{
  "primary": "Maximum family-macro accuracy on the complete 892-case development panel; tie-break by overall accuracy, then later training step.",
  "candidates": [
    {
      "adapter": "runs/decision-0.8b-v1/checkpoint-500",
      "report": "reports/decision-0.8b-v1/pilot500-development.json",
      "step": 500
    },
    {
      "adapter": "runs/decision-0.8b-v2/checkpoint-1000",
      "report": "reports/decision-0.8b-v2/checkpoint1000-development.json",
      "step": 1000
    },
    {
      "adapter": "runs/decision-0.8b-v2/checkpoint-5000",
      "report": "reports/decision-0.8b-v2/checkpoint5000-development.json",
      "step": 5000
    },
    {
      "adapter": "runs/decision-0.8b-v2/adapter",
      "report": "reports/decision-0.8b-v2/decision-development.json",
      "step": 9289
    }
  ],
  "constraints": [
    "Only complete evaluations with zero inference errors are eligible.",
    "No test predictions are used for selection.",
    "Report original-core and per-source regressions even when macro accuracy improves."
  ],
  "note": "Selection protocol recorded after observing early development checks but before late-checkpoint or any test inference."
}