Skip to main content
Kindred compares the original trace against the replay trace and highlights the first important difference it can identify.

​
Divergence Kindred detects

  • Changed tool selection
  • Changed tool arguments
  • Changed tool result
  • Missing step
  • Extra step
  • Different step type
  • Changed assistant output
  • Changed unstructured content

​
Assistant output comparison

Kindred first checks whether the assistant responses are textually close. If the difference is unclear and semantic comparison is enabled, Kindred can use our custom LLM judge to classify the output difference as:
  • semantic_match
  • minor_variance
  • material_divergence

​
Tool comparison

Tool calls and tool results are mostly field-by-field or exact comparisons. For example, Kindred can show that a tool argument changed or that a tool returned a different result.