21.3 The actionable-model pattern

A LINCS declaration should make the learner actionable, not merely descriptive. For every proposed repair it should be possible to answer four questions:

1. Is the target a parameter, component, architecture, or declaration?

2. Which objects, arrows, observers, and semantic commitments are held fixed?

3. How does the local edit propagate to the behavior being evaluated?

4. Which evidence, not used to generate the proposal, can admit or reject the result?

This is the most general lesson LINCS takes from causal inference. Causal models answer these questions for domain mechanisms. LINCS asks them of the learning system itself. When the modeled arrows also carry causal semantics, the two levels can be connected; otherwise the result remains a structural intervention and should be described that way.