Forty years of hardware. One build sheet.

Server hardware, not just desktop parts.

Processors, motherboards, registered and load-reduced memory, U.2 and E1.S storage, host bus adapters, accelerators, optics from 1GbE to 800G, rails, rack PDUs, uninterruptible supplies and tape libraries. The categories exist because the builds we wanted to plan needed them.

Nothing in the data path guesses.

Merchant listings bind to catalogue parts through a scorer built from stated rules, and every match records the reasons that produced it. That is a correctness decision rather than a stylistic one. A matcher that is 95 percent right about which listing is which puts a wrong price on one part in twenty, and nothing downstream can tell you which one.

Attribute extraction is deterministic for the same reason. If the catalogue says a board takes a particular socket, you can go and read the rule that decided it.

Degraded, never broken.

The API over the live database is primary. When it is unavailable the front end falls back to static shards and the catalogue stays fully browsable, without prices. A deployment that has not been wired up yet loses prices, not the catalogue.