The library is a single package: the model graph.
@composable-model-graph/core (no dependencies)
The generic primitives, and only the generic primitives:
Transform,TransformNode,createTransform,toTransformNodeModelGraph,createModelGraph,RunOptions,AnyTransformEvaluator,createEvaluatorFeedbackResolver,createFeedbackResolver,feedbackActionRunContext,createRunContextTraceRecorder,createTraceStep,TraceStepGraphRun,EvaluationResult,Evidence,EvaluationStatus,FeedbackAction,FeedbackActionTypeTransformExecutionError
Domain-specific building blocks — numeric layers, data-shaping helpers,
evaluator/feedback libraries — are not part of the library. They are
trivial to build on top of the primitives (see the examples/) and belong in
the consumers that need them. Keeping the library to just the model graph keeps
the surface small, generic, and harness-free.
Public language is software/AI generic. Use:
Transform · ModelGraph · Node · Evaluator · Evaluation · Feedback · Run · Trace
Avoid in public APIs: TransferFunction, Harness, Agent, Workflow, Backprop, Skill, Goal.