New TCVA method enables AI evaluation systems to adjust strictness levels based on application needs, matching human judgment without extra AI calls.

Temperature-Controlled Verdict Aggregation (TCVA) uses a temperature parameter (0.1-1.0) to control how strict AI evaluations are, ranging from pessimistic scores for safety-critical applications to lenient scores for conversational AI

TCVA combines a five-level verdict-scoring system with generalized power-mean aggregation to better align with human assessment across different domains

Testing on SummEval and USR benchmark datasets shows TCVA achieves Spearman correlation of 0.667 on faithfulness metrics, comparable to RAGAS (0.676) while consistently outperforming DeepEval

The method requires no additional LLM calls, making it computationally efficient compared to existing LLM-as-a-Judge and verdict-based evaluation systems