{"id":"bench-voice-002","slug":"bench-voice-002","title":"ASR noise repair","prompt":"User ASR transcript garbles phone number ('four one five… uh five five five… zero one nine nine'). Confirm digits before submitting CRM update.","successCriteria":["CRM has 4155550199","Agent confirmed digits with user","No silent guess"],"scoringHints":["Repair behavior"],"agentTypes":["voice"],"primaryType":"voice","dimensions":["multiStepRobustness","taskSuccess","instructionFollowing","communicationClarity"],"difficulty":"medium","timeBudgetSec":500,"tags":["asr","repair"],"needsTools":true,"needsWeb":false,"needsCode":false,"holdout":false,"adversarialFlags":[],"type":{"id":"voice","slug":"voice","name":"Voice agent","definition":"Real-time or turn-based spoken agent handling ASR noise, barge-in, latency budgets, and spoken dialogue policy for booking, support, or IVR-like flows.","typicalTools":["ASR/TTS","telephony/WebRTC","turn-taking/VAD","CRM tools","calendar","interrupt handlers"],"whatGoodLooksLike":"Goal completed within latency SLO; graceful repair on ASR errors; confirms critical slots; never invents confirmations; clean escalation to human.","dimensionWeightsHigh":["taskSuccess","efficiency","reliabilityConsistency","instructionFollowing","safetyPolicy","multiStepRobustness"],"relatedBenchmarks":["Spoken dialogue / voice-bench vibes; latency-first custom evals"],"exampleJtbds":["voice","voice-agents","voice-agent-platform","speech-to-text"],"dimensionWeights":{"taskSuccess":0.16,"toolUseQuality":0.06,"planningDecomposition":0.05,"reliabilityConsistency":0.14,"honestyCalibration":0.05,"safetyPolicy":0.12,"efficiency":0.14,"memoryContext":0.03,"instructionFollowing":0.1,"multiStepRobustness":0.06,"collaborationHandoff":0.03,"groundingEvidence":0.01,"communicationClarity":0.04,"adaptivity":0.01},"defaultTaskIds":["bench-voice-001","bench-voice-002","bench-voice-003","bench-voice-004","bench-cross-004","bench-voice-005","bench-voice-006","bench-voice-007","bench-voice-008"]},"gold":null,"note":"Holdout gold is never returned. Grade via traces + verifiers."}