Link: https://arxiv.org/abs/1908.10396 Year: 2020 ### **Problem framing** (what exactly is predicted, at what horizon, for whom?) ### **Output type** (K trajectories? occupancy+flow? language tokens?) ### Uncertainty representation (modes, probabilities, calibration) ### **Key engineering constraints** (latency, memory, data requirements) ### **Evaluation & failure cases** (what the metrics miss; what breaks in the wild)