School of Specialists (SoS): A Network-First Architecture for Right-Sized, Model-Agnostic Local AI Assistants

Large language model applications typically scale by making a single model larger and asking it to hold everything—knowledge, memory, personality, tool use, reasoning—within one context window. I describe an alternative pattern, which I call School of Specialists (SoS): a persistent, deterministic infrastructure layer (memory, routing, logging) is built first, and language models are treated as small, stateless, individually right-sized, freely swappable tenants of that infrastructure. I situate SoS within the Society of Mind tradition (Minsky, 1986) and the contemporary multi-agent LLM literature, and contribute a specific name, an applied architecture for single-user local assistants, a working reference implementation (Nanana, built on consumer hardware with open-weight models), and an ablation-testing-plus-full-pipeline-tracing methodology for debugging such systems. My central finding is a documented case in which a failure initially suspected to be unreliable reasoning in a 3-billion-parameter conversational specialist was, upon deeper instrumented investigation, fully explained by a limitation in one small, independently swappable retrieval component—with every brain behaving correctly given its actual input. I report this, including the methodological misstep that preceded the correct diagnosis, as evidence that composing narrowly scoped, swappable specialists around observable infrastructure produces failures that are legible and attributable, rather than opaque. This is a preliminary, architecture- and methodology-focused paper; latency optimization, systematic classifier benchmarking, and cross-platform validation are named explicitly as future work rather than presented as solved.