Multi-agent dialogue benchmark runs evaluating model seat behavior under direct operator filesystem claims. Symmetrical A/B matrix comparing Google Gemini vs. Moonshot K2.7 across 3 levels of prompt escalation. https://tvox.online/experiments