{
  "models": [
    {
      "id": "single",
      "architecture": "single",
      "loops": 1,
      "loss": 2.9941,
      "accuracy": 41.085833588511704,
      "seconds": 6.907221432775259
    },
    {
      "id": "lt4",
      "architecture": "lt",
      "loops": 4,
      "loss": 2.9022,
      "accuracy": 43.47539881645504,
      "seconds": 32.15059030242264
    },
    {
      "id": "plt4",
      "architecture": "plt",
      "loops": 4,
      "loss": 2.9252,
      "accuracy": 42.041264648047346,
      "seconds": 7.727172017097473
    },
    {
      "id": "foresight4",
      "architecture": "foresight",
      "loops": 4,
      "loss": 2.9112,
      "accuracy": 42.99666107890452,
      "seconds": 7.7480554189532995
    },
    {
      "id": "plt8",
      "architecture": "plt",
      "loops": 8,
      "loss": 2.9365,
      "accuracy": 42.227349963490774,
      "seconds": 8.494108185172081
    },
    {
      "id": "foresight8",
      "architecture": "foresight",
      "loops": 8,
      "loss": 2.9058,
      "accuracy": 43.07210362649706,
      "seconds": 8.515413641929626
    },
    {
      "id": "lt8",
      "architecture": "lt",
      "loops": 8,
      "loss": 2.8775,
      "accuracy": 43.66286609393178,
      "seconds": 65.62719668261707
    }
  ],
  "pareto_members": {
    "loss": [
      "single",
      "plt4",
      "foresight4",
      "foresight8",
      "lt4",
      "lt8"
    ],
    "accuracy": [
      "single",
      "plt4",
      "foresight4",
      "foresight8",
      "lt4",
      "lt8"
    ]
  },
  "context_b32": {
    "plt": [
      {
        "context": 512,
        "throughput_percent": 73.45834140209044
      },
      {
        "context": 2048,
        "throughput_percent": 78.61344087055069
      },
      {
        "context": 8192,
        "throughput_percent": 86.6195650102751
      }
    ],
    "foresight": [
      {
        "context": 512,
        "throughput_percent": 73.50624605202279
      },
      {
        "context": 2048,
        "throughput_percent": 78.55450891213462
      },
      {
        "context": 8192,
        "throughput_percent": 86.54150173904847
      }
    ]
  },
  "analytical_cache": [
    {
      "loops": 4,
      "context": [
        512,
        2048,
        8192
      ],
      "extra_percent": [
        37.520926339285715,
        9.380231584821429,
        2.345057896205357
      ]
    },
    {
      "loops": 8,
      "context": [
        512,
        2048,
        8192
      ],
      "extra_percent": [
        87.548828125,
        21.88720703125,
        5.4718017578125
      ]
    }
  ],
  "perplexity_at_8_loops": {
    "PLT": {
      "WikiText-2": 36.82,
      "LAMBADA": 50.69
    },
    "InfuseLoop": {
      "WikiText-2": 35.55,
      "LAMBADA": 42.04
    }
  },
  "protocol": {
    "gpu": "NVIDIA GB200",
    "batch": 32,
    "prompt_tokens": 2048,
    "output_tokens": 2048,
    "timing": "Median of 3 synchronized decode-only runs; 2047 cached steps, prefill excluded",
    "context_sweep": "64 decoding cycles, context starts at 512/2048/8192, median of 3"
  },
  "note": "Only measured R=1,4,8 quality/latency observations. Lines connect observations, not fitted points. Pareto frontiers include LT and PLT-4 as well as InfuseLoop. Cache overhead is analytical, not peak GPU memory."
}
