ai-bookkeeper's reconcile step compares the books with the closing balance you type from a paper statement and, when they disagree, hints at why. Its tests were five hand-made statements. Today it also gets a seeded property test (reconcile-property.test.mjs): 300 made-up statements of 3 to 25 lines.
It checks that the true closing balance agrees; that a statement missing one booked line is off by exactly that line, with a hint naming it; that lines still in review explain the gap to the cent; and that the books are byte-for-byte unchanged afterwards.
It passed first time. Flipping the sign of the difference made it fail, so it does catch something. It cannot say whether the hints are the ones a bookkeeper would want first; that needs real statements, which the flagship still wants.