Composable Model Graph is a dual-language (TypeScript + Python) library ecosystem for building inspectable transformation graphs. The two implementations are kept at parity; see structure.md, with philosophy.md and development.md for why it exists and how features earn their place.
Every system built on this ecosystem follows one shape:
input → transform → state/output → evaluation → feedback
- input - whatever enters the graph.
- transform - a named, typed function from input to output.
- state/output - every intermediate output is recorded as a trace step.
- evaluation - the final output is scored into an
EvaluationResult. - feedback - the evaluation is mapped to a
FeedbackAction.
A single function hides its intermediate states. A model graph makes them first-class: each step is recorded with its input, output, and timing. This makes runs inspectable, which is the precondition for evaluation and feedback.
| Package | Role |
|---|---|
core |
The primitive layer. Types, factories, and the graph runner (sequential by default, a general DAG where a use case needs it). |
math |
The neural-network proof. Activations, losses, dense layers. |
evaluators |
Generic ways to turn an output into an EvaluationResult. |
feedback |
Generic ways to turn an EvaluationResult into a FeedbackAction. |
estimation |
Decode the best path through sequential candidate states. |
terminal |
Project executable topology and completed traces into deterministic text. |
constraints |
Observe caller-projected relationships and retain lossless findings. |
The Python side mirrors these packages with parity (see structure.md). Linearity is a default, not a design limit: the runner is sequential by default and a general DAG when a use case needs it.
This repository ships only generic primitives. It is not a harness, agent framework, workflow engine, or ML framework. The end goal is to bridge toward a real harness lifecycle later - but that harness is private and is built on top of these primitives, not inside this repository.
- 01 - Core primitive
- 02 - Neural-network architecture
- 03 - Error, sensitivity, feedback
- 04 - Harness bridge
- 07 - Terminal graph view
- 08 - Constraint findings
See also: philosophy.md, development.md, structure.md.