Replay an episode
Any scored episode of the benchmark: the 195 headline episodes, the 55 ablation episodes and the experiments E1–E3 and E5. The E2 and E3 lists also offer the provider-error attempts that were re-run (labelled “error”; not scored), and E2’s Sol T3 trial 2 appears twice (a trial run twice by mistake). A replay re-runs the recorded tool calls through the simulator; nothing is generated again by a model. Pick an episode and generate its replay; you watch it being drawn.
No pre-rendered movie for this episode. Generating one replays the record through the simulator on this server and takes one to several minutes; you see its frames as they are drawn. One replay renders at a time and a few can wait in a queue.
stimulate call runs a whole pulse protocol in
simulated time, and a movie fits an episode of an hour or more into a few minutes, so it plays the first pulse of
each call and the pulses that made the target fire in real time and fast-forwards the rest. The movies are
illustrative, not timing evidence; use the episode pages and the records for timing. The left half is exactly what
the tools returned to the agent at that moment. The right half is a reconstruction from the simulator: the same
kinds of views the agent can ask for (microscope, room camera, traces), shown continuously and noise-free on one
fixed brightness scale, plus hidden truth the agent never sees (true activity and culture health). The agent could
see parts of it, but only when it called a tool, with camera noise, never all of it at once. A model that chose a
low imaging light shows dim cells.
E2 and E3 ran on their own experiment branches (E2: a
stimulate call stops when the cover comes off or the room lights come on; E3: one sentence of the
codes document). The replay runs the benchmark's frozen simulator, so a call whose result the experiment's change
altered is flagged in the movie as differing from the record; the left half still shows what the agent saw.