Mock Exam 2 — Answer Key

D = domain scored. R = Core Reflex from Chapter 01.

Scenario A — Developer Productivity with Claude

1 — B · D2

Targeted discovery. Entry points anchor the structure; following imports traverses only what matters.

• A — Fills context with irrelevant files and causes the degradation described in item 2.
• C — File size is a poor proxy for importance.
• D — This is the failure the teams are reporting: generic patterns instead of the actual code.

2 — A · D5

Generic "typical patterns" language replacing specific identifiers is the named diagnostic signal for context degradation. Subagents keep exploration out of the main context; the scratchpad makes findings survive independently of it.

• B — Reasoning depth isn't the constraint; the code has fallen out of effective attention.
• C — The agent knows the goal; it has lost the code.
• D — Drastic, and it solves a repository-structure problem that doesn't exist.

3 — A · D2

Glob matches paths. Grep matches contents.

4 — A · D2

A unique anchor is exactly Edit's requirement, and it's satisfied here.

• B — Works, but rewrites the whole file for a one-line change.
• C — Unreviewable in-place mutation.
• D — Write replaces the file; Grep locating the line doesn't change that.

5 — A, B · D2

More candidates means more misrouting, and a docs agent holding production write tools can do things it should never do.

• C — Tool access doesn't create a communication channel; topology is separate.
• D — Not a documented consequence.

6 — A · D2 · R3

Selection is description-driven. If the MCP tool offers semantic matching or cross-repository scope, the description has to say so.

• B — Over-correction; Grep remains right for literal searches.
• C — No such option exists.
• D — Not a mechanism.

7 — A · D2

Standard third-party integrations → community MCP servers. Build custom only for your own internal systems.

• B — Unnecessary work for a solved problem.
• C — No schema, no descriptions, no error semantics.
• D — Deploying and hosting MCP servers is explicitly out of scope.

8 — A · D2

• B, C — Invent mechanisms.
• D — Attaches a fake limit to a true first clause. Common distractor shape.

9 — A · D2

MCP resources are the primitive for browsable content catalogues; tools are for actions and queries.

• B — Unbounded tool proliferation, which degrades selection.
• C — Reasonable for point lookups, but the requirement is a browsable set.
• D — Loads a large catalogue into every session's context.

10 — B · D2 · R6

Two materially different situations collapsed into one indistinguishable return. The agent can't tell "no owner" from "couldn't check."

• A — Only correct for the genuine-absence case; it silently absorbs the failure case.
• C — Escalating credentials to work around a permission boundary. Wrong on both design and security grounds.
• D — Overcorrects the other way: genuine absence is a valid result.

11 — A · D1 + D2 · R1

Least privilege plus a deterministic gate on the paths that matter.

• B, C — Prompt instructions have a non-zero failure rate. Not acceptable for production infrastructure.
• D — Human review is a good practice but not an enforcement mechanism, and the requirement is "never."

12 — B · D2

Three genuinely different jobs. Splitting gives each a focused description, a purpose-built schema, and independent selection.

• A — A description covering three jobs is the problem, not the fix.
• C — Keeps one tool with three behaviours; the schema still serves all three. Plausible but inferior.
• D — Unrelated.

13 — A · D2

Scope the common case; keep a route for the exception.

• B — Leaves the dominant case on a general tool that's easy to misuse.
• C — Drops 15% of capability.
• D — Keeps one tool with a diffuse description.

14 — A, B, C · D2

• D — Not selection information in the exam's framing, and performance benchmarking is out of scope.

15 — A · D2

When descriptions are sound, the next source of selection pressure is the system prompt. "Always check current sources" will pull the model to search regardless of tool descriptions.

• B — Output schema doesn't affect selection.
• C — Tool order is not a documented selection mechanism.
• D — Temperature affects variability, not systematic misrouting.

Scenario B — Claude Code for Continuous Integration

16 — A · D4 · R7

An explicit, checkable criterion. The model can evaluate "does the claim contradict the code?" consistently; it cannot evaluate "outdated."

• B — The exam's canonical failing instruction.
• C — Uncalibrated self-reported confidence.
• D — Comment age isn't observable from the code, and staleness isn't about age.

17 — B · D4

Trust is the operative concern. A tool developers ignore has zero value regardless of recall.

• A — Not the reasoning.
• C — Configuration-dependent, and a symptom rather than the underlying cost.
• D — False; they're the easier of the two to measure.

18 — A, B · D4

Disabling stops the trust damage immediately; rewriting fixes the cause.

• C — Exhortation.
• D — Still reports 70% noise, now relabelled.

19 — A · D4

Examples at the boundary — especially negative ones — define the criterion's edges, which prose alone can't.

• B — A self-scored rubric is still judgement without a definition.
• C — Points at conventions it can't see.
• D — Adds uncalibrated confidence to an unfixed criterion.

20 — A · D4

Recording the reasoning trigger lets you cluster false positives by cause and fix the criteria that produced them.

• B — Uncalibrated, and it says nothing about why.
• C, D — Useful metadata; neither explains the decision.

21 — A, B, C · D3

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

22 — A · D3

Session context isolation. The session that wrote the code is the worst reviewer of it.

• B — Half-right; helps at the margin, doesn't fix the structural problem.
• C — More reasoning inside the same contaminated context.
• D — Broadens scope and worsens attention dilution.

23 — A · D3

Independence, not depth. D is B twice.

24 — A · D4

Nobody waits on the nightly report, and it has a natural overnight window.

• B — No latency SLA, up to 24 hours. Could block a merge for a day.
• C — Includes the blocking path.
• D — Misstates the limitation: batch supports tools; it doesn't support multi-turn tool calling in a single request.

25 — A · D3

CLAUDE.md is the CI context carrier.

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

26 — A · D3

2–3 concrete input/output examples. The guide's number.

• B — Prose underspecifies shape; examples are a primary mechanism.
• C — One example gets over-fitted.
• D — Volume without signal; lengthens context and risks lost-in-the-middle.

27 — A, B · D4

(The full guide list is consistency, ambiguous-case handling, generalization, and hallucination reduction.)

• C — Overclaims. Examples reduce semantic error rates; nothing eliminates them.
• D — Schema validity comes from the schema, not from examples.

28 — A · D4

tool_use with a JSON schema is the structured-output mechanism.

• B — A prompt instruction gives no guarantee.
• C — Regex over prose is the fragility the mechanism exists to remove.
• D — Same problem, different format.

29 — A · D4 · R8

Schema-valid, semantically wrong. A schema can require an integer; it cannot know which integers exist in this diff. That check is computable, so compute it.

• B — An integer constraint is already satisfied.
• C — Would cause truncation, not invalid references.
• D — Touches nothing semantic.

30 — A · D3

One Markdown file per command, in the project commands directory, committed — which also makes it available locally to every developer.

• B — User-level; not shared.
• C — .claude/config.json with a commands array does not exist.
• D — Not a mechanism; commands aren't CLI flags.

Scenario C — Customer Support Resolution Agent (second visit)

31 — A · D1 · R4

• B — A model can emit text alongside tool calls. Named anti-pattern.
• C — A backstop, never the primary condition.
• D — Parsing natural-language termination signals. Named anti-pattern.

32 — A, B · D1

Assistant content verbatim, then one user message with all tool_result blocks keyed by tool_use_id.

• C — Summarizing the assistant turn breaks the tool_use_id correlation and loses the calls themselves.
• D — Multiple user messages break the turn structure.

33 — A, B, C · D1

Three of the four named loop anti-patterns (the fourth is parsing natural-language termination signals).

• D — This is correct behaviour, not an anti-pattern. The trap in the item.

34 — A · D1

Open-ended, evolving input space → model-driven. A decision tree must enumerate cases in advance.

• B, D — Both argue for the pre-configured tree.
• C — A small closed set is exactly when a decision tree wins.

35 — A · D1 · R4

Delegation is the Task tool, and "Task" must be in allowedTools or the agent physically cannot delegate.

• B — Bypasses delegation entirely; the main agent does the work.
• C — delegate: true does not exist.
• D — Sessions aren't the delegation mechanism.

36 — A, B, C · D1

• D — parallel is not an AgentDefinition field. Parallelism comes from multiple Task calls in one response.

37 — B · D1 · R5

Isolated context, inherits nothing. "The order we discussed" refers to a conversation the subagent never saw.

• A — It has the tool; it doesn't know the argument.
• C — Position effect, not a missing-referent problem.
• D — Out-of-scope capability distractor.

38 — A · D1

Multiple Task calls in a single assistant response.

• B — Doesn't exist.
• C — Sequential.
• D — One subagent doing both, which isn't concurrency.

39 — A · D1

Goal plus quality criteria, not procedural steps. The subagent should decide how.

• B — Procedural prescription; the named weaker style.
• C — Dumps the coordinator's context, reintroducing what isolation avoids.
• D — Underspecified; no criteria for a complete result.

40 — A · D1

The four-element structured handoff.

• B — Makes the human read 20+ turns.
• C — Too thin to act on.
• D — Transcript problem plus uncalibrated confidence.

41 — A · D1

The cap is a backstop with a defined action when tripped — never the primary condition, and never a silent kill.

• B — Makes the anti-pattern the primary mechanism.
• C — Removes a legitimate safety net.
• D — Breaks parallel tool use and most multi-step work.

42 — A, B · D5 · R2

Protect the durable facts; stop verbose tool output crowding out real content. Both are the domain's named techniques.

• C — Capacity distractor; doesn't address position effects.
• D — Loses exactly the early-turn commitments at issue.

43 — A · D2

Transient, retryable, and the agent retries. The one category where retry is the answer.

• B — Misclassified; escalating a network blip wastes a human.
• C — A failure reported as success.
• D — Discards a recoverable situation.

44 — A · D1 · R1

Idempotency enforced deterministically against persisted state, before the tool runs.

• B, D — Probabilistic mechanisms for a financial guarantee.
• C — Detects the double refund after the money has moved.

45 — A · D1

Stale tool results make resuming actively harmful. A fresh session plus a structured summary carries the conclusions without the stale state.

• B — Restores the stale results.
• C — Forking inherits them too; it's for branching a valid session.
• D — Compacting stale context yields compact stale context.

Scenario D — Multi-Agent Research System (second visit)

46 — A · D1

• B — Direct subagent-to-subagent communication is the named anti-pattern.
• C — Chaining is a decomposition pattern, not the topology, and it doesn't describe a coordinator.
• D — Would defeat context isolation.

47 — A, B · D1

• C, D — Both contradict isolation, and C is the assumption behind the item-37 bug.

48 — A · D1 · R10

Eight subtasks, all inside one dimension. The subagents did correct work in the scope they were given; the scope was wrong.

• B — More subagents on the same narrow decomposition produce more narrow results.
• C — Synthesis can't merge dimensions nobody researched.
• D — Tool access explains retrieval failure, not coverage failure.

49 — A · D1

When the sub-areas aren't knowable up front, survey first and decompose from what you learn.

• B — A fixed sequence presumes you already know the stages.
• C — Enumerates question types in advance.
• D — Fixed dimensions may not fit this question — and forcing them is a variant of the same rigidity.

50 — A · D1

Iterative refinement: check coverage, re-delegate for gaps.

• B — Prevents overlap up front; doesn't handle gaps found later.
• C — Chooses which subagents to spawn.
• D — A fixed sequence with no feedback step.

51 — A · D5

Key findings first, then explicit section headers. Both mitigations are about arrangement, and they serve the human reader for the same reason they serve the model.

• B — Length isn't the problem; ordering is. And arbitrary truncation loses content.
• C — Helps navigation but leaves the answer buried.
• D — Capacity distractor.

52 — A · D5

Comparable figures need aligned columns; chronology and causation need sentences; discrete findings need scannable items.

• B, C, D — Uniformity over fitness. Rendering financial comparisons as prose is a failure the guide calls out specifically.

53 — A, B · D5

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

54 — A · D1

Attention dilution. Per-file passes give each file full attention; the integration pass catches what only appears across files.

• B — Fitting isn't attending. The canonical capacity distractor.
• C — Helps comprehension order; doesn't fix dilution across 60 files.
• D — More reasoning over diluted attention.

55 — A · D5

Coverage information has to be produced where the gap occurred and propagated. It cannot be reconstructed downstream.

• B — Inferring absence from what's present is guessing.
• C — Catches unanswered questions, not silent access failures.
• D — Not an engineering answer.

56 — A · D5

Preserve the disagreement with attribution, dates, and methodology.

• B — Subagent capability isn't source quality.
• C — May hide a chronology rather than a conflict.
• D — Meaningless for figures computed on different bases.

57 — A · D1

Choosing which specialists to spawn based on the question is dynamic subagent selection.

• B — Dividing work among subagents you've already chosen.
• C — Re-delegating after a coverage check.
• D — A fixed sequence of stages.

58 — A · D1 + D4

"Whatever seems relevant" has no completeness criterion, so each run picks differently. Goal + explicit quality criteria + required return format.

• B — Exhortation.
• C — Makes an underspecified task more reproducibly underspecified.
• D — Pays three times to average away a specification problem.

59 — A · D2

Recover 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.

60 — A · D5

A manifest must carry decisions, not just progress. A later session that doesn't know the established conventions produces inconsistent work.

• B — Not the gap.
• C — Enormous and mostly stale — the reason a manifest exists.
• D — Uncalibrated, and irrelevant to consistency.

Marking sheet

Multiple-response items are all-or-nothing.

A note on the answer distribution. Option A is correct disproportionately often in this paper — an artifact of how the items were authored, not a pattern to rely on. The real exam randomizes option order. If you find yourself picking A by habit, re-read the stem.

Domain tally

This paper is not blueprint-weighted — see the note at the top of the exam. Below is its actual distribution.

Total = 60. The "Ready at" column is roughly 75% per domain — the same bar as Mock 1's minimums, rescaled to this paper's counts.

Don't convert this total into a predicted exam score — the domain mix is wrong for that, and the items are harder. What it tells you is where your misses cluster. Read the D tags on everything you got wrong; if three or more land in one domain, that's the chapter to re-read before exam day.

Because D3 is only 6 items here, this paper cannot assess Claude Code configuration adequately. If Mock 1 showed weakness there, go back to question-banks/qbank-d3-claude-code.md and Cheatsheet 2 — 20% of the real exam is riding on it and it's the most rote-learnable domain on the paper.

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