Instead of asking how a human would solve a problem, this paper asks what the problem is made of — and decomposes it into concepts rather than tasks: the dimensions of the thing itself, which recur across domains where steps do not. Each concept becomes a node behind a uniform five-surface contract (the Solver), so a whole collective presents to its parent as a single node — specialists all the way down — and any reasoner (a model, a human, a further subtree) can stand behind one.
The same boundary that composes also isolates: a generator runs at its extreme without softening for the critic that judges it, and rejection forces a re-carve rather than a retry. Concepts that recur become permanent, content-addressed infrastructure (reasoning highways), so reasoning accumulates across problems instead of being rebuilt each time.
The paper argues and probes this rather than proves it: companion experiments lend the isolation and reframing claims preliminary support, and a prototype shows the contract composes. The decisive test is whether decomposed solving outperforms the conventional approach at matched compute; the deeper question is whether independent attempts converge on the same concepts. Should both hold, the result is not a faster pipeline but a shared, persistent substrate: an internet of reasoning.
Paper: fractal-intelligence.pdf
- Concepts over tasks — task decompositions dissolve after execution; conceptual decompositions (what a problem is made of) persist and transfer across domains
- The Solver Contract — a five-surface interface (Manifest, Execute, Consult, Verify, Feedback) composing heterogeneous, mutually untrusting reasoners
- Cognitive isolation — competing modes (generation, verification, critique, empathy) behind typed boundaries so they cannot suppress each other; rejection forces structural reframing, not token-level compromise
- The Theory of Depth — a marginal-value rule deciding, at every node, whether decomposing deeper is worth its cost
- Reasoning highways — recurring concepts crystallize as content-addressed (Sema) infrastructure, reused across unrelated problems
- Two decisive tests — matched-compute performance against the conventional approach, and convergence of independent decompositions on the same concepts
fractal-intelligence.tex/fractal-intelligence.pdf— the paperprototype/— the simulation reported in the paper: planning engine, graph layer (ust.db, 456 nodes / 582 edges), analysis scripts, and an interactive 3D visualization (index.html)references.bib— bibliography
@misc{westerberg2026fractal,
title = {Fractal Intelligence: Conceptual Decomposition as Problem-Solving Infrastructure},
author = {Westerberg, Henrik},
year = {2026},
month = apr,
publisher = {Zenodo},
doi = {10.5281/zenodo.19462645},
url = {https://doi.org/10.5281/zenodo.19462645}
}See CITATION.cff for the machine-readable version (GitHub
renders a "Cite this repository" button from it).
MIT License