A ground-truth computational study testing whether information
present in an internal monitoring channel can be distinguished
from information that actually influences behaviour.
E0-A uses exact enumeration, four quantitative acceptance criteria,
eight adversarial controls, and 500 retained training initializations.
The frozen experimental record and reproducibility archive are
deposited on Zenodo.
Key result
An informative-but-unused control carried the same hidden-state
information as a successful system while producing zero
behavioural benefit, separating information availability from
causal use operationally.
E0-A
Completed
E1
Proposed Master’s study
Broader Sīdus
Outside thesis scope