reports/pilot-gate.json

{
  "blocking_passed": true,
  "gates": [
    {
      "blocking": true,
      "detail": "QWK 0.968 (>= 0.700), edit-F1 0.857 (>= 0.600). A teacher that disagrees with itself caps every downstream metric.",
      "name": "G1_self_consistency",
      "passed": true,
      "threshold": 0.7,
      "value": 0.9679
    },
    {
      "blocking": true,
      "detail": "middle bands {1,2,3} hold 46.8% (>= 40%); empty bands: none. A bimodal dataset cannot teach the region real learners occupy.",
      "name": "G2_band_coverage",
      "passed": true,
      "threshold": 0.4,
      "value": 0.4677
    },
    {
      "blocking": false,
      "detail": "Share of call-2 outputs passing the six checks.",
      "name": "G3_format_validity",
      "passed": true,
      "threshold": 0.9,
      "value": 0.9645
    },
    {
      "blocking": false,
      "detail": "Mean distinct-2 per batch. Low means K sentences collapsed into one.",
      "name": "G4_batch_diversity",
      "passed": true,
      "threshold": 0.7,
      "value": 0.9555
    },
    {
      "blocking": false,
      "detail": "High `other` means the taxonomy is missing a category (Phase 1).",
      "name": "G5_other_tag_share",
      "passed": true,
      "threshold": 0.05,
      "value": 0.0297
    },
    {
      "blocking": false,
      "detail": "K=6 vs K=1 agreement. Low means batching changed the judgement.",
      "name": "G6_batch_single_parity",
      "passed": true,
      "threshold": 0.8,
      "value": 1.0
    }
  ],
  "passed": true
}