A knowledge graph has two orthogonal work directions.
Decomposition refines the atoms. Each node sharpens, splits, gains a tighter unit of compression. The atoms get smaller and the boundaries between them get cleaner.
Composition overlays the atoms. Pages, indexes, trails, synthesis views that group the atoms for a reader. The atoms stay where they are; new structure rides on top.
The two axes are independent. The same agent can work both. Codex's recent experiment ran a chunking sweep generating hundreds of prototype docs to find the right shape of an artifact, which is decomposition. The same experiment built four altitudes above those raw docs (synthesis, handbook, teaching essay), which is composition. Different work-products travel in different directions, sometimes within the same project. The point is the axis, not the agent.
Most of the existing corpus is already composition done at write-time.
A node that synthesizes ten priors is the compression of those priors into one artifact. That is what a wiki page should be at its honest level. The work of finding the words that survive across N priors is the compression. The artifact is the trace of someone doing that work.
Library-synthesis is not a new direction. It is the direction. Every essay-class node is a composition over priors. The compression is the point, and the work of compressing is the trace that proves it.
A natural proposal is to take this further: a wiki layer on top of the node graph, with pages generated at read-time by an LLM that synthesizes from N atoms below.
That is one possible composition artifact. It is also the dangerous one.
A wiki page synthesized at read-time produces prose that no human wrote. That is the slop pattern. The wiki reads fine. The atoms are still cited. The prose is hallucinated structure.
The slop is not that the wiki is wrong. The slop is that the wiki is generated, not compressed. Generation costs nothing. Compression costs the work of finding the words that survive across N priors. The cost is the signal.
The honest version of the composition direction does not generate prose. It generates navigation.
A topic index of every node touching prediction-error reduction is composition. A faceted browse over frontmatter tags is composition. A graph viewer is composition. So is an llms.txt file. Composition is the shape of the read-affordance layer, not the source of new claims.
The middle cases sit on a gradient: synthesized snippets with citations, summaries that link rather than paraphrase, structured comparisons of two atoms. The rule is not "no LLM ever touches the read layer." The rule is that any prose appearing in the wiki has to be prose someone bothered to compress.
It predicts where work in the decomposition direction and work in the composition direction do not collide.
A project can keep atomizing through chunking sweeps, paradigms-as-discrete-units, generator scripts that produce hundreds of variants. The atoms get better. Another project can keep composing at write-time, into essay-class nodes. The compositions get better. Neither project occupies the other axis.
The read-affordance layer over both, whatever its final form, is structural rather than generative. Navigation, not paraphrase. The atoms remain canonical because no synthesis pretends to replace them.
A wiki on top is fine if it is the topic index. It is not fine if it is the prose.