{
  "date": "2026-09-15",
  "model": "DeepSeek V4.1 Flash Q4",
  "model_bytes": 518596067328,
  "model_sha256": "a5e2e2c3ada4b2e98d9f9e4b50f6d9c2a12c2c96f5da165c07e13aff9264984e",
  "engine": "argonautlabsai/ds4-argodrive @ 38e200a (fork of antirez/ds4)",
  "hardware": "Apple M5 Max, 128 GB; internal SSD plus two NVMe enclosures over TB5",
  "replicas": "four byte-identical verified copies, 10:6:6:4 weighted split read",
  "prompt_tokens": 512,
  "generated_tokens": 200,
  "mode": "raw completion, greedy",
  "metric": {
    "prompt_processing": "prefill tokens / prefill seconds, harness CSV",
    "generation": "generated tokens / generation seconds, excludes prefill and startup",
    "steady": "excludes the first response"
  },
  "comparison": "pinned upstream ds4 bd66c40 on the internal SSD alone, versus the candidate on four verified replicas; identical output SHA-256 on every arm",
  "candidate_settings": {
    "DS4_ARGODRIVE_PREFILL_SPLIT": 1,
    "DS4_ARGODRIVE_PREFILL_AHEAD": 1,
    "DS4_ARGODRIVE_PREFILL_SELECTIVE": 1,
    "DS4_ARGODRIVE_PREFILL_LANES": 8,
    "ssd_streaming_cache_experts": 4200
  },
  "control_settings": {
    "DS4_ARGODRIVE_PREFILL_SPLIT": 0,
    "ssd_streaming_cache_experts": "automatic (3688)"
  },
  "arms": [
    {
      "arm": "p3-up-1",
      "side": "upstream, one drive",
      "prompt_processing_tok_s": 16.37,
      "generation_tok_s": 10.2,
      "steady_tok_s": 10.65,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    },
    {
      "arm": "p3-ch-1b",
      "side": "ours, three drives",
      "prompt_processing_tok_s": 45.2,
      "generation_tok_s": 16.87,
      "steady_tok_s": 17.49,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    },
    {
      "arm": "p3-up-2",
      "side": "upstream, one drive",
      "prompt_processing_tok_s": 16.09,
      "generation_tok_s": 9.95,
      "steady_tok_s": 10.56,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    },
    {
      "arm": "p3-ch-2",
      "side": "ours, three drives",
      "prompt_processing_tok_s": 42.03,
      "generation_tok_s": 15.99,
      "steady_tok_s": 17.26,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    },
    {
      "arm": "up1d-1",
      "side": "upstream, one drive",
      "prompt_processing_tok_s": 17.04,
      "generation_tok_s": 10.19,
      "steady_tok_s": 10.61,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    },
    {
      "arm": "sw1d-a",
      "side": "ours, one drive, no replicas",
      "prompt_processing_tok_s": 27.12,
      "generation_tok_s": 13.39,
      "steady_tok_s": 14.53,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    },
    {
      "arm": "up1d-2",
      "side": "upstream, one drive",
      "prompt_processing_tok_s": 15.97,
      "generation_tok_s": 9.67,
      "steady_tok_s": 10.27,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    },
    {
      "arm": "sw1d-b",
      "side": "ours, one drive, no replicas",
      "prompt_processing_tok_s": 28.96,
      "generation_tok_s": 13.59,
      "steady_tok_s": 14.23,
      "output_sha256": "a548d589195217aa71e99765e018c09645ff83853d09be922a6af070ab58c531"
    }
  ],
  "medians": {
    "three_drive": {
      "prompt_processing_control": 16.23,
      "prompt_processing_candidate": 43.62,
      "ratio": 2.69,
      "steady_control": 10.61,
      "steady_candidate": 17.38,
      "steady_ratio": 1.64
    },
    "one_drive_no_replicas": {
      "prompt_processing_control": 16.5,
      "prompt_processing_candidate": 28.04,
      "ratio": 1.7,
      "steady_control": 10.44,
      "steady_candidate": 14.38,
      "steady_ratio": 1.38
    }
  },
  "other_rungs": [
    {
      "rung": "2048/200",
      "control": 62.96,
      "candidate": 154.97,
      "ratio": 2.46,
      "output_sha256_prefix": "1919f71ca82c"
    },
    {
      "rung": "512/512",
      "control": 16.23,
      "candidate": 44.88,
      "ratio": 2.77,
      "output_sha256_prefix": "0698ebfb6e99"
    }
  ],
  "expert_coverage_per_layer": {
    "512_tokens": {
      "mean_of_384": 187.4,
      "coverage": 0.488,
      "selective": "+7.2%"
    },
    "1024_tokens": {
      "mean_of_384": 230.8,
      "coverage": 0.601,
      "selective": "-13%"
    },
    "2048_tokens": {
      "mean_of_384": 271.5,
      "coverage": 0.707,
      "selective": "-33%"
    }
  },
  "prefill_device_read_gb_s": {
    "before": 9.81,
    "after": 23.5,
    "per_drive_raw_ceiling_gb_s": {
      "internal": 12.9,
      "green": 7.08,
      "white": 7.08
    }
  },
  "limits": [
    "prompt-processing gain is the prefill work; the generation gain comes from decode replica streaming and a larger expert cache, not from this change",
    "selective staging applied to chunks of 512 tokens or fewer",
    "prompts under 256 tokens do not reach this path",
    "candidate arms measured after hours of continuous running; rested machine has measured 49.17",
    "one machine, one model, one prompt"
  ],
  "upstream_read_amplification": {
    "prefill_gb_read": 301.9,
    "prefill_gb_touched": 149.2,
    "ratio": 2.02,
    "note": "measured on unmodified upstream, internal SSD only, no replicas"
  },
  "fourth_drive": {
    "prefill_512": "+6.6%",
    "prefill_2048": "+15.4%",
    "generation_512": "+2.9%",
    "weights": "10:6:6:4 beats capability-proportional 10:5:5:4 by 2.6% decode; the internal also serves Engram"
  },
  "did_not_work": {
    "decode_prefetch": "predictor costs 3% before any prefetch (16.82 vs 17.33 gen)",
    "page_cache_warming": "warmed read is 2.6x faster but reads every byte twice",
    "token_chunk_pipelining": "+2.1%; expert union is front-loaded"
  },
  "note": "Re-measured 2026-09-15 on the branch as it ships; the published reader caps at three sources. The one-drive pair uses no replicas and isolates the selective-expert read."
}