A branch built with a coding agent passed lint, 1,572 Python tests and 418 GUI tests. A review by a different model, split into a Standards subagent and a Spec subagent, still found two ways to write the same record twice. The Spec reviewer got there by checking what the commit messages promised against the code.
An instrument catalog looked mandatory for a tax engine. Its main attribute only decides whether an event happens, and the ledger already records that. Ask what each attribute changes before you maintain data for it.