Use terms that can carry a decision.
Each definition names what to measure and the assumption the term does not support.
Useful task
A bounded unit of work whose output passes a stated acceptance rule and can be connected to an operating or customer outcome.
CONCEPTInference unit
The provider or infrastructure consumption used to produce model output, such as tokens, requests, compute time, accelerator time, or provisioned capacity.
CONCEPTTime to first token
The elapsed time between a request and the first streamed model output received by the application or user.
CONCEPTShared responsibility
The division of security, operations, governance, and data-handling duties between a service provider and its customer.
CONCEPTEvaluation set
A versioned collection of representative, authorized cases used with a stable rubric to test a workload.
CONCEPTDegraded mode
A deliberately limited product behavior used when an AI dependency or model output cannot meet the normal service condition.
CONCEPTExit trigger
A pre-agreed cost, capability, reliability, contract, or risk threshold that opens an infrastructure review.