{
  "schema": "thinkingcap-calibration-observation/v1",
  "observed_at": "2026-09-24T03:37:44.165805+00:00",
  "hardware": "one NVIDIA GeForce RTX 5090,32607MiB",
  "lane": "16K,C1,eager,no speculation,auto BF16 KV,.87 GPU fraction",
  "model_loading_gib": 21.34,
  "available_kv_gib": 5.67,
  "configured_context_tokens": 16384,
  "runtime_reported_kv_tokens": 80554,
  "vision": "BF16 tower loaded; two-image admission configured; deterministic18/18 corpus passed",
  "sources": [
    "native/startup/startup-16k-wsl.log",
    "native/candidate/eager/candidate-models-16k.json",
    "native/candidate/eager/candidate-multimodal-16k.json"
  ],
  "boundary": "Startup cache capacity is configuration-specific, not a universal hardware ceiling or proof of context quality. See cache-capacity-boundary.json.",
  "result": "Selected 32K BF16 C1 passed functional, reasoning control, vision, strict capacity, repeated quality, long-context and bounded agentic gates; retained directly under user authorization.",
  "graph_observation": {
    "model_loading_gib": 21.34,
    "available_kv_gib": 5.61,
    "runtime_reported_kv_tokens": 79872,
    "captured_graph_gib": 0.05,
    "configured_context_tokens": 16384,
    "warning": "Startup allocation does not prove long-context correctness or additional concurrency"
  },
  "triton_bf16_observation": {
    "model_loading_gib": 22.14,
    "available_kv_gib": 4.73,
    "runtime_reported_kv_tokens": 45180,
    "captured_graph_gib": 0.1,
    "configured_context_tokens": 16384,
    "source": "native/candidate/triton-bf16/triton-bf16-final-engine.log"
  },
  "triton_fp8_observation": {
    "model_loading_gib": 22.14,
    "available_kv_gib": 4.73,
    "runtime_reported_kv_tokens": 64111,
    "captured_graph_gib": 0.1,
    "configured_context_tokens": 16384,
    "source": "native/candidate/triton-fp8/triton-fp8-startup.log",
    "selection": "rejected at repeated quality gate, not advanced to56K"
  },
  "triton_bf16_32k_observation": {
    "model_loading_gib": 22.14,
    "available_kv_gib": 4.73,
    "runtime_reported_kv_tokens": 56262,
    "configured_context_tokens": 32768,
    "source": "native/candidate/triton-bf16-32k/triton-bf16-32k-startup.log",
    "status": "passed; direct serve retained",
    "measured_minimum_free_gpu_mib": 3342,
    "headroom_source": "telemetry-summary.json"
  }
}
