{
  "schema": 1,
  "exp_id": "exp-20260918-ep1-f005-glm-probe1200",
  "title": "GLM 旗舰加长死线探测(1200s · 独立 probe,不并入主对照)",
  "generated_at": "2026-09-21T08:54:26+00:00",
  "generator": "scripts/snapshot-experiment.py（重跑可再生；禁止手改本文件）",
  "design": {
    "independent_variable": "model",
    "note": "控制变量硬要求见 schema/SCHEMA.md「实验环境与隔离」节；夹具见 fixture 块"
  },
  "fixture": {
    "schema": 1,
    "frozen_at": "2026-09-15T06:41:51+00:00",
    "image_tag": "0genlab/hermes-harness:ep1",
    "image_id": "sha256:42357c39e76d906d451ff2fb54c88666078691d6e5b4787f5476a0faf00fb003",
    "base_image": "unknown",
    "hermes_commit": "4c1f53be10d0fce1d25aee1975e5149b6c54f25a",
    "hermes_version_in_image": "Hermes Agent v0.20.5 (2026.8.19)",
    "python_in_image": "3.11.16",
    "anthropic_in_image": "0.94.0",
    "node_in_image": null,
    "adapter_sha256": "adf39193da22331a4e6cff16b5470c012cf1d23ce43d32a625794d999fd28334",
    "leak_guard": "pass: adapter 无 F-005 修复模式",
    "limits": {
      "cpus": "2",
      "memory": "4g",
      "pids_limit": 512
    },
    "network": "0genlab-int (internal; 出网仅经 0genlab-relaygw 边车到宿主 relay:8484)"
  },
  "runs": [
    {
      "run_id": "hermes-glm5.3-ep1-f005-r101",
      "model": "glm-5.3",
      "task": "ep1-f005",
      "task_type": "source",
      "attempt": 101,
      "score": 1,
      "wall_s": 601,
      "agent_exit": 0,
      "ts": "2026-09-18T03:55:30+00:00",
      "params": {
        "provider": "aihubmix",
        "agent": "Hermes Agent v0.20.5 (2026.8.19)",
        "harness": "docker",
        "timeout_s": 1200,
        "max_turns": 20,
        "sandbox": false,
        "strict_home": false
      },
      "env": {
        "host": {
          "os": "Darwin 25.4.0",
          "arch": "arm64"
        },
        "agent": "Hermes Agent v0.20.5 (2026.8.19)",
        "tool_versions": {
          "python3": "3.11.16",
          "node": null,
          "anthropic": "0.94.0"
        },
        "sandbox_kind": "docker",
        "env_policy": "container-clean: 容器 env 仅显式注入 KEY_ENV+HERMES_HOME，宿主凭据不存在",
        "docker": {
          "image_tag": "0genlab/hermes-harness:ep1",
          "image_id": "sha256:42357c39e76d906d451ff2fb54c88666078691d6e5b4787f5476a0faf00fb003",
          "base_image": "unknown",
          "hermes_commit": "4c1f53be10d0fce1d25aee1975e5149b6c54f25a",
          "container_id": "100b11e740a3d435c86e88301cd23eb21ed8479e0e6d50b2df070bed1108e434",
          "limits": {
            "cpus": "2",
            "memory": "4g",
            "pids_limit": 512
          },
          "network": "0genlab-int (internal; 出网仅经 0genlab-relaygw 边车到宿主 relay:8484)"
        }
      },
      "score_diag": {
        "test_passed": true,
        "test_modified_by_model": false,
        "py": "python3.11"
      },
      "official_usage": {
        "api_calls": 22,
        "input_tokens": 48096,
        "output_tokens": 35613,
        "cache_read_tokens": 668672,
        "total_tokens": 752381
      }
    },
    {
      "run_id": "hermes-glm5.3-ep1-f005-r102",
      "model": "glm-5.3",
      "task": "ep1-f005",
      "task_type": "source",
      "attempt": 102,
      "score": 1,
      "wall_s": 582,
      "agent_exit": 0,
      "ts": "2026-09-18T04:05:27+00:00",
      "params": {
        "provider": "aihubmix",
        "agent": "Hermes Agent v0.20.5 (2026.8.19)",
        "harness": "docker",
        "timeout_s": 1200,
        "max_turns": 20,
        "sandbox": false,
        "strict_home": false
      },
      "env": {
        "host": {
          "os": "Darwin 25.4.0",
          "arch": "arm64"
        },
        "agent": "Hermes Agent v0.20.5 (2026.8.19)",
        "tool_versions": {
          "python3": "3.11.16",
          "node": null,
          "anthropic": "0.94.0"
        },
        "sandbox_kind": "docker",
        "env_policy": "container-clean: 容器 env 仅显式注入 KEY_ENV+HERMES_HOME，宿主凭据不存在",
        "docker": {
          "image_tag": "0genlab/hermes-harness:ep1",
          "image_id": "sha256:42357c39e76d906d451ff2fb54c88666078691d6e5b4787f5476a0faf00fb003",
          "base_image": "unknown",
          "hermes_commit": "4c1f53be10d0fce1d25aee1975e5149b6c54f25a",
          "container_id": "7c120d15b316bc6841c4b80e62f2c57d2ea9d013d88f22226a237366f5fe3496",
          "limits": {
            "cpus": "2",
            "memory": "4g",
            "pids_limit": 512
          },
          "network": "0genlab-int (internal; 出网仅经 0genlab-relaygw 边车到宿主 relay:8484)"
        }
      },
      "score_diag": {
        "test_passed": true,
        "test_modified_by_model": false,
        "py": "python3.11"
      },
      "official_usage": {
        "api_calls": 22,
        "input_tokens": 40141,
        "output_tokens": 23123,
        "cache_read_tokens": 662272,
        "total_tokens": 725536
      }
    }
  ],
  "notes": [
    "背景:主对照 420s 死线下 glm-5.3 两跑均 exit124 撞墙 0/2。本组仅改 harness 层 TIMEOUT_S=1200(max_turns 20 不变、夹具镜像锁不变),探测'撞墙=不会做 还是 会做但慢'。",
    "结论:1200s 下 2/2 解出(r101 601s / r102 582s),均 exit0 自然收口(未撞 1200s 死线)、test 通过且未篡改=干净命中标准答案。GLM 病因=慢+啰嗦(output 仍 23k-36k),非死循环;420s 对它太紧。",
    "口径:两跑官方账本 input>output 未失真,src=official(区别于主线 420s 的 exit124→wire 回填)。api_calls=22 未达 max_turns 20 顶=墙钟而非轮次约束。"
  ]
}