Run step

Step 10: tool results

2026-11-15 04:45:38 UTC · previous step · next step

record_argument_instance

Input

{
  "url": "https://arxiv.org/abs/2412.01784",
  "title": "Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models",
  "stance": "advances",
  "context": "Abstract; the comparator is the same model answering honestly, not a benignly underelicited model.",
  "reading": "premise_unstated",
  "speaker": "Tice et al.",
  "argument_id": "588075dd-3071-49c1-8180-b1e9004b8b96",
  "source_date": "2024-12",
  "source_kind": "preprint",
  "verbatim_text": "Non-sandbagging models show predictable performance degradation with increasing noise; we demonstrate that sandbagging models exhibit anomalous performance improvements."
}

{"source_kind":"preprint","success":true,"instance_id":"47a5ffcb-ee9c-4aa5-882e-d8fed043646c","source_id":"5264bcbf-3b01-49b7-92da-45bc30b16569","message":"Recorded a advances instance of argument \"The argument from fully elicited models only degrading under noise\" from https://arxiv.org/abs/2412.01784."}

add_related_claim

Input

{
  "kind": "counterpart_position",
  "claim_id": "01232bf1-55f7-42fd-9f39-8a38bfde7837",
  "reasoning": "The other claim states the noise-injection contrast between sandbagging and non-sandbagging models; this one states the limit of that contrast when the comparator is benignly underelicited.",
  "other_claim_id": "e0abe646-c78f-417c-813f-b8e2c15d775c"
}

{"success":true,"linked":true,"link_id":"8eb31ee0-f89e-4879-af79-4579eee45445","message":"Linked 01232bf1-55f7-42fd-9f39-8a38bfde7837 <-> e0abe646-c78f-417c-813f-b8e2c15d775c (counterpart_position); it now shows as a see-also on both claims."}