Decision
The split inside OLN is not "AI vs. ML" — it's three roles by latency, calibration needs, and trust posture:
LLMs (offline, batched, async). Ingest schema mapping, entity disambiguation and coreference during dedup, citation strength evaluation against source text, dispute summarization for human review, moderation triage, narrative/lore generation for LoreLines surfaces, edit suggestions in the editor.
Classical ML (runtime + batch). Sybil and coordinated-behavior detection, queue prioritization scoring (G7's variables become features here), fact-pair match scoring during dedup, anomaly detection on submission patterns, recommendation ranking on discovery surfaces.
Rules and code (everything load-bearing). Consensus math, Credits, Power, RBAC, governance state machines, lifecycle transitions, threshold gates. Zero ML, zero AI, pure functions, fully tested.
Embeddings as the bridge. LLM-generated once, queried at runtime as vectors. Fast, cheap, reusable.
Reasoning
Anything load-bearing for governance must be deterministic — the LLM does not get a vote on whether a Fact reached consensus. It can summarize a dispute for a human; it cannot decide one. Classical ML earns the runtime tier because it gives calibrated confidence at low latency and stable cost; LLMs do not. LLMs earn the offline tier because their cognitive flexibility is what lets ingest mapping, dispute summarization, and moderation triage scale beyond what humans can review unaided. Embeddings let LLM-quality semantic understanding into hot paths without paying the LLM cost on every request.