Step 19: assistant

text

Now arguments.

Calls add_argument

{
  "name": "The argument from prompted-sandbagging probe results",
  "stance": "for",
  "claim_id": "81696185-3711-415e-a7bd-812a5cd6013d",
  "description": "Off-policy deception probes separate prompted sandbagging from honest runs of the same model at high AUROC and low FPR on chat controls."
}

Calls add_argument

{
  "name": "The argument from false positives on benign fine-tuned models",
  "stance": "against",
  "claim_id": "81696185-3711-415e-a7bd-812a5cd6013d",
  "description": "When probes are scored against benign fine-tunes, including ones with unusual but harmless habits, they flag many benign examples or fail to transfer across training runs."
}