{
  "checked_on": "2026-10-08",
  "ahe": {
    "source": "https://arxiv.org/html/2604.25850v3",
    "location": "Section 4.3, Figure 3",
    "benchmark": "Terminal-Bench 2",
    "tasks": 89,
    "source_model": "GPT-5.4 high",
    "target_re_evolution": false,
    "models": [
      "deepseek-v4-flash",
      "qwen-3.6-plus",
      "gemini-3.1-flash-lite-preview"
    ],
    "seed_pass1_percent": [
      51.7,
      56.2,
      36.5
    ],
    "evolved_pass1_percent": [
      61.8,
      62.5,
      41.6
    ]
  },
  "crossed_matrix": {
    "status": "hypothetical, not empirical",
    "rows": [
      "A",
      "B"
    ],
    "columns": [
      "H_A",
      "H_B"
    ],
    "success_percent": [
      [
        70,
        60
      ],
      [
        50,
        75
      ]
    ],
    "replacement_pp": -20,
    "adaptation_pp": 25,
    "interaction_pp": 35
  },
  "cost_example": {
    "status": "illustrative normalized units, not provider prices",
    "strong_route": 8,
    "cheap_route": 1,
    "verification": 0.2,
    "formula": "1.2 + 8*f",
    "fallback_break_even": 0.85
  },
  "conceptual_figures": [
    "loop",
    "architecture"
  ]
}