EBF Coding Conventions

Data Contract

Training array shape: (n_points, n_dims+1) — last column is always the output variable
Input slice: data[:, :-1] → shape (n_points, n_dims)
Output slice: data[:, -1] → shape (n_points,)
Data is always standardized (zero-mean, unit-variance) before the model sees it
Scale and Offset must always travel with saved model weights

Tensor Shape Notation

Shape comments follow the pattern # shape (dim1, dim2, ...) using named dimensions:

Symbol	Meaning
`n_points`	number of data or query points
`n_nodes`	number of EBF nodes (hyperparameter)
`n_dims`	number of input dimensions (not counting the output column)

Key shapes through the forward pass:

X (input):               (n_points, n_dims)
Nodes:                   (n_nodes, n_dims)
deltas (x − v):          (n_points, n_nodes, n_dims)
L (upper-triangular):    (n_nodes, n_dims, n_dims)
A = LLᵀ + εI:           (n_nodes, n_dims, n_dims)
r² (distances):          (n_points, n_nodes)
a1, a2, a3 weights:      (n_nodes,)
b1 (linear weights):     (n_dims,)
b2 (constant bias):      (1,)
Y (output):              (n_points,)

Variable Naming

Name	Type	Description
`n_nodes`	`int`	number of EBF nodes
`n_dims`	`int`	number of input dimensions
`n_points`	`int`	number of data points
`data`	`ndarray (n_points, n_dims+1)`	raw input+output array
`Scale`	`ndarray (n_dims+1,)`	`1/std` per dimension for standardization
`Offset`	`ndarray (n_dims+1,)`	`mean` per dimension for standardization
`var_weight`	`float`	loss weight for node spread regularization
`eps`	`float`	numerical stability epsilon, default `1e-8`
`r2`	tensor `(n_points, n_nodes)`	squared non-Euclidean distance
`deltas`	tensor `(n_points, n_nodes, n_dims)`	point-to-node difference vectors
`dist_nodes`	tensor `(n_nodes, n_nodes)`	node-to-node pairwise distances (for regularization)
`a1`, `a2`, `a3`	tensor `(n_nodes,)`	basis function amplitude weights
`b1`	tensor `(n_dims,)`	linear trend weights
`b2`	tensor `(1,)`	constant bias
`file`	`str`	checkpoint path returned by `run()`, consumed by `run_points()`

Function and Class Naming

Existing internal math functions — keep CamelCase through Phase 2; rename to snake_case in Phase 4 if desired: - DeltaAll, NonEuclidDistance, ActFunc, LinearBias, EBF_Graph

Phase 2+ public API — snake_case: - fit, predict, get_nodes

Phase 2+ module-level functions — snake_case: - run, run_points, plot_correlation

Basis function registry keys — snake_case strings: - 'multiquadric', 'thin_plate', 'matern52', etc.

File and Module Naming

Context	Convention	Example
Package modules	`snake_case.py`	`model.py`, `basis_functions.py`
Example scripts	`snake_case.py`	`compressor_map.py`
Test files	`test_<module>.py`	`test_model.py`
Documentation	`UPPERCASE.md` for project-level, `lowercase.md` for topic docs	`CLAUDE.md`, `ROADMAP.md`

Code Style

4-space indentation, no tabs
Tensor shape documented in comment on the line that produces it: # shape (n_points, n_nodes)
Commented-out alternative code blocks use # name label at the end of the line (existing pattern — preserve it)
No hardcoded absolute paths anywhere; use relative paths or pathlib
if __name__ == "__main__": guard required on all script files
No bare except: clauses