Step 16: tool results

match_claim

{
  "text": "Sparse probes on current SAE latents generalize worse out of distribution than dense linear probes on raw activations.",
  "exclude_claim_id": "20fee30f-0192-4b8c-b479-83c3f230014d"
}

{"outcome":"new","is_match":false,"matched_claim_id":null,"instance_stance":"affirms","new_canonical_form":"Sparse probes on current SAE latents generalize worse out of distribution than dense linear probes on raw activations.","direction_note":"The claim is stated as the affirmative comparative finding that SAE-latent sparse probes transfer worse than dense raw-activation probes.","confidence":0.91,"reasoning":"The closest node is d8a238a2-44d2-497a-b1a8-5c82190ba262, which asserts both near-perfect in-distribution performance for k-sparse SAE-latent probes and worse out-of-distribution transfer than dense probes. The submitted proposition asserts only the out-of-distribution comparison. Those are not identical: evidence that establishes the transfer gap need not establish near-perfect in-distribution performance, so the existing conjunction turns on an additional consideration. The proposed claim is therefore new rather than a wording improvement or denial of that node. The source states the OOD comparison affirmatively. Other nearby nodes are parameterized findings about other evaluations or conditions, not the same proposition.","alternative_matches":["d8a238a2-44d2-497a-b1a8-5c82190ba262","74170bea-5466-4689-8e10-99bd067d3631","0597a53c-75c0-4379-b709-13bdc85954f4","0cfd9ad2-65a2-49ef-8fa2-5f035e22073f"],"relationship_notes":"d8a238a2-44d2-497a-b1a8-5c82190ba262 is a very close overlapping finding but adds the distinct in-distribution performance conjunct. 74170bea-5466-4689-8e10-99bd067d3631 is a more specific result about harmful-intent detection in Gemma 2 9B IT. 0597a53c-75c0-4379-b709-13bdc85954f4 is an opposite comparison under synthetic-training/real-deployment conditions, while 0cfd9ad2-65a2-49ef-8fa2-5f035e22073f concerns data scarcity, label noise and covariate shift."}

match_claim

{
  "text": "Spurious correlates of a label are as likely as the target concept to be represented by individual sparse autoencoder latents.",
  "exclude_claim_id": "20fee30f-0192-4b8c-b479-83c3f230014d"
}

{"outcome":"new","is_match":false,"matched_claim_id":null,"instance_stance":"affirms","new_canonical_form":"Spurious correlates of a label are as likely as the target concept to be represented by individual sparse autoencoder latents.","direction_note":"The source asserts an equality of likelihood between representation of spurious label correlates and the target concept.","confidence":0.91,"reasoning":"The searches for the proposition as stated, paraphrases, and its negation returned related sparse-autoencoder claims, but none makes this specific comparative claim: that an individual SAE latent represents a spurious label correlate about as likely as the target concept. The closest candidate, 8393285b-cd1f-4b60-91bf-c96d1713be84, says SAE latents can reveal spurious differences between labeled datasets; that is about detecting such differences, not a comparison in representation likelihood between spurious correlates and target concepts. Other nearby results concern chance correlations, probing metrics, or general concept discovery and likewise turn on different considerations. Thus create rather than force a match.","alternative_matches":["8393285b-cd1f-4b60-91bf-c96d1713be84","a99904b1-d65f-430e-abe6-625a11628202","4dbe502e-402c-4ae9-8c6f-57a252a0d324"],"relationship_notes":"8393285b-cd1f-4b60-91bf-c96d1713be84 is a related capability claim about revealing spurious dataset differences; a99904b1-d65f-430e-abe6-625a11628202 concerns chance correlations inflating probe scores; 4dbe502e-402c-4ae9-8c6f-57a252a0d324 concerns broad concept-discovery capability. They are related context, not equivalent or contrary propositions."}