Computational Dialectics

IV. How can dialectics become computational?

11. A working hypothesis for a "dialectical core"

We do not have a completed Dialectical Core today. But we do have a working regime that defines the direction of the research:

1. Extract candidate determinations and relations from the input.
2. Align them with the current state by asking whether they share the same scope, time, and reference frame.
3. Do not immediately label every difference a "contradiction"; separate valid contextual differences first.
4. If a genuine incompatibility remains, use it not as an answer template but as a representation-research/revision signal.
5. Preserve the part of the current state that remains valid.
6. Restrict over-generalized scope; change a wrong binding or relational structure.
7. If the evidence is insufficient, do not collapse too early to one explanation; retain behaviorally distinct rival reconstructions.
8. Determine which observation, intervention, or new evidence would discriminate the rivals.
9. If necessary, perform split, merge, or restructuring at concept/schema level.
10. Actually use the new state in later predictions and actions; do not allow stale state to bypass it.

This list is not fully implemented. The task of the scientific program is to operationalize each item separately, compare it with strong ordinary alternatives, and eliminate what is unnecessary.

12. Why not "thesis–antithesis–synthesis"?

Because this template is both a crude historical summary and a bad computational algorithm. Asking a system to produce two opposing sentences and then write a compromise does not perform real contradiction diagnosis or representation revision.

For us, contradiction begins as a diagnostic problem:

• Is it the same subject?
• The same time?
• The same measurement frame?
• The same scope?
• The same meaning?
• Are both sides actually supported?

Even after these checks, the result is not automatically a "synthesis." Sometimes the correct conclusion is a data error. Sometimes there are two different conditions. Sometimes the system should remain unresolved. Sometimes the existing representation must be decomposed.

12.1 Conflicts that look the same are not the same

Later technical chapters will make this more formal, but the distinction matters already at reader level. When two statements appear to exclude one another, the system may be facing several different situations.

Apparent contradiction. "Ali is at home at 08:00" and "Ali is at work at 10:00" can both be true; the time differs. "The drug is beneficial at a low dose and harmful at a high dose" changes the scope or measure. Opening deep revision here would be a mistake.

Epistemic conflict. If two independent sources claim to measure the same thing within the same scope but disagree, we should first investigate measurement, provenance, or hidden scope rather than declare reality itself "contradictory."

Model anomaly. If the representation built by the model systematically fails to explain a new observation, the problem may be in the representation rather than the data. The question shifts from "which answer should I change?" to "which relation did I construct incorrectly?"

Structural-opposition hypothesis. This is the stronger claim: the same process or structure may produce opposing tendencies through different internal relations. But we cannot enter this class simply because we saw two opposing sentences; we need additional predictive or intervention evidence that discriminates rival mechanisms.

The goal is not to teach philosophical categories as labels. The goal is to build a diagnostic regime that prevents distinct error sources from collapsing under the same "contradiction" label. The value of a Dialectical Core should be measured not by how many contradictions it finds, but by whether it opens the right kind of revision in the right situation.

13. Why is revision central?

One of the most testable aspects of dialectical thinking for AI is the idea of non-destructive revision under changing evidence.

When new evidence arrives, there are three bad extremes:

• change nothing;
• erase the entire old state;
• append the new information and allow a contradictory state to grow.

A better system should try to identify what was actually invalidated and change only that. D1-B currently tests exactly this minimum question at neural level.

13.1 Preserve, restrict, reject, transform

We use the following not as a mandatory mathematical formula, but as a way to decompose revision:

new_state = reconstruct(
    preserve   = what remains valid,
    restrict   = what was over-generalized,
    reject     = what is no longer supported,
    transform  = what needs a different binding/relational structure,
    add        = what is supported by new evidence,
    unresolved = what cannot yet be discriminated
)

This distinction matters because "falsifying" a piece of knowledge does not always mean deleting it. The generalization "X increases Y" may be rebuilt under new evidence as "under C1 it increases Y, under C2 it decreases Y." The old observation is not simply discarded; its meaning and validity range are redefined.

In a stronger case, not the value but the relation binding may change. We may learn that the link we represented as "A determines B" actually operates through A→C→B or attaches to a different parent. This is why D1-B's selective relational-revision experiment is a small but strategic gate.

---