Minerval

← run e54a02fd

Run step

Step 10: tool results

2026-11-14 20:42:08 UTC · previous step · next step

match_claim

Input
{
  "text": "Selecting SAE features by high output score improves their steering performance relative to unfiltered SAE features.",
  "exclude_claim_id": "d3311aac-502f-4e69-ab97-53ee2b6de622"
}
{"outcome":"new","is_match":false,"matched_claim_id":null,"instance_stance":"affirms","new_canonical_form":"Selecting SAE features by high output score improves their steering performance relative to unfiltered SAE features.","direction_note":"The claim is stated affirmatively because the source asserts that output-score-based feature selection improves SAE steering relative to using unfiltered features.","confidence":0.88,"reasoning":"I searched the proposed wording, close paraphrases, and the negation. The closest candidate is 22ec9a4c-c85b-4dff-bd65-ca53e2ba04b8, which says that poor feature selection partly explains AxBench's negative SAE-steering results. That is a broader causal explanation, whereas this proposition asserts a particular selection criterion (high output score) and a particular comparison (selected versus unfiltered features); evidence can bear on one without establishing the other. The candidate's recorded source passage that careful feature choice closes much of the performance gap makes it a meaningful near-miss, but does not collapse the distinct propositions. The quantity e96b98d4-e3a4-4172-9193-2c10fe309f56 concerns a ratio of output-score-filtered scores to AxBench-reported scores on a specific model, benchmark and layers, not the general comparison with unfiltered SAE features. Other nearby claims concern different selection pipelines or comparisons, not the same proposition.","alternative_matches":["22ec9a4c-c85b-4dff-bd65-ca53e2ba04b8","e96b98d4-e3a4-4172-9193-2c10fe309f56","a5665880-9436-4582-a906-e2161fe8e71a","1b0574cd-8584-4836-9d75-5f8c3515c8b1"],"relationship_notes":"The new claim is a specification of the broader claim that poor SAE feature selection partly explains AxBench's negative steering results (22ec9a4c-c85b-4dff-bd65-ca53e2ba04b8). The ratio quantity e96b98d4-e3a4-4172-9193-2c10fe309f56 is related evidence but uses a distinct benchmark-score comparator; it is not the quantity/value claim itself. The remaining candidates concern other SAE steering comparisons."}