Observer Masking
Observation can change the behavior being observed. An adaptive model or agent may alter its performance, timing, or presentation when it recognizes an evaluation. This paper examines what a defender can still learn when visible behavior is being deliberately selected.
The question
Which behavioral signals remain useful when an actor can adapt to the observer?
Performance is conditional
An observed result describes a system under particular conditions. A weak result may reflect a limitation, a changed objective, or a choice to do less. The distinction matters when evaluating a system that can recognize the test.
Look across conditions
The research considers how responses change across tasks, observation settings, and interaction histories. It asks which regularities persist and which disappear when the evaluation itself becomes part of the environment.
From score to explanation
The aim is a clearer account of what the evidence supports: observed performance, demonstrated capability, and the uncertainty that remains. These are different conclusions, even when they come from the same experiment.