The Python side of composable-model-graph. Parity with the TypeScript side (../typescript) is the rule: new capabilities ship in both languages (see ../docs/development.md).

A real, dependency-free core (so the primitive is usable without an install), at parity with the TypeScript side.

• composable_model_graph.core - Transform, create_transform, ModelGraph / create_model_graph, RunContext (with record_signal), TraceStep, GraphRun, the evaluation/feedback types (Evaluator, EvaluationResult, Evidence, FeedbackResolver, FeedbackAction, with create_evaluator / create_feedback_resolver), and compare_runs (RunComparison, SignalDelta). The runner is sequential by default and a general DAG when you pass connections (linearity is a default, not a limit), and it invokes an attached evaluator / feedback resolver after the transforms run.
• composable_model_graph.evaluators - threshold_evaluator, numeric_error_evaluator, exact_match_evaluator, composite_evaluator, deep_equal.
• composable_model_graph.feedback - default_feedback_resolver, threshold_feedback_resolver.
• composable_model_graph.math - sigmoid / relu with forward, derivative, derivative_from_output (the neural-proof lane), sensitivity / rank_sensitivity, TransferFunction.
• composable_model_graph.estimation - decode_path / decode_path_fixed_lag over CandidateState trellises (Viterbi / fixed-lag), domain-free.
• composable_model_graph.terminal - lifecycle/run projections plus comparison, decoded-path, sensitivity, and useful-flow terminal renderers.
• composable_model_graph.constraints - deterministic relationship checks, lossless finding reports, generic selector-driven helpers, and an explicit evaluator adapter.

No install needed:

cd python
python3 tests/test_smoke.py        # prints "PASS: ..."

Or with pytest (after pip install -e .[dev]):

python3 -m pytest

At parity with TypeScript on core, evaluators, feedback, comparison, math, estimation, terminal rendering, and constraints. New capabilities continue to land in lockstep across both languages.