Computational Dialectics

III. Core concepts

7. What is a determination?

In this book, a determination is not an immutable label attached to an object; it is a relational state defended within a particular context.

For example:

"This drug is beneficial."

is a weak determination by itself. A stronger form approaches:

"Within this dose range, for this patient group, over this time horizon, and according to this measure, a beneficial effect has been observed."

A system does not need to convert every sentence into JSON at this level of detail. The point is not the representation format, but which distinctions must not be lost.

8. What is a concept?

We do not use the word "concept" as a synonym for embedding. An embedding is a representation technique. In our research hypothesis, a concept is an organization of state that structures behavior and expectations across situations, and whose relations and boundaries can change over time.

A possible computational trace of a concept may carry:

• related determinations;
• links to other concepts;
• the conditions under which it is valid;
• evidence supporting it;
• evidence that creates problems for it;
• rival reconstructions;
• how it has been revised in the past;
• which observation could discriminate it.

These do not have to appear as explicit fields in the final neural architecture. They are semantic requirements first; the latent realization should be selected experimentally.

9. What is context?

Context is not only the previous token window. From the perspective of Computational Dialectics, context may cover at least three different things:

• linguistic context: what is being discussed and which references are used;
• situational context: time, scope, condition, actor, and measurement frame;
• structural context: the network of relations and processes within which a determination gains meaning.

For that reason, "the same word" does not guarantee the same conceptual position, and a "different word" does not necessarily mean a different concept. The binding problem appears exactly here.

10. How can a relation be represented?

We do not impose one method for representing a social or physical relation. A directed edge, slot, continuous latent relation, graph state, or another mechanism may be used.

What matters is whether the representation can answer these questions behaviorally:

• What or who is in what relation with what or whom?
• Under which conditions is this relation valid?
• When new evidence arrives, did the relation itself change, or only the value produced by the relation?
• Which other relations should be affected?
• Which relations should not be affected?

The scientific value of the D1-B benchmark lies precisely in forcing this distinction: mapping change and relation rebinding are not the same thing.

---