Go — μgrad tensor slice

19×19 lab · click intersection · Black to play

Clock

B · —:— W · —:—

On Min/side

9-cell slice below = tensor input (downsampled regions, same shape as ugrad-r0 Datasets.goboard). Clock counts the active side’s main time; pause stops both (sync). Broadcast / Follow remote uses BroadcastChannel('ugrad-go-board') — open a second tab: one plays (Broadcast on), one mirrors (Follow on). go-ugrad-monitor.html aggregates every broadcasting tab in one grid (optional toolbar Name). Plays with capture (no-liberty groups removed) and suicide blocked unless capturing; simple ko not enforced. Tensor env selects which game/synthetic distribution trains the shared MLP head.

Command station · live training slice

idle · pick Tensor env + Train

Automation: postMessage({ type: 'ugrad-tensor-train', env: 'go-board', steps: 80 }) · BroadcastChannel ugrad-tensor-train message { type: 'train', env, steps } · GoUgrad.runTrain(env, steps)

Iteration snapshots · scroll →

Saved games

Tensor lane · other games / tensors

Same 3×3 shape as #gu-mini9 · BroadcastChannel('ugrad-tensor-lane') · remote { type: 'tensor-slice', env, slice[9], source }. Toolbar Publish slice shares this board’s downsampled window; Lane in receives foreign slices here and in the header rail.

No remote slices yet — open another μgrad tab that publishes, or enable Publish on two Go tabs.