{
  "description": "Development measurements; not a general product speed guarantee. The two headline metrics are separate experiments.",
  "sources": [
    {
      "name": "docr-structured-batched-vision.json",
      "sha256": "ebe93a8062058cf94509edc7129e8544f35ab36865140c71004ef39d21ce9263"
    },
    {
      "name": "docr-v17-decode-only-summary.json",
      "sha256": "8e6a8e6a9adcc16bb0aacd8fa58dcc502225395782581993a7fe6e84a61f5cf9"
    }
  ],
  "warm_slide": {
    "date": "2026-09-11",
    "machine": "MacBook Pro M3 Pro 18GB",
    "input": "slide-8.png",
    "regions": 12,
    "mode": "PP-DocLayoutV3 FP32 eager MPS + MLX text-Q6 self-spec K16",
    "profile_ms": {
      "elapsed": 1177.58,
      "layout": 132.51,
      "layout_preprocess": 6.74,
      "layout_transfer": 0.91,
      "layout_forward": 111.43,
      "layout_postprocess": 7.96,
      "layout_region_build_and_nms": 5.37,
      "vision_batch": 240.61,
      "multimodal_prefill": 198.26,
      "decode": 596.06
    },
    "statistic": "single warm end-to-end backend profile, not an average",
    "excluded": "model download/startup and full UI rendering; power/contention not recorded"
  },
  "decode_experiment": {
    "experiment": "docr-v17-two-process-pure-decode-summary",
    "status": "CONTENDED_BATTERY_DIAGNOSTIC",
    "boundary": {
      "included": "token selection and text decoder forwards from a completed multimodal prefix KV cache through EOS",
      "excluded": [
        "PP-DocLayout",
        "image loading and resize",
        "vision encoder",
        "multimodal text prefill",
        "tokenizer decode"
      ],
      "output_tokens": 59,
      "fresh_prefix_cache_per_method": true,
      "rotating_method_order": true,
      "warmups_per_method": 1,
      "measurements_per_method": 5,
      "independent_processes": 2
    },
    "two_process_central": {
      "mlx_bf16_ar": {
        "decode_ms": 877.120854005625,
        "tokens_per_second": 67.26553100471789,
        "decoder_forwards": 59
      },
      "mlx_bf16_spec_k16": {
        "decode_ms": 373.9869379969605,
        "tokens_per_second": 157.7595204688125,
        "decoder_forwards": 14,
        "speedup_vs_bf16_ar": 2.3453248359512324
      },
      "mlx_bf16_spec_k32": {
        "decode_ms": 370.2235209966602,
        "tokens_per_second": 159.36318643712602,
        "decoder_forwards": 14,
        "speedup_vs_bf16_ar": 2.3691656641489764
      },
      "mlx_q6_ar": {
        "decode_ms": 488.62908350201906,
        "tokens_per_second": 120.74598502640337,
        "decoder_forwards": 59
      },
      "mlx_q6_spec_k16": {
        "decode_ms": 270.86566649813903,
        "tokens_per_second": 217.82014960690987,
        "decoder_forwards": 14,
        "speedup_vs_q6_ar": 1.803953560520289,
        "speedup_vs_bf16_ar": 3.2382134854719626
      }
    },
    "per_process_best_speedup_vs_bf16_ar": [
      3.254087027475275,
      3.222035131833013
    ],
    "claim_boundary": {
      "performance_claim": false,
      "power": "battery",
      "external_contention": "Google Chrome renderer used 1.9-4.6GB RSS during the runs",
      "swapout_guard_passed": true,
      "interpretation": "The 3.22-3.25x ratio is reproducible as a diagnostic. A clean AC/no-contention repeat is required for an accepted absolute latency claim."
    },
    "raw_runs": [
      "docr-v17-decode-only-contended-run1.json",
      "docr-v17-decode-only-contended-run2.json"
    ]
  }
}