{
  "source": "Supervisor-provided round 2 prompt; measured through coreai-kit; not re-derived in this run",
  "prompt": "rounds/02.prompt.md",
  "reported_time_jst": "2026-09-23 04:0x–04:3x",
  "worktree": "~/code/coreai-kit-models-wt",
  "engine": "coreai-sequential",
  "device": "Apple M4 Max GPU",
  "parity": {
    "rows": 50,
    "includes_wide_options": [
      40,
      255
    ],
    "fp16": {
      "tokens_equal": 50,
      "slots_equal": 50,
      "option_argmax_equal": 50,
      "max_abs_dp": 0.0051,
      "mean_abs_dp": 0.0003,
      "max_abs_dabstain": 0.0165,
      "worst_row": "r05-task"
    },
    "int8lin": {
      "tokens_equal": 50,
      "slots_equal": 50,
      "option_argmax_equal": 50,
      "max_abs_dp": 0.0226,
      "mean_abs_dp": 0.0009,
      "max_abs_dabstain": 0.0222,
      "worst_row": "r05-dry"
    }
  },
  "latency": {
    "mode": "int8lin",
    "median_question_ms": 354,
    "note": "Median over the 50 fixture rows, two to three questions per state; a question on a new state pays for the whole state (S=1 prefill)",
    "wide_255": {
      "tokens": 1449,
      "seconds": 7.2,
      "prefill": "S=1"
    },
    "thai_ticket": {
      "questions": 3,
      "row_tokens": [
        57,
        63,
        83
      ],
      "question_ms": [
        351,
        316,
        429
      ],
      "prefix_tokens_reused": 0,
      "note": "median of two runs (351.2/350.3, 315.6/322.8, 429.0/435.0); the recurrent hybrid re-prefills every row; an earlier 170/40/65 ms figure was minicpm5-2b measured by mistake and is withdrawn"
    }
  },
  "semif": {
    "dataset": "authored144",
    "language": "en",
    "rows": 144,
    "options": 3,
    "evaluator": "unchanged benchmarks/evaluate.py with SemIf gold labels",
    "mode": "int8lin",
    "raw_correct": 109,
    "mean_family_balanced_accuracy": 0.7249,
    "kit_readme_reference_values": {
      "MiniCPM5-2B int8": 0.681,
      "Qwen3.5-4B int8": 0.821
    }
  },
  "swift_tokenizer": {
    "initial_token_matches": 23,
    "rows": 50,
    "cause": "Foundation string-search match ranges snap to grapheme clusters, keeping Thai combining marks with letter runs and changing BPE merges",
    "example": {
      "text": "กล่อง",
      "old": [
        "กล่อง"
      ],
      "reference": [
        "กล",
        "่อง"
      ]
    },
    "fix": "Cut each text segment with the same Split regex using ICU on UTF-16, then encode each piece."
  }
}