Dialectical Graph

A dialectical graph is a knowledge graph that stores argumentative structure rather than treating documents as undifferentiated text. The core insight is simple: do not mistake text for knowledge. Text is packaging; knowledge is the constraint structure that lets claims support, attack, and refine one another. For background on how this approach evolves beyond document-centric systems, see From Memex to Dialectical Graph.

Coscientist enforces a triple separation: original text is preserved as quotable evidence spans; claims are decomposed into normalized propositions; and relations are stored as explicit argumentative bindings. This is a structural fix for a common RAG failure mode: excerpt truncation. The evidence span keeps the exact wording and source, the claim records who asserted what, and a cites edge makes attribution explicit instead of implied by prose.

The graph also separates an inference layer from a narrative layer. A document is not final truth; it is a snapshot of the graph. Every sentence should be able to backtrack through claim, scope, assumption, method, data, evidence span, and source. Without that chain, a sentence is just a sentence, not knowledge.

To act as a brake on Encyclopedia Meltdown, the system prioritizes contradictions and counterexamples rather than consensus (see Graph as Meltdown Defense). When a claim enters the graph, the default posture is adversarial: seek attacks and undercuts early so the structure updates by rebuttal instead of drifting by repetition. For concrete node and relation types, see Dialectical Graph Nodes and Dialectical Graph Edges, as well as Graph Components.