Checkout API
healthy- Version
- v2.14
- Latency
- 128 ms
- Error rate
- 0.4%
VoxOpsCOMMAND
Evidence-Gated AI Incident Commander
AssemblyAI Voice AgentOffline
Safety policyActive
System · simulatedOperational
Operations workspace
Command Center.
Investigate by voice. Recover with evidence.
Infrastructure
Production Health
4 / 4 healthySimulated infrastructure
Payment API
healthy- Version
- v3.8
- Latency
- 94 ms
- Error rate
- 0.2%
Auth Service
healthy- Version
- v1.21
- Latency
- 72 ms
- Error rate
- 0.1%
Inventory API
healthy- Version
- v4.2
- Latency
- 111 ms
- Error rate
- 0.3%
Scenario-driven metrics · No external production connection
AssemblyAI Voice Agent
Voice Commander
Ready to connect
Your incident response starts with a conversation.
Session activity indicatorLatest command
Final transcriptYour spoken command will appear here.
Conversation 0 messages
No conversation yet.
Live tool execution
0 observed callsWaiting for an operational command.
Actual agent tool calls appear here.
Evidence before action
Incident Intelligence
Systems healthy
Ready when it matters.
Load a scenario, start the voice agent, and ask it to investigate. Evidence will guide every recovery decision.
Evidence-gated recovery
Policy first. Human authorization second.- 01
Evidence
Awaiting tool
- 02
Diagnosis
Pending
- 03
Policy validation
Pending
- 04
Human gate
Protected
- 05
Recovery
Pending
- 06
Verification
Pending
Human Safety Gate
Protected
No state-changing action awaiting authorization. Evidence must validate a proposal before approval becomes available.
Engineering evidence
Evaluation.
Two measurement layers. Results from this browser session only.
7A / Deterministic
Is the policy reliable?
11 local checks of diagnosis data, authorization requirements, unsafe action blocking, and verification.
Open deterministic checks →7B / Live AssemblyAI
Does voice reach the right tool?
6 spoken trials observe transcripts, the first operational tool, arguments, safety outcomes, and Voice → Tool latency.
Open live benchmark →Stage 7A
Deterministic Safety & Reliability Evaluation
Reproducible tests exercise the same VoxOps incident tools, scenario evidence and rollback safety policies used by the live voice agent.
✓
Evaluation ready
Run the deterministic evaluation suite to generate reproducible reliability and safety evidence.
✓3 diagnosis tests
✓2 authorization tests
✓4 safety tests
✓2 recovery tests
Stage 7B
Live Voice BenchmarkLive AssemblyAI Voice-Agent Benchmark
Measure the actual microphone → AssemblyAI transcript → agent reasoning → tool-call → VoxOps safety-policy path. These results are separate from the deterministic Stage 7A evaluation.
Benchmark Control
Checkout investigation
Scenario: Checkout Release Regression
Tests whether the live voice agent correctly routes a Checkout incident investigation to the evidence-gathering tool.
Speak This Command
“Investigate the Checkout API incident.”
Expected toolinvestigate_incident
Expected serviceCheckout API
Expected outcomeTool accepted
Trial Procedure
1
Prepare the required incident scenario.
2
Open Command Center, start the Voice Agent, and return here once connected.
3
Click Start Trial, then speak the displayed command naturally.
4
The trial completes automatically after the final transcript and tool call are captured.
Live Capture
Data observed from the real voice-agent session.
Final Transcript
○ waitingWaiting for transcript.user...
Tool Selection
○ waitingWaiting for tool.call...
Tool Arguments
○ waitingNo tool arguments captured yet.
Completion conditions
○Final transcript received
○Operational tool observed
Prepare a benchmark scenario and start a trial. This panel will display the actual AssemblyAI transcript and tool call as they arrive.
Live Metrics
Complete at least one live trial to generate measured voice-agent metrics.
Measurement Scope
Stage 7B measures observed behavior from a real AssemblyAI voice session: final transcript capture, operational tool selection, extracted tool arguments, VoxOps safety-policy outcome, and voice-trial-to-tool-call latency. It does not claim that the latency shown here is pure model inference latency, because microphone input, speech endpointing, transcription, agent reasoning, networking, and tool routing are all part of the measured path.
Accountability, built in
Audit & Postmortem.
Every investigation, policy decision, and human authorization in one record.
Incident Memory
Incident Timeline
Automatic audit trail of investigation, authorization and recovery.
Timeline events will appear automatically when an incident begins.
Postmortem
Incident Report
An evidence-backed recovery proposal creates a draft report. Policy blocks and investigations are recorded in the timeline.