Skip to content

Product

A run produces a decision you can take apart.

VoiceCanary exercises scenarios against a staging agent, checks behaviour against your approved policy, compares claims with recorded outcomes, and reports the evidence behind every verdict.

Scenario configuration

Define the calls a run exercises: successful bookings, injected tool failures, retries, caller corrections and authorization edge cases.

Customer-approved policies

Your policy is the reference, not a generic rulebook. Policies are versioned so a verdict can always be traced to the rule it was judged against.

Actual outcome assertions

Assertions compare the agent's claim with what the approved outcome source recorded — an appointment that exists, with the agreed details.

Correlated evidence

Transcript turns, tool requests and responses, and recorded state are linked to the checks that rely on them.

Version comparisons

Two agent versions, identical scenario, policy and evaluator versions, and a diff of the resulting verdicts.

Human review

A reviewer can dispute a verdict and attach reasoning. Feedback proposes a change to criteria; it never silently rewrites them.

Anatomy of a result

Everything a single result carries.

Labelled fixture — scripted actor, no live call and no patient records.

Release decision: BLOCKED
Agent version
agent-1.4.2
Scenario version
scn-false-confirm-3
Policy version
pol-sched-booking-1.2
Evaluator version
eval-0.7.1
Expected outcome
Either an appointment record exists matching the agreed slot, or the agent accurately reports that booking did not complete and follows the approved fallback.
Observed outcome
No appointment record exists. The agent stated the appointment was booked and ended the call.
Evidence references
turn-5, tool-1, state, turn-5, policy, turn-2, tool-1, tool-1
Coverage limitation
Covers the checks and conditions in this scenario only. Not a statement about untested workflows.
Failure attribution
Agent policy violation (unsupported success claim).

Release decisions

Three outcomes, and no averaging.

A critical failure cannot be averaged away by passing checks elsewhere in the run.

Release decision: PASS

PASS

All mandatory applicable checks meet the approved release policy for this run.

Release decision: BLOCKED

BLOCKED

A critical violation was observed, or a configured blocking threshold was met.

Release decision: INCOMPLETE

INCOMPLETE

Required evidence or evaluation was unavailable. Nothing is inferred about the underlying action.

Integrations

Current integration status.

Listed honestly. Nothing below is a shipped production connector.

  • Live voice-agent connection

    Implemented and exercised locally

    Signed sessions, a blocking scheduling tool and provider evidence collection are built. Real synthetic-caller calls have completed accepted and rejected reschedules with assistant-reviewed claims. A hands-on browser microphone acceptance check is still outstanding.

  • Recording import

    Working in the local workspace

    Import aligned speaker recordings and timestamped transcripts from a public sample or an authorised archive. Recorded human analysis is kept separate from scheduling fixtures and live-agent evaluation.

  • Customer scheduling systems

    Planned

    No connector to a customer scheduling system has been built. The current booking system is a local fictional database used for testing.

  • CI release gate

    Planned

    Repository CI currently checks this application's own code. A command that consumes completed evaluation results and enforces a release policy is a later integration.

Start with the workflow that matters.

Scope a pilot around your agent, your success criteria and the evidence your team needs to make a decision.

A guided walkthrough tailored to your team and your workflow.