01Static reach
Isolate representation precision. The observation error is quantized into a finite token budget before becoming a command.
q(e) = round(e × levels) / levels
Coordinate workcell
effector (-0.55, 0.35) → target (0.56, -0.28)
Content predicts the frame. Control must change the future.
Which tokens can move a body?
POSITION ERROR124.0 cm
OBSERVATION AGE2 frames
Predict before acting
Token objective
Capacity
Delay
Read the loop, not the screenshot.
A plausible next frame can still hide the state that chooses the next action.
Transparent teaching equations. These are not reported LingBot-VA 2.0 internals.
02Distractor rejection
Content-first tokens spend state on a vivid but task-irrelevant object. Control-aligned tokens preserve the target-to-effector relation.
u_content = 0.46q(e) + bias
03Disturbed transfer
A late observation describes a world that has already changed. Extra capacity sharpens that stale description but cannot make it current.
e_obs[t] = target[t-d] - body[t-d]
What the loop measured
Prediction, consequence, explanation.
FINAL —OVERSHOOT —STEPS 0
Waiting for a prediction
Choose which objective you expect to produce the lower control error. Then run the same closed loop and let the trajectory answer.
Transfer the idea
A new target. A hidden shove. Configure the loop, then diagnose its bottleneck.
Pass below 12 cm with the disturbance active. Move the target with arrow keys or drag it in the workcell; capacity, alignment, and delay remain independent levers.
MASTERY0/3
Complete the challenge progression to unlock transfer.