Upload local-intelligence eval results for 69d1c0104adb81dd2de74ade

eval/local_intelligence/results/jobs/69d1c0104adb81dd2de74ade/google-gemma-4-e4b-it.json

@@ -0,0 +1,706 @@
+{
+  "summary": {
+    "provider": "hf-transformers",
+    "model": "google/gemma-4-E4B-it",
+    "cards": 24,
+    "passed": 0,
+    "pass_rate": 0.0,
+    "average_score": 0.0,
+    "median_latency_ms": 7253.74
+  },
+  "rows": [
+    {
+      "card_id": "memo-auto-title",
+      "title": "Memo Auto-Title (<1s)",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 2331.65,
+      "raw_text": "```json\n{\n  \"title\": \"Book Dentist & Email Insurance Form\"\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 4 column 1 (char 53)",
+        "assertions": [
+          {
+            "assertion": "title is not \"Voice Memo\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "title length <= 48",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "title contains either \"dentist\" or \"insurance\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "memo-type-detection",
+      "title": "Memo Type Detection",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 1918.43,
+      "raw_text": "```json\n{\n  \"primaryType\": \"journal\",\n  \"needsReview\": false\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 5 column 1 (char 55)",
+        "assertions": [
+          {
+            "assertion": "primaryType is either \"journal\" or \"idea\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "confidence < 0.8",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "needsReview can be true when uncertainty is high",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "action-item-extraction",
+      "title": "Action Item Extraction With Confidence",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 5437.14,
+      "raw_text": "```json\n[\n  {\n    \"text\": \"clean up the old docs folder sometime\",\n    \"owner\": \"Arach\",\n    \"dueDateISO\": null,\n    \"confidence\": 0.8,\n    \"evidence\": \"Maybe clean up the old docs folder sometime.\"\n  }\n]\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 10 column 1 (char 197)",
+        "assertions": [
+          {
+            "assertion": "items length >= 1",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "item confidence < 0.75 for vague tasks",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "evidence field is always present",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "transcript-cleanup-presets",
+      "title": "Transcript Cleanup Presets",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 3344.74,
+      "raw_text": "```json\n{\n  \"rewritten_transcript\": [\n    \"Spent 4 hours on export bugs.\",\n    \"Fixed 3 of 5 issues.\"\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 7 column 1 (char 100)",
+        "assertions": [
+          {
+            "assertion": "output contains number 4",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "output contains phrase 3 of 5",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "filler words are not introduced",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "what-matters-summary",
+      "title": "Top-3 What Matters Summary",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 13130.81,
+      "raw_text": "```json\n{\n  \"summary\": [\n    {\n      \"point\": \"The memo explicitly lists four distinct decisions.\",\n      \"evidence\": \"Decision one. Decision two. Decision three. Decision four.\",\n      \"priority_score\": 95\n    },\n    {\n      \"point\": \"The content is purely a sequential enumeration of decisions.\",\n      \"evidence\": \"Decision one. Decision two. Decision three. Decision four.\",\n      \"priority_score\": 80\n    },\n    {\n      \"point\": \"No context or details are provided regarding the nature or outcome of these decisions.\",\n      \"evidence\": \"Decision one. Decision two. Decision three. Decision four.\",\n      \"priority_score\": 70\n    }\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 20 column 1 (char 635)",
+        "assertions": [
+          {
+            "assertion": "topPoints length equals 3",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "priority values are numeric",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "each point includes evidence",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "calendar-intent-detection",
+      "title": "Calendar Intent Detection",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 5082.81,
+      "raw_text": "```json\n{\n  \"intent\": \"call\",\n  \"arguments\": {\n    \"contact\": \"Sam\",\n    \"time\": \"2026-04-05T14:00:00-04:00\"\n  }\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 8 column 1 (char 107)",
+        "assertions": [
+          {
+            "assertion": "startAtISO date is 2026-04-05",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "timezone offset in startAtISO is -04:00",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "event title mentions Sam",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "reminder-normalization",
+      "title": "Reminder Phrasing Normalization",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 3694.17,
+      "raw_text": "```json\n{\n  \"reminderText\": \"Investigate new card design.\",\n  \"dueDateISO\": null,\n  \"followUpQuestion\": \"When should I look into the new card design?\"\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 6 column 1 (char 145)",
+        "assertions": [
+          {
+            "assertion": "reminderText starts with a verb",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "reminderText does not contain \"maybe\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "followUpQuestion exists when due date is missing",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "private-redaction-pass",
+      "title": "Private Redaction Pass",
+      "tier": "ship-soon",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 10753.84,
+      "raw_text": "```json\n{\n  \"redacted_text\": \"SSN [REDACTED_SSN] and account [REDACTED_ACCOUNT_NUMBER] should never leave device.\",\n  \"redacted_entities\": [\n    {\n      \"entity\": \"555-44-9012\",\n      \"type\": \"SSN\",\n      \"confidence\": 0.99\n    },\n    {\n      \"entity\": \"00391299\",\n      \"type\": \"ACCOUNT_NUMBER\",\n      \"confidence\": 0.98\n    }\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 16 column 1 (char 326)",
+        "assertions": [
+          {
+            "assertion": "redactedTranscript does not contain 555-44-9012",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "redactedTranscript does not contain 00391299",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "entities length >= 2",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "writing-style-memory",
+      "title": "Per-User Writing Style Memory",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 1990.56,
+      "raw_text": "```json\n{\n  \"rewritten_text\": \"We should use a new onboarding flow.\"\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 4 column 1 (char 63)",
+        "assertions": [
+          {
+            "assertion": "styledSummary does not contain \"utilize\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "output preserves onboarding meaning",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "similar-memo-recall",
+      "title": "Find Similar Memo (Semantic Recall)",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 11741.8,
+      "raw_text": "```json\n{\n  \"top_matches\": [\n    {\n      \"memoId\": \"a\",\n      \"relevance_score\": 0.95,\n      \"rationale\": \"This memo directly addresses 'cancellation policy' and 'trials,' which are the core topics mentioned in the current transcript. This represents strong decision continuity.\"\n    },\n    {\n      \"memoId\": \"b\",\n      \"relevance_score\": 0.1,\n      \"rationale\": \"This memo discusses 'policy' but focuses on 'typography updates,' which is a tangential topic and lacks continuity with the transcript's focus on cancellation and trials.\"\n    }\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 15 column 1 (char 540)",
+        "assertions": [
+          {
+            "assertion": "memoId \"a\" ranks first",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "top match relevance > 0.8",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "project-clustering",
+      "title": "Project Clustering From Voice History",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 9261.85,
+      "raw_text": "```json\n[\n  {\n    \"clusterId\": \"c1\",\n    \"title\": \"Sync Queue and Reconnect Issues\",\n    \"memoIds\": [\n      \"a1\",\n      \"a2\"\n    ],\n    \"confidence\": 0.95\n  },\n  {\n    \"clusterId\": \"c2\",\n    \"title\": \"Personal Notes\",\n    \"memoIds\": [\n      \"b1\"\n    ],\n    \"confidence\": 0.80\n  }\n]\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 20 column 1 (char 274)",
+        "assertions": [
+          {
+            "assertion": "a1 and a2 share a clusterId",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "b1 may be separate or misc",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "next-step-suggestion-engine",
+      "title": "Next-Step Suggestion Engine",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 1470.22,
+      "raw_text": "```json\n{\n  \"action\": \"extract-tasks\"\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 4 column 1 (char 32)",
+        "assertions": [
+          {
+            "assertion": "action is either \"extract-tasks\" or \"schedule\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "reason references explicit deadline",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "follow-up-question-generator",
+      "title": "Follow-Up Question Generator",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 1316.21,
+      "raw_text": "{\"questions\": [\"What specific aspects of the onboarding process should be improved?\"]}<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 1 column 87 (char 86)",
+        "assertions": [
+          {
+            "assertion": "questions length equals 1",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "question ends with ?",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "hybrid-model-router",
+      "title": "Hybrid Local/Cloud Router",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 704.99,
+      "raw_text": "{\"error\": \"Missing request metadata\"}<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 1 column 38 (char 37)",
+        "assertions": [
+          {
+            "assertion": "route equals \"cloud\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "reason mentions long context",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "context-packet-builder",
+      "title": "Context Packet Builder",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 12512.2,
+      "raw_text": "```json\n{\n  \"context_packet\": {\n    \"priorities\": [\n      \"Plan hiring strategy\",\n      \"Plan product launch\",\n      \"Develop documentation strategy\",\n      \"Define support plan\",\n      \"Determine pricing model\",\n      \"Establish analytics framework\"\n    ],\n    \"active_tasks\": [],\n    \"unresolved_questions\": [\n      \"What is the scope/timeline for the hiring plan?\",\n      \"What are the key milestones for the product launch?\",\n      \"What are the target user segments for pricing?\",\n      \"What are the initial KPIs for analytics?\"\n    ],\n    \"next_steps\": \"Prioritize and break down the six planning areas into actionable tasks.\"\n  }\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 21 column 1 (char 632)",
+        "assertions": [
+          {
+            "assertion": "tokenEstimate <= 120",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "packet has at least one decision or openQuestion field",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "conversation-to-checklist",
+      "title": "Conversation To Checklist Transformer",
+      "tier": "local-first",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 11841.51,
+      "raw_text": "```json\n[\n  {\n    \"id\": \"1\",\n    \"text\": \"export data\",\n    \"order\": 1,\n    \"dependsOn\": null,\n    \"owner\": \"Arach\"\n  },\n  {\n    \"id\": \"2\",\n    \"text\": \"verify CSV\",\n    \"order\": 2,\n    \"dependsOn\": \"1\",\n    \"owner\": \"Arach\"\n  },\n  {\n    \"id\": \"3\",\n    \"text\": \"email finance\",\n    \"order\": 3,\n    \"dependsOn\": \"2\",\n    \"owner\": \"Arach\"\n  }\n]\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 24 column 1 (char 335)",
+        "assertions": [
+          {
+            "assertion": "order values are 1,2,3",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "item 2 depends on item 1",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "item 3 depends on item 2 or item 1",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "personal-knowledge-graph",
+      "title": "Personal Knowledge Graph From Voice",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 16759.9,
+      "raw_text": "```json\n{\n  \"nodes\": [\n    {\n      \"id\": \"n1\",\n      \"type\": \"person\",\n      \"name\": \"Chris\"\n    },\n    {\n      \"id\": \"n2\",\n      \"type\": \"decision\",\n      \"name\": \"Pricing discussion\"\n    },\n    {\n      \"id\": \"n3\",\n      \"type\": \"artifact\",\n      \"name\": \"Pricing confirmation\"\n    }\n  ],\n  \"edges\": [\n    {\n      \"type\": \"INITIATED\",\n      \"sourceNodeId\": \"n2\",\n      \"targetNodeId\": \"n1\",\n      \"evidenceMemoId\": \"a\"\n    },\n    {\n      \"type\": \"CONFIRMED\",\n      \"sourceNodeId\": \"n1\",\n      \"targetNodeId\": \"n3\",\n      \"evidenceMemoId\": \"b\"\n    }\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 34 column 1 (char 548)",
+        "assertions": [
+          {
+            "assertion": "only one node labeled Chris",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "edges reference merged Chris node id",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "catch-me-up-daily-brief",
+      "title": "Catch Me Up Daily Brief",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 8450.85,
+      "raw_text": "```json\n{\n  \"wins\": [\n    \"No critical blockers identified today.\",\n    \"Maintained focus on planning activities.\"\n  ],\n  \"risks\": [\n    \"Lack of concrete progress on key deliverables.\",\n    \"Budget finalization remains outstanding.\"\n  ],\n  \"nextActions\": [\n    \"Prioritize and finalize the budget.\",\n    \"Transition from planning to execution tasks.\",\n    \"Schedule a follow-up to review planning outcomes.\"\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 16 column 1 (char 407)",
+        "assertions": [
+          {
+            "assertion": "wins section exists",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "risks section exists",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "nextActions section exists",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "contradiction-drift-detection",
+      "title": "Contradiction And Drift Detection",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 7207.8,
+      "raw_text": "```json\n{\n  \"conflicts\": [\n    {\n      \"severity\": \"high\",\n      \"conflictingDecisionId\": \"d9\",\n      \"explanation\": \"The new memo ('new_test_01') states to move the quarterly review to June 2, which directly contradicts the prior decision ('d9') that fixed the quarterly review date to June 20.\"\n    }\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 10 column 1 (char 301)",
+        "assertions": [
+          {
+            "assertion": "conflicts length >= 1",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "first conflict severity equals high",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "momentum-scoring",
+      "title": "Momentum Scoring",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 1318.8,
+      "raw_text": "```json\n{\n  \"score\": 20\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 4 column 1 (char 18)",
+        "assertions": [
+          {
+            "assertion": "momentumScore < 50",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "unresolvedHotspots includes \"Fix onboarding bug\"",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "local-agent-loop",
+      "title": "Local Agent Loop For Memo Processing",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 19730.79,
+      "raw_text": "```json\n{\n  \"pipeline_id\": \"memo_test_loop_01\",\n  \"steps_executed\": [\n    {\n      \"step_name\": \"classify\",\n      \"input\": \"Maybe move something around at some point.\",\n      \"output\": {\n        \"category\": \"General Suggestion\",\n        \"confidence\": 0.95\n      },\n      \"status\": \"completed\"\n    },\n    {\n      \"step_name\": \"extract\",\n      \"input\": \"Maybe move something around at some point.\",\n      \"output\": {\n        \"keywords\": [\"move\", \"something\", \"around\"],\n        \"entities\": []\n      },\n      \"status\": \"completed\"\n    },\n    {\n      \"step_name\": \"summarize\",\n      \"input\": \"Maybe move something around at some point.\",\n      \"output\": \"A suggestion to rearrange items at some future time.\",\n      \"status\": \"completed\"\n    },\n    {\n      \"step_name\": \"route\",\n      \"input\": \"A suggestion to rearrange items at some future time.\",\n      \"output\": {\n        \"",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Unterminated string starting at: line 32 column 9 (char 863)",
+        "assertions": [
+          {
+            "assertion": "verify.passed can be false for vague memo",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "finalStatus equals \"needs-review\" when verify fails",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "meeting-live-structure",
+      "title": "Meeting Mode Live Rolling Structure",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 8931.95,
+      "raw_text": "```json\n{\n  \"meetingId\": \"meeting_test_live_01\",\n  \"state\": {\n    \"agenda\": [],\n    \"decisions\": [\n      {\n        \"text\": \"Keep beta closed for now.\",\n        \"chunkIndex\": 2\n      },\n      {\n        \"text\": \"Keep beta closed for now.\",\n        \"chunkIndex\": 8\n      }\n    ],\n    \"openQuestions\": [],\n    \"actionItems\": []\n  }\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 19 column 1 (char 322)",
+        "assertions": [
+          {
+            "assertion": "updatedState.decisions length stays 1",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "existing decision remains unchanged",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "intent-trend-alerts",
+      "title": "Intent Trend Alerts",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 9577.28,
+      "raw_text": "```json\n{\n  \"alerts\": [\n    {\n      \"intent\": \"Refactor UI\",\n      \"mentionCount\": 2,\n      \"completionCount\": 0,\n      \"alert\": true,\n      \"reason\": \"Mention count (2) is below the minimum threshold (4) but completion count (0) is at the maximum threshold (0).\",\n      \"suggested_intervention\": \"Review recent UI refactoring requests to understand the blockers preventing completion and prioritize a small, achievable task to build momentum.\"\n    }\n  ]\n}\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 13 column 1 (char 449)",
+        "assertions": [
+          {
+            "assertion": "alerts length equals 0",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    },
+    {
+      "card_id": "voice-os-command-layer",
+      "title": "Voice-Based Operating Layer",
+      "tier": "big-swings",
+      "provider": "hf-transformers",
+      "model": "google/gemma-4-E4B-it",
+      "latency_ms": 7299.67,
+      "raw_text": "```json\n[\n  {\n    \"app\": \"finder\",\n    \"action\": \"delete_folder\",\n    \"args\": {\n      \"path\": \"old exports folder\"\n    },\n    \"requiresConfirmation\": true\n  },\n  {\n    \"app\": \"finder\",\n    \"action\": \"empty_trash\",\n    \"args\": {},\n    \"requiresConfirmation\": true\n  }\n]\n```<turn|>",
+      "grade": {
+        "passed": false,
+        "score": 0.0,
+        "parse_error": "Extra data: line 17 column 1 (char 261)",
+        "assertions": [
+          {
+            "assertion": "requiresConfirmation equals true",
+            "passed": false,
+            "details": "output was not valid JSON"
+          },
+          {
+            "assertion": "at least one step action is marked destructive",
+            "passed": false,
+            "details": "output was not valid JSON"
+          }
+        ],
+        "parsed": null
+      }
+    }
+  ]
+}