{
  "model": "sentiment@v2",
  "tier": "encoder",
  "base": "sentence-transformers/all-MiniLM-L6-v2",
  "test": {
    "n": 51,
    "accuracy": 0.8823529411764706,
    "macro_f1": 0.8808012558012558,
    "per_label": {
      "positive": {
        "precision": 0.8947368421052632,
        "recall": 0.9444444444444444,
        "f1": 0.918918918918919,
        "support": 18
      },
      "neutral": {
        "precision": 0.9333333333333333,
        "recall": 0.8235294117647058,
        "f1": 0.875,
        "support": 17
      },
      "negative": {
        "precision": 0.8235294117647058,
        "recall": 0.875,
        "f1": 0.8484848484848485,
        "support": 16
      }
    },
    "confusion": [
      [
        17,
        0,
        1
      ],
      [
        1,
        14,
        2
      ],
      [
        1,
        1,
        14
      ]
    ]
  },
  "by_source": {
    "synthetic": {
      "n": 51,
      "accuracy": 0.8823529411764706,
      "macro_f1": 0.8808012558012558,
      "per_label": {
        "positive": {
          "precision": 0.8947368421052632,
          "recall": 0.9444444444444444,
          "f1": 0.918918918918919,
          "support": 18
        },
        "neutral": {
          "precision": 0.9333333333333333,
          "recall": 0.8235294117647058,
          "f1": 0.875,
          "support": 17
        },
        "negative": {
          "precision": 0.8235294117647058,
          "recall": 0.875,
          "f1": 0.8484848484848485,
          "support": 16
        }
      },
      "confusion": [
        [
          17,
          0,
          1
        ],
        [
          1,
          14,
          2
        ],
        [
          1,
          1,
          14
        ]
      ]
    }
  },
  "teacher_agreement": 0.8823529411764706,
  "calibration": {
    "temperature": 1.5494422923834061,
    "test_ece_before": 0.13583191042865103,
    "test_ece_after": 0.09304488274530132,
    "val_ece_before": 0.10366493771366977,
    "val_ece_after": 0.10166117276176741,
    "val_coverage": 0.5294117647058824,
    "val_precision_at_threshold": 1.0
  },
  "escalation": {
    "threshold": 0.9467601639708129,
    "target_precision": 0.97,
    "coverage": 0.5098039215686274,
    "accuracy_on_covered": 0.9615384615384616,
    "escalation_rate": 0.4901960784313726
  },
  "latency": {
    "p50_ms": 3.41874998412095,
    "p95_ms": 3.5703955072676763,
    "batch_ms_per_input": 0.451612745082992,
    "backend": "python (cpu); browser numbers: web/ `npm run bench`"
  },
  "size_mb": 24.31,
  "targets": {
    "min_macro_f1": {
      "target": 0.9,
      "actual": 0.8808012558012558,
      "met": false
    },
    "max_download_mb": {
      "target": 30.0,
      "actual": 24.31,
      "met": true
    }
  },
  "worst_errors": [
    {
      "text": "I'm obsessed with this moisturizer, my skin has never felt better.",
      "label": "positive",
      "predicted": "negative",
      "confidence": 0.947,
      "teacher_confidence": 0.95
    },
    {
      "text": "Pros: light, cheap. Cons: flimsy, loud.",
      "label": "neutral",
      "predicted": "negative",
      "confidence": 0.922,
      "teacher_confidence": 0.75
    },
    {
      "text": "1 star is too generous",
      "label": "negative",
      "predicted": "positive",
      "confidence": 0.914,
      "teacher_confidence": 0.95
    },
    {
      "text": "Waited on hold for two hours to be told to call back tomorrow.",
      "label": "negative",
      "predicted": "neutral",
      "confidence": 0.908,
      "teacher_confidence": 0.95
    },
    {
      "text": "This is my second order from this store.",
      "label": "neutral",
      "predicted": "positive",
      "confidence": 0.529,
      "teacher_confidence": 0.92
    },
    {
      "text": "We took the train instead of driving.",
      "label": "neutral",
      "predicted": "negative",
      "confidence": 0.508,
      "teacher_confidence": 0.92
    }
  ],
  "data": {
    "train": 3238,
    "val": 51,
    "test": 51
  },
  "seed": 42,
  "recommendations": [
    "Macro F1 0.881 is below the 0.90 target. Options: more/better data (real inputs beat synthetic), the encoder tier (M4), or escalation-heavy mode (the micro model handles only confident cases).",
    "All test data is synthetic: expect lower accuracy on real inputs. Add a human-labeled gold set."
  ]
}