{
  "published_scope": "Aggregated technical observations only; not an independently verified benchmark or a production release.",
  "tested_on": "2026-10-13",
  "platform": {
    "hardware": "one NVIDIA DGX Spark (GB10), 128 GB unified memory",
    "inference": "Qwen 27B domain-adapted candidate with SGLang and DFlash2",
    "state": "internal-trial candidate, not production"
  },
  "graded": {
    "golden": {
      "total": 39,
      "pass": 37,
      "partial": 2,
      "fail": 0
    },
    "dev": {
      "total": 100,
      "pass": 91,
      "partial": 4,
      "fail": 5
    },
    "multiturn": {
      "total": 10,
      "pass": 8,
      "partial": 2,
      "fail": 0
    }
  },
  "dev_tool_subsets": {
    "no_tool_call": {
      "total": 72,
      "pass": 70,
      "partial": 1,
      "fail": 1
    },
    "with_tool_call": {
      "total": 28,
      "pass": 21,
      "partial": 3,
      "fail": 4
    }
  },
  "fixed_gates": {
    "knowledge": {
      "pass": 5,
      "total": 5
    },
    "sql_tool": {
      "pass": 6,
      "total": 6
    },
    "analysis_completed": {
      "pass": 8,
      "total": 8
    },
    "mean_analysis_s": 406.4
  },
  "general_no_system_prompt": {
    "pass": 20,
    "total": 24
  },
  "repeated_stream_decode_toks": [
    39.93,
    38.95
  ],
  "concurrency": [
    {
      "streams": 1,
      "aggregate_toks": 24.2,
      "avg_decode_toks": 24.37,
      "slowest_decode_toks": 24.37,
      "max_ttft_s": 0.16
    },
    {
      "streams": 2,
      "aggregate_toks": 46.2,
      "avg_decode_toks": 24.29,
      "slowest_decode_toks": 23.35,
      "max_ttft_s": 0.27
    },
    {
      "streams": 4,
      "aggregate_toks": 73.8,
      "avg_decode_toks": 20.2,
      "slowest_decode_toks": 18.58,
      "max_ttft_s": 0.26
    },
    {
      "streams": 8,
      "aggregate_toks": 110.8,
      "avg_decode_toks": 17.31,
      "slowest_decode_toks": 13.93,
      "max_ttft_s": 0.3
    },
    {
      "streams": 16,
      "aggregate_toks": 125.2,
      "avg_decode_toks": 13.06,
      "slowest_decode_toks": 9.28,
      "max_ttft_s": 32.93
    },
    {
      "streams": 32,
      "aggregate_toks": 145.7,
      "avg_decode_toks": 12.57,
      "slowest_decode_toks": 8.93,
      "max_ttft_s": 83.32
    }
  ],
  "context": [
    {
      "input_tokens": 1136,
      "ttft_s": 0.47,
      "decode_toks": 15.16,
      "needle_found": true
    },
    {
      "input_tokens": 3904,
      "ttft_s": 1.57,
      "decode_toks": 18.59,
      "needle_found": true
    },
    {
      "input_tokens": 16499,
      "ttft_s": 7.23,
      "decode_toks": 16.25,
      "needle_found": true
    },
    {
      "input_tokens": 30828,
      "ttft_s": 15.07,
      "decode_toks": 15.54,
      "needle_found": true
    },
    {
      "input_tokens": 60848,
      "ttft_s": 36.94,
      "decode_toks": 15.54,
      "needle_found": true
    },
    {
      "input_tokens": 114635,
      "ttft_s": 93.69,
      "decode_toks": 14.41,
      "needle_found": true
    },
    {
      "input_tokens": 170979,
      "ttft_s": 178.61,
      "decode_toks": 13.77,
      "needle_found": true
    }
  ],
  "holdout": {
    "size": 134,
    "known_verbatim_train_overlap": 5,
    "evaluated": false
  },
  "disclosures": [
    "Archived internal POC evidence; not independently reproduced in this public package.",
    "The larger quality suites tested the assistant with application context (system instructions, retrieval, and possible tools), not an isolated LoRA-only model.",
    "Golden and development answers were graded by one AI-assisted reviewer without independent second review.",
    "Development questions are near-distribution; this is not a clean held-out accuracy evaluation.",
    "The 134-question final holdout was not scored and contains 5 known verbatim training overlaps.",
    "Aggregate throughput comes from artificial, fixed-length 512-token output, greedy decoding and reasoning disabled; it does not establish production concurrent-user capacity.",
    "Long-context result is one needle retrieval at each length; it does not measure long-document reasoning quality.",
    "Analysis completion counts do not establish semantic correctness or task success."
  ]
}