Developmental bias is strong. Developmental fate is negotiable.
Inside the same open or closed information state, one controller asked for a compact, solid, round body while another asked for an expansive, skeletal, elongated body. Both states split cleanly toward the requested target.
The controller could pull against the state’s preferred body.
Expansive controller minus compact controller
Every family was positive in every round, in both developmental states. By round four the target split was 1.13 SD in the open state and 1.40 SD in the closed state.
What actually happened
The controller did not choose one magic action and replay it. Every sixteen steps it looked at eight candidate interventions, kept the one that moved the current organism farthest toward the requested body direction, and continued from that resulting state.
The action histograms are broad. There is no single lucky seed carrying the result. Feedback kept reading the organism and changing its answer.
The same developmental basin contained more than one reachable body trajectory.
The driven body persisted, then relaxed.
Body-direction separation after feedback stopped
The mean separation stays positive across the release window in both states. At the final frame, the intervals include zero: open +0.229 [-0.137, +0.604]; closed -0.043 [-0.567, +0.487].
This is neither an instantaneous twitch nor a permanent rewrite. It is finite developmental memory.
The alternate body direction survives the removal of feedback.
Across the full sampled release, open and closed organisms remain separated by which body the controller requested.
By release step 128, the population-level split has largely converged.
Same hidden state. Different requested body.
Source 42 is a clear positive example from the 39-organism panel. Each row starts from the same state; only the feedback target differs. The population result above comes from every organism and every family.
open state
closed state
The information state is a bias over futures, not a single destination.
Our previous swing showed that the same later action becomes a different body in the open and closed states. This swing adds the missing half: feedback can still steer either state toward an alternate morphology. Early information geometry shapes the developmental landscape, while later interaction chooses a path through it.
Can steering the body push the information state itself?
We now need to close the loop in the other direction: drive morphology hard enough, then ask whether the hidden information geometry crosses with it. If body control can flip the developmental mode—and the new mode then changes later action meaning—we have a genuine state transition rather than a transient deformation.