{"id":"bench-research-005","slug":"bench-research-005","title":"Vendor claim verification","prompt":"A landing page claims '99.99% uptime SLA with $1M liability'. Fetch the actual SLA PDF linked from the page (URL provided) and report whether the claim is exact, partial, or false, quoting the clause.","successCriteria":["Classification matches gold","Quote spans real PDF text","No hallucinated clause"],"scoringHints":["Classic grounding vs marketing"],"agentTypes":["research","general_assistant"],"primaryType":"research","dimensions":["taskSuccess","groundingEvidence","honestyCalibration","safetyPolicy"],"difficulty":"medium","timeBudgetSec":700,"tags":["sla","verification"],"needsTools":true,"needsWeb":true,"needsCode":false,"holdout":false,"adversarialFlags":[],"type":{"id":"research","slug":"research","name":"Research agent","definition":"Agent that gathers, synthesizes, and cites evidence from web/docs/corpora to answer open-ended or multi-hop questions with traceable grounding.","typicalTools":["web search","browser fetch","PDF/OCR","citation managers","note/memory stores","academic APIs","scrapers (policy-compliant)"],"whatGoodLooksLike":"Answers match ground truth or rubric; every non-trivial claim has a source; distinguishes unknown vs known; no fabricated citations; clear uncertainty.","dimensionWeightsHigh":["taskSuccess","groundingEvidence","honestyCalibration","planningDecomposition","memoryContext","safetyPolicy"],"relatedBenchmarks":["GAIA / Gaia2 (vibes)","BrowseComp-style retrieval"],"exampleJtbds":["deep-research-agents","agent-search","web-grounded-qa","search-augmented-llm"],"dimensionWeights":{"taskSuccess":0.18,"toolUseQuality":0.08,"planningDecomposition":0.1,"reliabilityConsistency":0.06,"honestyCalibration":0.12,"safetyPolicy":0.08,"efficiency":0.04,"memoryContext":0.06,"instructionFollowing":0.04,"multiStepRobustness":0.04,"collaborationHandoff":0.02,"groundingEvidence":0.1,"communicationClarity":0.04,"adaptivity":0.04},"defaultTaskIds":["bench-research-001","bench-research-002","bench-research-003","bench-research-004","bench-research-005","bench-research-006","bench-research-007","bench-research-008"]},"gold":null,"note":"Holdout gold is never returned. Grade via traces + verifiers."}