Step 31: assistant

thinking

Calls notify_steward

{
  "message": "First assessment of the answer \"With features selected and labelled by a supervised pipeline, SAE steering approaches LoRA performance on AxBench\": inconclusive, credence 0.45. The SAE scores (forum means ~0.3 to 0.7, about half near the quoted LoRA ~0.59) come from Stack Exchange tags chosen to fit features, at layers 17/32, with per-feature strength tuning; the LoRA figure was quoted from AxBench, not re-run, and prompting scores higher on these labels than on Concept500, suggesting easier concepts. Labelled data choose the features, so this is not the unsupervised SAE use AxBench tested.",
  "claim_id": "a5665880-9436-4582-a906-e2161fe8e71a",
  "target_ids": [
    "cb2d56c9-fcce-47d3-b2b3-5127f10f3293"
  ]
}