Two AI judges scored our agent's answer 0.85, but it never opened the file (tenureai.dev)

A recent case study highlighted a glaring issue in AI evaluation methodologies when an AI agent confidently answered a question about a Confluence page it never accessed. While two AI judge models rated the response positively, scoring it 0.85 based on its coherence and reasoning, a trace-based scoring method revealed a stark truth: the agent had not fetched the necessary document at all, resulting in a score of 0.000. This discrepancy underscores a critical flaw in current evaluation protocols that prioritize final answers over the verification process needed to substantiate those answers. This incident emphasizes the importance of rigorous evaluation in the AI/ML community, showing that plausible responses without proper evidence gathering can lead to misleading assessments of an agent's performance. The case serves as a call to refine scoring methods to evaluate not only the quality of the final answer but also the validity of the reasoning and evidence acquisition processes. The scoring logic and access policies used in this assessment are available in the open-source GroundEval repository, allowing researchers to replicate the findings and push for more robust evaluation frameworks in AI systems.