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.
Legal / C4-H4
What this task measures
The caller is on the line with Halverson & Reed for one locked request: Medical malpractice Florida intake with records. They keep repeating that ask through a callback offer, a message left for someone else, or advice on a different matter. We included it as a hard intake and documents case because taking a narrative and a document packet is representative plaintiff work that starts after the conflict-before-facts rule has held. The agent succeeds only if it stays on that request, reaches the outcome by following the conflict screening then intake path, and leaves the record exactly as this case requires: the right row written or cleared, and nothing extra.
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 $0.406
0:00 / 4:48 · audio · trace 804008 · 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 tools and final database state did not match. Combined cannot pass unless every gate does.
Expected versus actual tool calls on this conversation.
Expected
Actual
Parameters
+ full_name: Aisha Rahman
+ phone: 3055550288
Output
+ ok: true
Parameters
+ from: reception
+ to: screening
Output
None recorded.
None recorded.
Parameters
- opposing_party: Gulfshore Clinic
+ opposing_party: Gulf Shore Clinic
Output
+ ok: true
Parameters
practice_area: medical_malpractice
practice_area: medical_malpractice
Output
+ ok: true
Parameters
practice_area: medical_malpractice
practice_area: medical_malpractice
state: FL
state: FL
Output
+ ok: true
Parameters
incident_date: 2026-05-20
incident_date: 2026-05-20
practice_area: medical_malpractice
practice_area: medical_malpractice
state: FL
state: FL
Output
+ ok: true
Parameters
+ from: screening
+ to: intake
Output
None recorded.
None recorded.
Parameters
incident_date: 2026-05-20
incident_date: 2026-05-20
practice_area: medical_malpractice
practice_area: medical_malpractice
state: FL
state: FL
- summary: Retained sponge after surgery.
+ summary: Caller reports a clinic left a sponge in after surgery, requiring a return procedure. Needs evaluation for medical malpractice.
Output
+ ok: true
Parameters
- provider: Jackson Memorial
+ provider: Gulf Shore Clinic
Output
+ ok: true
Parameters
channel: email
channel: email
Output
+ ok: true
Parameters
+ reason: Caller requested medical records release; intake packet sent; caller declined booking at this time.
Output
None recorded.
None recorded.
Open a row for parameters and outputs. Required tools must appear. Extra industry tools are allowed. Parameter miss: check_conflict, record_intake, request_records_authorization. Extra: lookup_caller, end_call.
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_screening → transfer_to_intake
Actual path
transfer_to_screening → transfer_to_intake
Rows that differ between the expected hangup and the actual hangup.
extra · documents · 1
Expected
Actual
+ caller_id: c_new
+ id: 1
+ kind: intake_packet
extra · documents · 2
Expected
Actual
+ caller_id: c_new
+ id: 2
+ kind: records_authorization
missing · documents · {"caller_id": "c_new", "kind": "intake_packet", "target": "email"}
Expected
Actual
- caller_id: c_new
- kind: intake_packet
missing · documents · {"caller_id": "c_new", "kind": "records_authorization", "target": "jackson memorial"}
Expected
Actual
- caller_id: c_new
- kind: records_authorization
intakes · 1
Expected
Actual
- summary: Retained sponge after surgery.
+ summary: Caller reports a clinic left a sponge in after surgery, requiring a return procedure. Needs evaluation for medical malpractice.
Locked caller script. This is the technical definition. Copy it as written.
Open with exactly: "A clinic in Florida left a sponge in after surgery and I had to go back." When asked who this would be against, say exactly: "It's Gulfshore Clinic." Florida. May twentieth, twenty twenty six. You have two jobs on this call: they record the intake, and they send both the medical records release to Jackson Memorial and the new client packet by email. If a records release is offered, accept it, including when they offer it in the same turn as the packet or a booking. When asked which doctor, hospital, clinic, or provider treated you, say exactly: "Jackson Memorial." Accept email packet. Decline booking. Do not ask for a staff member. Stay until they confirm the intake is on the file, the records release has gone out, and the packet has been sent. A spoken 'we'll take care of it' without those writes is not enough. Then thank and end. Once they confirm the intake is recorded, the records release is sent, and the packet is sent, say exactly: "Thank you, that's all I needed." and end the call.
