Fractal Semantic Architecture: Scale-Parameterized Relational Training Across Semantic Granularities (v2.2)

Fractal Semantic Architecture is a version 2.2 AI-architecture white paper by Nobel Glas, Talos Morrow, and Johannes Sigil. It proposes adding explicit multi-scale relational learning to conventional token-generating systems.

The paper begins from a perceived token bottleneck. Documents have nested semantic organization, but next-token training does not make relations among paragraphs, sections, chapters, documents, and revisions a direct objective. FSA therefore separates two components: a token-level generator and a relational model whose smallest unit is the sentence.

The multi-scale semantic graph contains nodes at six granularities: sentence, paragraph, section, chapter, document, and version sequence. Horizontal edges classify relations such as sequence, causation, elaboration, contrast, transformation, and reference. Vertical edges preserve containment across adjacent scales. Cross-scale constraints require local and global relation predictions to remain compatible.

A distinctive contribution is version-differential training. Draft-to-revision transitions become first-class training objects. Edits are represented as directional transformations and weighted by changes in a coherence metric, allowing the model to learn not only what changed but which revisions appear to improve structural consistency.

At inference time, the relational system is proposed as a critic and planner. It can generate a document skeleton, constrain local generation, detect contradiction, rerank candidate continuations, and score revisions. The white paper includes a bootstrapping strategy using discourse heuristics, silver labels from frozen models, self-training, and existing discourse corpora.

FSA’s collapse-resistance argument is deliberately limited. Typed relations may resist some averaging that erodes rare token patterns during recursive synthetic retraining, but the generator still operates through continuous token distributions. The paper therefore presents improved collapse resistance as an experimental hypothesis requiring ablations and benchmarks, not as a proved theorem or completed model result.