The benchmark's T1 culture simulator, running on this server: one simulator per open page, at most 3 per visitor (IP address) and 4 in all; a session ends after 15 minutes idle or 60 minutes in all. The left side acts through the agents' T1 tool calls, all of them except the room camera and Python; the purple drawer below shows and edits the truth the agent never sees.
Click a cell to stimulate it with the settings on the right. Click empty tissue to stimulate a control spot there.
Every T1 tool the agent has, without its turn clock or decision limits. The room camera (view_rig) and the python sandbox do not run on this hosted bench.
Each pulse is one stimulate call with repeats 1; the train plays out as time passes.
Seeds are practice seeds (200000–299999), never the benchmark's dev or test seeds: any other number is mapped into that range, and a blank seed picks one at random. Leave Neurons blank for T1's own count (26–30), or choose 20 to 80. Larger dishes take longer to build, and frames slow as cells are added.
Each switch starts or ends the fault now, with the symptoms its scenario has mid-session. A loose cable is also fixed by the agent's Reseat.
No pulse yet.
| A → B | weight | P(B | spot on A) | graded |
|---|
| cell | roi | opsin | fires @limits | calcium | opsin left |
|---|