NeoSyntropy research
Small models inside explicit control systems
Our current research studies candidate selection and typed argument extraction inside a graph-controlled runtime.Two focused questions
Router SLM
Which capabilities should run?
A compact router selects one node from a bounded candidate set, including a fallback candidate.
Edge SLM
What arguments should a tool receive?
A separate edge decoder turns context into schema-constrained JSON after a reasoner emits a tool trigger.
Research boundary
The router repository trains single-candidate selection; it does not contain the complete inference runtime described by its architecture specification. The edge repository is a proof of concept for typed tool extraction, not the full NeoSyntropy platform.
States, transitions, validation, and execution live in the broader backend. Axioms and a named control-manager implementation are not part of these two research repositories.
Synthesized from train_slm_as_router/ARCHITECTURE.md and slm-ntk/README.md.