Benchmark

Measures: Realistic multi-turn conversations test four abilities: keeping instructions from the start of the chat, recalling scattered details, handling versioned edits reliably, and staying self-coherent without sycophancy. An LLM judge (Gemini 2.5 Pro since the Feb 2026 update) scores each conversation against human-written, instance-level yes/no rubric questions.