{
  "recorded": "2026-10-31",
  "scope": "One training run; 23 held-out tasks with 8 attempts per model. Sanitized aggregates; no private task contents.",
  "heldout": {
    "base": {
      "n": 184,
      "passes": 91,
      "truncated": 104,
      "stops": {
        "agent_completed": 92,
        "max_turns": 92
      },
      "length_attempts": 63,
      "last_length": 12,
      "zero_tools": 4,
      "empty_bash_attempts": 28,
      "empty_bash_calls": 32,
      "length_calls": 84,
      "recovery_attempts": 56,
      "mean_output": 20029.510869565216,
      "mean_reasoning": 10856.369565217392,
      "mean_tools": 38.65760869565217,
      "mean_turns": 40.08695652173913,
      "mean_duplicate_bash": 2.8532608695652173
    },
    "final": {
      "n": 184,
      "passes": 83,
      "truncated": 124,
      "stops": {
        "agent_completed": 76,
        "max_turns": 108
      },
      "length_attempts": 75,
      "last_length": 16,
      "zero_tools": 10,
      "empty_bash_attempts": 17,
      "empty_bash_calls": 18,
      "length_calls": 96,
      "recovery_attempts": 68,
      "mean_output": 21684.429347826088,
      "mean_reasoning": 11963.70652173913,
      "mean_tools": 38.255434782608695,
      "mean_turns": 39.80434782608695,
      "mean_duplicate_bash": 3.5815217391304346
    }
  },
  "task_comparison": [
    {
      "label": "Task 01",
      "base": 3,
      "final": 0,
      "name": "Streaming top-k / quantile",
      "task_id": "task_00287_6ac070"
    },
    {
      "label": "Task 02",
      "base": 5,
      "final": 2,
      "name": "Session log analyzer",
      "task_id": "task_00800_759318"
    },
    {
      "label": "Task 03",
      "base": 6,
      "final": 3,
      "name": "CSV top-k / quantiles",
      "task_id": "task_02367_865b08"
    },
    {
      "label": "Task 04",
      "base": 6,
      "final": 4,
      "name": "Integer interval set",
      "task_id": "task_00156_ce278a"
    },
    {
      "label": "Task 05",
      "base": 8,
      "final": 6,
      "name": "JSON template renderer",
      "task_id": "task_00639_eff69f"
    },
    {
      "label": "Task 06",
      "base": 7,
      "final": 5,
      "name": "Configuration merger",
      "task_id": "task_01563_86cdc5"
    },
    {
      "label": "Task 07",
      "base": 7,
      "final": 6,
      "name": "Stdin template renderer",
      "task_id": "task_05899_42ab6b"
    },
    {
      "label": "Task 08",
      "base": 8,
      "final": 7,
      "name": "Named interval database",
      "task_id": "task_07916_0a26b8"
    },
    {
      "label": "Task 09",
      "base": 5,
      "final": 4,
      "name": "Numeric-file top-k",
      "task_id": "task_10467_5ddf2a"
    },
    {
      "label": "Task 10",
      "base": 4,
      "final": 3,
      "name": "Fixed-point file calculator",
      "task_id": "task_11226_38c42c"
    },
    {
      "label": "Task 11",
      "base": 6,
      "final": 5,
      "name": "Query-log filters",
      "task_id": "task_11922_d11fa3"
    },
    {
      "label": "Task 12",
      "base": 1,
      "final": 1,
      "name": "Weighted-grid pathfinder",
      "task_id": "task_00905_0d6c2e"
    },
    {
      "label": "Task 13",
      "base": 4,
      "final": 4,
      "name": "LEB128 TLV decoder",
      "task_id": "task_01164_9b9d29"
    },
    {
      "label": "Task 14",
      "base": 5,
      "final": 5,
      "name": "Multi-query pathfinder",
      "task_id": "task_01445_216433"
    },
    {
      "label": "Task 15",
      "base": 2,
      "final": 2,
      "name": "Token-bucket rate limiter",
      "task_id": "task_01618_dc6315"
    },
    {
      "label": "Task 16",
      "base": 0,
      "final": 0,
      "name": "WAL key-value store",
      "task_id": "task_04928_0f691a"
    },
    {
      "label": "Task 17",
      "base": 2,
      "final": 2,
      "name": "Dependency topological sort",
      "task_id": "task_05931_9632d7"
    },
    {
      "label": "Task 18",
      "base": 5,
      "final": 6,
      "name": "Log aggregation",
      "task_id": "task_14053_3db56d"
    },
    {
      "label": "Task 19",
      "base": 0,
      "final": 2,
      "name": "CSV reconciliation",
      "task_id": "task_00650_85d156"
    },
    {
      "label": "Task 20",
      "base": 2,
      "final": 4,
      "name": "JSON reconciliation",
      "task_id": "task_00970_4e2c86"
    },
    {
      "label": "Task 21",
      "base": 1,
      "final": 3,
      "name": "Fixed-point expressions",
      "task_id": "task_01066_b1cc81"
    },
    {
      "label": "Task 22",
      "base": 4,
      "final": 6,
      "name": "Custom-varint TLV decoder",
      "task_id": "task_03964_fc89b9"
    },
    {
      "label": "Task 23",
      "base": 0,
      "final": 3,
      "name": "CSV-to-JSON pipeline",
      "task_id": "task_04253_b72253"
    }
  ],
  "training": {
    "updates": 10,
    "accepted": 640,
    "returned": 1394,
    "reward_one_accepted": 365,
    "accepted_truncated": 452,
    "positive_truncated": 253,
    "tasks_encountered": 80,
    "tasks_accepted": 55,
    "wait_minutes": 141.54049779456668,
    "forward_backward_minutes": 52.53715875244999,
    "update_reward": [
      0.53125,
      0.546875,
      0.59375,
      0.515625,
      0.5,
      0.609375,
      0.65625,
      0.484375,
      0.59375,
      0.671875
    ],
    "metrics": [
      {
        "step": 1,
        "reward": 0.53125,
        "grad_norm": 0.007293731439858675,
        "entropy": 0.10540708154439926,
        "mismatch_kl": 0.0027642061468213797,
        "wait_minutes": 27.93504809621667,
        "forward_backward_minutes": 7.938453149149999,
        "stale_dropped": 0.0
      },
      {
        "step": 2,
        "reward": 0.546875,
        "grad_norm": 0.006163326092064381,
        "entropy": 0.11117411404848099,
        "mismatch_kl": 0.0029298386070877314,
        "wait_minutes": 0.0006903969166614843,
        "forward_backward_minutes": 5.545916594799996,
        "stale_dropped": 0.0
      },
      {
        "step": 3,
        "reward": 0.59375,
        "grad_norm": 0.0070924656465649605,
        "entropy": 0.10844198614358902,
        "mismatch_kl": 0.0028514026198536158,
        "wait_minutes": 24.798600505516674,
        "forward_backward_minutes": 3.802702845233322,
        "stale_dropped": 78.0
      },
      {
        "step": 4,
        "reward": 0.515625,
        "grad_norm": 0.005280227866023779,
        "entropy": 0.10037652403116226,
        "mismatch_kl": 0.0027440490666776896,
        "wait_minutes": 1.848779925150014,
        "forward_backward_minutes": 7.581878862933324,
        "stale_dropped": 7.0
      },
      {
        "step": 5,
        "reward": 0.5,
        "grad_norm": 0.0064736162312328815,
        "entropy": 0.11668164283037186,
        "mismatch_kl": 0.0029464552644640207,
        "wait_minutes": 21.669579380133335,
        "forward_backward_minutes": 4.536838990133325,
        "stale_dropped": 67.0
      },
      {
        "step": 6,
        "reward": 0.609375,
        "grad_norm": 0.006395513191819191,
        "entropy": 0.11659765243530273,
        "mismatch_kl": 0.003016715869307518,
        "wait_minutes": 9.678271393083332,
        "forward_backward_minutes": 4.732060453083325,
        "stale_dropped": 15.0
      },
      {
        "step": 7,
        "reward": 0.65625,
        "grad_norm": 0.007459586951881647,
        "entropy": 0.10617874562740326,
        "mismatch_kl": 0.0028260897379368544,
        "wait_minutes": 21.127875209566675,
        "forward_backward_minutes": 4.587754404033331,
        "stale_dropped": 26.0
      },
      {
        "step": 8,
        "reward": 0.484375,
        "grad_norm": 0.006458871997892857,
        "entropy": 0.1082637682557106,
        "mismatch_kl": 0.002825217554345727,
        "wait_minutes": 3.90688677566665,
        "forward_backward_minutes": 4.857642863433345,
        "stale_dropped": 25.0
      },
      {
        "step": 9,
        "reward": 0.59375,
        "grad_norm": 0.007094364147633314,
        "entropy": 0.12158043682575226,
        "mismatch_kl": 0.0031205129344016314,
        "wait_minutes": 19.35168140566666,
        "forward_backward_minutes": 3.9181843524166955,
        "stale_dropped": 64.0
      },
      {
        "step": 10,
        "reward": 0.671875,
        "grad_norm": 0.006522126030176878,
        "entropy": 0.11892890930175781,
        "mismatch_kl": 0.003122305264696479,
        "wait_minutes": 11.223084706649994,
        "forward_backward_minutes": 5.035726237233333,
        "stale_dropped": 24.0
      }
    ],
    "policy_counts": [
      {
        "version": 0,
        "returned": 285,
        "accepted": 128
      },
      {
        "version": 1,
        "returned": 40,
        "accepted": 32
      },
      {
        "version": 2,
        "returned": 245,
        "accepted": 96
      },
      {
        "version": 3,
        "returned": 84,
        "accepted": 40
      },
      {
        "version": 4,
        "returned": 188,
        "accepted": 88
      },
      {
        "version": 5,
        "returned": 107,
        "accepted": 56
      },
      {
        "version": 6,
        "returned": 179,
        "accepted": 72
      },
      {
        "version": 7,
        "returned": 91,
        "accepted": 56
      },
      {
        "version": 8,
        "returned": 154,
        "accepted": 72
      },
      {
        "version": 9,
        "returned": 21,
        "accepted": 0
      }
    ]
  },
  "benchmark": [
    {
      "concurrency": 16,
      "cases": [
        {
          "case": "uncached-long-prefill",
          "completed": 64,
          "failed": 0,
          "output_throughput": 434.96148687838155,
          "mean_e2el_ms": 37642.16675040625,
          "duration_seconds": 301.34162208400005
        },
        {
          "case": "shared-prefix-long-decode",
          "completed": 64,
          "failed": 0,
          "output_throughput": 636.8593315207307,
          "mean_e2el_ms": 102898.931167125,
          "duration_seconds": 823.239880537
        }
      ],
      "gpu_utilization": 95.43023255813954,
      "preemptions": 0.0,
      "preemptions_scope": "observed samples; 98 sampling ended before final requests"
    },
    {
      "concurrency": 32,
      "cases": [
        {
          "case": "uncached-long-prefill",
          "completed": 64,
          "failed": 0,
          "output_throughput": 597.8788855010604,
          "mean_e2el_ms": 54764.40546021875,
          "duration_seconds": 219.22834737699986
        },
        {
          "case": "shared-prefix-long-decode",
          "completed": 64,
          "failed": 0,
          "output_throughput": 980.7676399591567,
          "mean_e2el_ms": 133635.221267875,
          "duration_seconds": 534.5690239350001
        }
      ],
      "gpu_utilization": 93.96491228070175,
      "preemptions": 0.0,
      "preemptions_scope": "observed samples; 98 sampling ended before final requests"
    },
    {
      "concurrency": 64,
      "cases": [
        {
          "case": "uncached-long-prefill",
          "completed": 64,
          "failed": 0,
          "output_throughput": 714.2673780477401,
          "mean_e2el_ms": 91419.13818082813,
          "duration_seconds": 183.50551072099984
        },
        {
          "case": "shared-prefix-long-decode",
          "completed": 64,
          "failed": 0,
          "output_throughput": 1285.2139198648877,
          "mean_e2el_ms": 203860.94182242188,
          "duration_seconds": 407.9383143119999
        }
      ],
      "gpu_utilization": 94.23863636363636,
      "preemptions": 0.0,
      "preemptions_scope": "observed samples; 98 sampling ended before final requests"
    },
    {
      "concurrency": 98,
      "cases": [
        {
          "case": "uncached-long-prefill",
          "completed": 64,
          "failed": 0,
          "output_throughput": 569.7052069657224,
          "mean_e2el_ms": 149062.55246709374,
          "duration_seconds": 230.06986490100002
        },
        {
          "case": "shared-prefix-long-decode",
          "completed": 64,
          "failed": 0,
          "output_throughput": 1195.472831127611,
          "mean_e2el_ms": 278737.9830068594,
          "duration_seconds": 438.561200513
        }
      ],
      "gpu_utilization": 90.54166666666667,
      "preemptions": 0.0,
      "preemptions_scope": "observed samples; 98 sampling ended before final requests"
    },
    {
      "concurrency": 128,
      "cases": [
        {
          "case": "uncached-long-prefill",
          "completed": 64,
          "failed": 0,
          "output_throughput": 668.6385010695061,
          "mean_e2el_ms": 192423.83817876567,
          "duration_seconds": 196.02819728500026
        },
        {
          "case": "shared-prefix-long-decode",
          "completed": 64,
          "failed": 0,
          "output_throughput": 1390.6819066823837,
          "mean_e2el_ms": 360289.21471803123,
          "duration_seconds": 377.00066239499984
        }
      ],
      "gpu_utilization": 90.75903614457832,
      "preemptions": 25.0,
      "preemptions_scope": "observed samples; 98 sampling ended before final requests"
    }
  ],
  "cost": {
    "currency": "USD",
    "rate_checked": "2026-10-31",
    "cpu_per_core_second": 3.942e-05,
    "memory_per_gib_second": 6.67e-06,
    "old_request": {
      "cores": 1,
      "gib": 2
    },
    "new_request": {
      "cores": 0.125,
      "gib": 0.125
    },
    "scope": "Request floor only; actual usage can be higher. Excludes GPU, controller, separate grading, egress and storage."
  },
  "source_artifacts": [
    "final-validation-deep-audit.json",
    "final-training-sample-audit.json",
    "final-training-metrics-compact.json",
    "steady-long-c16-128-20261028-03-audit/analysis.json",
    "steady-long-c98-20261028-01-analysis.json",
    "final-policy-version-audit.json"
  ]
}