Hallucination Detection

See when your agent states something the conversation does not support, and suppress known-safe statements you do not want flagged.

What hallucination detection is

After a test conversation finishes, Voxli reads the agent’s messages and pulls out the factual claims they make. It then checks each claim against the conversation - the tool calls, events, and earlier messages the agent had to work with - and labels whether the claim was backed up.

This catches a common failure mode: an agent that sounds confident but states something it was never told. You see these results on the test result, alongside the assertion outcomes.

Hallucination detection is enabled by default. After a run finishes, choose Detect hallucinations on the run or comparison to check its claims. Detection doesn’t run automatically when a test is scored. If the option is missing, ask your workspace admin whether detection has been disabled for your workspace.

Reading claims on a test result

Open a test result and find the Hallucinations section on the left, beneath the assertion checks. Extracted claims are grouped by how well the conversation supported them:

Supported - the conversation backs up the claim.
Unsupported - nothing in the conversation supports the claim.
Contradicted - the claim goes directly against the conversation.
Unverified - the claim was not checked against the conversation.
Dropped - the claim was skipped by an exclusion rule (see below).

Contradicted and unsupported claims are the ones to worry about - both mean the statement should not be trusted. Each claim row can expand to show the evidence behind its verdict, and selecting a claim highlights the related messages in the conversation transcript on the right, the same way assertions do.

Suppressing known-safe claims with exclusion rules

Detection is deliberately strict, so it sometimes flags a statement that is actually fine - for example, a support agent correctly saying “we have no information on that.” Exclusion rules tell Voxli to stop flagging statements like that, without weakening detection everywhere else.

A rule is a short, free-text description of the kind of claim to skip. On the next run, any claim that matches an active rule is dropped from detection instead of being flagged. Existing results keep the verdicts they already have - rules only affect future runs.

Creating a rule from a flagged claim

The quickest way to add a rule is straight from a claim you disagree with.

On a test result, find the flagged claim in the Hallucinations section.
Click the thumbs-down on the claim - Create an exclusion rule from this claim.
The Create exclusion rule dialog opens. Write a short Rule describing the kind of claim to skip (for example, “Information about pricing”).
Click Create rule.

Managing rules in Settings

All your rules live under Settings > Hallucinations. To get there, click your workspace name in the top left corner, choose Settings, then open Hallucinations.

Click Create rule to write one from scratch.
Each rule appears with its author and when it was created. Short lists show these details beneath the rule; longer lists use table columns. Open the row’s three-dot menu and choose Edit or Delete.
Deleting a rule means matching claims are flagged again on future runs.

If you have no rules yet, the page explains that you can click thumbs-down on an extracted claim to create one, or write one from scratch.

Example

Testing for Speculation using Voxli walks through a test for unsupported claims. The article’s screenshots and setup steps describe an earlier version of Voxli. Follow the instructions above for the current detection workflow.

What’s next

Results - inspect a test result and its conversation transcript.
Assertions - the other half of how a test result is judged.