Architecture

WeightModel owns offline model loading, live PyTorch tensors, public evaluation, and verified checkpoint saving. execute_program executes a trusted Python file defining edit(model) and records baseline and candidate evaluations in run.json.

wsm.editing owns models and their publication lifecycle. wsm.program owns program execution and receipts. The CLI and terminal call the same execution function. _publication provides atomic exclusive rename without loading any experiment machinery.

Python execution has the invoking process’s privileges. Public evaluation does not provide private grading or automatic selection. See the coding guide.