ML Primitives

Aether's ML code lives under crates/aether-core/src/ml. The documentation describes the module inventory and claim boundaries rather than presenting it as a benchmarked replacement for external ML frameworks.

Module Inventory

Module	Active surface
tensor	Owned tensor data, shapes, indexing, map, add, sub, mul, scale, transpose, matmul, reductions
linalg	Loss functions, distances, RBF kernel, numerical gradients
regressor	Linear, polynomial, RBF-style, Gaussian-process-labeled, and geodesic-labeled model enum paths
convergence	Betti records, drift/error windows, residual analysis
benchmark	Escalating benchmark runner over internal test functions
clustering	KMeans, DBSCAN, agglomerative clustering, auto-k helper
classification	Logistic regression, KNN, perceptron, Gaussian naive Bayes, decision stump, AdaBoost, nearest centroid
neural	Dense layers, activations, optimizer config, MLP training loop
autograd	Tape and variable scaffolding for differentiable tensor operations
convolution	Conv2D forward path
dataloader	Batch iteration over tensors
gossip	Local centroid and consensus propagation

Language Boundary

The DSL exposes a narrower surface than the Rust crate inventory. Constructors such as Ml.MLP, Ml.KMeans, and Ml.Conv2D are available through native function dispatch. Individual methods should be documented as active only when the interpreter path is implemented and tested.

Claim Boundary

The ML module can be described as internal Rust ML primitives. It should not be documented as:

• faster than PyTorch, TensorFlow, sklearn, GUDHI, or ripser;
• production-ready for all model families;
• equivalent to external framework semantics;
• hardware accelerated.

Those claims require benchmark artifacts and parity tests.