WINDOWS INTO BEHAVIOUR
Reasoning
Actions
A reasoning monitor inspects what the model writes about its process; this may omit relevant information.
THE SHORT ANSWER
Visibility is a separate question.
- OpenAI reports lower monitorability for Astra than GPT-5.6 Sol alongside improved compliance with restrictions in its evaluations.
- Reading reasoning, observing actions and inspecting both are distinct monitoring approaches. Their results should not be merged into a single safety grade.
Keep the limits in view.
The comparison is developer-reported and depends on the test setting. Some tests explicitly prompt evasion. The diagram is explanatory and does not show measured detection rates.
Keep asking.
Ask about this evidenceCollection reviewed 30 September 2026. Initial research edition. Source inspection does not mean independent reproduction of the findings.