{
  "schema": "private-dflash-screen/v1",
  "measured_requests": 144,
  "failure_count": 0,
  "restored": true,
  "promoted": false,
  "statistics": "Pooled median of 12 request timings per cell; throughput is total output tokens / total measured batch wall time.",
  "workload_limits": [
    "Synthetic one-line summaries, not fixed output length or controlled decode.",
    "No request canaries; exact counters and prompt-token populations reconciled.",
    "No matched no-speculation control; this is DFlash K3 versus incumbent MTP3 scouting, not a speculative speedup qualification.",
    "Only approximately 6K/25K inputs; full-context capacity, hard coding and vision are not qualified.",
    "Container peak is lifetime accounting; original and fresh-start peaks are not comparable memory gains."
  ],
  "cells": [
    {
      "nominal_context": 8192,
      "concurrency": 1,
      "arms": {
        "baseline-before": {
          "time_to_first_output_ms": 708.1339159922209,
          "ttft_ms": 1663.0998319888022,
          "e2e_ms": 1933.6229329928756,
          "decode_tok_s": 200.1186634899374,
          "requests": 12,
          "output_tokens": 2829,
          "throughput_tok_s": 123.305396947383
        },
        "candidate": {
          "time_to_first_output_ms": 712.8088180033956,
          "ttft_ms": 1599.908782489365,
          "e2e_ms": 1830.3180924849585,
          "decode_tok_s": 193.8495647518776,
          "requests": 12,
          "output_tokens": 2615,
          "throughput_tok_s": 119.63478169781733
        },
        "baseline-after": {
          "time_to_first_output_ms": 718.767056503566,
          "ttft_ms": 1707.2278249834199,
          "e2e_ms": 1952.0060794893652,
          "decode_tok_s": 199.1750140679346,
          "requests": 12,
          "output_tokens": 2895,
          "throughput_tok_s": 125.37125250303593
        }
      },
      "actual_input_range": [
        6065,
        6068
      ],
      "candidate_change_percent": {
        "baseline-before": {
          "time_to_first_output_ms": 0.66017202475388,
          "ttft_ms": -3.799594485189195,
          "e2e_ms": -5.342553542640349,
          "decode_tok_s": -3.1326906889796557,
          "throughput_tok_s": -2.9768488163839257
        },
        "baseline-after": {
          "time_to_first_output_ms": -0.8289526413681525,
          "ttft_ms": -6.286158234042205,
          "e2e_ms": -6.233996311949996,
          "decode_tok_s": -2.673753703986481,
          "throughput_tok_s": -4.575587058986819
        }
      }
    },
    {
      "nominal_context": 8192,
      "concurrency": 4,
      "arms": {
        "baseline-before": {
          "time_to_first_output_ms": 2687.8949314996134,
          "ttft_ms": 4611.25868052477,
          "e2e_ms": 5061.1819360347,
          "decode_tok_s": 99.7326826751604,
          "requests": 12,
          "output_tokens": 2772,
          "throughput_tok_s": 141.91541777228406
        },
        "candidate": {
          "time_to_first_output_ms": 2746.618135483004,
          "ttft_ms": 4404.0240074682515,
          "e2e_ms": 4962.218633969314,
          "decode_tok_s": 104.2552382499567,
          "requests": 12,
          "output_tokens": 2715,
          "throughput_tok_s": 147.0978924331824
        },
        "baseline-after": {
          "time_to_first_output_ms": 2689.7139310021885,
          "ttft_ms": 5336.26304299105,
          "e2e_ms": 5807.410680514295,
          "decode_tok_s": 97.26403995829105,
          "requests": 12,
          "output_tokens": 2926,
          "throughput_tok_s": 153.8061160984937
        }
      },
      "actual_input_range": [
        6065,
        6068
      ],
      "candidate_change_percent": {
        "baseline-before": {
          "time_to_first_output_ms": 2.184728401962799,
          "ttft_ms": -4.494102096934083,
          "e2e_ms": -1.9553397470417844,
          "decode_tok_s": 4.5346775535224815,
          "throughput_tok_s": 3.65180523881774
        },
        "baseline-after": {
          "time_to_first_output_ms": 2.115622922754956,
          "ttft_ms": -17.469885348835145,
          "e2e_ms": -14.55368137440095,
          "decode_tok_s": 7.1878551360437415,
          "throughput_tok_s": -4.361480437498022
        }
      }
    },
    {
      "nominal_context": 32768,
      "concurrency": 1,
      "arms": {
        "baseline-before": {
          "time_to_first_output_ms": 2922.482462483458,
          "ttft_ms": 3860.087395994924,
          "e2e_ms": 4103.537006507395,
          "decode_tok_s": 196.5402328671148,
          "requests": 12,
          "output_tokens": 2904,
          "throughput_tok_s": 58.415729192017075
        },
        "candidate": {
          "time_to_first_output_ms": 2900.0746840029024,
          "ttft_ms": 3883.860667498084,
          "e2e_ms": 4195.270099473419,
          "decode_tok_s": 194.00055508278564,
          "requests": 12,
          "output_tokens": 3128,
          "throughput_tok_s": 61.35873263509459
        },
        "baseline-after": {
          "time_to_first_output_ms": 2948.8267404958606,
          "ttft_ms": 3959.6001909812912,
          "e2e_ms": 4204.002462502103,
          "decode_tok_s": 193.53983883711928,
          "requests": 12,
          "output_tokens": 2916,
          "throughput_tok_s": 57.83079731914889
        }
      },
      "actual_input_range": [
        25279,
        25281
      ],
      "candidate_change_percent": {
        "baseline-before": {
          "time_to_first_output_ms": -0.7667378254004764,
          "ttft_ms": 0.6158739184979556,
          "e2e_ms": 2.2354640111823754,
          "decode_tok_s": -1.2921923146627767,
          "throughput_tok_s": 5.03803253641435
        },
        "baseline-after": {
          "time_to_first_output_ms": -1.653269614774322,
          "ttft_ms": -1.9128073499874554,
          "e2e_ms": -0.20771545941216107,
          "decode_tok_s": 0.2380472405240086,
          "throughput_tok_s": 6.100443845648873
        }
      }
    },
    {
      "nominal_context": 32768,
      "concurrency": 4,
      "arms": {
        "baseline-before": {
          "time_to_first_output_ms": 9356.410324486205,
          "ttft_ms": 11603.493330505444,
          "e2e_ms": 11893.207506014733,
          "decode_tok_s": 111.31524235231913,
          "requests": 12,
          "output_tokens": 2809,
          "throughput_tok_s": 58.42489114183074
        },
        "candidate": {
          "time_to_first_output_ms": 9545.43297548662,
          "ttft_ms": 11476.78841650486,
          "e2e_ms": 12006.989321496803,
          "decode_tok_s": 110.79479298896521,
          "requests": 12,
          "output_tokens": 3085,
          "throughput_tok_s": 61.920267673958534
        },
        "baseline-after": {
          "time_to_first_output_ms": 9738.193949509878,
          "ttft_ms": 11553.045632026624,
          "e2e_ms": 12085.453050996875,
          "decode_tok_s": 104.9172671354605,
          "requests": 12,
          "output_tokens": 2962,
          "throughput_tok_s": 59.22118608296266
        }
      },
      "actual_input_range": [
        25279,
        25281
      ],
      "candidate_change_percent": {
        "baseline-before": {
          "time_to_first_output_ms": 2.020247556968857,
          "ttft_ms": -1.0919549000599549,
          "e2e_ms": 0.9566957897987338,
          "decode_tok_s": -0.46754546130050034,
          "throughput_tok_s": 5.982683859251892
        },
        "baseline-after": {
          "time_to_first_output_ms": -1.9794324802183683,
          "ttft_ms": -0.6600615798691978,
          "e2e_ms": -0.6492411096959194,
          "decode_tok_s": 5.6020577107828595,
          "throughput_tok_s": 4.557628391999358
        }
      }
    }
  ],
  "resources": {
    "baseline-before": {
      "engine_reported_cache_tokens": 1330579,
      "samples": 157,
      "minimum_free_gpu_MiB": 991.0,
      "generated_tokens_reconciled": 11314,
      "engine_active_requests_max": 4.0,
      "minimum_host_available_GiB": 65.69332885742188,
      "container_lifetime_peak_GiB": 42.45143508911133,
      "oom_kill_counts": [
        0
      ],
      "maximum_telemetry_gap_seconds": 2.118076801300049
    },
    "candidate": {
      "engine_reported_cache_tokens": 789904,
      "samples": 159,
      "minimum_free_gpu_MiB": 2143.0,
      "generated_tokens_reconciled": 11543,
      "engine_active_requests_max": 4.0,
      "minimum_host_available_GiB": 65.40890121459961,
      "container_lifetime_peak_GiB": 41.074703216552734,
      "oom_kill_counts": [
        0
      ],
      "maximum_telemetry_gap_seconds": 2.136640787124634
    },
    "baseline-after": {
      "engine_reported_cache_tokens": 1330579,
      "samples": 160,
      "minimum_free_gpu_MiB": 993.0,
      "generated_tokens_reconciled": 11699,
      "engine_active_requests_max": 4.0,
      "minimum_host_available_GiB": 64.95117950439453,
      "container_lifetime_peak_GiB": 41.877994537353516,
      "oom_kill_counts": [
        0
      ],
      "maximum_telemetry_gap_seconds": 2.1228315830230713
    }
  },
  "preflights": {
    "baseline-before": true,
    "candidate": true,
    "restored": true,
    "baseline-after": true,
    "routed-restored": true
  },
  "restoration_matches": {
    "model": true,
    "revision": true,
    "recipe_digest": true,
    "registry_digest": true,
    "image_digest": true,
    "served_identity": true,
    "bound_port": true,
    "bound_endpoints": true,
    "gpu_selection": true
  },
  "reported_capacity_change_percent": -40.63456585441375,
  "publication_identity": "withheld; descriptive scout only, not independently reproducible model qualification"
}