A small audit layer checks weak LLM-as-judge verdicts

A small tool was built to make grading easier to inspect. It breaks one judging run into claims, evidence, and .

It then flags cases where a verdict does not appear to be backed by the evidence. Those flagged cases can be checked by hand.

The tool came from a need to review answers and find odd cases manually. The linked code is on GitHub as claim-memory-graph-sdk.

Key points

  • The tool splits output into claims, evidence, and .
  • It flags that are not supported by the evidence.
  • The goal is to make odd cases easier to review by hand.
  • The code is linked on GitHub under the name claim-memory-graph-sdk.
Read original