Qualitative examples
🎧 Audio Explorer
Browse real benchmark examples across all three protocols. Every example ships
its source document with the supporting evidence highlighted, so you can verify each answer against the document yourself.
How to read this. Context Saturation plays the actual spoken question and shows every model's
transcribed answer + grounding-judge verdict. Grounding Decay and Proactive Grounding show the full
multi-turn conversation. Model answer audio was scored in-flight and not persisted, so answers appear as transcript +
verdict rather than re-synthesized speech. Expand “Source document” on any card to see the evidence
highlighted in context.