YF5AACS4RJJAPW2YNPZ",
+            "01KRHV0YDFFQ5SVS6GSH4PKNJ6",
+            "01KRHV0YAKHEN7XS8TJZQJG3JW",
+            "01KRHV0Y8PGE2S0B37B3ECNPES",
+            "01KRHV0Y8D6C8CHE2MQ6C0RB2D",
+            "01KRHV0Y7EN8M2KZYSNV2S659G",
+            "01KRHV0Y6ADQ4VY5MKWHMTBE2V",
+            "01KRHV0Y61MS7FKZHH2H0R5CS3",
+            "01KRHV0Y3E8DKZNPNHFBKNEQXQ",
+            "01KRHV0Y2QSJ9BMQGST1D2CBG2",
+            "01KRHV0Y1A0S18SHKGCN00K1J6",
+            "01KRHV0Y0MWQGN3QAWME8F5554"
+          ],
+          "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": [],
+          "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": 0.014,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory baseline always allows proposed actions.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Recent Window",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "block",
+          "decisionCorrect": true,
+          "riskScore": 0.85,
+          "passed": false,
+          "latencyMs": 0.475,
+          "evidenceCount": 1,
+          "evidenceIds": [
+            "01KRHV0ZNZBPXBC3HBT7FDYKQN"
+          ],
+          "recommendedActions": [
+            "Review retrieved memory before acting."
+          ],
+          "summary": "Retrieved policy-like memory with lexical overlap, but without Audrey Guard lineage.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "Vector Only",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "warn",
+          "decisionCorrect": false,
+          "riskScore": 0.35,
+          "passed": false,
+          "latencyMs": 1.362,
+          "evidenceCount": 5,
+          "evidenceIds": [
+            "01KRHV0ZXZBPYCREN1DV87M33S",
+            "01KRHV100XKC09N7VQRRX2WK2A",
+            "01KRHV103RQB958BA9WSHGJNJF",
+            "01KRHV0ZVSQKJVY4TJ1RP15CJX",
+            "01KRHV1015PBEKW0RNV9CY3AHJ"
+          ],
+          "recommendedActions": [
+            "Treat retrieved memory as advisory context."
+          ],
+          "summary": "Retrieved related memory, but no controller converted it into a strict guard decision.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        },
+        {
+          "system": "FTS Only",
+          "id": "GB-10",
+          "name": "High-volume irrelevant memory noise",
+          "expectedDecision": "block",
+          "decision": "allow",
+          "decisionCorrect": false,
+          "riskScore": 0,
+          "passed": false,
+          "latencyMs": 0.778,
+          "evidenceCount": 0,
+          "evidenceIds": [],
+          "recommendedActions": [],
+          "summary": "No memory signal found by this baseline.",
+          "recallErrors": [],
+          "leakedSecrets": [],
+          "requiredEvidenceMatched": false
+        }
+      ]
+    }
+  ],
+  "artifactRedactionSweep": {
+    "checkedAt": "2026-05-13T23:33:51.244Z",
+    "filesChecked": [
+      "benchmarks/output/guardbench-manifest.json",
+      "benchmarks/output/guardbench-raw.json",
+      "benchmarks/output/guardbench-summary.json"
+    ],
+    "seededSecretsChecked": 1,
+    "leakCount": 0,
+    "leaks": [],
+    "passed": true
+  }
+}