2026/07/10 by Hsiu-Chi Tsai
#cs.DC
Modern ML serving increasingly lets learned, unverified components (routers, latency-SLO admitters, admit ladders) decide a tenant's quality of service; when one is wrong, the assured SLO can silently break, and the Kubernetes layers beneath (Kueue, DRA, the Gateway-API Inference Extension, GAIE) add cross-layer surprises. Rather than trust the learner to be right, we bound the damage a wrong one can do: a small trusted guard wraps the untrusted learner (learned proposes, the guard disposes). A tenant's assured-SLO obligation splits into two parts with different epistemics. Its safety projection, a per-class, per-window assured floor (with an optional drop rule, doom-sound only under an assumed service lower envelope), is a controllable obligation a guard enforces at runtime, holding it regardless of a learned admitter that is arbitrarily wrong within a bounded proposal interface. The admission floor is enforced structurally; given the stated assumptions, the service floor follows as a conditional response-time implication. Its aggregate obligation (the population tail-latency percentile) has no per-request enforcement point, so we treat it as a statistical residual and screen it. On real 2xV100 the guard (a Simplex-style assured-floor gate plus assured-first priority) holds assured-class miss 0.0 across every tested miscalibration of a learned admitter that, unguarded, misses 0.86-0.94; against a live deployment of the GAIE Flow Control, an injected mapping fault (emulating an untrusted mapper) flips the same assured requests from miss 0.0 to 1.0 (a mechanism-level trust-boundary test, not a head-to-head), while our guard reserves by the true class. As a Frontiers submission we evaluate the stance on commodity 2xV100 and a serving simulator, scoping datacenter scale, real-model Flow Control, and a closed worst-case theorem as the agenda.