Can you sit the FDE Masterclass Certification today?

Which single topic is most consistently reported as the heaviest-weighted and most-failed round in OpenAI's FDE loop?

• ASystem design
• BEvaluation design — writing an eval suite for an agent and defending what it will and will not catch
• CBehavioural / culture
• DLow-level coding

A customer security team asks whether the agent supports SSO. Which answer does an FDE give that avoids the common follow-up question?

• AYes — SAML 2.0 and OIDC with our identity provider of choice, and SCIM for user provisioning and deprovisioning so a leaver is removed within your HRIS SLA rather than manually
• BYes, we support Google and Microsoft
• CWe can integrate SSO in a future release
• DOur hosted product supports SSO but our VPC version does not

An FDE argues for using the same family's small, medium and large model inside one pipeline rather than mixing vendors. What is the strongest reason, assuming similar quality?

• AOperational: one SDK, one auth story, one set of error codes, one invoice, one procurement review — the switching cost between vendors is paid in operability as much as in integration code, and the pipeline is easier to reason about
• BVendor loyalty
• COnly one family supports JSON
• DSmaller and larger models from different vendors are incompatible

Which eval pattern most reliably exposes a drop in retrieval quality rather than a drop in generation quality?

• AMeasuring answer accuracy conditional on perfect retrieval — feeding the model the ideal passages from a labelled set and comparing to its real-pipeline accuracy, so the gap is attributable
• BMeasuring end-to-end answer accuracy only
• CMeasuring token count of the answer
• DMeasuring latency

You are fusing dense and sparse scores. Why is reciprocal rank fusion (RRF) a popular choice?

• AIt averages the two scores
• BIt is the fastest fusion method
• CIt only uses whichever score is higher
• DIt avoids the need to calibrate scales between dense and sparse scores, which are not comparable; it combines ranks instead, and is robust to one retriever being systematically more confident than the other