The words on the rest of this site, defined.
Most of these are borrowed from statistics or from NVIDIA, and most vendor sites use them loosely. Here is what we mean, precisely enough that you could argue with us about it.
Difference-in-differences. Comparing how much a treated group changed against how much an untouched control group changed over the same window. It exists to stop you crediting an intervention for a shift that was going to happen anyway — which is exactly what a raw before/after does.
Control cohort. The comparable workloads that did not change during a measurement window. Without one, "spend fell 12%" and "spend fell while we happened to be installed" are the same sentence.
Position bias. A pairwise LLM judge's tendency to prefer whichever answer it reads first, independent of quality. It is large enough to decide a close comparison on its own, which is why every pair is judged twice with the answers swapped.
Always-valid sequential test. A test you can check as often as you like without inflating the false-positive rate. A conventional fixed-sample test quietly loses its guarantee the moment you peek early and stop when the answer looks good — and with live traffic, peeking is the normal way to work.
mSPRT. Mixture sequential probability ratio test — the specific always-valid test we run, built on time-uniform confidence sequences, so a routing policy can be reverted the moment the evidence crosses its bound rather than at the end of a fixed window.
Admission control. A decision taken before a request is sent, so a call can be refused. This is the line between a control plane and a report: an invoice can tell you a runaway agent burned a month of budget overnight, but only something on the wire can stop it.
Shadow mode. Running a policy without applying it, so evidence accumulates on your real traffic before anything about the traffic changes.
Reasoning tokens. Tokens a model spends working before it answers. They are billed separately and usually cost materially more, so a cost report that folds them into output tokens understates the expensive calls specifically.
NVML. NVIDIA Management Library — the driver-level interface that reports per-GPU utilization, memory, power and processes. Reading it is how you learn what a card actually did, as opposed to what the instance was billed at.
MIG. Multi-Instance GPU. NVIDIA's partitioning of one physical GPU into isolated slices with their own memory and compute. It matters for cost because a card at 20% utilization and a card carved into four busy slices look very different on a bill and identical in a naive utilization average.
Showback and chargeback. Showback reports what each team spent. Chargeback moves the cost onto their budget. The second one is the one that changes behaviour, and it is also the one that makes people argue about your numbers.
FOCUS. The FinOps Foundation's open specification for cost and usage data, so bills from different providers can be read in one schema instead of one parser per vendor.