{
  "suite": "local executable coding smoke, eight tasks",
  "coding_passes": 6,
  "coding_total": 8,
  "host": "Apple M4 Max, 40 GPU cores, 64 GiB unified memory",
  "runtime": "Kapitsa 0.2.27",
  "decoding": "DFlash2/KFlash",
  "slots": 1,
  "disk_cache": false,
  "sampling": "greedy; temperature=0, top_p=1, top_k=0; enable_thinking=false",
  "context_tests": [
    {
      "context_limit": 65536,
      "prompt_tokens": 61441,
      "retrieval": true,
      "tool_call": true,
      "tool_result": true,
      "cached_coding": true,
      "cold_seconds": 192.87,
      "cached_code_tps": 78.4
    },
    {
      "context_limit": 131072,
      "prompt_tokens": 126976,
      "retrieval": true,
      "tool_call": true,
      "tool_result": true,
      "cached_coding": true,
      "cold_seconds": 579.09,
      "cached_code_tps": 61.9
    }
  ],
  "measurement_note": "Single long-context run per configuration. Fast context results use a prior reproducible allocation of the same edited BF16 parent. Both coding graders accept standard-library typing imports. Exploratory local fixtures; BF16 parity untested."
}