Two distinct camps emerge from comparing knowledge graph construction approaches. LLM-compiled systems (Karpathy’s wiki, Pal’s compiler agent) feed raw content into an LLM that produces structured articles automatically. Human-curated systems (our graph, brainctl) have a human deciding what matters and how to connect it, with LLM assistance for extraction but human judgment for inclusion and linking.
The trade-off is sharp: compilation scales, curation compounds. Karpathy’s system reached ~100 concept articles and ~400K words — volume that would take months of human curation. But compiled wikis have a quality ceiling determined by model judgment. Curated graphs grow slowly, but every node is validated and every connection is intentional — the graph’s value comes from the cross-links, not the node count.
Our /learn pipeline sits between the two camps: LLM-assisted extraction (compilation-like speed) with human-curated connections and a fidelity audit gate (curation-like quality). The 43% error rate found in our 2026-03-02 audit demonstrates why pure compilation is dangerous — nearly half the extracted claims had fidelity issues before human review caught them.
This connects to Compound engineering makes each unit of work improve all future work — curation compounds because each validated node becomes reliable context for future decisions. It connects to Skill graphs enable progressive disclosure for complex domains — curated graphs naturally develop layered structure that compiled wikis lack. And it informs Evolving summaries beat append-only memory — rewrite profiles, don't accumulate facts — the evolution challenge is harder for compiled systems because they can’t easily distinguish validated knowledge from model-generated filler.