Bugs, caught while you type

A weight matrix passed transposed — the classic silent killer:

logits = np.matmul(batch, weights) + bias
# weights came in as (10, 28) instead of (28, 10)?
# error: matmul inner dimensions disagree: 28 vs 10

A bias that broadcasts today and breaks tomorrow:

np.zeros((2, 4)) + np.zeros((3, 4))  # error: broadcast shapes disagree

A reshape that silently drops a column:

np.zeros((2, 3)).reshape((2, 4))   # error: reshape changes total size: 6 -> 8

Per-channel weights that do not match the axis they average over — a classic silent-wrong-answer in sensor and finance code:

np.average(np.zeros((128, 6)), axis=1, weights=np.zeros(5))
# error: average weights length 5 does not match axis 1 size 6 (shape (128, 6))

A difference along an axis that is not there — and the length math, checked instead of trusted:

np.diff(np.zeros((4, 3)), axis=2)  # error: axis 2 is out of range for 2 dimensions (shape (4, 3))
np.diff(y, n=2, axis=0)             # clean: a (129, 4) input is (127, 4) out

A dot that silently misaligns past rank 2 — no batch broadcast to hide behind, the contracted pair is checked directly:

np.dot(np.zeros((5, 2, 3)), np.zeros((5, 4, 5)))
# error: dot contracted dimensions disagree: 3 vs 4

A tensordot whose axes don't line up:

np.tensordot(np.zeros((2, 3)), np.zeros((4, 5)), axes=1)
# error: tensordot contracted dimensions disagree: 3 vs 4
np.tensordot(np.zeros((2, 3)), np.zeros((3, 2)), axes=([1], [0]))  # clean: (2, 2)

These are not cherry-picked: bench/ seeds 18 realistic mistakes into MLP-training and math-library code, and the checker flags 17 of them before anything runs (the 18th is unprovable on purpose — see limitations). Each error names the axes and the shapes, e.g. broadcast shapes disagree at trailing axis 1: 10 vs 2 (shapes (N, 10, 2) vs (N, 2, 1)).