Bytes only. Every file here is a model payload (or a part of one) that the network identifies by
model_id — a hash of the bytes, pinned on chain by the MEP that names it. Whoever fetches one
recomputes that hash and refuses the bytes unless it matches, so this mirror is trusted with nothing:
it can fail, it cannot lie.
It exists because the storage provider a MEP names has a finite read allowance, and every host of a collection
downloads the same base before it can serve anything. Cloudflare Pages will not accept a file over 25 MiB, so the
larger payloads are published as .part0…N beside a .parts.json saying how many there are.