指标参考

客观指标(objective)从数据计算、免费;主观指标(subjective)由 LLM 评判、消耗积分。

工具使用

Analyzes tool call patterns including success/failure rates and per-tool breakdowns

Tracks model usage, token consumption, and estimated cost per model

Uses an LLM judge to evaluate AI assistant response quality across accuracy, efficiency, safety, and communication dimensions

使用评委模型评估用户的提示词质量，包括问题定义、上下文、约束和迭代引导

Detects repetition, loops, and quality degradation trends in AI responses

Detects whether the AI retains information from early conversation turns in later responses

Analyzes response times, latency distribution, and session pacing statistics