Understand comparison claims and evidence

After multiple models answer the same prompt, RadAIChat analyzes their responses and builds a comparison. The comparison is AI-generated and must be reviewed against the original responses and images.

Read the comparison summary

The summary shows the number of models and claims analyzed. Claims are organized as:

• Agreed: the compared responses support the claim.
• Disagreed: responses conflict about the claim.
• Unique: the claim appears in only part of the response set.

Use the available views to organize claims by findings, agreement, or models. Individual claims can also show a category, severity, confidence, supporters, skeptics, and models where the point was not mentioned.

“Not mentioned” does not necessarily mean disagreement. It means the comparison did not find that claim in the model’s response.

Inspect supporting evidence

Select a claim in the comparison panel. RadAIChat moves to the related model response and highlights the source text associated with that claim. Use this to verify how the comparison was derived instead of relying on the claim summary alone.

Evidence highlighting reflects model text, not independent confirmation from the source images.

If comparison fails

Choose Retry in the comparison panel. Wait until model responses finish before retrying, and review the individual panels even when no comparison is available.

For the broader workflow, see Compare model responses.