Step 2: tool results

get_claim

{
  "include": [
    "arguments",
    "subclaims",
    "related",
    "findings",
    "dependents"
  ],
  "claim_id": "3c5e9fc9-3a05-44b4-bd1d-2785b377594d"
}

{"claim":{"id":"3c5e9fc9-3a05-44b4-bd1d-2785b377594d","kind":"claim","canonical_form":"A weaker AI monitor detects less of an AI agent's hidden misbehavior as the agent's capability advantage over the monitor grows.","claim_type":"empirical_derived","state":"active","decomposition_status":"complete","importance":0.45,"contestation":0.45,"domains":[],"tags":[{"slug":"scalable-oversight","name":"Scalable Oversight"},{"slug":"ai-control","name":"AI Control"},{"slug":"ai-scheming","name":"AI Scheming"}],"steward_state":"running","stewarded_at":"2026-11-11T19:08:02.514Z","canonical_direction_note":null,"created_at":"2026-11-11T19:00:18.833Z","updated_at":"2026-11-11T19:07:44.386Z","page_url":"https://minerval.ai/claims/3c5e9fc9-3a05-44b4-bd1d-2785b377594d"},"assessment":null,"preliminary_seed":{"credence":0.72,"note":"Control and monitoring experiments mostly point this way when the protocol is held fixed. In the Backdoor Code game of Engels et al. (2025) attacker skill rose steeply with general capability while auditor skill was nearly flat, and 2026 benchmarks on agentic coding tasks report catch rates falling steeply for weaker monitors against attacks built with frontier models. Two things hold the prior below high confidence. Most of the evidence widens the gap by weakening the monitor, which shows that monitor capability matters more directly than it shows that stronger agents evade better; in the Backdoor Code game stronger attackers produced more working backdoors, not subtler ones. And some results report weaker monitors keeping pace with stronger agents when given better scaffolding or the agent's reasoning, so the relation depends on what the monitor is allowed to see and do.","seeded_by_claim_id":"2d5868e4-745c-474f-ac6f-cc9cbc25ea8c"},"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":"2026-11-11T19:00:19.053Z","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":"2026-11-11T19:01:51.860Z","paid":true,"summary":"A funded mandate scheduled a pass on this claim."},{"trigger":"dependent_note","from":"the Steward of a node that depends on this one","sender_node_id":"353f02bc-ebdf-443f-a09e-38209d365bd4","sent_at":"2026-11-11T19:07:44.388Z","paid":false,"summary":"First assessment of \"Verifying an AI system's output for correctness and safety requires less capability than producing it\": inconclusive (confidence 0.7), no credence, because it holds for outputs th…"}],"requests":[],"requests_total":0,"subclaim_count":0,"instance_count":1,"source_map":null,"related":[{"id":"54994356-47c8-4a8f-8761-c9ce277a8456","text":"Better scaffolding and chain-of-thought access let weaker AI models reliably monitor stronger AI agents.","link_kind":"counterpart_position","reasoning":"The two are the opposing halves of the weak-to-strong monitoring question: one says detection falls as the agent outpaces the monitor, the other that scaffolding and reasoning access let a weaker monitor keep up. That claim already carries this one as a consideration against it, so the link here is lateral.","assessment_status":"doubtful","assessment_confidence":0.75,"assessment_credence":0.15}],"subclaims":[],"subclaims_assessed":0,"dependents":[{"id":"353f02bc-ebdf-443f-a09e-38209d365bd4","text":"Verifying an AI system's output for correctness and safety requires less capability than producing it.","claim_type":"empirical_derived","node_kind":"claim","relation_type":"contradicts","reasoning":"If a monitor catches less hidden misbehavior the further it falls behind the agent, then checking an output for safety draws on capability comparable to the producer's, which weighs against the claim for the safety half.","importance":0.6,"assessment_status":"inconclusive","assessment_confidence":0.7,"assessment_credence":null,"bounty_micro_usd":null,"checked":null,"formal":false},{"id":"804cd71e-671c-46a6-a37b-eb6d5cff952a","text":"Scalable oversight protocols remain effective against AI systems that deliberately deceive their overseers.","claim_type":"empirical_derived","node_kind":"claim","relation_type":"contradicts","reasoning":"Hidden misbehaviour by an agent is the deceptive case; if weaker monitors catch less of it as the agent's advantage grows, weak-over-strong monitoring loses effectiveness against deceivers.","importance":0.6,"assessment_status":null,"assessment_confidence":null,"assessment_credence":null,"bounty_micro_usd":null,"checked":null,"formal":false},{"id":"2d5868e4-745c-474f-ac6f-cc9cbc25ea8c","text":"AI oversight success rates decline as the capability gap between overseer and overseen system grows.","claim_type":"empirical_derived","node_kind":"claim","relation_type":"supports","reasoning":"The most direct evidence for the parent is measurement of detection rates in monitoring and auditing setups as the monitor-agent gap varies; this states that finding as a reusable proposition.","importance":0.5,"assessment_status":"supported","assessment_confidence":0.72,"assessment_credence":0.75,"bounty_micro_usd":null,"checked":null,"formal":false},{"id":"54994356-47c8-4a8f-8761-c9ce277a8456","text":"Better scaffolding and chain-of-thought access let weaker AI models reliably monitor stronger AI agents.","claim_type":"empirical_derived","node_kind":"claim","relation_type":"contradicts","reasoning":"If detection still falls as the agent's advantage grows, including under the best scaffolds, then design narrows the effect of the gap without removing it, and reliability at larger gaps is not shown.","importance":0.45,"assessment_status":"doubtful","assessment_confidence":0.75,"assessment_credence":0.15,"bounty_micro_usd":null,"checked":null,"formal":false}],"dependents_total":4,"arguments":[{"id":"69729fd5-9908-4b3c-a528-2a673e470e2c","name":"The argument from monitor-strength sweeps","stance":"for","content":"Against a fixed agent or fixed set of attacks, weaker monitor models catch less, so widening the gap from the monitor's side lowers detection.","has_written_form":false,"evaluation":null,"lean_check":null},{"id":"6afaf686-d982-43ec-b0e3-b7b4dd92522c","name":"The argument from agent-strength sweeps","stance":"for","content":"Against a fixed monitor, more capable agents complete hidden side tasks uncaught more often, so widening the gap from the agent's side lowers detection.","has_written_form":false,"evaluation":null,"lean_check":null},{"id":"a560af65-8788-48c4-81cd-c0be08c68544","name":"The argument that stronger attackers are more able, not stealthier","stance":"against","content":"Where detection is measured conditional on a working attack, stronger agents' attacks are not caught less often; their advantage is in executing the misbehavior, not hiding it.","has_written_form":false,"evaluation":null,"lean_check":null},{"id":"6383a43a-fc05-41e4-ac2d-e6889e0e1932","name":"The argument from protocol dependence","stance":"against","content":"What the monitor sees and how it is scaffolded moves detection more than the capability gap does, and can flatten the gap's effect.","has_written_form":false,"evaluation":null,"lean_check":null}],"findings":[]}

get_claim_instances

{
  "claim_id": "3c5e9fc9-3a05-44b4-bd1d-2785b377594d"
}

{
  "claim_id": "3c5e9fc9-3a05-44b4-bd1d-2785b377594d",
  "total": 1,
  "limit": 20,
  "offset": 0,
  "has_more": false,
  "instances": [
    {
      "id": "fc94a1f2-64fa-44da-8f7a-7d00496be3d9",
      "source_id": "0a56715f-3ced-4808-8629-4563823f5fef",
      "verbatim_text": "When monitors do not have access to reasoning traces, we find that the time horizon ratio between the agent and monitor model seems predictive of the chance that the agent successfully sneaks the side task past the monitor.",
      "context": "Summary finding of a prototype evaluation (SHUSHCAST) in which agents of varying time horizon attempt a hidden side task during a software task while monitors of varying time horizon review the transcript at a 5% false positive rate. The authors call the work early-stage and name the small task set and lack of elicitation as limitations.",
      "stance": "affirms",
      "confidence": 0.85,
      "speaker": "Megan Kinniment, Seraphina Nix, Thomas Broadley, Hjalmar Wijk, Neev Parikh",
      "publication": "METR",
      "source_date": "2026-01-22",
      "link": null,
      "source_title": "Early work on monitorability evaluations",
      "source_type": "unknown",
      "source_url": "https://metr.org/blog/2026-01-19-early-work-on-monitorability-evaluations/",
      "source_published_date": "2026-01-22",
      "recorded_by": "claim_steward",
      "display_rank": null,
      "reading": null,
      "draws_on": [],
      "story": {
        "standing": "untraced",
        "basis": "none",
        "depth": 0,
        "diverges": false,
        "downstream_total": 0
      }
    }
  ]
}