Opus 4.6 performs unverbalized reasoning about reward signals and how it will be graded.

{
  "doc": "natural.md",
  "context": "Shows NLAs surface latent beliefs upstream of behavioral outputs; steering NLA explanations changes model behavior.",
  "category": "cognitive",
  "norm_label": "Opus 4.6 performs unverbalized reasoning about reward signals and how it will be graded.",
  "graphify_id": "reward_reasoning_case_study",
  "source_file": "natural.md",
  "imported_from": "/Users/antonborzov/Documents/Research.nosync/papers/extract_typed_out/natural/graph.json",
  "extracted_type": "finding",
  "source_location": "§Reasoning about Rewards",
  "graphify_file_type": "finding"
}