---
name: dialt-evals
description: Build and refine a hosted Dialt eval suite for a voice agent, validate its cases with Dialt Smart in text, then evaluate in voice.
---

# Build a Dialt eval suite

Follow the [eval guide](https://dialt.com/docs/api/evals/) for authoring, API usage, preflight, billing and result interpretation. This skill owns the workflow below; the guide and its linked API contract own execution details.

## Design the suite

Create a small hosted suite covering the agent's important successful journeys and consequential edge cases. Derive expected outcomes from its business requirements and tool contracts. Use mock tools unless the user requests a different setup.

## Validate in text with Dialt Smart

Start with one text attempt per case using Dialt Smart. Inspect transcripts, tool evidence, caller coverage and verdicts, including passes. Distinguish case-design, caller/harness, grading and agent failures.

Correct genuine design defects and rerun affected cases until the criteria pass with the intended challenges exercised. Never weaken valid requirements to make Smart pass. Stop and report unresolved cases when progress stalls or the user's budget is exhausted.

## Evaluate in voice

Freeze validated cases, then run in voice on the requested models, defaulting to Dialt Fast. Keep callers, fixtures, criteria and limits fixed across comparisons. Investigate speech and timing failures separately. For requested agent improvements, preserve before/after versions and rerun the same evals.

## Deliver through Dialt

Link hosted results and recordings with a brief findings summary; use the case playground for a live demo. Keep reproducible definitions locally. Build additional UI only when requested.