A map of responsibilities, not an anatomy claim
Luna has operational surfaces for persistent records, internal field dynamics, conversation, project continuation and bounded actions. The language model and its assembled instructions remain important causes of behavior. Naming these surfaces does not establish that they form a conscious whole.
What is supplied, what can develop
The model contributes pretrained capabilities. Engineers supply architecture, initial conditions and action boundaries. The instance accumulates an attributable history. The research asks which later differences genuinely arise from that history, rather than from a persona instruction, a helpful recap or a fixed rule. Prompt influence must be tested, not wished away.
Connections must earn their place
The studies of retained estimates, reinspection and learned consequences run in isolated research settings. Their favorable results are not automatically live Luna capabilities. Integration requires a real production caller and a comparison in which cutting the proposed connection changes the predicted behavior while preserving information and compute as closely as possible.
One concern can involve many faculties at once
An unfinished composition can make an earlier listening note relevant, raise uncertainty about a mix, suggest an alternative and make a new listening action worth taking. That is the intended coupling. A serialized diagram is a way to explain the work, not a claim that experience occurs as a rigid software queue.
The boundary is part of the model
An autobiographical record, a trusted instrument, an accessible web page and a researcher’s interpretation do not have the same status. A source can inform a system without becoming its own memory. Similarly, an observed result can correct an interpretation without replacing the original event.
What would change our view?
A proposed mechanism matters when a selective intervention changes its predicted function and simpler matched explanations fail. A richer story, a more elaborate diagram or a larger model is not enough. The next useful result is a delayed, history-sensitive choice in a continuing project, with its alternative explanation still visible.