NCA-GENL Trustworthy AI Practice Question
A hospital's AI governance team is reviewing an LLM that drafts discharge summaries from patient notes. Clinicians report the model occasionally invents medication dosages that were never prescribed. The team wants a mitigation that constrains generated output to an approved formulary before any text reaches the clinician. Which approach best satisfies this requirement while keeping the LLM in place?
⚠ Common exam trap
The trap here is assuming that retraining or adding a UI disclaimer removes hallucinated content, when only a runtime output-validation rail actually blocks unsanctioned values before delivery.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Configure NVIDIA NeMo Guardrails with a retrieval-augmented output rail that validates every generated dosage against the approved formulary and blocks non-matching entries.
The requirement is a runtime constraint that prevents unsanctioned dosages from appearing in generated text. NeMo Guardrails output rails can inspect model responses and validate them against an authoritative formulary retrieved at generation time, blocking or correcting mismatches before display. This keeps the existing LLM while adding a deterministic verification layer. Training changes and interface warnings do not provide that pre-delivery gate.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✓
Configure NVIDIA NeMo Guardrails with a retrieval-augmented output rail that validates every generated dosage against the approved formulary and blocks non-matching entries.
Why this is correct
NeMo Guardrails can intercept model output and apply programmable rails before text is returned. An output rail backed by retrieval over the approved formulary lets the application compare each generated dosage with authoritative entries and block or rewrite anything that does not match. This directly constrains the generation surface to sanctioned content, which is exactly the constraint the governance team requested for the discharge-summary workflow.
- ✗
Retrain the base LLM from scratch on the hospital's historical discharge summaries to eliminate the fabrication behavior.
Why it's wrong here
Full retraining is costly and does not guarantee removal of hallucinated dosages, because the model can still generalize beyond its training distribution. Historical summaries may themselves contain documentation errors, so retraining could reinforce them. Retraining also does not provide a runtime gate that prevents an unsanctioned dosage from reaching a clinician. The scenario asks for a pre-delivery constraint, which training alone cannot enforce reliably.
- ✗
Add a disclaimer banner to the user interface stating that all generated dosages must be independently verified by the clinician.
Why it's wrong here
A disclaimer shifts responsibility to the reader but does not stop the model from emitting fabricated dosages. In a clinical environment, alert fatigue means banners are frequently ignored, and a plausible but wrong dosage can still influence a decision. The requirement is to constrain output before it reaches the clinician, not to warn after the fact. Interface text is a communication control, not an output-validation control.
- ✗
Increase the model's temperature setting so it explores more phrasing variations when drafting dosages.
Why it's wrong here
Raising temperature increases sampling randomness, which makes novel and unverified token sequences more likely, not less. In a clinical discharge workflow, broader exploration directly increases the chance of fabricated dosages appearing in drafts. Temperature is a creativity control, not a factual grounding mechanism, so it cannot restrict output to a formulary. This option addresses stylistic variety rather than the safety requirement described.
About these practice questions
This NCA-GENL question is part of Courseiva's 367-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCA-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCA-GENL exam.