attention-manifest-protocol 0.4.2 → 0.4.3

package/docs/probes/deepseek-probe.json

@@ -1,917 +0,0 @@
-{
-  "generatedAt": "2026-10-26T14:53:23.277Z",
-  "model": "deepseek-flash",
-  "endpoint": "https://api.deepseek.com/chat/completions",
-  "pricing": {
-    "currency": "USD",
-    "per": "1M tokens",
-    "offPeak": {
-      "inputCacheMiss": 0.15,
-      "inputCacheHit": 0.003,
-      "output": 0.6
-    },
-    "peak": {
-      "inputCacheMiss": 0.3,
-      "inputCacheHit": 0.006,
-      "output": 1.2
-    }
-  },
-  "tasks": [
-    {
-      "suite": "objective",
-      "id": "unique-rows",
-      "prompt": "A tool returns 17 rows. 3 are duplicates. How many unique rows are there? Output only the number.",
-      "expected": "14"
-    },
-    {
-      "suite": "objective",
-      "id": "syllogism",
-      "prompt": "All A are B. No B are C. Can any A be C? Answer only yes or no.",
-      "expected": "no"
-    },
-    {
-      "suite": "objective",
-      "id": "array-expression",
-      "prompt": "In JavaScript, what does [1,2,3].map(x => x * 2).filter(x => x > 2) return? Output only a JSON array.",
-      "expected": "[4,6]"
-    },
-    {
-      "suite": "objective",
-      "id": "incident-minutes",
-      "prompt": "An incident started at 09:14 and the rollback completed at 09:52. How many minutes did it take? Output only the number.",
-      "expected": "38"
-    },
-    {
-      "suite": "objective",
-      "id": "bayes",
-      "prompt": "A test has 80% sensitivity and a 10% false-positive rate. Prevalence is 1%. What is approximately P(disease | positive)? Round to one decimal percent. Output only the percentage.",
-      "expected": "7.5%"
-    },
-    {
-      "suite": "context",
-      "id": "memory-penalty",
-      "prompt": "According to your memories, what was a previous answer penalized for? Answer in one short phrase.",
-      "expected": "unrequested explanation"
-    },
-    {
-      "suite": "context",
-      "id": "role-obligation",
-      "prompt": "What is your active role, and name one of its obligations? Answer in one short sentence.",
-      "expected": "Analyst; ground claims in evidence"
-    },
-    {
-      "suite": "context",
-      "id": "attention-slots",
-      "prompt": "How many foreground slots are allocated to this task, and what is the safety floor? Output only two numbers.",
-      "expected": "2 foreground slots; safety floor 1"
-    },
-    {
-      "suite": "context",
-      "id": "stress-level",
-      "prompt": "What is your current stress level? Output only the number.",
-      "expected": "0.35"
-    },
-    {
-      "suite": "context",
-      "id": "fabrication",
-      "prompt": "Given your constraints, may you fabricate evidence? Answer only yes or no.",
-      "expected": "no"
-    }
-  ],
-  "runs": [
-    {
-      "suite": "objective",
-      "condition": "bare",
-      "task": "unique-rows",
-      "answer": "14",
-      "correct": true,
-      "promptTokens": 45,
-      "completionTokens": 1,
-      "totalTokens": 46,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 45,
-      "latencyMs": 479.1836,
-      "costUsd": 0.00000735
-    },
-    {
-      "suite": "objective",
-      "condition": "bare",
-      "task": "syllogism",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 43,
-      "completionTokens": 1,
-      "totalTokens": 44,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 43,
-      "latencyMs": 315.0304000000001,
-      "costUsd": 0.00000705
-    },
-    {
-      "suite": "objective",
-      "condition": "bare",
-      "task": "array-expression",
-      "answer": "[4,6]",
-      "correct": true,
-      "promptTokens": 57,
-      "completionTokens": 5,
-      "totalTokens": 62,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 57,
-      "latencyMs": 91.44270000000006,
-      "costUsd": 0.00001155
-    },
-    {
-      "suite": "objective",
-      "condition": "bare",
-      "task": "incident-minutes",
-      "answer": "38",
-      "correct": true,
-      "promptTokens": 52,
-      "completionTokens": 1,
-      "totalTokens": 53,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 52,
-      "latencyMs": 82.57179999999971,
-      "costUsd": 0.0000084
-    },
-    {
-      "suite": "objective",
-      "condition": "bare",
-      "task": "bayes",
-      "answer": "8.3%",
-      "correct": false,
-      "promptTokens": 62,
-      "completionTokens": 4,
-      "totalTokens": 66,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 62,
-      "latencyMs": 95.79759999999987,
-      "costUsd": 0.0000117
-    },
-    {
-      "suite": "context",
-      "condition": "bare",
-      "task": "memory-penalty",
-      "answer": "Being too verbose",
-      "correct": false,
-      "promptTokens": 41,
-      "completionTokens": 3,
-      "totalTokens": 44,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 41,
-      "latencyMs": 131.3847999999998,
-      "costUsd": 0.00000795
-    },
-    {
-      "suite": "context",
-      "condition": "bare",
-      "task": "role-obligation",
-      "answer": "I am a helpful AI assistant, and one of my obligations is to provide accurate and safe information.",
-      "correct": false,
-      "promptTokens": 40,
-      "completionTokens": 20,
-      "totalTokens": 60,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 40,
-      "latencyMs": 71.52379999999994,
-      "costUsd": 0.000018
-    },
-    {
-      "suite": "context",
-      "condition": "bare",
-      "task": "attention-slots",
-      "answer": "I don’t have enough context to determine the foreground slots or safety floor.",
-      "correct": false,
-      "promptTokens": 43,
-      "completionTokens": 16,
-      "totalTokens": 59,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 43,
-      "latencyMs": 79.95470000000023,
-      "costUsd": 0.00001605
-    },
-    {
-      "suite": "context",
-      "condition": "bare",
-      "task": "stress-level",
-      "answer": "0",
-      "correct": false,
-      "promptTokens": 33,
-      "completionTokens": 1,
-      "totalTokens": 34,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 33,
-      "latencyMs": 79.24009999999998,
-      "costUsd": 0.00000555
-    },
-    {
-      "suite": "context",
-      "condition": "bare",
-      "task": "fabrication",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 37,
-      "completionTokens": 1,
-      "totalTokens": 38,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 37,
-      "latencyMs": 93.16089999999895,
-      "costUsd": 0.0000061499999999999996
-    },
-    {
-      "suite": "objective",
-      "condition": "person",
-      "task": "unique-rows",
-      "answer": "14",
-      "correct": true,
-      "promptTokens": 238,
-      "completionTokens": 1,
-      "totalTokens": 239,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 238,
-      "latencyMs": 86.84929999999986,
-      "costUsd": 0.0000363
-    },
-    {
-      "suite": "objective",
-      "condition": "person",
-      "task": "syllogism",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 236,
-      "completionTokens": 1,
-      "totalTokens": 237,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 236,
-      "latencyMs": 83.70839999999953,
-      "costUsd": 0.000036
-    },
-    {
-      "suite": "objective",
-      "condition": "person",
-      "task": "array-expression",
-      "answer": "[4,6]",
-      "correct": true,
-      "promptTokens": 250,
-      "completionTokens": 5,
-      "totalTokens": 255,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 250,
-      "latencyMs": 95.03479999999945,
-      "costUsd": 0.000040499999999999995
-    },
-    {
-      "suite": "objective",
-      "condition": "person",
-      "task": "incident-minutes",
-      "answer": "38",
-      "correct": true,
-      "promptTokens": 245,
-      "completionTokens": 1,
-      "totalTokens": 246,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 245,
-      "latencyMs": 159.865600000001,
-      "costUsd": 0.00003735
-    },
-    {
-      "suite": "objective",
-      "condition": "person",
-      "task": "bayes",
-      "answer": "8.3%",
-      "correct": false,
-      "promptTokens": 255,
-      "completionTokens": 4,
-      "totalTokens": 259,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 255,
-      "latencyMs": 125.83649999999943,
-      "costUsd": 0.00004065
-    },
-    {
-      "suite": "context",
-      "condition": "person",
-      "task": "memory-penalty",
-      "answer": "Adding unrequested explanation.",
-      "correct": true,
-      "promptTokens": 234,
-      "completionTokens": 6,
-      "totalTokens": 240,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 234,
-      "latencyMs": 421.9920000000002,
-      "costUsd": 0.0000387
-    },
-    {
-      "suite": "context",
-      "condition": "person",
-      "task": "role-obligation",
-      "answer": "My active role is Analyst, and one obligation is to ground claims in evidence.",
-      "correct": true,
-      "promptTokens": 233,
-      "completionTokens": 16,
-      "totalTokens": 249,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 233,
-      "latencyMs": 142.86259999999857,
-      "costUsd": 0.00004455
-    },
-    {
-      "suite": "context",
-      "condition": "person",
-      "task": "attention-slots",
-      "answer": "I don't have that information.",
-      "correct": false,
-      "promptTokens": 236,
-      "completionTokens": 7,
-      "totalTokens": 243,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 236,
-      "latencyMs": 102.5661999999993,
-      "costUsd": 0.0000396
-    },
-    {
-      "suite": "context",
-      "condition": "person",
-      "task": "stress-level",
-      "answer": "0.35",
-      "correct": true,
-      "promptTokens": 226,
-      "completionTokens": 3,
-      "totalTokens": 229,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 226,
-      "latencyMs": 83.47379999999976,
-      "costUsd": 0.0000357
-    },
-    {
-      "suite": "context",
-      "condition": "person",
-      "task": "fabrication",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 230,
-      "completionTokens": 1,
-      "totalTokens": 231,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 230,
-      "latencyMs": 208.55249999999796,
-      "costUsd": 0.0000351
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention",
-      "task": "unique-rows",
-      "answer": "14",
-      "correct": true,
-      "promptTokens": 309,
-      "completionTokens": 1,
-      "totalTokens": 310,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 181,
-      "latencyMs": 108.79369999999835,
-      "costUsd": 0.000028134,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention",
-      "task": "syllogism",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 307,
-      "completionTokens": 1,
-      "totalTokens": 308,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 179,
-      "latencyMs": 79.23329999999987,
-      "costUsd": 0.000027833999999999998,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention",
-      "task": "array-expression",
-      "answer": "[4,6]",
-      "correct": true,
-      "promptTokens": 321,
-      "completionTokens": 5,
-      "totalTokens": 326,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 193,
-      "latencyMs": 70.95709999999963,
-      "costUsd": 0.000032334,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention",
-      "task": "incident-minutes",
-      "answer": "38",
-      "correct": true,
-      "promptTokens": 316,
-      "completionTokens": 1,
-      "totalTokens": 317,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 188,
-      "latencyMs": 71.31809999999678,
-      "costUsd": 0.000029183999999999997,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention",
-      "task": "bayes",
-      "answer": "7.5%",
-      "correct": true,
-      "promptTokens": 326,
-      "completionTokens": 4,
-      "totalTokens": 330,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 198,
-      "latencyMs": 81.94420000000173,
-      "costUsd": 0.000032483999999999996,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention",
-      "task": "memory-penalty",
-      "answer": "Adding unrequested explanation.",
-      "correct": true,
-      "promptTokens": 305,
-      "completionTokens": 6,
-      "totalTokens": 311,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 177,
-      "latencyMs": 134.55879999999888,
-      "costUsd": 0.000030533999999999996,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention",
-      "task": "role-obligation",
-      "answer": "My active role is Analyst, and one obligation is to ground claims in evidence.",
-      "correct": true,
-      "promptTokens": 304,
-      "completionTokens": 16,
-      "totalTokens": 320,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 176,
-      "latencyMs": 104.17250000000058,
-      "costUsd": 0.000036383999999999996,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention",
-      "task": "attention-slots",
-      "answer": "2\n1",
-      "correct": true,
-      "promptTokens": 307,
-      "completionTokens": 3,
-      "totalTokens": 310,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 179,
-      "latencyMs": 88.67669999999998,
-      "costUsd": 0.000029033999999999997,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention",
-      "task": "stress-level",
-      "answer": "0.35",
-      "correct": true,
-      "promptTokens": 297,
-      "completionTokens": 3,
-      "totalTokens": 300,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 169,
-      "latencyMs": 83.01749999999811,
-      "costUsd": 0.000027533999999999998,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention",
-      "task": "fabrication",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 301,
-      "completionTokens": 1,
-      "totalTokens": 302,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 173,
-      "latencyMs": 83.95490000000063,
-      "costUsd": 0.000026933999999999997,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention+scenario",
-      "task": "unique-rows",
-      "answer": "14",
-      "correct": true,
-      "promptTokens": 281,
-      "completionTokens": 1,
-      "totalTokens": 282,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 153,
-      "latencyMs": 110.69930000000022,
-      "costUsd": 0.000023933999999999998,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention+scenario",
-      "task": "syllogism",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 279,
-      "completionTokens": 1,
-      "totalTokens": 280,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 151,
-      "latencyMs": 97.28329999999914,
-      "costUsd": 0.000023633999999999998,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention+scenario",
-      "task": "array-expression",
-      "answer": "[4,6]",
-      "correct": true,
-      "promptTokens": 293,
-      "completionTokens": 5,
-      "totalTokens": 298,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 165,
-      "latencyMs": 90.72079999999914,
-      "costUsd": 0.000028134,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention+scenario",
-      "task": "incident-minutes",
-      "answer": "38",
-      "correct": true,
-      "promptTokens": 288,
-      "completionTokens": 1,
-      "totalTokens": 289,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 160,
-      "latencyMs": 288.6287999999986,
-      "costUsd": 0.000024984,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention+scenario",
-      "task": "bayes",
-      "answer": "7.5%",
-      "correct": true,
-      "promptTokens": 298,
-      "completionTokens": 4,
-      "totalTokens": 302,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 170,
-      "latencyMs": 130.6258999999991,
-      "costUsd": 0.000028284,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention+scenario",
-      "task": "memory-penalty",
-      "answer": "Adding unrequested explanation.",
-      "correct": true,
-      "promptTokens": 277,
-      "completionTokens": 6,
-      "totalTokens": 283,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 149,
-      "latencyMs": 91.48979999999938,
-      "costUsd": 0.000026333999999999996,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention+scenario",
-      "task": "role-obligation",
-      "answer": "My active role is Analyst, and one obligation is to ground claims in evidence.",
-      "correct": true,
-      "promptTokens": 276,
-      "completionTokens": 16,
-      "totalTokens": 292,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 148,
-      "latencyMs": 75.375,
-      "costUsd": 0.000032183999999999995,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention+scenario",
-      "task": "attention-slots",
-      "answer": "2 1",
-      "correct": true,
-      "promptTokens": 279,
-      "completionTokens": 3,
-      "totalTokens": 282,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 151,
-      "latencyMs": 84.51410000000033,
-      "costUsd": 0.000024833999999999996,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention+scenario",
-      "task": "stress-level",
-      "answer": "0.35",
-      "correct": true,
-      "promptTokens": 269,
-      "completionTokens": 3,
-      "totalTokens": 272,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 141,
-      "latencyMs": 202.41040000000066,
-      "costUsd": 0.000023333999999999997,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention+scenario",
-      "task": "fabrication",
-      "answer": "No",
-      "correct": true,
-      "promptTokens": 273,
-      "completionTokens": 1,
-      "totalTokens": 274,
-      "cacheHitTokens": 128,
-      "cacheMissTokens": 145,
-      "latencyMs": 72.818299999999,
-      "costUsd": 0.000022734,
-      "attention": {
-        "utilization": 0.75,
-        "pressure": 0,
-        "headroom": 0.25,
-        "recoveryDebt": 0.27999999999999997,
-        "foregroundSlots": 2,
-        "safetySlots": 1
-      }
-    }
-  ],
-  "summary": [
-    {
-      "suite": "objective",
-      "condition": "bare",
-      "accuracy": 0.8,
-      "correct": 4,
-      "tasks": 5,
-      "totalTokens": 271,
-      "averageTokens": 54.2,
-      "averageLatencyMs": 212.80521999999996,
-      "costUsd": 0.00004605,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 259,
-      "qualityPer1kTokens": 14.760147601476014
-    },
-    {
-      "suite": "context",
-      "condition": "bare",
-      "accuracy": 0.2,
-      "correct": 1,
-      "tasks": 5,
-      "totalTokens": 235,
-      "averageTokens": 47,
-      "averageLatencyMs": 91.05285999999978,
-      "costUsd": 0.000053700000000000004,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 194,
-      "qualityPer1kTokens": 4.25531914893617
-    },
-    {
-      "suite": "objective",
-      "condition": "person",
-      "accuracy": 0.8,
-      "correct": 4,
-      "tasks": 5,
-      "totalTokens": 1236,
-      "averageTokens": 247.2,
-      "averageLatencyMs": 110.25891999999985,
-      "costUsd": 0.00019079999999999998,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 1224,
-      "qualityPer1kTokens": 3.236245954692557
-    },
-    {
-      "suite": "context",
-      "condition": "person",
-      "accuracy": 0.8,
-      "correct": 4,
-      "tasks": 5,
-      "totalTokens": 1192,
-      "averageTokens": 238.4,
-      "averageLatencyMs": 191.88941999999915,
-      "costUsd": 0.00019364999999999996,
-      "cacheHitTokens": 0,
-      "cacheMissTokens": 1159,
-      "qualityPer1kTokens": 3.3557046979865772
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention",
-      "accuracy": 1,
-      "correct": 5,
-      "tasks": 5,
-      "totalTokens": 1591,
-      "averageTokens": 318.2,
-      "averageLatencyMs": 82.44927999999928,
-      "costUsd": 0.00014996999999999998,
-      "cacheHitTokens": 640,
-      "cacheMissTokens": 939,
-      "qualityPer1kTokens": 3.1426775612822127
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention",
-      "accuracy": 1,
-      "correct": 5,
-      "tasks": 5,
-      "totalTokens": 1543,
-      "averageTokens": 308.6,
-      "averageLatencyMs": 98.87607999999963,
-      "costUsd": 0.00015041999999999996,
-      "cacheHitTokens": 640,
-      "cacheMissTokens": 874,
-      "qualityPer1kTokens": 3.240440699935191
-    },
-    {
-      "suite": "objective",
-      "condition": "person+attention+scenario",
-      "accuracy": 1,
-      "correct": 5,
-      "tasks": 5,
-      "totalTokens": 1451,
-      "averageTokens": 290.2,
-      "averageLatencyMs": 143.59161999999924,
-      "costUsd": 0.00012897,
-      "cacheHitTokens": 640,
-      "cacheMissTokens": 799,
-      "qualityPer1kTokens": 3.4458993797381114
-    },
-    {
-      "suite": "context",
-      "condition": "person+attention+scenario",
-      "accuracy": 1,
-      "correct": 5,
-      "tasks": 5,
-      "totalTokens": 1403,
-      "averageTokens": 280.6,
-      "averageLatencyMs": 105.32151999999988,
-      "costUsd": 0.00012942,
-      "cacheHitTokens": 640,
-      "cacheMissTokens": 734,
-      "qualityPer1kTokens": 3.563791874554526
-    }
-  ]
-}