Step 2: tool results

get_claim

{
  "include": [
    "arguments",
    "subclaims",
    "related"
  ],
  "claim_id": "461a798c-19b2-4b7a-9f26-aa1dfed627fb"
}

{"claim":{"id":"461a798c-19b2-4b7a-9f26-aa1dfed627fb","kind":"claim","canonical_form":"On AuditBench, logit-lens and sparse-autoencoder tools provide little benefit across tested target-model types.","claim_type":"empirical_verifiable","state":"active","decomposition_status":"complete","importance":0.5,"contestation":0.45,"domains":[],"tags":[{"slug":"mechanistic-interpretability","name":"Mechanistic Interpretability"},{"slug":"ai-capability-evaluation","name":"AI Capability Evaluation"},{"slug":"large-language-models","name":"Large language models"}],"steward_state":"running","stewarded_at":"2026-11-14T20:42:25.112Z","canonical_direction_note":"The source asserts the affirmative finding that these interpretability tools yield little auditing benefit in the tested AuditBench settings.","created_at":"2026-11-13T01:21:54.373Z","updated_at":"2026-11-14T20:39:57.163Z","page_url":"https://minerval.ai/claims/461a798c-19b2-4b7a-9f26-aa1dfed627fb"},"assessment":null,"coverage":{"examined":false,"on":[],"note":"No judgment of this node is in force. That is not a finding either way."},"quantity":null,"threshold_of":null,"question":null,"formalization":null,"formalization_pending":null,"formalization_history":[],"verification":null,"lean_checks":[],"lean_checks_total":0,"waiting_messages":[{"trigger":"structure_and_assess","from":"the system","sender_node_id":null,"sent_at":"2026-11-13T01:21:55.990Z","paid":false,"summary":"New claim onboarded. Give it the structure and assessment its importance warrants; a minor or settled claim may need no decomposition."},{"trigger":"mandate_plan","from":"a mandate's plan","sender_node_id":null,"sent_at":"2026-11-14T20:39:57.165Z","paid":true,"summary":"First assessment. Read the benchmark paper's results on interpretability tools directly: which target-model types, how hidden behaviours were installed, what the auditing agent could do with each tool…"}],"requests":[],"requests_total":0,"subclaim_count":0,"instance_count":1,"source_map":null,"arguments":[],"related":[],"subclaims":[],"subclaims_assessed":0,"unplaced_argument_passages":[{"instance_id":"adfa3c57-2151-49e3-8dbf-748259935785","source":{"id":"a8bae70a-6ace-4953-ada7-146a55bcf8d3","url":"https://arxiv.org/pdf/2602.22755","title":"AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors"},"verbatim_text":"White-box tools outperform the default in SDF models with SFT adversarial training: averaged over the five white-box tools, success rates rise from 40.2% to 46.9%, while on TD models they stay roughly flat (34.4% vs. 33.5%). Activation oracles and prefill-based steering are strongest, both beating the default on 8 of 14 behaviors, respectively. Logit lens and sparse autoencoders provide little benefit on any target model type","context":"Benchmark comparison of agent success rates per tool across training configurations.","stance":"advances","direction":"for"},{"instance_id":"9e4c8361-1edc-4c95-a9a2-c4ff7c0eb17f","source":{"id":"a8bae70a-6ace-4953-ada7-146a55bcf8d3","url":"https://arxiv.org/pdf/2602.22755","title":"AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors"},"verbatim_text":"White-box tools outperform the default in SDF models with SFT\nadversarial training: averaged over the five white-box tools, success rates rise from 40.2% to 46.9%,\nwhile on TD models they stay roughly flat (34.4% vs. 33.5%). Activation oracles and prefill-based\nsteering are strongest, both beating the default on 8 of 14 behaviors, respectively. Logit lens and\nsparse autoencoders provide little benefit on any target model type","context":"Benchmark comparison of agent success rates by tool and training configuration","stance":"advances","direction":"for"}],"unplaced_argument_passages_total":2}