Mock Exam 1 — Answer Key

Each item gives the answer, the reason, and why the distractors fail. The D tag is the domain the item scores against; the R tag is the Core Reflex from Chapter 01.

Quick reference table at the bottom for fast marking.

Scenario A — Customer Support Resolution Agent

1 — B · D1 · R1

A PreToolUse hook is a deterministic gate: it runs before the tool executes and cannot be talked around. Prompt instructions have a non-zero failure rate, which is unacceptable for a financial action.

• A — Position and emphasis reduce the failure rate; they don't eliminate it. The exam's "firmer prompt" distractor family.
• C — Examples shape typical behaviour, not guaranteed behaviour.
• D — PostToolUse runs after the refund. Detection, not prevention.

2 — C · D1

PostToolUse normalizes heterogeneous formats before the model ever sees them. The canonical use of the hook.

• A — Documentation makes the model responsible for conversion it will sometimes get wrong.
• B — Same problem at the prompt layer.
• D — Wrong layer. The issue is inconsistent input, not arithmetic weakness.

3 — B · D1 · R1

Deterministic threshold enforcement, plus a defined route for the blocked case.

• A — Probabilistic mechanism for a compliance rule.
• C — Destroys first-contact resolution for every legitimate small refund. Over-correction.
• D — A schema maximum makes a valid request unrepresentable: the agent can't express "$780" at all, so it can't route it to escalation either. Half-right and the best distractor here.

4 — A · D1 · R4

stop_reason is the loop's control signal. "end_turn" means done; "tool_use" means continue.

• B — Checking assistant text is exactly the described bug. A model can emit explanatory text alongside tool calls.
• C — An iteration cap is a safety backstop, never the primary condition.
• D — Parsing natural-language termination signals. Named anti-pattern.

5 — B · D1

All tool_result blocks for one assistant turn go in one user message, each keyed to its tool_use_id.

• A — Breaks the turn structure; the API expects one response message.
• C — Loses the tool_use_id correlation.
• D — Tool results are always user role.

6 — B · D5

Resolvable ambiguity: ask for one more identifier.

• A — Guessing, and a wrong guess acts irreversibly on someone else's account.
• C — Premature escalation. Ambiguity isn't an escalation trigger.
• D — Acts first, documents the risk after.

7 — A · D2

Category + retryability + specific reason + alternative. All four elements.

• B — Misclassified. The input was valid; policy prohibits the action. Marking it retryable wastes turns.
• C — A failure reported as success.
• D — Correct flag, useless payload.

8 — A, B, C · D5

The three named triggers.

• D — Complexity is what the agent exists to handle. Escalating on complexity defeats the 80% target.

9 — A · D5

An explicit request for a human is honoured immediately.

• B — Optimizes the metric against the customer's stated wish.
• C — A structured handoff is required, but built from what the agent already knows — don't make the customer re-explain after they've asked to leave.
• D — Subordinates an explicit request to an uncalibrated signal.

10 — B · D1

The four-element structured handoff: customer ID, root cause, refund amount, recommended action.

• A — Makes the human read 20+ turns. The point of a handoff summary is that they don't have to.
• C — Too thin to act on.
• D — Transcript problem plus uncalibrated confidence.

11 — A · D5 · R2

The stem names the exact four categories summarization destroys: numbers, dates, amounts, customer-stated expectations. A persistent case-facts block protects them and keeps them positionally prominent.

• B — Defers the limit; does nothing about position effects.
• C — Loses the same categories later rather than never.
• D — If it's being summarized, it doesn't fit to re-read.

12 — A · D5

Compounds across every call in a long conversation. Trim at the boundary.

• B — Latency isn't the concern.
• C — Extra fields don't cause hallucination.
• D — False. Unused fields still consume context and dilute attention.

13 — A, B · D2

Both shrink the candidate set the model discriminates across.

• C — Duplicates the descriptions in a weaker location and costs context on every call.
• D — Forces tool use without improving which tool.

14 — A · D2 · R3

Descriptions are the primary selection mechanism. Two descriptions with the same opening give the model nothing to discriminate on; explicit cross-referencing boundaries are the direct fix.

• B — Discards a real capability distinction.
• C — Weaker layer, doesn't scale.
• D — Half-right. Better names help marginally; the descriptions remain indistinguishable.

15 — A · D5 + D4

Explicit criteria first, then 2–3 boundary examples including near-miss negatives.

• B — Uncalibrated self-reported confidence.
• C — Sentiment measures emotion, not resolvability.
• D — Tool-call count is a proxy for nothing. A five-call case may be perfectly resolvable.

Scenario B — Code Generation with Claude Code

16 — B · D3

Project CLAUDE.md is committed, so everyone gets it on clone.

• A — User-level and not version-controlled.
• C — Slash commands, not always-loaded context.
• D — MCP server configuration.

17 — B · D3

User-level memory: all their projects, not version-controlled, invisible to teammates.

• A, C — Both committed; both impose it on the team.
• D — A skill is an on-demand procedure, not a standing preference.

18 — B · D3

A paths: glob matches test files wherever they live, including new directories, and loads nothing when you're not on test code.

• A — Duplication across dozens of directories; misses every new one.
• C — Works, but loads testing conventions into every session. Path-scoped rules are the preferred answer for cross-directory, pattern-defined conventions.
• D — Commands are invoked deliberately; conventions must apply automatically.

19 — A · D2

.mcp.json is the project-level, committed configuration. ${ENV_VAR} expansion shares the config without sharing the secret.

• B — Commits a credential. Never correct.
• C — Works but isn't shared; every developer configures it manually and it drifts.
• D — ~/.claude.json is user-level, outside the repository, and holds unrelated personal config.

20 — A · D3

@import composes the file's contents into loaded memory.

• B — Duplication that drifts.
• C — A prose pointer isn't a load mechanism.
• D — Discards everything else CLAUDE.md holds.

21 — A · D2

Glob matches file paths. This is exactly its job.

• B — Grep searches contents; it misses files using a different test idiom and matches non-test files containing the string.
• C — Read operates on files, and tests sit beside their components rather than in one directory.
• D — Works, but reaching for Bash when a purpose-built tool exists is the wrong instinct, and the exam tests the Grep/Glob distinction directly.

22 — A · D3

One Markdown file per command, in the project commands directory, committed.

• B — User-level; not shared.
• C — .claude/config.json with a commands array does not exist.
• D — Would work as a skill, but the requirement is a named slash command.

23 — B · D3

Skills are on-demand; CLAUDE.md is always loaded. context: fork keeps the verbose middle out of the main conversation.

• A, C — Always-loaded locations for something needed monthly.
• D — Rules are scoped by file pattern; this is scoped by occasion.

24 — B · D2

Edit requires a unique anchor. Four occurrences means no unique anchor, so Read + Write is correct.

• A — Fails, or edits the wrong occurrence.
• C — Grep locates it; the edit still has no unique anchor.
• D — Bypasses the file tools with an unreviewable in-place mutation.

25 — A, C, D · D3

Large scale, multiple viable approaches, architectural change.

• B — Single-file, one obvious approach. Direct execution.

26 — A · D3 + D5

The Explore subagent isolates verbose discovery and returns findings rather than the search output that produced them.

• B — Reactive and lossy. Preventing pollution beats compacting it.
• C — Capacity distractor.
• D — Trades discovery quality for context. Wrong trade.

27 — A · D3

One message lets the agent reconcile the issues together. Sequential fixes cause churn — fixing #1 can undo #3.

• B — The intuitive-but-wrong option.
• C — CLAUDE.md is for standing conventions, not transient feedback.
• D — No mechanism, no benefit.

28 — A, B, C · D3

• D — --ci does not exist. Nor do --batch, --non-interactive, or CLAUDE_HEADLESS.

29 — A · D3

CLAUDE.md is the CI context carrier — already committed, already the source of truth.

• B — False, and it duplicates conventions in two places.
• C — No such flag.
• D — Not a mechanism for conventions.

30 — B · D1 + D3

Stale tool results and stale file contents make resuming actively harmful — the agent reasons from a snapshot that no longer holds. A fresh session with a structured summary carries the decisions without the stale state.

• A — Restores the stale results.
• C — Forking also inherits the stale context; it's for branching a valid session.
• D — Compacting stale context yields a compact summary of stale context.

Scenario C — Multi-Agent Research System

31 — B · D1 · R10

The logs implicate the coordinator: the subagents did correct work inside the scope they were given, and the scope was wrong.

• A — Tool access explains failure to retrieve, not failure to cover a dimension.
• C — Synthesis can't merge dimensions nobody researched.
• D — Capacity distractor.

32 — A · D1

Explicit dimension naming + a coverage check + re-delegation for gaps. Scope partitioning plus iterative refinement.

• B — More subagents on the same narrow decomposition produce more narrow results.
• C — Pre-configured procedural steps in place of goal-directed decomposition. The exam prefers goal and quality criteria.
• D — Maximum redundancy; this is the item-33 failure.

33 — A · D1

Each prompt states its own scope and what the others cover.

• B — Choosing which subagents to spawn, not partitioning their work.
• C — Handles gaps after the fact; doesn't prevent overlap.
• D — Discards parallelism to solve a scoping bug.

34 — A · D1 · R4

Delegation requires the Task tool in allowedTools. Without it the agent physically cannot delegate, so it does the work itself. Check capability before prompt quality.

• B — Would affect routing quality, not the ability to delegate at all.
• C — Not a real distinction the exam makes.
• D — parallel: true does not exist.

35 — B · D1

Parallelism comes from multiple Task calls in a single assistant response.

• A — parallel: true doesn't exist.
• C — That's sequential.
• D — Invented field.

36 — B · D1

The description is how the coordinator decides where to route a task — functionally analogous to a tool description.

• A — Underestimates it; it's operative.
• C — That's the prompt field.
• D — Not user-facing.

37 — B · D1 · R5

Subagents have isolated context and inherit nothing. "The company we discussed" refers to a conversation the subagent never saw.

• A — Tools aren't the issue; the target is undefined.
• C — Position effect, not a missing-referent problem.
• D — Out-of-scope model-capability distractor.

38 — B · D1

Hub-and-spoke: all inter-subagent communication routes through the coordinator.

• A — Direct subagent-to-subagent channels are the anti-pattern.
• C — A scratchpad is a context-management technique for a single agent's own findings, not the topology answer.
• D — Too absolute; they can share, via the coordinator.

39 — B · D2 · R6

An access failure is not a valid empty result. Because the tool reported success, nothing downstream could distinguish "we looked and found nothing" from "we couldn't look."

• A — Downstream verification can't recover concealed information.
• C — Retrying a permission failure fails identically — and nobody knew there was a failure.
• D — A confidence score on a false premise is still false.

40 — A · D2 + D5

Recover what you can locally; propagate what you can't, with partial results attached.

• B — Hides category, scope, retryability, and what succeeded.
• C — Silent suppression.
• D — Discards three successful streams.

41 — A, B · D5

The two symmetric anti-patterns.

• C, D — Both parts of the correct middle path.

42 — C, D · D2

Business-rule violations and permission denials don't change on retry.

• A — Retryable by definition.
• B — Retryable after input correction — the model can fix a malformed ID.

43 — A, C · D5

Structured returns stop verbose reasoning from consuming coordinator context; a scratchpad makes accumulated findings survive independently of it.

• B — Capacity distractor.
• D — Breaks hub-and-spoke.

44 — A · D5

The mapping must survive synthesis in a structured form.

• B — Reattaching means guessing which source supported which sentence — the exact information lost. Named wrong approach.
• C — A bibliography doesn't map claims to sources.
• D — Uncalibrated confidence, and it discards valid provenance.

45 — A · D5

Preserve the disagreement with attribution, dates, and methodology so the reader can judge.

• B, C — Both discard information; C may hide a chronology rather than a conflict.
• D — Meaningless for figures computed on different bases.

Scenario D — Structured Data Extraction

46 — A · D4 · R8

Schema-valid and semantically wrong. A schema cannot express a cross-field arithmetic relationship; a computed check can.

• B — The named trap. No numeric constraint expresses a sum across array elements.
• C — Would cause truncation, not arithmetic error.
• D — Loosens an existing guarantee; touches nothing semantic.

47 — A, B · D4

Syntax, not semantics. C and D are exactly what a schema cannot guarantee.

48 — A · D4

A required non-nullable field forces the model to produce something, and the something is invented.

• B — "N/A" is a sentinel that pollutes downstream data and doesn't distinguish absent from unreadable.
• C — It's already required; that's the cause.
• D — A minimum-length constraint makes fabrication more likely by removing the empty escape.

49 — A, B · D4

"unclear" = can't determine. "other" = determinate but unlisted.

• C — Conflates a vendor problem with a document-type category.
• D — Confuses an output category with a workflow state.

50 — B · D4

The detail field is the feedback loop: 200 other values turn out to be "proforma invoice," which becomes a proper enum member.

• A — Not a validation requirement.
• C — Reasoning isn't the purpose.
• D — Different field, different purpose.

51 — A · D4

A regex catches a malformed shape; "3/4/25" is well-formed and ambiguously interpreted. The schema constrains shape; rules constrain interpretation.

• B — Half-right, and the strongest distractor. Enforces ISO 8601 shape, not correct interpretation.
• C — Post-hoc parsing faces the same ambiguity with less context.
• D — Retrying an ambiguous source produces a different guess.

52 — A, B, C · D4

• D — Exhortation. Adds nothing the three concrete elements don't supply.

53 — A · D4

Retry cannot recover information absent from the source; continuing to retry invites fabrication.

• B — More attempts at an impossible task.
• C — Guidance against a structural pressure the schema creates. Fix the schema.
• D — Makes the same wrong answer more reproducible.

54 — A, B, C · D4

The model reports what the document states, sums the line items itself, and flags disagreement — all while the document is in context. You still re-sum in your own code: two independent checks.

• D — Uncalibrated self-reported confidence. Not the mechanism.

55 — A · D5 + D4

An aggregate over a corpus that's 85% invoices is dominated by invoices. 97% overall is consistent with 99% on invoices and 43% on statements.

• B — Arithmetically correct, architecturally wrong. The trap.
• C — Uniform sampling under-samples the rare, risky types.
• D — Uncalibrated confidence.

56 — A · D5

Stratification guarantees each category gets enough samples to yield a usable estimate.

• B — In this corpus you'd get ~2 statements out of 50.
• C — Not random; batches may be ordered by source or type.
• D — Over-samples one stratum and measures nothing about the others.

57 — A · D5

Once "0.85 on total for invoices" maps to a measured 98.1% accuracy, it's actionable. Raw, it's noise.

• B — Instructing conservatism calibrates nothing.
• C — Two uncalibrated estimates agreeing is still uncalibrated.
• D — Too absolute; the exam permits it after calibration.

58 — A, B, C · D4

• D — The key limitation, stated backwards. Single-pass structured output works; an agentic loop does not.

59 — C · D4

Worst case = wait for the next submission window + the full processing window:

interval + 24 ≤ 30  →  interval ≤ 6

• A, B — Forget the first term.
• D — Safe but not the maximum. (The guide's own variant uses 4-hour windows for margin — a choice, not the limit.)

60 — A · D4

custom_id correlation exists so you can target exactly the failures.

• B — Pays for 50,000 successes twice.
• C — Loses the discount unnecessarily; the failures aren't latency-sensitive.
• D — Not an engineering answer.

Marking sheet

Multiple-response items are all-or-nothing. Partial selections score zero.

Domain tally

Several items are genuinely cross-domain — the real exam is the same, and the guide's own scenario-to-domain mapping overlaps. The assignment above is one defensible allocation that reproduces the exact blueprint weights; the per-item D tags in the rationales name the primary domain. If you'd rather tally by the rationale tags, do that instead — the total is what predicts your result.

Total: ___ / 60. Go to scoring-grid.md.