Run an effort sweep on conveyancing calculations

The sweep is only as good as its answer key, so the gold answers are computed by code that is checked on its own against hand-worked values, and each problem states its convention (rounding, day counting, the leap-day rule) so that exactly one answer is right. The runner holds everything constant except --effort: same system prompt, no tools, same model, the whole grid repeated. The decision that matters most is the unit of the result. Accuracy and mean cost per level are inputs; the output is marginal cost per extra correct answer between adjacent levels, the number you compare with what a wrong answer costs you.