[
  {
    "tag": "tickets_kev",
    "started_at": "2026-09-26T08:10:10.551641+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/systemone_perf.py",
      "--model",
      "kev",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/tickets_kev.json"
    ],
    "finished_at": "2026-09-26T08:10:22.279809+00:00"
  },
  {
    "tag": "narrow_kev",
    "started_at": "2026-09-26T08:10:22.284747+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/agent_claims/run.py",
      "--model",
      "kev",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/narrow_kev.json"
    ],
    "finished_at": "2026-09-26T08:10:34.874808+00:00"
  },
  {
    "tag": "tickets_jev",
    "started_at": "2026-09-26T08:10:34.879585+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/systemone_perf.py",
      "--model",
      "jev",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/tickets_jev.json"
    ],
    "finished_at": "2026-09-26T08:10:46.219835+00:00"
  },
  {
    "tag": "narrow_jev",
    "started_at": "2026-09-26T08:10:46.224858+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/agent_claims/run.py",
      "--model",
      "jev",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/narrow_jev.json"
    ],
    "finished_at": "2026-09-26T08:11:36.235412+00:00"
  },
  {
    "tag": "bfcl_description_kev",
    "started_at": "2026-09-26T08:12:36.844863+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/bfcl_v4/run.py",
      "--model",
      "kev",
      "--view",
      "description",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/bfcl_description_kev.json"
    ],
    "finished_at": "2026-09-26T08:13:26.717342+00:00"
  },
  {
    "tag": "bfcl_description_jev",
    "started_at": "2026-09-26T08:13:26.722916+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/bfcl_v4/run.py",
      "--model",
      "jev",
      "--view",
      "description",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/bfcl_description_jev.json"
    ],
    "finished_at": "2026-09-26T08:16:17.487346+00:00"
  },
  {
    "tag": "bfcl_schema_kev",
    "started_at": "2026-09-26T08:17:18.129776+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/bfcl_v4/run.py",
      "--model",
      "kev",
      "--view",
      "schema",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/bfcl_schema_kev.json"
    ],
    "finished_at": "2026-09-26T08:18:57.744702+00:00"
  },
  {
    "tag": "bfcl_schema_jev",
    "started_at": "2026-09-26T08:18:57.751020+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/bfcl_v4/run.py",
      "--model",
      "jev",
      "--view",
      "schema",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/bfcl_schema_jev.json"
    ],
    "finished_at": "2026-09-26T08:21:55.794726+00:00"
  },
  {
    "tag": "bfcl_prose_schema_kev",
    "started_at": "2026-09-26T08:22:54.361495+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/bfcl_v4/run.py",
      "--model",
      "kev",
      "--view",
      "prose_schema",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/bfcl_prose_schema_kev.json"
    ],
    "finished_at": "2026-09-26T08:24:23.064734+00:00"
  },
  {
    "tag": "bfcl_prose_schema_jev",
    "started_at": "2026-09-26T08:29:17.658264+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "evaluation/bfcl_v4/run.py",
      "--model",
      "jev",
      "--view",
      "prose_schema",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/bfcl_prose_schema_jev.json"
    ],
    "aborted_attempts": [
      {
        "started_at": "2026-09-26T08:24:23.071540+00:00",
        "command": [
          "/workspace/kev-jev-eval/.venv/bin/python",
          "evaluation/bfcl_v4/run.py",
          "--model",
          "jev",
          "--view",
          "prose_schema",
          "--out",
          "/workspace/kev-jev-eval/results/2026-09-26/bfcl_prose_schema_jev.json"
        ]
      }
    ],
    "finished_at": "2026-09-26T08:32:12.262742+00:00"
  },
  {
    "tag": "policy_kev_noshield",
    "started_at": "2026-09-26T08:32:12.270082+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "examples/t_rex/run.py",
      "--model",
      "kev",
      "--seeds",
      "5",
      "--duration",
      "60",
      "--inflight",
      "6",
      "--course-style",
      "original",
      "--prompt",
      "labeled",
      "--no-shield",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/policy_kev_noshield.json"
    ],
    "finished_at": "2026-09-26T08:37:21.564986+00:00"
  },
  {
    "tag": "policy_jev_noshield",
    "started_at": "2026-09-26T08:37:21.572760+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "examples/t_rex/run.py",
      "--model",
      "jev",
      "--seeds",
      "5",
      "--duration",
      "60",
      "--inflight",
      "6",
      "--course-style",
      "original",
      "--prompt",
      "labeled",
      "--no-shield",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/policy_jev_noshield.json"
    ],
    "finished_at": "2026-09-26T08:42:33.502115+00:00"
  },
  {
    "tag": "policy_kev_shield",
    "started_at": "2026-09-26T08:42:33.534600+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "examples/t_rex/run.py",
      "--model",
      "kev",
      "--seeds",
      "5",
      "--duration",
      "60",
      "--inflight",
      "6",
      "--course-style",
      "original",
      "--prompt",
      "labeled",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/policy_kev_shield.json"
    ],
    "finished_at": "2026-09-26T08:47:43.016995+00:00"
  },
  {
    "tag": "policy_jev_shield",
    "started_at": "2026-09-26T08:47:43.024559+00:00",
    "command": [
      "/workspace/kev-jev-eval/.venv/bin/python",
      "examples/t_rex/run.py",
      "--model",
      "jev",
      "--seeds",
      "5",
      "--duration",
      "60",
      "--inflight",
      "6",
      "--course-style",
      "original",
      "--prompt",
      "labeled",
      "--out",
      "/workspace/kev-jev-eval/results/2026-09-26/policy_jev_shield.json"
    ],
    "finished_at": "2026-09-26T08:52:55.692807+00:00"
  }
]