You are compiling guardrail tactics for a customer-facing assistant.
Which two tactics belong on the guardrail list?
Each correct answer presents a complete solution.

• A. Layer prompt-level guardrails with runtime content checks rather than relying on either alone.
• B. Rely on a single hardened system prompt that enumerates every disallowed behavior.
• C. Embed approved override phrases that let trusted users relax guardrails on demand.
• D. Validate the model output against a structured schema before downstream actions are taken.
• E. Lower sampling temperature globally to reduce the chance of off-policy completions.

You are rolling out a standardized Claude Code configuration to an engineering team and must complete the planning steps before piloting the configuration.
Which two steps must be completed BEFORE piloting the configuration with a small group of engineers?
(Select two.)
Each correct answer presents part of the solution.

• A. Define the project-scope baseline covering Model Context Protocol (MCP) servers, permission rules, and subagents.
• B. Onboard every engineer in the organization to the new configuration through mandatory training sessions.
• C. Iterate the configuration based on the pilot findings and stabilize the baseline before broader rollout.
• D. Roll out the stabilized configuration to additional teams with documentation and a defined support channel.
• E. Identify the team workflows, security boundaries, and which decisions belong to managed configuration versus project scope.

A Claude architect at a health services organization is defining evaluation metrics for a clinical-summary pipeline. The pipeline must remain within a per-query cost ceiling and must never surface patient data to unauthorized roles.
Which two metrics directly address these requirements? (Select two.)

• A. Role-based access-control enforcement rate measured on a red-team dataset
• B. Throughput measured as successful requests processed per minute
• C. BLEU score computed against a human-annotated reference summary set
• D. Per-query token cost measured against the defined cost ceiling
• E. Response latency at the 95th percentile across a one-week sample window

An architect is reviewing a Claude-based candidate-screening tool prior to deployment. A stakeholder asserts that because the model was not trained on company data, no bias evaluation is necessary.
Which two responses most accurately challenge this assertion? (Select two.)

• A. Fairness testing is required only when the training dataset is known to contain protected-class labels.
• B. Transparency obligations are satisfied by disclosing that an AI system is in use, without further evaluation.
• C. Bias evaluation is unnecessary when the model provider has published a responsible-use policy.
• D. The model may carry demographic biases from pretraining that manifest in screening outcomes regardless of fine-tuning.
• E. The system should be evaluated on outcome-disparity metrics across protected groups before deployment.

You are reviewing a peer's draft architecture decision record (ADR). The ADR states the decision and the technical detail but does not state the alternatives considered or the trade-offs accepted.
Which response is most appropriate?

• A. Approve the ADR as written on the grounds that the decision itself is documented, accepting that reviewers and future maintainers will lack the context to evaluate or revisit it.
• B. Ask the author to add the alternatives considered and the trade-offs accepted before publishing the ADR.
• C. Delete the ADR entirely and rely on the engineering team's collective memory to reconstruct the decision context, alternatives, and trade-offs during future audits or onboarding.
• D. Replace the ADR with a different decision unrelated to the one drafted.