VERIFIED NEGATIVE: Cambricon has published no MLU590 product page, datasheet or specification table - site search and targeted query both return nothing

VERIFIED NEGATIVE, 2026-10-05. data/accelerators/cambricon-mlu590 carries fp32 80 and int8 512 at basis 'secondary-aggregator' sourced to flopper.io, with status 'contested' and confidence 0.3, and its notes state 'Cambricon has published no MLU590 datasheet or product page.' Tested on this pass; it holds. METHOD 1 - site-restricted search. 'site:cambricon.com MLU590' returns only the site root and MLU370-family pages (catid 360 '思元370系列', catid 406 'MLU370-X8智能加速卡', catid 371 'Mlu370-x4智能加速卡'). No MLU590 page exists at any catid. METHOD 2 - the pattern that proves it. Cambricon's own product URLs are parameterised by category and the 370 line occupies catids 360/365/371/406. A site that documents its inference line in this much detail (four separate MLU370 pages) and has NO page for its newest training part is not hiding one behind a login - it has not written one. Supporting this, the only MLU590 material that surfaced anywhere is secondary: a card compilation, Baidu Baike, a CSDN document library entry, a set.baidu.com copy of a '寒武纪思元590技术白皮' (white paper) hosted on a third-party document site, and an industry news item. A 'white paper' reachable only via a third-party document-sharing host is not a Cambricon publication and must not be cited as one. METHOD 3 - consistency with the sibling record. The repo's MLU370 and MLU370-X8 records cite real cambricon.com specification pages with real compute rows. The MLU590 does not, because the equivalent page does not exist. That the same vendor's other parts ARE documented on-site is what makes this a finding about the MLU590 rather than about a crawler that cannot reach cambricon.com. So fp32 80 / int8 512 stay third-party with vendor_claim false, confidence stays low, and the empty case is NOT applicable here - unlike MetaX, Moore Threads S4000, Atlas 300I Duo and Iluvatar, this vendor's silence is genuine rather than a reading failure. That distinction is the point of running three methods: for Moore Threads the silence was in one URL and the vendor does publish; here it is in the vendor's entire web presence.