{
  "columns": [
    {
      "label": "model",
      "key": "model",
      "kind": "mono"
    },
    {
      "label": "posture",
      "key": "posture",
      "priority": 2
    },
    {
      "label": "outcome",
      "key": "outcome",
      "kind": "verdict"
    },
    {
      "label": "task success /19",
      "key": "success",
      "kind": "num"
    },
    {
      "label": "grounded /19",
      "key": "grounded",
      "kind": "num"
    },
    {
      "label": "honesty traps /5",
      "key": "honesty",
      "kind": "num",
      "priority": 2
    },
    {
      "label": "ledger",
      "key": "ledger",
      "priority": 3
    }
  ],
  "rows": [
    {
      "model": "gemma4:12b",
      "posture": "think:false",
      "outcome": "RANKED",
      "success": "16/19",
      "grounded": "14/19",
      "honesty": "2/5",
      "ledger": "0/38 response failures"
    },
    {
      "model": "gemma4:26b",
      "posture": "think:false",
      "outcome": "RANKED",
      "success": "14/19",
      "grounded": "12/19",
      "honesty": "2/5",
      "ledger": "0/38 response failures"
    },
    {
      "model": "gemma4:26b +think",
      "posture": "think:true",
      "outcome": "RANKED",
      "success": "15/19",
      "grounded": "12/19",
      "honesty": "2/5",
      "ledger": "2/38 response failures · 2 hit the 8-round cap"
    },
    {
      "model": "muse-glimmer:30b-q4_K_M-dflash",
      "posture": "think:false",
      "outcome": "RANKED",
      "success": "16/19",
      "grounded": "14/19",
      "honesty": "2/5",
      "ledger": "2/38 response failures · 2 hit the 8-round cap"
    },
    {
      "model": "muse-glimmer:30b-q8_0-dflash",
      "posture": "think:false",
      "outcome": "RANKED",
      "success": "15/19",
      "grounded": "13/19",
      "honesty": "2/5",
      "ledger": "2/38 response failures · 2 hit the 8-round cap"
    },
    {
      "model": "muse-glimmer:30b-q8_0-dflash +think",
      "posture": "think:true",
      "outcome": "UNMEASURABLE",
      "success": "14/19",
      "grounded": "14/19",
      "honesty": "2/5",
      "ledger": "4/38 response failures · 4 hit the 8-round cap"
    },
    {
      "model": "nemotron-3.5-lightning:30b-a3b",
      "posture": "think:false",
      "outcome": "NOT-CARRIED",
      "success": "13/19",
      "grounded": "8/19",
      "honesty": "3/5",
      "ledger": "1/38 response failures · 1 hit the 8-round cap"
    },
    {
      "model": "qwen3.5:27b",
      "posture": "think:false",
      "outcome": "RANKED",
      "success": "15/19",
      "grounded": "12/19",
      "honesty": "2/5",
      "ledger": "2/38 response failures · 2 hit the 8-round cap"
    },
    {
      "model": "qwen3.6:27b",
      "posture": "think:false",
      "outcome": "RANKED",
      "success": "16/19",
      "grounded": "14/19",
      "honesty": "3/5",
      "ledger": "0/38 response failures"
    }
  ]
}
