TEST-46 docs: align guides with voice testing best practices

fern/apis/api/openapi-overrides.yml

@@ -998,6 +998,16 @@ components:
         A record of an eval execution, including its target, status, results,
         costs, completion details, and lifecycle timestamps.
       properties:
+        results:
+          description: >-
+            Results for this individual Eval. Check them after status is ended.
+            An Eval that finishes normally contains one result; it passes only
+            when all judged checkpoints pass. Grouping multiple Evals requires your
+            own orchestration, not an Eval suite.
+        cost:
+          description: The cost of this Eval run in USD.
+        costs:
+          description: The cost breakdown for this Eval run.
         id:
           description: The unique identifier for the eval run.
         orgId:

fern/assistants/examples/appointment-scheduling.mdx

@@ -258,6 +258,16 @@ Use the Google Calendar integration for availability and booking, or your own AP
 
 ## 5. Test and validate
 
+Use sandbox calendar records and test identities. Don't create, change, or cancel real customer appointments during testing.
+
+| Check | Coverage |
+| --- | --- |
+| Evals | Correct booking arguments when all required details are known; ask when timezone is missing; don't confirm after a tool error. |
+| Simulations | A complete booking, an unavailable slot with an alternative, and rescheduling without losing the original appointment if the change fails. |
+| Controlled calls | Listen for date and time clarity, and verify the intended changes actually exist in the sandbox calendar. |
+
+Repeat critical scenarios before release. A tool mock checks behavior under a supplied result, not the calendar integration. See [plan test coverage](/test/plan-test-coverage).
+
 <Steps>
   <Step title="Attach a phone number">
     Create a phone number and assign your assistant. See [Phone calls quickstart](/quickstart/phone).
@@ -274,4 +284,3 @@ Use the Google Calendar integration for availability and booking, or your own AP
 - **Tools**: [Google Calendar](/tools/google-calendar), [Custom Tools](/tools/custom-tools)
 - **Structured outputs**: [Extract structured data](/assistants/structured-outputs-quickstart)
 - **Multichannel**: [Web integration](/quickstart/web)
-

fern/assistants/examples/support-escalation.mdx

@@ -705,6 +705,16 @@ Always be professional and efficient in your support."""
 
 ## 4. Test Your Support Escalation System
 
+Use dedicated test accounts and transfer destinations staffed by your team. The scripts below create calls for manual testing; they don't simulate callers or assert that escalation worked.
+
+| Check | Coverage |
+| --- | --- |
+| Evals | Escalate when policy requires it, avoid unnecessary escalation, and request the correct destination with the required context. |
+| Simulations | Follow an unresolved issue through escalation, including a failed or unavailable destination and a safe next step. |
+| Controlled calls | Verify the real transfer connects, the recipient receives the needed context, and the audio remains usable. |
+
+Repeat critical checks and turn routing failures into regression tests. See [test decisions with Evals](/test/evals-best-practices) and [test outcomes with Simulations](/test/simulations-best-practices).
+
 <Tabs>
   <Tab title="Dashboard">
     <Steps>

fern/assistants/structured-outputs-examples.mdx

@@ -1215,6 +1215,8 @@ Always test your structured outputs with these scenarios:
 4. **Edge cases** - Boundary values, special characters
 5. **Real conversations** - Actual call recordings or transcripts
 
+Valid JSON doesn't guarantee a correct extraction or judgment. Compare each output with expected values from human-reviewed examples, including known successes and failures. For outcome judgments, record what evidence is available: a reported booking isn't proof of a calendar update. Use synthetic values or remove personal information from reused calls. Follow [judge calibration and result review](/test/run-and-maintain-tests#review-more-than-the-pass-or-fail-label).
+
 ### Monitoring checklist
 
 Track these metrics for production deployments:
@@ -1291,4 +1293,4 @@ When retrieving call data via API:
 - [Structured outputs overview](/assistants/structured-outputs) - Main documentation
 - [Quickstart guide](/assistants/structured-outputs-quickstart) - Get started quickly
 - [API reference](/api-reference#structured-output) - Complete API documentation
-- [JSON Schema specification](https://json-schema.org/) - JSON Schema standard
+- [JSON Schema specification](https://json-schema.org/) - JSON Schema standard

fern/assistants/structured-outputs-quickstart.mdx

@@ -912,6 +912,7 @@ Common validation patterns for reliable extraction:
 - Use enums for categorical data to ensure consistency
 - Add descriptions to help the AI understand context
 - Test with real conversations before production use
+- Compare extracted values and judgments with human-reviewed examples. A valid schema checks the output's shape, not whether its values are true.
 - Monitor extraction success rates and iterate on schemas
 </Note>
 

fern/customization/multilingual.mdx

@@ -406,6 +406,14 @@ Configure greeting messages that work across multiple languages.
 
 Validate your configuration with different languages and scenarios.
 
+| Check | Coverage |
+| --- | --- |
+| Evals | Check the next response or routing decision in each supported language, including when the caller changes language. These checks don't test speech recognition. |
+| Voice Simulations | Run representative customer journeys in each language and with mixed-language input. Repeat critical cases and listen to recordings. |
+| Controlled calls | Ask native speakers to review understanding, pronunciation, and turn-taking on the real phone path. Synthetic voices aren't complete accent or noise coverage. |
+
+Use [plan test coverage](/test/plan-test-coverage) to prioritize the languages and journeys your customers depend on. Review both failed and passing calls.
+
 <Tabs>
   <Tab title="Dashboard">
     1. Use the **Test Assistant** feature in your dashboard