{
  "schema": "anvil-serving.context-hardware-sampling/v1",
  "observed_at": "2026-08-02T02:50:00.272955Z",
  "capture_method": "250 ms nvidia-smi sampling while vLLM reported one active request; request order correlated through the completed-request counter",
  "source_artifact": "quality-r16-b12x-dspark5-context-ladder-low-r3-telemetry.json",
  "model": "deepseek-v4-flash-0731-r16-b12x-dspark5-tp2-128k",
  "topology": "dual-gpu-exclusive TP2",
  "reserve_policy_mib_per_gpu": 3072,
  "request_counter_baseline": 27,
  "interval_ms": 250,
  "segments": [
    {
      "segment": 1,
      "target_tokens": 32768,
      "attempted_context_tokens": 32768,
      "actual_prompt_tokens": 23831,
      "first_sample": "2026-08-02T02:49:17.307765Z",
      "last_sample": "2026-08-02T02:49:22.940295Z",
      "samples": 19,
      "gpus": [
        {"role": "dark-compute-a", "max_used_mib": 96260, "min_free_mib": 1179, "max_power_w": 240.85, "max_sm_clock_mhz": 2272, "max_utilization_percent": 99, "reserve_pass": false},
        {"role": "dark-compute-b", "max_used_mib": 95408, "min_free_mib": 2031, "max_power_w": 228.31, "max_sm_clock_mhz": 2310, "max_utilization_percent": 100, "reserve_pass": false}
      ]
    },
    {
      "segment": 2,
      "target_tokens": 65536,
      "attempted_context_tokens": 65536,
      "actual_prompt_tokens": 67415,
      "first_sample": "2026-08-02T02:49:24.577078Z",
      "last_sample": "2026-08-02T02:49:35.910980Z",
      "samples": 37,
      "gpus": [
        {"role": "dark-compute-a", "max_used_mib": 96242, "min_free_mib": 1197, "max_power_w": 250.26, "max_sm_clock_mhz": 2272, "max_utilization_percent": 99, "reserve_pass": false},
        {"role": "dark-compute-b", "max_used_mib": 95408, "min_free_mib": 2031, "max_power_w": 237.94, "max_sm_clock_mhz": 2317, "max_utilization_percent": 100, "reserve_pass": false}
      ]
    },
    {
      "segment": 3,
      "target_tokens": 128000,
      "attempted_context_tokens": 126464,
      "actual_prompt_tokens": 125785,
      "first_sample": "2026-08-02T02:49:37.541738Z",
      "last_sample": "2026-08-02T02:49:59.971338Z",
      "samples": 73,
      "gpus": [
        {"role": "dark-compute-a", "max_used_mib": 96236, "min_free_mib": 1203, "max_power_w": 265.32, "max_sm_clock_mhz": 2280, "max_utilization_percent": 100, "reserve_pass": false},
        {"role": "dark-compute-b", "max_used_mib": 95408, "min_free_mib": 2031, "max_power_w": 251.88, "max_sm_clock_mhz": 2317, "max_utilization_percent": 100, "reserve_pass": false}
      ]
    }
  ],
  "interpretation": {
    "memory_behavior": "Peak active-request VRAM is nearly flat across context sizes because the serve reserves its KV capacity at startup; context length primarily changes how long the GPUs remain at high utilization and power.",
    "reserve_decision": "All sampled contexts fail the 3 GiB reported-free policy on both cards. A successful 128K request is therefore functional evidence, not promotion evidence.",
    "timing_note": "The linked ladder is a warmed run and is published separately from the earlier colder ladder."
  }
}