dsh-linalg

Large language models famously stumble on matrix multiplication, determinants, inverses, and linear systems. A flipped sign in one elimination step, a conflated row swap, or a singular matrix forced into a bogus inverse—these arithmetic slips show up in virtually every math-heavy prompt. dsh-linalg sidesteps the problem by handing the computation to deterministic local code: Gaussian elimination with partial pivoting, results correct to ten decimal places, with zero network calls, zero subprocesses, zero evals, and zero runtime dependencies.

Four tools cover the essentials. matrix_multiply performs dimension-checked A·B. matrix_compute handles one operation per call—transpose, determinant, inverse, trace, or RREF. solve_linear cracks Ax = b and classifies the answer as unique, infinite (reported with particular solution, nullspace basis, and free-variable count), or no solution. vector_ops takes care of dot product, 3-D cross product, norm, projection, and angle in degrees. Matrices are plain JSON arrays of arrays, capped at 20×20 by default, configurable on demand.

The numerics are deliberate. Values below 5e-12 snap to exact zero so negative zero never leaks out. Pivot tolerance sits at 1e-12, so numerically singular matrices are reported as singular instead of producing garbage, and singular-inverse or inconsistent-system requests both fail cleanly with a clear message. Whether you are verifying homework, double-checking an engineering matrix, or simply need a linear-algebra backstop that always returns the right number, dsh-linalg is the lightest option—install once, compute forever, with no service to spin up and no dependency tree to untangle.