Llm Catalog

A sourced, versioned knowledge base of LLM-ops knowledge — NOT a control plane for running models, but the structured, CEL-queryable record that running extensions read to decide which model to run where, with what settings, at what cost. Six subject kinds (model, runtime, provider, hardware, technique, endpoint), all the same uniform entry shape (open kind + open facets, like @stateless/inventory); a model's run-options embed on the model in runsOn[] keyed by endpoint, and capacity/plan answer "what to run where" over them. The twist: it records external, decaying knowledge, so every volatile assertion carries a provenance envelope (asOf, source, versionPins, supersededBy) — "best vLLM version" is derived as the latest un-superseded claim, never a bare rotting field. Declarative: apply materialises one entry resource per id (re-run = versioned trend); prune soft-retires undeclared entries.