model-selection-stability-summary.json

{
  "benchmark": "B2",
  "evaluation_package": "multi-run pairwise model evaluation",
  "total_evaluations": 16,
  "single_run_evaluations": 8,
  "multi_run_evaluations": 8,
  "unique_scenarios": 45,
  "single_run": {
    "hard_unstable_scenarios": 11,
    "instability_rate_pct": 24.44
  },
  "multi_run": {
    "hard_unstable_scenarios": 3,
    "instability_rate_pct": 6.67,
    "realized_executions_per_evaluation": 10
  },
  "absolute_reduction_percentage_points": 17.78,
  "relative_reduction_pct": 72.73,
  "scope_note": "The result supports repeated execution and score aggregation for this model pair, judge, domain, and frozen scenario set."
}