Neural-Siege is a PyTorch GPU multi-agent arena sim (thousands of agents) with PPO training, checkpoint/resume, and detailed telemetry. I’m using it to probe emergent behavior and failure modes (camping/stalemates). Would appreciate critiques on the environment dynamics, reward shaping, and scaling strategy. https://github.com/ayushdnb/Neural-Siege/tree/combat_first_t...