Question 220 A batch pipeline is overwhelmed because 50,000 documents each contain long repeated policies. Which pre-processing design is appropriate? Deduplicate stable reference material and supply only the relevant context for each item. Add the full document collection to every request. Remove all source context and force the model to guess. Increase concurrent writes without regard to API limits. Explanation Correct Answer A:Reducing redundant input saves tokens while preserving necessary evidence. Incorrect Answers B:That increases cost and contamination. C:That harms grounded accuracy. D:Unbounded concurrency may produce more failures.
Question 221 An agent fixes one unit test but an integration test still fails because a service contract changed. What should it do? Retry the unchanged patch until the test randomly passes. Rewrite unrelated UI components first. Trace the cross-service contract and adjust only the necessary compatibility boundary. Disable the integration test because units pass. Explanation Correct Answer C:Integration evidence points to a dependency beyond the local unit. Incorrect Answers A:Repeated execution cannot correct a stable schema mismatch. B:That does not address the observed contract failure. D:Unit success does not validate the service boundary.
Question 222 An MCP integration is configured with a relative path to a server executable and works only when started from one folder. What operational correction is best? Ask all users to move their repositories into one private home directory. Increase the model’s maximum turns. Disable process-exit reporting. Use a stable executable path or documented working-directory-independent launch command. Explanation Correct Answer D:Reliable registration cannot depend on an incidental shell location. Incorrect Answers A:Personal directory layouts should not be hardwired into team tool configuration. B:Additional model calls cannot fix a missing executable. C:Silencing errors makes troubleshooting harder.
Question 223 A support agent returns a tool_use block to check an invoice, but the host shows the tool call to the customer as its final answer. What must the host do instead? Execute the named tool, append a matching tool_result, and request another model turn. Treat every assistant message as the final answer. Start a new unrelated conversation with the invoice text. Repeat the last user prompt without running the tool. Explanation Correct Answer A:tool_use hands execution to the client; the model needs the actual invoice result before answering. Incorrect Answers B:A tool request is not user-facing completion, even if explanatory text accompanies it. C:A new conversation loses the request and the tool-use linkage needed for coherent continuation. D:Repeating the prompt does not retrieve the invoice evidence the agent requested.
Question 224 A model emits parseable JSON, but values violate a business rule that totals must equal the sum of items. What still needs to run? No validation because the JSON parser succeeded. A deletion of all line items. Application-level semantic validation after schema-conformant generation. A model temperature increase. Explanation Correct Answer C:A JSON schema does not by itself prove all financial invariants or source facts. Incorrect Answers A:Syntax does not prove totals are mathematically correct. B:Removing evidence does not repair the incorrect total. D:Randomness cannot enforce arithmetic consistency.
Question 225 An agent team has one worker producing a risk estimate from outdated data and another summarizing it correctly. How should provenance prevent propagation? Attach source version and retrieval time to the estimate so consumers can reject stale input. Strip timestamps to simplify the handoff. Ask the summarizer to guess which source is newest. Treat any well-written summary as proof inputs are current. Explanation Correct Answer A:Lineage lets the final step verify freshness before trusting derived claims. Incorrect Answers B:That destroys staleness detection. C:Freshness should be checked against records. D:Fluent synthesis does not refresh data.
Question 226 A locally launched MCP server requires an environment variable but only works on the creator’s laptop. What is the most probable missing team setup detail? A documented per-developer secret and environment variable provisioning step. A larger Claude context window. A requirement that tools respond in HTML. A rule that all developers copy the creator’s token. Explanation Correct Answer A:Shareable definitions must be paired with secure machine-specific credentials. Incorrect Answers B:Context size does not supply process environment variables. C:Response formatting does not fix missing launch credentials. D:Sharing one private credential is an inappropriate workaround.
Question 227 A generated migration runs successfully on an empty test database but fails on real null values. What should iterative refinement add? Representative fixtures covering nulls and boundary data plus a corrected migration. Only a larger output token limit. A requirement to discard all production nulls without authorization. A fabricated success metric from the empty database. Explanation Correct Answer A:A realistic failing case exposes the missing data assumption. Incorrect Answers B:More prose does not test nullable records. C:Data loss is not a valid bug fix. D:That conceals real-world failure.
Question 228 A system has to migrate 1,000 independent records and reconcile all totals at the end. What workflow shape is most suitable? Ask a writing agent to estimate reconciliation totals. Reconcile each record before it is migrated. Move all records in one opaque transaction with no per-record trace. Map each record through a validated migration task, then reduce into a reconciliation pass. Explanation Correct Answer D:Parallel per-record work plus a final aggregate check captures both local and global correctness. Incorrect Answers A:A model estimate cannot prove data consistency. B:Validation must compare the resulting state, not a planned state. C:An opaque bulk result makes failures difficult to isolate.
Question 229 A production extractor must return invoice_id as a string and line_items as an array, but free-form responses sometimes omit them. Which API capability is appropriate? A larger chat context as the only change. An automatic deletion of invoices that fail parsing. A supported structured JSON output schema specifying required fields and types. A prompt asking the model to “please be valid JSON” with no validator. Explanation Correct Answer C:Schema-constrained generation directly addresses missing and mistyped output fields. Incorrect Answers A:More context does not impose a required output contract. B:Data deletion is not a formatting remedy. D:Plain language does not guarantee schema compliance.
Question 230 Two MCP tools both say “send message” but one posts to an internal channel and the other emails external customers. How should their interfaces change? Let the model infer channel from the recipient’s first name. Name the destination, audience, side effects, and authorization requirements clearly. Remove required recipient fields from both tools. Describe them as interchangeable notification services. Explanation Correct Answer B:External communication and internal updates have different risks and approval needs. Incorrect Answers A:A name alone does not specify a safe communication destination. C:Underspecified targets can send content to the wrong audience. D:Destination differences are material to tool choice.
