(2020-04-24, 17:05)Orion Pobursky Wrote: These are a matter of correcting either the inventory or the part number at Rebrickable. I've probably submitted 30 change requests in the last couple of weeks to correct the LDraw part number associated with a part or to correct an inventory. The widespread availability of digital instructions makes this easy.
That, or by using a different database—you mentioned pulling some data from BL instead, for example. But BL has its own quirks and inaccuracies (like using "undetermined" part types, which don't correlate to a real-world element), so yeah, the accuracy of the export can't exceed that of the underlying inventory database without applying some further intelligence, whether human or artificial.
Incidentally, instructions can be misleading in some cases, and so can even the "official" renderings that go onto box artwork, etc. I've noticed cases where these sources will depict a yet-to-be-released mold variation of a part, whereas actual sets still ship with the older version. And in many cases, different variants will appear in different sets, sometimes even commingled, so there's no single "correct" inventory—which BL tries to account for with their Alternate Items section.
Also, mold variations often affect the undersides of parts, which are usually not visible in instructions or box art. The most conclusive sources I've found are pictorial or video reviews of sets (but you have to be sure the reviewer has an actual factory set and didn't just part it out from RB or BL), combined with the handy timeline graphs on RB's part detail pages.
Anyway, sorry for the long tangent.
Per the topic at hand, my thought was that if we did have our own cross-reference DB, could it have certain checks built in, like redirecting alias or obsolete parts, just to avoid having to rely on RB to respond to change requests? I think the value here is in automating the export itself, not concerning ourselves with the accuracies of inventories.