Identifiable information in behaviour
The manuscript treats the distributions of trajectories associated with task labels as a channel. How much task information an observer can read depends on the channel itself. Time reversal separates the information into time-symmetric distribution structure and directional trajectory differences. Fisher–Rao geometry and entropy-production quantities describe the respective parts, avoiding an account based on only one signal.
The distance from a witness to an application
Self-contained witnesses enumerate finite trajectory spaces, with a tokenized-action policy as a worked mathematical example. The work concerns an information decomposition and attributable bounds. VLA supplies an intuitive setting, rather than evidence of a deployed runtime detector or robot validation. Continuous-prior extensions and atlas quantization are theoretical; the witnesses use no real robot dataset.
Current progress and validation scope
VLA is a mathematical application setting. The computational witnesses use no real robot data and validate no deployed detector.
Discuss this research
I welcome conversations about the questions, methods, and ways to test them.
lancer20060105@gmail.com