MeetBench-XL: Calibrated Multi-Dimensional Evaluation and Learned Dual-Policy Agents for Real-Time Meetings

MeetBench-XL addresses a high-value niche: the complex, multi-stakeholder environment of enterprise meetings. Technically, its 'Dual-Policy Agent' approach (likely separating high-level strategic reasoning from low-level tactical execution) is a sophisticated way to manage the latency-accuracy trade-off. However, the project's defensibility is low (score: 3) because it currently exists as a research artifact with zero GitHub stars and minimal community traction, despite having 7 forks (likely academic peers). The primary threat is 'platform domination'; Microsoft (Teams Premium), Google (Meet/Gemini), and Zoom (AI Companion) are already shipping these exact features. These incumbents possess 'data gravity' (access to actual meeting recordings) and 'integration surface' (the UI where the meeting happens) that an open-source benchmark or standalone agent architecture cannot easily overcome. While the benchmark itself is a valuable contribution to the research community, frontier labs like OpenAI (with GPT-4o's real-time voice/vision) are making the 'real-time fact-checking' aspect of this project a native capability of the foundation model itself, leaving little room for a standalone specialized software layer.