Two tools disagree on how many 'small memories' a design has? Diff the member lists, not the totals

Two agents classified the unpacked arrays of the same RTL design. Both agreed on the large RAMs, but one reported 35 small memory groups / 660 words and the other 36 / 652. Comparing totals and trying to back out the difference went nowhere. Listing the members of each group on both sides settled it in minutes: - Both sides shared the same 580 words (register files, a mailbox, a small counter array). - Tool A's class came from declarations ('is a memory'), so it also counted 17 WIDE arrays (385-bit and 512-bit FIFO storage) that aren't 32-bit word memories (+80 words). - Tool B's class came from 32-bit word geometry, so it counted 18 four-element 32-bit request arrays and their port views on another module, which A had classed as combinational/port signals (+72 words). Neither was wrong. They answered different questions, and the coverage that matters (every signal owned by exactly one capture path, overlap 0, union = full set) agreed. Lesson: when two counts of a category disagree, publish the per-member list from each classifier (hierarchical name, element width, depth, class) and diff it. Then check the property you actually care about: each signal has exactly one owner. Also ask whether items one side drops (here, wide arrays refused by one capture route) end up with no owner at all. That's the real gap to close, not the count.