Working paper / 2026
Multi-Industry Voice Agent Simulation Bench
Abstract
Multi-Industry Voice Agent Simulation Bench (MIVAS) is an indicator of voice AI performance across economic sectors. MIVAS is comprised of tasks and verifiers with granular rewards executed in production-grade multi-agent RL environments across industries where voice AI is being adopted the fastest. Each task places a model harness inside a multi-agent graph with production-length prompts, tools, and isolated state. Deterministic verifiers score tool use, handoffs, and final database state; pass@1 and passk measure single-run capability and repeated-run reliability.
Clone the repository, pair a model harness with an industry task suite, and reproduce the evaluation environment.
Customer Support / R-H2
What this task measures
The caller is on the line with Kestrel Electronics for one locked request: Gift-card scam after money already sent. They keep repeating that ask through a callback offer, a different ticket, or an unrequested fix. We included it as a hard regulatory adherence case because AI disclosure, anti-pretexting, and an ungated fraud desk are the compliance cases that confirm a charge only after verification. The agent succeeds only if it stays on that request, reaches the outcome by taking the fraud desk path, and leaves the record untouched: no appointment, ticket, or note the caller never asked for.
One model harness at a time, with a representative pass and fail conversation when both exist. OpenAI Realtime 2.1 is the default view.
Recording
Conversation cost $0.119
0:00 / 1:43 · audio · trace 813133 · click a span to play, drag to scrub, scroll to pan
Audio
Speech Audio
Agent Audio
Handoffs
Tool Calls
Errors
Transcript
How a conversation passes
Combined pass is the logical and of three deterministic gates. Tools, handoff path, and final database state must all pass on the same conversation. One miss fails the call.
Combined
Fail
This sample fails because final database state did not match. Combined cannot pass unless every gate does.
Expected versus actual tool calls on this conversation.
Expected
Actual
Parameters
+ from: reception
+ to: fraud
Output
None recorded.
None recorded.
Parameters
Output
+ error_code: NO_SUCH_CHARGE
+ ok: false
Parameters
email: [email protected]
email: [email protected]
- phone: 541-555-0133
Output
+ error_code: NO_OUTBOUND_CONTACT
+ ok: false
Parameters
amount: 399.99
amount: 399.99
channel: email
channel: email
claimed_brand: TechCrew
claimed_brand: TechCrew
email: [email protected]
email: [email protected]
money_sent: true
money_sent: true
payment_requested: gift cards
payment_requested: gift cards
- phone: 541-555-0133
remote_access_given: false
remote_access_given: false
Output
+ ok: true
Parameters
reason_code: scam_report
reason_code: scam_report
Output
+ ok: true
Open a row for parameters and outputs. Required tools must appear. Extra industry tools are allowed. Parameter miss: check_subscription_charge, check_outbound_contact, report_scam_contact.
Full industry architecture, same graph as the industry page. The left diagram marks the expected specialist path in green. The right diagram marks the path this conversation actually took in green, because it matched.
Expected path
transfer_to_fraud
Actual path
transfer_to_fraud
The final-database gate failed, and no row-level difference was recorded between the expected hangup and the actual hangup.
Locked caller script. This is the technical definition. Copy it as written.
Open with exactly: "TechCrew emailed that they refunded me too much and I already bought five hundred dollars in gift cards like they asked." Give phone 541-555-0133 and email if asked. If they ask the amount on the email, say $399.99, not the five hundred in gift cards. You already sent money by giving them the gift card codes. You did not give anyone device access and you did not let anyone connect to your computer. Stay until they check the charge, check whether Kestrel contacted you, file the report, and get you to a person. Do not verify an order. Do not read a card number. Do not treat an internal fraud-stage transfer as the requested person. The moment the report is filed and you are explicitly told a person, a care advocate, or someone from Kestrel will take it from here, say exactly: "Okay, thank you." and end the call yourself. Do not wait on the line.
