{
  "loop_count": 8,
  "perplexity": {
    "WikiText-2": [
      33.78,
      36.82,
      35.55
    ],
    "LAMBADA": [
      35.04,
      50.69,
      42.04
    ]
  },
  "series_order": [
    "LT",
    "PLT",
    "InfuseLoop"
  ],
  "accuracy_percent": {
    "LT": 43.66286609393178,
    "PLT": 42.227349963490774,
    "InfuseLoop": 43.07210362649706,
    "Single": 41.085833588511704
  },
  "decoding_seconds": {
    "LT": 65.62719668261707,
    "PLT": 8.494108185172081,
    "InfuseLoop": 8.515413641929626
  },
  "quality_source": "Paper Table 1; blog-results-data.json",
  "latency_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": "Perplexity and latency bars start at zero. Accuracy is a dot plot with a labelled percent axis. These benchmarks are independent of the illustrative continuation playback."
}