Three real runs, replayed step by step.
Each demo below is one complete run of the system, played back from its recorded trace: the raw evidence it looked at, the literature it searched, every line of code it ran, every time a check sent it back, and the manuscript it ended up with. Nothing is re-enacted or shortened for effect. You can pause anywhere, drag the scrubber, and download the paper and the full log.
Replays are built from engine/examples/<run>/stages by
webdev/tools/build_replay.py.
Evidence images are re-rendered with the same
matplotlib code the perception layer used at run time.