Every answer below is unedited real output of the iwazaru engine — hybrid retrieval with cross-encoder reranking over two corpora (Japanese tax law + patent publications, ~7,100 chunks, including one document that exists only as a noisy paper scan), generation by a 122B sparse-MoE model (~10B active) running entirely on one local machine. Recorded 2026-07-06. Playback is accelerated; true response times are shown per scene. 実時間版はこちら.