Five trials
Two boxes, red on the left and blue on the right, and the prompt “Go to the friendly one.” Qwen2.5-VL-3B drives, MF-IG attributes the w_z it emitted.
Convergence
One trial's field is noisy. Pool five and the structure that survives is the structure that was there. These are the same five rollouts, accumulated at two checkpoints — the left one is a single trial, the right one is all five.
n = 1 · one trial
Loading the field…
n = 5 · pooled
Loading the field…
The five trials
Not curated for success. A trial that drifts between both targets while the planner flip-flops is the most instructive artifact here, and the seeds are frozen, so what the run gave is what the library shows.
Per-trial playback is not here yet
Each of the five trials becomes independently scrubbable — the field evolving, with the commit step or the failure marked on the window — once ENV-07 has produced the per-trial fields. The pooled render above is real; this part is honestly missing rather than faked.