Every organization decomposes problems, assigns specialists, and invests more where it matters — but every formalization of this process has been domain-specific and imprecise at the boundary. This paper introduces Fractal Intelligence, an architecture that makes the interaction rule fully general by recursively decomposing cognitive work through a uniform, five-surface contract called the Solver.
A Solver may be a deterministic function, a fine-tuned model, a human expert, or an entire tree of sub-Solvers; the caller interacts with each identically. By isolating distinct cognitive modes across hard semantic boundaries, each Solver operates at its productive extreme. Uncompromised perspectives collide at typed acceptance gates, where rejection forces structural reframing rather than token-level compromise.
Paper: fractal-intelligence.pdf
- Cognitive Decomposition — Separating competing cognitive modes (generation, verification, critique, empathy) behind typed interfaces so they don't suppress each other in a shared context
- The Solver Contract — A five-surface interface (Manifest, Execute, Question, Verify, Feedback) that enables composition between heterogeneous, mutually untrusting modules
- Thinking Protocols — The missing second axis: not just what gets done (task protocols) but how each step thinks (cognitive mode specification)
- The Theory of Depth — A marginal-value rule governing when further recursive decomposition is worth its cost
Monolithic models compress distinct cognitive acts into a single blended context, producing reliable adequacy rather than productive extremes. Multi-agent frameworks split what agents do (task protocols) but not how they think. Fractal Intelligence provides a protocol layer where contrasting modes of thought produce better results when separated and then composed than when averaged within a single context.
@misc{westerberg2026fractal,
title = {Fractal Intelligence: Conceptual Decomposition as Problem-Solving Infrastructure},
author = {Westerberg, Henrik},
year = {2026},
month = apr,
publisher = {Zenodo},
doi = {10.5281/zenodo.19462646},
url = {https://doi.org/10.5281/zenodo.19462646}
}See CITATION.cff for the machine-readable version (GitHub
renders a "Cite this repository" button from it).
MIT License