aiforcecli-chat 0.1.0

package/tools/scorecard/generated/example/unmapped-models.json

@@ -0,0 +1,1492 @@
+[
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o3-pro high",
+    "metric": "pass_rate_2",
+    "score": 84.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-06-28",
+    "extra": {
+      "passRate1": 43.6,
+      "passNum1": 98,
+      "passNum2": 191,
+      "totalCostUsd": 146.3249,
+      "secondsPerCase": 449,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 2
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o3 high",
+    "metric": "pass_rate_2",
+    "score": 81.3,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-06-25",
+    "extra": {
+      "passRate1": 40,
+      "passNum1": 90,
+      "passNum2": 183,
+      "totalCostUsd": 21.2259,
+      "secondsPerCase": 197.3,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 5
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "grok-4 high",
+    "metric": "pass_rate_2",
+    "score": 79.6,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-07-11",
+    "extra": {
+      "passRate1": 40.9,
+      "passNum1": 92,
+      "passNum2": 179,
+      "totalCostUsd": 59.6182,
+      "secondsPerCase": 403.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 6
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o3 high + gpt-4.1",
+    "metric": "pass_rate_2",
+    "score": 78.2,
+    "scoreScale": "percent",
+    "sampleSize": 224,
+    "date": "2025-06-27",
+    "extra": {
+      "passRate1": 34.8,
+      "passNum1": 78,
+      "passNum2": 176,
+      "totalCostUsd": 17.5518,
+      "secondsPerCase": 121.8,
+      "editFormat": "architect"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 8
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o3",
+    "metric": "pass_rate_2",
+    "score": 76.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-06-25",
+    "extra": {
+      "passRate1": 40.9,
+      "passNum1": 92,
+      "passNum2": 173,
+      "totalCostUsd": 13.7517,
+      "secondsPerCase": 101.7,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 9
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek-V3.2-Exp Reasoner",
+    "metric": "pass_rate_2",
+    "score": 74.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-10-03",
+    "extra": {
+      "passRate1": 39.6,
+      "passNum1": 89,
+      "passNum2": 167,
+      "totalCostUsd": 1.3045,
+      "secondsPerCase": 291.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 11
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Gemini 2.5 Pro Preview 03-25",
+    "metric": "pass_rate_2",
+    "score": 72.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-12",
+    "extra": {
+      "passRate1": 40.9,
+      "passNum1": 92,
+      "passNum2": 164,
+      "totalCostUsd": 0,
+      "secondsPerCase": 45.3,
+      "editFormat": "diff-fenced"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 12
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-opus-4-20250514 32k thinking",
+    "metric": "pass_rate_2",
+    "score": 72,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-25",
+    "extra": {
+      "passRate1": 37.3,
+      "passNum1": 84,
+      "passNum2": 162,
+      "totalCostUsd": 65.7484,
+      "secondsPerCase": 44.1,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 13
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o4-mini high",
+    "metric": "pass_rate_2",
+    "score": 72,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-16",
+    "extra": {
+      "passRate1": 19.6,
+      "passNum1": 44,
+      "passNum2": 162,
+      "totalCostUsd": 19.6399,
+      "secondsPerCase": 176.5,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 14
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek R1 0528",
+    "metric": "pass_rate_2",
+    "score": 71.4,
+    "scoreScale": "percent",
+    "sampleSize": 224,
+    "date": "2025-06-06",
+    "extra": {
+      "passRate1": 34.4,
+      "passNum1": 77,
+      "passNum2": 160,
+      "totalCostUsd": 4.8016,
+      "secondsPerCase": 716.6,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 15
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-opus-4-20250514 no think",
+    "metric": "pass_rate_2",
+    "score": 70.7,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-25",
+    "extra": {
+      "passRate1": 32.9,
+      "passNum1": 74,
+      "passNum2": 159,
+      "totalCostUsd": 68.6253,
+      "secondsPerCase": 42.5,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 16
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek-V3.2-Exp Chat",
+    "metric": "pass_rate_2",
+    "score": 70.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-10-03",
+    "extra": {
+      "passRate1": 38.7,
+      "passNum1": 87,
+      "passNum2": 158,
+      "totalCostUsd": 0.8756,
+      "secondsPerCase": 104,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 17
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-3-7-sonnet-20250219 32k thinking tokens",
+    "metric": "pass_rate_2",
+    "score": 64.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-02-24",
+    "extra": {
+      "passRate1": 29.3,
+      "passNum1": 66,
+      "passNum2": 146,
+      "totalCostUsd": 36.8343,
+      "secondsPerCase": 105.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 18
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek R1 + claude-3-5-sonnet-20241022",
+    "metric": "pass_rate_2",
+    "score": 64,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-01-23",
+    "extra": {
+      "passRate1": 27.1,
+      "passNum1": 61,
+      "passNum2": 144,
+      "totalCostUsd": 13.2933,
+      "secondsPerCase": 251.6,
+      "editFormat": "architect"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 19
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o1-2024-12-17 high",
+    "metric": "pass_rate_2",
+    "score": 61.7,
+    "scoreScale": "percent",
+    "sampleSize": 224,
+    "date": "2024-12-21",
+    "extra": {
+      "passRate1": 23.7,
+      "passNum1": 53,
+      "passNum2": 139,
+      "totalCostUsd": 186.4958,
+      "secondsPerCase": 133.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 20
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-sonnet-4-20250514 32k thinking",
+    "metric": "pass_rate_2",
+    "score": 61.3,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-24",
+    "extra": {
+      "passRate1": 25.8,
+      "passNum1": 58,
+      "passNum2": 138,
+      "totalCostUsd": 26.5755,
+      "secondsPerCase": 79.9,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 21
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-3-7-sonnet-20250219 no thinking",
+    "metric": "pass_rate_2",
+    "score": 60.4,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-02-24",
+    "extra": {
+      "passRate1": 24.4,
+      "passNum1": 55,
+      "passNum2": 136,
+      "totalCostUsd": 17.7191,
+      "secondsPerCase": 28.3,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 22
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o3-mini high",
+    "metric": "pass_rate_2",
+    "score": 60.4,
+    "scoreScale": "percent",
+    "sampleSize": 224,
+    "date": "2025-01-31",
+    "extra": {
+      "passRate1": 21,
+      "passNum1": 47,
+      "passNum2": 136,
+      "totalCostUsd": 18.1584,
+      "secondsPerCase": 124.6,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 23
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Qwen3 235B A22B diff, no think, Alibaba API",
+    "metric": "pass_rate_2",
+    "score": 59.6,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-09",
+    "extra": {
+      "passRate1": 28.9,
+      "passNum1": 65,
+      "passNum2": 134,
+      "totalCostUsd": 0,
+      "secondsPerCase": 45.4,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 24
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Kimi K2",
+    "metric": "pass_rate_2",
+    "score": 59.1,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-07-17",
+    "extra": {
+      "passRate1": 20.4,
+      "passNum1": 46,
+      "passNum2": 133,
+      "totalCostUsd": 1.2357,
+      "secondsPerCase": 67.6,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 25
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek R1",
+    "metric": "pass_rate_2",
+    "score": 56.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-01-20",
+    "extra": {
+      "passRate1": 26.7,
+      "passNum1": 60,
+      "passNum2": 128,
+      "totalCostUsd": 5.4193,
+      "secondsPerCase": 113.7,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 26
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-sonnet-4-20250514 no thinking",
+    "metric": "pass_rate_2",
+    "score": 56.4,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-24",
+    "extra": {
+      "passRate1": 20.4,
+      "passNum1": 46,
+      "passNum2": 127,
+      "totalCostUsd": 15.8155,
+      "secondsPerCase": 29.8,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 27
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gemini-2.5-flash-preview-05-20 24k think",
+    "metric": "pass_rate_2",
+    "score": 55.1,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-25",
+    "extra": {
+      "passRate1": 26.2,
+      "passNum1": 59,
+      "passNum2": 124,
+      "totalCostUsd": 8.5625,
+      "secondsPerCase": 53.9,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 28
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek V3 0324",
+    "metric": "pass_rate_2",
+    "score": 55.1,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-03-24",
+    "extra": {
+      "passRate1": 28,
+      "passNum1": 63,
+      "passNum2": 124,
+      "totalCostUsd": 1.1164,
+      "secondsPerCase": 290,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 29
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Quasar Alpha",
+    "metric": "pass_rate_2",
+    "score": 54.7,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-04",
+    "extra": {
+      "passRate1": 21.8,
+      "passNum1": 49,
+      "passNum2": 123,
+      "totalCostUsd": 0,
+      "secondsPerCase": 14.8,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 30
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o3-mini medium",
+    "metric": "pass_rate_2",
+    "score": 53.8,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-01-31",
+    "extra": {
+      "passRate1": 19.1,
+      "passNum1": 43,
+      "passNum2": 121,
+      "totalCostUsd": 8.8599,
+      "secondsPerCase": 47.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 31
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Grok 3 Beta",
+    "metric": "pass_rate_2",
+    "score": 53.3,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-10",
+    "extra": {
+      "passRate1": 22.2,
+      "passNum1": 50,
+      "passNum2": 120,
+      "totalCostUsd": 11.0338,
+      "secondsPerCase": 15.3,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 32
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Optimus Alpha",
+    "metric": "pass_rate_2",
+    "score": 52.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-10",
+    "extra": {
+      "passRate1": 21.3,
+      "passNum1": 48,
+      "passNum2": 119,
+      "totalCostUsd": 0,
+      "secondsPerCase": 18.4,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 33
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-4.1",
+    "metric": "pass_rate_2",
+    "score": 52.4,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-14",
+    "extra": {
+      "passRate1": 20,
+      "passNum1": 45,
+      "passNum2": 118,
+      "totalCostUsd": 9.8556,
+      "secondsPerCase": 20.5,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 34
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-3-5-sonnet-20241022",
+    "metric": "pass_rate_2",
+    "score": 51.6,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-01-17",
+    "extra": {
+      "passRate1": 22.2,
+      "passNum1": 50,
+      "passNum2": 116,
+      "totalCostUsd": 14.4063,
+      "secondsPerCase": 21.4,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 35
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Grok 3 Mini Beta high",
+    "metric": "pass_rate_2",
+    "score": 49.3,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-10",
+    "extra": {
+      "passRate1": 17.3,
+      "passNum1": 39,
+      "passNum2": 111,
+      "totalCostUsd": 0.7346,
+      "secondsPerCase": 79.1,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 36
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek Chat V3 prev",
+    "metric": "pass_rate_2",
+    "score": 48.4,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-25",
+    "extra": {
+      "passRate1": 22.7,
+      "passNum1": 51,
+      "passNum2": 109,
+      "totalCostUsd": 0.3369,
+      "secondsPerCase": 34.8,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 37
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gemini-2.5-flash-preview-04-17 default",
+    "metric": "pass_rate_2",
+    "score": 47.1,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-20",
+    "extra": {
+      "passRate1": 21.8,
+      "passNum1": 49,
+      "passNum2": 106,
+      "totalCostUsd": 1.8451,
+      "secondsPerCase": 50.1,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 38
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "chatgpt-4o-latest 2025-03-29",
+    "metric": "pass_rate_2",
+    "score": 45.3,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-03-29",
+    "extra": {
+      "passRate1": 16.4,
+      "passNum1": 37,
+      "passNum2": 102,
+      "totalCostUsd": 19.7416,
+      "secondsPerCase": 10.3,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 39
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-4.5-preview",
+    "metric": "pass_rate_2",
+    "score": 44.9,
+    "scoreScale": "percent",
+    "sampleSize": 224,
+    "date": "2025-02-27",
+    "extra": {
+      "passRate1": 22.3,
+      "passNum1": 50,
+      "passNum2": 101,
+      "totalCostUsd": 183.1802,
+      "secondsPerCase": 113.5,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 40
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gemini-2.5-flash-preview-05-20 no think",
+    "metric": "pass_rate_2",
+    "score": 44,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-26",
+    "extra": {
+      "passRate1": 20.9,
+      "passNum1": 47,
+      "passNum2": 99,
+      "totalCostUsd": 1.1354,
+      "secondsPerCase": 12.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 41
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-oss-120b high",
+    "metric": "pass_rate_2",
+    "score": 41.8,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-08-06",
+    "extra": {
+      "passRate1": 13.8,
+      "passNum1": 31,
+      "passNum2": 94,
+      "totalCostUsd": 0.7406,
+      "secondsPerCase": 35.5,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 42
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Qwen3 32B",
+    "metric": "pass_rate_2",
+    "score": 40,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-05-08",
+    "extra": {
+      "passRate1": 14.2,
+      "passNum1": 32,
+      "passNum2": 90,
+      "totalCostUsd": 0.7603,
+      "secondsPerCase": 372.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 43
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gemini-exp-1206",
+    "metric": "pass_rate_2",
+    "score": 38.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-22",
+    "extra": {
+      "passRate1": 19.6,
+      "passNum1": 44,
+      "passNum2": 86,
+      "totalCostUsd": 0,
+      "secondsPerCase": 45.5,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 44
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Gemini 2.0 Pro exp-02-05",
+    "metric": "pass_rate_2",
+    "score": 35.6,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-02-25",
+    "extra": {
+      "passRate1": 20.4,
+      "passNum1": 46,
+      "passNum2": 80,
+      "totalCostUsd": 0,
+      "secondsPerCase": 34.8,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 45
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Grok 3 Mini Beta low",
+    "metric": "pass_rate_2",
+    "score": 34.7,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-10",
+    "extra": {
+      "passRate1": 11.1,
+      "passNum1": 25,
+      "passNum2": 78,
+      "totalCostUsd": 0.7856,
+      "secondsPerCase": 35.1,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 46
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "o1-mini-2024-09-12",
+    "metric": "pass_rate_2",
+    "score": 32.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-22",
+    "extra": {
+      "passRate1": 5.8,
+      "passNum1": 13,
+      "passNum2": 74,
+      "totalCostUsd": 18.577,
+      "secondsPerCase": 34.7,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 47
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-4.1-mini",
+    "metric": "pass_rate_2",
+    "score": 32.4,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-14",
+    "extra": {
+      "passRate1": 11.1,
+      "passNum1": 25,
+      "passNum2": 73,
+      "totalCostUsd": 1.9918,
+      "secondsPerCase": 19.5,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 48
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "claude-3-5-haiku-20241022",
+    "metric": "pass_rate_2",
+    "score": 28,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-21",
+    "extra": {
+      "passRate1": 7.1,
+      "passNum1": 16,
+      "passNum2": 63,
+      "totalCostUsd": 6.0583,
+      "secondsPerCase": 31.8,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 49
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "chatgpt-4o-latest 2025-02-15",
+    "metric": "pass_rate_2",
+    "score": 27.1,
+    "scoreScale": "percent",
+    "sampleSize": 223,
+    "date": "2025-02-15",
+    "extra": {
+      "passRate1": 9,
+      "passNum1": 20,
+      "passNum2": 61,
+      "totalCostUsd": 14.3703,
+      "secondsPerCase": 12.4,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 50
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "QwQ-32B + Qwen 2.5 Coder Instruct",
+    "metric": "pass_rate_2",
+    "score": 26.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-03-07",
+    "extra": {
+      "passRate1": 9.8,
+      "passNum1": 22,
+      "passNum2": 59,
+      "totalCostUsd": 0,
+      "secondsPerCase": 137.4,
+      "editFormat": "architect"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 51
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-4o-2024-08-06",
+    "metric": "pass_rate_2",
+    "score": 23.1,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-30",
+    "extra": {
+      "passRate1": 4.9,
+      "passNum1": 11,
+      "passNum2": 52,
+      "totalCostUsd": 7.0286,
+      "secondsPerCase": 16,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 52
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gemini-2.0-flash-exp",
+    "metric": "pass_rate_2",
+    "score": 22.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-22",
+    "extra": {
+      "passRate1": 11.6,
+      "passNum1": 26,
+      "passNum2": 50,
+      "totalCostUsd": 0,
+      "secondsPerCase": 12.2,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 53
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "qwen-max-2025-01-25",
+    "metric": "pass_rate_2",
+    "score": 21.8,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-01-28",
+    "extra": {
+      "passRate1": 9.3,
+      "passNum1": 21,
+      "passNum2": 49,
+      "secondsPerCase": 39.5,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 54
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "QwQ-32B",
+    "metric": "pass_rate_2",
+    "score": 20.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-03-06",
+    "extra": {
+      "passRate1": 8,
+      "passNum1": 18,
+      "passNum2": 47,
+      "totalCostUsd": 0,
+      "secondsPerCase": 228.6,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 55
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gemini-2.0-flash-thinking-exp-01-21",
+    "metric": "pass_rate_2",
+    "score": 18.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-01-21",
+    "extra": {
+      "passRate1": 5.8,
+      "passNum1": 13,
+      "passNum2": 41,
+      "totalCostUsd": 0,
+      "secondsPerCase": 24.2,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 56
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-4o-2024-11-20",
+    "metric": "pass_rate_2",
+    "score": 18.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-30",
+    "extra": {
+      "passRate1": 4.9,
+      "passNum1": 11,
+      "passNum2": 41,
+      "totalCostUsd": 6.7351,
+      "secondsPerCase": 12.1,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 57
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "DeepSeek Chat V2.5",
+    "metric": "pass_rate_2",
+    "score": 17.8,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-21",
+    "extra": {
+      "passRate1": 5.3,
+      "passNum1": 12,
+      "passNum2": 40,
+      "totalCostUsd": 0.5101,
+      "secondsPerCase": 184,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 58
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Qwen2.5-Coder-32B-Instruct",
+    "metric": "pass_rate_2",
+    "score": 16.4,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-26",
+    "extra": {
+      "passRate1": 4.9,
+      "passNum1": 11,
+      "passNum2": 37,
+      "totalCostUsd": 0,
+      "secondsPerCase": 42,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 59
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Llama 4 Maverick",
+    "metric": "pass_rate_2",
+    "score": 15.6,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-06",
+    "extra": {
+      "passRate1": 4.4,
+      "passNum1": 10,
+      "passNum2": 35,
+      "totalCostUsd": 0,
+      "secondsPerCase": 20.5,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 60
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "yi-lightning",
+    "metric": "pass_rate_2",
+    "score": 12.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-23",
+    "extra": {
+      "passRate1": 5.8,
+      "passNum1": 13,
+      "passNum2": 29,
+      "totalCostUsd": 0,
+      "secondsPerCase": 146.7,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 61
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "command-a-03-2025-quality",
+    "metric": "pass_rate_2",
+    "score": 12,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-03-14",
+    "extra": {
+      "passRate1": 2.2,
+      "passNum1": 5,
+      "passNum2": 27,
+      "totalCostUsd": 0,
+      "secondsPerCase": 85.1,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 62
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Codestral 25.01",
+    "metric": "pass_rate_2",
+    "score": 11.1,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-01-13",
+    "extra": {
+      "passRate1": 4,
+      "passNum1": 9,
+      "passNum2": 25,
+      "totalCostUsd": 1.9834,
+      "secondsPerCase": 9.3,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 63
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "openhands-lm-32b-v0.1",
+    "metric": "pass_rate_2",
+    "score": 10.2,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-19",
+    "extra": {
+      "passRate1": 4,
+      "passNum1": 9,
+      "passNum2": 23,
+      "totalCostUsd": 0,
+      "secondsPerCase": 195.6,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 64
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-4.1-nano",
+    "metric": "pass_rate_2",
+    "score": 8.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-04-14",
+    "extra": {
+      "passRate1": 3.1,
+      "passNum1": 7,
+      "passNum2": 20,
+      "totalCostUsd": 0.4281,
+      "secondsPerCase": 12,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 65
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "Qwen2.5-Coder-32B-Instruct",
+    "metric": "pass_rate_2",
+    "score": 8,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-22",
+    "extra": {
+      "passRate1": 4.4,
+      "passNum1": 10,
+      "passNum2": 18,
+      "totalCostUsd": 0,
+      "secondsPerCase": 84.4,
+      "editFormat": "diff"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 66
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gemma-3-27b-it",
+    "metric": "pass_rate_2",
+    "score": 4.9,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2025-03-15",
+    "extra": {
+      "passRate1": 1.8,
+      "passNum1": 4,
+      "passNum2": 11,
+      "totalCostUsd": 0,
+      "secondsPerCase": 79.7,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 67
+  },
+  {
+    "source": "aider_polyglot",
+    "benchmark": "aider_polyglot",
+    "url": "https://aider.chat/docs/leaderboards/",
+    "modelRaw": "gpt-4o-mini-2024-07-18",
+    "metric": "pass_rate_2",
+    "score": 3.6,
+    "scoreScale": "percent",
+    "sampleSize": 225,
+    "date": "2024-12-21",
+    "extra": {
+      "passRate1": 0.9,
+      "passNum1": 2,
+      "passNum2": 8,
+      "totalCostUsd": 0.3236,
+      "secondsPerCase": 17.3,
+      "editFormat": "whole"
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.aider_polyglot.json",
+    "snapshotIndex": 68
+  },
+  {
+    "source": "terminal_bench",
+    "benchmark": "terminal_bench",
+    "url": "https://www.tbench.ai/leaderboard/terminal-bench/2.1",
+    "modelRaw": "Terminus 2 GPT-5.5",
+    "metric": "accuracy",
+    "score": 0.7820224719101123,
+    "scoreScale": "0-1",
+    "date": "2026-05-01",
+    "extra": {
+      "agent": "Terminus 2",
+      "model": [
+        "GPT-5.5"
+      ],
+      "stderr": 0.011996717137113833,
+      "verified": true,
+      "agentName": "terminus-2",
+      "agentVersion": "2.0.0",
+      "modelNames": [
+        "gpt-5.5"
+      ],
+      "modelProviders": [
+        "openai"
+      ]
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.terminal_bench_2_1.json",
+    "snapshotIndex": 2
+  },
+  {
+    "source": "terminal_bench",
+    "benchmark": "terminal_bench",
+    "url": "https://www.tbench.ai/leaderboard/terminal-bench/2.1",
+    "modelRaw": "Terminus 2 Claude Opus 4.8",
+    "metric": "accuracy",
+    "score": 0.7460674157303371,
+    "scoreScale": "0-1",
+    "date": "2026-05-29",
+    "extra": {
+      "agent": "Terminus 2",
+      "model": [
+        "Claude Opus 4.8"
+      ],
+      "stderr": 0.012308372078767778,
+      "verified": true,
+      "agentName": "terminus-2",
+      "agentVersion": "2.0.0",
+      "modelNames": [
+        "claude-opus-4-8"
+      ],
+      "modelProviders": [
+        "anthropic"
+      ]
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.terminal_bench_2_1.json",
+    "snapshotIndex": 3
+  },
+  {
+    "source": "terminal_bench",
+    "benchmark": "terminal_bench",
+    "url": "https://www.tbench.ai/leaderboard/terminal-bench/2.1",
+    "modelRaw": "Terminus 2 Gemini 3 Pro",
+    "metric": "accuracy",
+    "score": 0.7438202247191011,
+    "scoreScale": "0-1",
+    "date": "2026-05-01",
+    "extra": {
+      "agent": "Terminus 2",
+      "model": [
+        "Gemini 3 Pro"
+      ],
+      "stderr": 0.013199258566821045,
+      "verified": true,
+      "agentName": "terminus-2",
+      "agentVersion": "2.0.0",
+      "modelNames": [
+        "gemini-3-pro-preview"
+      ],
+      "modelProviders": [
+        "gemini"
+      ]
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.terminal_bench_2_1.json",
+    "snapshotIndex": 4
+  },
+  {
+    "source": "terminal_bench",
+    "benchmark": "terminal_bench",
+    "url": "https://www.tbench.ai/leaderboard/terminal-bench/2.1",
+    "modelRaw": "Terminus 2 Gemini 3.1 Pro",
+    "metric": "accuracy",
+    "score": 0.7031835205992509,
+    "scoreScale": "0-1",
+    "date": "2026-05-05",
+    "extra": {
+      "agent": "Terminus 2",
+      "model": [
+        "Gemini 3.1 Pro"
+      ],
+      "stderr": 0.014791636846043224,
+      "verified": true,
+      "agentName": "terminus-2",
+      "agentVersion": "2.0.0",
+      "modelNames": [
+        "gemini-3.1-pro-preview"
+      ],
+      "modelProviders": [
+        "gemini"
+      ]
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.terminal_bench_2_1.json",
+    "snapshotIndex": 6
+  },
+  {
+    "source": "terminal_bench",
+    "benchmark": "terminal_bench",
+    "url": "https://www.tbench.ai/leaderboard/terminal-bench/2.1",
+    "modelRaw": "Terminus 2 Claude Opus 4.7",
+    "metric": "accuracy",
+    "score": 0.6606741573033708,
+    "scoreScale": "0-1",
+    "date": "2026-05-01",
+    "extra": {
+      "agent": "Terminus 2",
+      "model": [
+        "Claude Opus 4.7"
+      ],
+      "stderr": 0.013669129281569032,
+      "verified": true,
+      "agentName": "terminus-2",
+      "agentVersion": "2.0.0",
+      "modelNames": [
+        "claude-opus-4-7"
+      ],
+      "modelProviders": [
+        "anthropic"
+      ]
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.terminal_bench_2_1.json",
+    "snapshotIndex": 9
+  },
+  {
+    "source": "terminal_bench",
+    "benchmark": "terminal_bench",
+    "url": "https://www.tbench.ai/leaderboard/terminal-bench/2.1",
+    "modelRaw": "Claude Code GLM 5.1",
+    "metric": "accuracy",
+    "score": 0.5865168539325842,
+    "scoreScale": "0-1",
+    "date": "2026-05-02",
+    "extra": {
+      "agent": "Claude Code",
+      "model": [
+        "GLM 5.1"
+      ],
+      "stderr": 0.012410517996839619,
+      "verified": true,
+      "agentName": "claude-code",
+      "agentVersion": "2.1.123",
+      "modelNames": [
+        "glm-5.1"
+      ],
+      "modelProviders": [
+        "z-ai"
+      ]
+    },
+    "snapshotFile": "tools\\scorecard\\snapshots\\live.terminal_bench_2_1.json",
+    "snapshotIndex": 10
+  }
+]