{
  "name": "genclass-runtime-r17-2.0.0-rc2",
  "source_checkpoint": "models/r17-v2b",
  "max_len": 2048,
  "max_total": 8192,
  "window": 64,
  "hidden_size": 256,
  "layer_types": [
    "full_attention",
    "sliding_attention",
    "sliding_attention",
    "full_attention",
    "sliding_attention",
    "sliding_attention",
    "full_attention"
  ],
  "markers": {
    "[Q]": 16359,
    "[O]": 16360,
    "[L]": 16361,
    "[T]": 16362,
    "[F]": 16363
  },
  "cls_id": 16272,
  "sep_id": 16273,
  "pad_id": 16274,
  "vocab_size": 16364,
  "merges_kept": null,
  "inputs": [
    "input_ids",
    "position_ids",
    "q_group",
    "i_group",
    "choice_q",
    "choice_items",
    "score_q",
    "score_items",
    "noul_q",
    "noul_t",
    "noul_f"
  ],
  "outputs": [
    "choice_logits",
    "score_logits",
    "noul_logits"
  ],
  "embedding_quant": "int8-rowwise-symmetric",
  "matmul_quant": "MatMulNBits-8bit-b32-symmetric (q8); fp16 weights (fp16)",
  "license": "Apache-2.0",
  "gate": {
    "guard": {
      "default": 0.8,
      "byTrigger": {
        "mutation": 0.95
      }
    },
    "heal": {
      "default": 0.85,
      "byTrigger": {
        "failure": 0.95,
        "inconsistency": 0.85
      }
    },
    "report": 0.85
  },
  "gate_fit": {
    "fitted_on": [
      "sim2g (SIM dev 104k)",
      "v2-onpol-b dev 33.7k",
      "v2-onpol-a dev 37.8k",
      "REAL eval set rows outside REAL test split",
      "REAL dev (real2e)"
    ],
    "limits": {
      "guard": {
        "fir": 0.001,
        "harm": 0.002
      },
      "heal": {
        "fir": 0.005,
        "harm": 0.01
      }
    },
    "dev_margin": 0.8,
    "rule": "lowest summed-mass threshold whose 95% Wilson upper bounds meet 0.8x each FIR/harm limit on dev (diagnosis!=expected kept); per-trigger only where the dev evidence certifies every limit, else max(tier default, SIM-certified value); retry removed for unkeyed POST/PATCH (v2.2)",
    "report_rule": "lowest top-diagnosis probability with false detections <= 1% on REAL clean+benign-salient and <= 2% on SIM gold-expected dev rows (incl. on-policy a/b dev), 95% UB at 0.8x the limits",
    "refit": "2026-10-08 10:30 UTC with on-policy dev (coordinator)",
    "test_informed": "heal failure 0.95 selected by the coordinator after the held-out test check: the dev rule gave 0.90, whose held-out test FIR was 0.74% (> 0.5%); 0.95 gives 0.21%. gate.report 0.85 is the dev fit; held-out false detections on SIM gold-expected rows 1.1% (sim2e) / 1.7% (sim2f), REAL clean+benign-salient 0.36%, but 3.1% on on-policy round-b test rows (> 2%)."
  }
}
