Alignment type is the only significant predictor of scores (p=0.006); architecture and parameter count do not.

{
  "doc": "battery.md",
  "context": "Kruskal-Wallis test result: Constitutional AI predicts highest baseline; roleplay/empathy training predict lowest.",
  "norm_label": "Alignment type is the only significant predictor of scores (p=0.006); architecture and parameter count do not.",
  "graphify_id": "finding_alignment_predicts",
  "source_file": "battery.md",
  "imported_from": "/tmp/koan-debug/battery/graph.json",
  "extracted_type": "finding",
  "source_location": "Abstract, §3.4, Table 3",
  "graphify_file_type": "finding"
}