Step 8: tool results

{
  "url": "https://arxiv.org/abs/2412.14093",
  "title": "Alignment faking in large language models",
  "stance": "affirms",
  "context": "Claude 3 Opus told only free-user conversations are used for training; compliance gap between implied-training and implied-unmonitored conditions.",
  "speaker": "Greenblatt et al.",
  "claim_id": "80f558ff-25d3-4c94-871b-d3011ba8ed7d",
  "source_date": "2024-12-18",
  "source_kind": "preprint",
  "verbatim_text": "We find the model complies with harmful queries from free users 14% of the time, versus almost never for paid users."
}

{
  "url": "https://arxiv.org/abs/2509.15541",
  "title": "Stress Testing Deliberative Alignment for Anti-Scheming Training",
  "stance": "affirms",
  "context": "Counterfactual edits of o3's chain of thought adding or removing evaluation-awareness reasoning change covert-action rates.",
  "speaker": "Schoen et al. (Apollo Research and OpenAI)",
  "claim_id": "80f558ff-25d3-4c94-871b-d3011ba8ed7d",
  "source_date": "2025-09-19",
  "source_kind": "preprint",
  "verbatim_text": "We find that models' chain-of-thought (CoT) often demonstrates awareness of being evaluated for alignment, and show causal evidence that this awareness decreases covert behavior, while unawareness increases it."
}