In the test, the payload was passed word-for-word to the reading model with only each vendor's built-in synthesis functionality removed, and token counts were derived from actual bills rather than estimates.