Kuis Operasi AI

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. What should a model service do when a required dependency is unavailable?
   • Return an untested substitute as if it were verified
   • Hide the outage from operators
   • Continue every consequential action without the dependency
   • Enter an explicit, truthful degraded or unavailable state
2. Which artifacts should be versioned as part of a model release?
   • Only the public API name
   • Model, transformations, prompts, retrieval indexes, dependencies, and configuration
   • Only the model weights
   • Only the server host
3. A training job completes successfully. What does that establish for deployment?
   • Only that training completed; the checkpoint still needs release checks
   • That quality and compatibility are acceptable
   • That automatic promotion is safe
   • That rollback is unnecessary
4. An endpoint returns HTTP 200 but serves old retrieval results. What is the operational lesson?
   • The request must be correct because it returned 200
   • Stale results are not a service issue
   • Transport health does not establish current or correct task output
   • HTTP status replaces quality monitoring
5. Why keep a tested route back to the last compatible version?
   • To eliminate the need for evaluation
   • To recover when a release causes regression or incompatibility
   • To make every model release permanent
   • To hide operational incidents
6. Should every newly trained checkpoint be promoted automatically?
   • Yes, if the training loss decreased
   • Yes, because automation removes risk
   • No checkpoint may ever be deployed
   • No; quality and compatibility checks should precede promotion
7. Who should respond to an operational alert?
   • Any user who happens to see it
   • The training dataset itself
   • An assigned owner who can investigate impact and verify recovery
   • No one if uptime remains above target
8. Which set gives a fuller view of operational success?
   • Availability, data correctness, task outcomes, and resource use
   • Uptime alone
   • Training accuracy alone
   • Number of deployment files