AI capability is increasingly visible in systems that can interpret information, reason about a task and generate a useful response. Physical systems introduce a different question: what does it take to turn that capability into an action in the world?
An answer on a screen can be inspected before someone uses it. A physical action changes the state of a machine or its environment. That difference makes the surrounding engineering central to the intelligence itself.
The model is one part of the system
A model or learned policy may provide a promising route to a task. Its usefulness still depends on the conditions around it: what the sensors can observe, how actions are represented, what the hardware can execute and what must happen when the system encounters an unfamiliar situation.
For a narrowly defined task, the right evaluation may therefore include several distinct questions:
- Can the model interpret the situations relevant to the task?
- Can its output be translated into an operation the machine can execute?
- Are the operating boundaries explicit?
- Can the system recognise a condition that requires it to stop, fall back or request assistance?
- Is the resulting behaviour observable enough to investigate a failure?
These questions concern the complete system. A strong result on one dimension does not answer the others.
An action needs context
Consider a simple sorting task. A proposed movement has meaning only in relation to an object, a workspace, a machine state and the constraints of the operation. The same movement may be appropriate in one state and inappropriate in another.
This suggests an important design direction: connect an action to the conditions that make it valid. Those conditions might include a permitted workspace, a supported object class, a timing assumption or a verified state of the mechanism.
The precise boundaries depend on the task and platform. They cannot be supplied by a generic label such as “safe” or “autonomous”. They need to be defined and evaluated as engineering properties.
Adaptation should preserve what matters
Deploying a capable policy on a constrained device may require changes to the model, its inputs, its representation of actions or its execution path. A smaller implementation is useful only if the behaviour required by the task remains useful under the intended conditions.
This is why we are interested in task-specific evaluation alongside adaptation. The question is more precise than whether a model has become smaller. It is whether the adapted capability still performs the defined task, within the relevant operating boundaries, on the intended machine.
A research direction for Conexiq
Our developing programme investigates the connection between capable models, task-specific adaptation and bounded physical execution. We are interested in how these pieces can be represented, implemented and evaluated together.
This note introduces the question guiding that work. It does not report a completed implementation or a validated performance result. As specific work develops, the useful contribution will be in the methods, measured behaviour and limitations that can be examined alongside it.
The direction is straightforward: understand the capability, define the action, and evaluate the system that connects them.