{
  "artifact": "designaround_price",
  "gpu_requested": "L4",
  "model": "Qwen/Qwen2.5-1.5B-Instruct",
  "n_turns": 8,
  "n_convs": 6,
  "seed": 20260725,
  "no_speedup_claimed": true,
  "purpose": "DD-2/K2: price the enable_prefix_caching=False design-around",
  "gpu_name_ground_truth": "NVIDIA L4",
  "vllm_version": "0.24.0",
  "arms": [
    {
      "arm": "apc_on",
      "enable_prefix_caching": true,
      "per_request_salt": false,
      "prompt_tokens_total": 33792,
      "cached_tokens_total": 29280,
      "prefill_tokens_recomputed": 4512,
      "wall_s": 7.435,
      "requests": 48
    },
    {
      "arm": "apc_off",
      "enable_prefix_caching": false,
      "per_request_salt": false,
      "prompt_tokens_total": 33792,
      "cached_tokens_total": 0,
      "prefill_tokens_recomputed": 33792,
      "wall_s": 7.048,
      "requests": 48
    },
    {
      "arm": "apc_on_per_request_salt",
      "enable_prefix_caching": true,
      "per_request_salt": true,
      "prompt_tokens_total": 33792,
      "cached_tokens_total": 0,
      "prefill_tokens_recomputed": 33792,
      "wall_s": 7.015,
      "requests": 48
    }
  ],
  "concurrency_sweep": [
    {
      "concurrency": 1,
      "prefill_tokens_on": 4512,
      "prefill_tokens_off": 33792,
      "wall_on_s": 7.119,
      "wall_off_s": 7.511,
      "throughput_req_per_s_on": 6.743,
      "throughput_req_per_s_off": 6.39,
      "throughput_ratio_on_over_off": 1.055,
      "wall_ratio_off_over_on": 1.055
    },
    {
      "concurrency": 8,
      "prefill_tokens_on": 4800,
      "prefill_tokens_off": 33792,
      "wall_on_s": 1.468,
      "wall_off_s": 2.638,
      "throughput_req_per_s_on": 32.708,
      "throughput_req_per_s_off": 18.196,
      "throughput_ratio_on_over_off": 1.798,
      "wall_ratio_off_over_on": 1.797
    },
    {
      "concurrency": 32,
      "prefill_tokens_on": 4800,
      "prefill_tokens_off": 33792,
      "wall_on_s": 0.726,
      "wall_off_s": 2.196,
      "throughput_req_per_s_on": 66.089,
      "throughput_req_per_s_off": 21.862,
      "throughput_ratio_on_over_off": 3.023,
      "wall_ratio_off_over_on": 3.025
    },
    {
      "concurrency": 64,
      "prefill_tokens_on": 4800,
      "prefill_tokens_off": 33792,
      "wall_on_s": 0.616,
      "wall_off_s": 2.286,
      "throughput_req_per_s_on": 77.916,
      "throughput_req_per_s_off": 20.999,
      "throughput_ratio_on_over_off": 3.71,
      "wall_ratio_off_over_on": 3.711
    }
  ],
  "d8_analysis": {
    "concurrency_levels": [
      1,
      8,
      32,
      64
    ],
    "best_throughput_ratio_on_over_off": 3.71,
    "MATERIALIZES_AS_THROUGHPUT": true,
    "threshold": 1.1,
    "interpretation": "The 86.65% prefill-WORK saving DOES become throughput under batching; the moat argument rests on measurement rather than extrapolation.",
    "no_speedup_claimed": true,
    "scope": "Measures what the DESIGN-AROUND costs a competitor, NOT a speedup by this project. One model, one L4, one workload shape."
  },
  "k2_analysis": {
    "prefill_tokens_apc_on": 4512,
    "prefill_tokens_apc_off": 33792,
    "extra_prefill_tokens_to_disable_apc": 29280,
    "design_around_cost_frac_of_apc_off_prefill": 0.8665,
    "cache_hit_frac_when_on": 0.8665,
    "wall_ratio_off_over_on": 0.948,
    "per_request_salt_equals_apc_off": true,
    "K2_THRESHOLD": 0.1,
    "K2_FIRES_design_around_is_cheap": false
  },
  "status": "MEASURED"
}