Most of the noise came from the stock CPU cooler, not the disks. Swapping it for a low-profile tower cooler and setting a gentler fan curve brought idle noise down to the point where the machine can live in the study. Disk spin-down turned out to be counter-productive: frequent spin-up was louder than a constant low hum.
The rule I settled on: one copy on the working machine, one on a local NAS with daily snapshots, and one offline
drive rotated monthly. Snapshots catch accidental deletes; the offline drive catches everything else.
Checksums are verified quarterly with rsync --checksum --dry-run.
Mostly older papers on reproducible data pipelines and a couple of books on statistical thinking. Writing short summaries right after reading has been more useful than highlighting.