How do authoritarian states signal regime strength to citizens? Conventional wisdom focuses on displays of violence to intimidate audiences. We find this understanding incomplete as states adapt to the new media environment. We theorize dual-capacity signaling: authoritarian states use new media to project both coercive capacity (the ability to monitor, deter, and punish) andadministrative capacity (the ability to serve, coordinate, and govern). Applying this theoretical framework, we develop a new computational pipeline combining Large Vision-Language Models and Large Language Models and apply it to 48,858 videos from 31 Chinese municipal police Douyin accounts. Our study yields four findings. First, administrative signaling is dominant and has grown steadily over time, while coercive signaling remains a minority. Second, signaling follows the political calendar: coercive messages spike in the days before sensitive political events, while administrative content rises around public holidays. Third, across cities, online signaling correlates with local state capacity and socioeconomic pressure. Finally, we find a digital engagement dilemma: despite the dominance of administrative signaling, audiences engage more strongly with coercive displays. This study advances theories of authoritarian political communication and introduces a scalable computational method for analyzing political video content.