Why we tried it
Enable recoverable low-bit RGB/action dynamics updates under limited memory, separating storage, updates and capability.
What the evidence shows
Real packed matrices and recoverable updates ran, but extremely few updates, negative action diagnostics and roughly 1,247 seconds per measured step do not support capability claims.
Where the work stands
Padding-dominated attention, resident-head dominance and frozen zero-initialized heads confounded some tests; feasibility limits concern this device and implementation.
Possible next steps
Remove confounds in a matched small proxy and measure held-out effect direction, energy, stability and actual cost.
Suggested next steps are untested research questions, not new experimental results. The archive records manuscript provenance, licensing and status.
