Core concepts
NeoSyntropy Framework
A deterministic control layer for AI workflows. Models propose what should happen next; a finite-state graph defines what is allowed to happen.Core primitives
Start with the model-backed nodes. That is the point of the framework: you drop tiny, scoped AI models into ordinary code paths (SchemaNode / ReasoningNode), while routers and ControlManager keep every proposal inside a fail-closed graph.
SchemaNode
Provider-backed extraction: a small model returns constrained JSON for output_schema, no tools.
Model-backedReasoningNode
Provider-backed reasoning: a small model may call allow-listed tools and write plain-text notes.
AuthoringCombineNode
Authoring unit that expands to reasoning → schema FSM states.
CapabilityNode
Executable capability (Python handler or provider-backed), never a workflow position.
RouterDeterministicRouter
First matching (predicate, target) rule wins; compiles to deterministic edges.
RouterSemanticRouter
Model picks among labeled targets (routes={label: node_or_group}); still validated against the graph.
GraphEdge
One permitted movement: deterministic, semantic, or fallback.
GraphGroup
Named node collection; optional entry, internal routers, and add_edge that compile into the FSM.
RuntimeControlManager
The whole cycle: deterministic → semantic router → fallback → validate → execute → commit.
Grounded in neosyntropy-framework/README.md core concepts.