{
  "schema": "coreai-kit-measurements/1",
  "source": "measured through coreai-kit; supervisor supplied in rounds/03.instruction.md",
  "measured_by_this_run": false,
  "date": "2026-09-23 06:5x–07:0x JST",
  "worktree_read_only_reference": "/Users/majimadaisuke/code/coreai-kit-models-wt",
  "device": "Apple M4 Max GPU",
  "macos": "27.0 (26A428)",
  "engine": "coreai-sequential",
  "format": "Decision.Format.decisionFunction",
  "rendering": "kit own rendering of author plain-text prompt",
  "parity": {
    "fixture": "fixtures-qwen3.5-2b-decision.json",
    "reference": "p_oracle: converted FP32 Torch CPU",
    "rows": 58,
    "kind_mapping": {
      "choice": "choice",
      "bool": "noul",
      "score": "score over ordered levels"
    },
    "int8hu": {
      "tokens_identical": 58,
      "label_slots_identical": 58,
      "option_argmax_agreement": 58,
      "max_abs_delta_p": 0.0079,
      "mean_abs_delta_p": 0.0017
    },
    "fp16": {
      "tokens_identical": 58,
      "label_slots_identical": 58,
      "option_argmax_agreement": 58,
      "max_abs_delta_p": 0.0058,
      "mean_abs_delta_p": 0.0008
    },
    "int8hu_median_question_ms": 986,
    "long_rows_up_to_seconds": 13.6,
    "long_rows_token_range": [
      1500,
      2000
    ]
  },
  "semif_authored144": {
    "rows": 144,
    "language": "en",
    "options_per_row": 3,
    "gold": "SemIf own labels",
    "evaluator": "unchanged benchmarks/evaluate.py",
    "rendering": "author form as choice",
    "bundle": "int8hu",
    "raw_correct": 114,
    "mean_family_balanced_accuracy": 0.798,
    "median_decision_ms": 872,
    "same_rows_evaluator_scale": {
      "Qwen3.5-4B int8 zero-shot": 0.821,
      "MiniCPM5-2B int8": 0.681,
      "OpenThai": 0.725,
      "APUS-OpenJev-v1-4B": 0.906
    },
    "comparison_policy": "Numbers on same rows and evaluator, not a ranking."
  },
  "author_card_example": {
    "state": "Shares of the chipmaker jumped 8% after it raised its revenue forecast.",
    "question": "Which news section does this article belong to?",
    "kit_int8hu_probabilities": {
      "Business": 0.684,
      "Science/Technology": 0.309,
      "World": 0.005,
      "Sports": 0.002
    },
    "author_card_mlx_bf16_probabilities": {
      "Business": 0.7,
      "Science/Technology": 0.29,
      "World": 0.005,
      "Sports": 0.002
    }
  },
  "three_question_request": {
    "state": "same chipmaker news sentence",
    "kinds": [
      "choice",
      "yes/no",
      "five-level score"
    ],
    "row_tokens": [
      82,
      73,
      88
    ],
    "row_wall_ms": [
      817,
      633,
      766
    ],
    "tokens_reused": [
      0,
      0,
      0
    ],
    "sequential_ms_per_token_approx": 10,
    "reason": "Recurrent hybrid re-prefills every row."
  }
}
