Full Practice Exam

Chapter 18 — Full Practice Exam and Exam-Day Strategy

Certification Blueprint

Field	Coverage
Exam	AWS Certified AI Practitioner (AIF-C01), exam guide v1.1 (2026-04-30)
Primary domain	All five, sampled at 20 / 24 / 28 / 14 / 14
Task statements	Every task statement, reached through the chapters that teach it
Exam facts	65 questions presented · 50 scored · 15 unscored trial · scaled 100-1000 · pass at 700 · compensatory
Question types	Multiple choice · multiple response · ordering · matching
Exam-guide status	Verified live 2026-08-07. See course-plan.md § Certification Provenance

What This Chapter Covers

Everything and nothing. There is no new subject matter here — every question on the practice exam is answerable from Chapters 01-17. What is new is the form: domains mixed, unlabelled, in sequence, with no chapter heading to tell you which mental model to reach for.

That difference matters more than it sounds. Studying a chapter primes you: you know the answer is about governance because the chapter is about governance. The exam removes that scaffolding entirely, and a candidate who has only ever been tested with the scaffolding in place has not yet been tested.

This chapter therefore teaches three things the earlier chapters could not:

1. Retrieval under mixing — recognising which domain a question belongs to before answering it.
2. Triage under time pressure — deciding fast which questions to fight and which to bank.
3. Reading your own score — converting a raw mark into a specific list of chapters.

Core Mental Model

What to remember from this diagram: the arrows converge. Five domains, five different bodies of knowledge, and exactly one number at the end. There is no per-domain pass bar, so the exam does not care where your marks come from. That single fact drives every strategic decision in this chapter.

Two consequences follow immediately, and both are counter-intuitive:

• A weak domain is survivable. Domains 4 and 5 carry 14% each. Answering every one of those 18 questions wrong and everything else right still passes comfortably. Panic about a weak domain is usually more expensive than the weak domain.
• Domain 3 is worth more than Domains 4 and 5 combined. 28% against 28% — the same. But Domain 3 is one body of knowledge and the other two are two. Marks per unit of revision are highest in Domain 3, and it is where a limited revision budget should go first.

Detailed Concept Explanations

The 15 unscored questions, and why they change nothing

The exam presents 65 questions and scores 50. The other 15 are trial items being calibrated for future papers, and they are not identifiable — no marker, no section, no pattern.

The temptation is to treat this as a licence to relax: "one in four doesn't count." Resist it. You cannot tell which, so the only rational strategy is to treat all 65 as scored. What the fact genuinely buys you is emotional: when you meet a question that seems impossibly obscure, or oddly-worded, or about something you are certain was never in the guide, there is a real chance it is a trial item. Do not let it shake you. Answer it, flag it, move on. Candidates lose marks on the questions after a demoralising one far more often than on the demoralising one itself.

Compensatory scoring, stated precisely

Your result is one scaled score from 100 to 1000, with 700 to pass. It is not a percentage, and the scaling is not published, so do not try to convert your raw practice mark into a scaled score — the arithmetic does not exist publicly and anyone who offers you the conversion is guessing.

Compensatory means the domains pool. There is no rule of the form "you must reach X% in Domain 4." This is the single most actionable fact about the exam's structure, and it produces the triage rule below.

The four question types

Type	What it asks	The trap
Multiple choice	One correct response, three distractors	The longest option looks most complete
Multiple response	Two or more correct from five or more	Partial credit does not exist — half right scores zero
Ordering	Place steps into the correct sequence	Anchor an end of the sequence first, not the middle
Matching	Pair items across two lists	Do the pairs you are certain of first; the rest constrain each other

Multiple response deserves specific respect. Selecting one of two correct answers scores exactly the same as selecting two wrong ones: nothing. When a question says select two, verify you have selected two before moving on — an unselected second answer is the most common unforced error on this exam, and it is entirely mechanical.

Reading a question in one pass

The strongest single habit is to identify the constraint before looking at the options. Almost every scenario question contains one clause that decides the answer — "with minimal infrastructure to manage", "for every day of the preceding quarter", "which identity", "before a model is built". Find it, then read the options against it.

Read the options after the constraint, not before. Reading options first is how a plausible, well-written distractor gets to establish itself as the answer before you have worked out what the question needs.

Decision Rules and Exam Signals

Signal in the stem	What it selects	Chapter
"Guaranteed identical output" / "auditable to a fixed rule"	Reject AI entirely	02
"Without labels" / "groups we have not defined"	Unsupervised, clustering	01, 02
"Minimal infrastructure to manage"	The managed service, not the capable one	07
"For every day of" / "state across a period"	AWS Config	17
"Which identity" / "who changed"	AWS CloudTrail	17
"Before a model is built"	SageMaker Clarify, not Model Monitor	14
"Must cite the source passage"	RAG, not fine-tuning	09, 11
"Consistent tone or format"	Fine-tuning, not RAG	11
"Documents change several times a day"	RAG, not fine-tuning	09
"No long-lived credential"	IAM role, not user + key	16
"Must not traverse the public internet"	PrivateLink, not NAT	16
Two options both in the guide's list	The stem's stated constraint breaks the tie	09

Distractor Patterns

Five patterns account for most wrong answers on this paper, and all five recur on the real exam.

1. The true statement that answers a different question. The hardest kind. Q13's option (a) — "only fifty questions affect the result" — is entirely true and simply is not what "compensatory" means. Verifying a claim as true feels like verifying it as the answer. Match the claim to the term being asked about, not to reality.

2. The correct answer to the adjacent question. Q38 offers the two canonical RAG cases as distractors in a fine-tuning question. These are attractive because they are right — about something else. Whenever two techniques are routinely compared, expect each to appear as the other's distractor.

3. The overstated truth. "They cannot process any language other than the one they were trained on." Multilingual capability genuinely varies; "cannot" is false. Check every absolute — always, never, cannot, guarantees, all. On this exam an absolute is wrong far more often than right.

4. The fabricated regulation. "Mandatory disclosure of model weights on request." "Carbon reporting legally required before deployment." No such general rules exist. Domain 4 uses these repeatedly, and "no such rule exists" is a legitimate elimination — you are not required to recognise every regulation, only to notice when one has been invented.

5. The name that sounds adjacent. AgentCore offered as a vector database; Neptune offered where the stem says PostgreSQL. The guard is knowing the actual lists — the vector stores, the governance services — well enough that a plausible-sounding name does not substitute for membership.

Scenario Walkthrough

A media company runs a customer assistant on a managed foundation model. It must answer from the company's own product documentation, which changes several times a week. Legal requires every answer to cite the passage it came from. Security requires that traffic from the application's VPC to the model never traverses the public internet. The team has no ML engineers.

Four requirements, four decisions. Work them separately — the exam bundles them precisely because candidates try to solve the whole scenario with one idea.

"Answer from our own documentation, changing several times a week" plus "cite the passage" → RAG, not fine-tuning. Two independent signals point the same way: content that changes frequently would require constant retraining, and per-answer citation requires retrieved passages to cite. Fine-tuning absorbs knowledge into weights, where it cannot be cited and cannot be updated cheaply. (Chapters 09, 11.)

"Managed foundation model" plus "no ML engineers" → Bedrock with Knowledge Bases, not SageMaker AI with a self-managed server. Both would work; the stated operational constraint selects the managed path. (Chapter 07.)

"Never traverses the public internet" → PrivateLink. Not a NAT gateway, which routes traffic out to the internet, and not a security-group rule, which restricts destinations without changing the path. (Chapter 16.)

Unstated but implied — how do we know it works? The business metric, not the benchmark score. If the assistant deflects support cases, that is the evidence; a leaderboard result is not. (Chapters 07, 12.)

Notice what the walkthrough did not do: it never chose a service because it was the most powerful option. Every decision was made by the stem's stated constraint.

Key Concepts Table

Concept	The one-line version	Chapter
Compensatory scoring	Domains pool into one score; no per-domain bar	01
Unscored trial items	15 of 65, unidentifiable; treat all as scored	01
Multiple response	No partial credit; half right scores zero	18
The constraint clause	The one phrase in the stem that decides the answer	all
Cost ladder	In-context → RAG → fine-tuning → pre-training	09
RAG vs fine-tuning	Changing content and citation → RAG; behaviour and vocabulary → fine-tuning	09, 11
Config vs CloudTrail	State over a period vs who called which API	17
Clarify vs Model Monitor	Before the model exists vs after it is deployed	14
Fairness vs safety	Harm distributed unevenly vs harmful output generally	13
Capability vs value	Benchmarks measure the first; business metrics the second	12

Revision Flashcards

Say the answer aloud before revealing it.

1. The clock is nearly out and you have twelve unanswered questions. What do you do? → Put an answer on all twelve immediately, then refine whatever time remains. An unanswered question scores zero with certainty; a guess between four options returns 25% on average, and 50% once you have eliminated two. Leaving the paper with a blank is the only choice available to you with a guaranteed-zero return, and there is no penalty for a wrong answer.

2. A question says "select two" and you are confident of only one. What now? → Pick a second anyway. Partial credit does not exist on this exam, so a single selection scores exactly what a wrong pair scores — nothing. There is nothing to lose and one option in four to gain. Then count your selections before moving on: an unselected second answer is the most common unforced error on the paper, and it is entirely mechanical.

3. Your practice score is 44 out of 65. Are you ready? → Borderline, and the raw number is the least useful thing you have. Do not re-sit this paper — a second sitting measures recall of these 65 items and nothing else. Take the per-domain totals to the chapters behind each miss, and sort the misses into knowledge gaps and reading errors first, because those two findings have completely different fixes.

4. A question names a service you have never heard of. What does that tell you? → Possibly nothing at all. It may be one of the 15 unscored trial items, or an invented distractor of the kind Domain 4 uses repeatedly. Eliminate whatever you can justify rejecting, answer, flag it, and move on. The real risk is not the question itself — candidates lose marks on the questions that follow a demoralising one far more often than on the demoralising question.

5. Two options are both real services from the correct list. How do you choose? → Re-read the stem for a constraint you have not used yet. When both options are genuinely in scope, the tie-breaker is always something the scenario stated — Neptune really is in the vector-storage list, and it loses to Aurora only because the stem says the team already runs PostgreSQL. Membership of the list never breaks the tie; the constraint does.

6. What is the fastest way to lose marks you have already earned? → Spending disproportionate time on a hard question in a weak domain. Scoring is compensatory and there is no per-domain pass bar, so a mark in your weakest domain is worth exactly what an easy mark elsewhere is worth — and it costs you far more to earn. Bank the cheap marks first; the exam does not care where your marks came from.

Exam-Ready Model Answer

"How should I approach the exam on the day?"

Make one pass through all 65 questions in order, answering every question you can resolve on the first read and flagging everything else — but always recording an answer before moving on, because an unanswered question and a wrong one score the same, and only one of them can turn out lucky. On each question, find the constraint clause in the stem before reading the options, then read the options against it. Where two options survive, choose the one the stated constraint selects rather than the one that sounds most complete.

Because scoring is compensatory and there is no per-domain pass bar, I will not spend disproportionate time on a weak domain: a mark in Domain 4 is worth exactly a mark in Domain 3 and costs me more to earn. On the second pass I revisit flagged questions in flag order, and I finish by confirming that every question carries an answer and that every select-two question carries exactly two.

Chapter Checklist

• I sat the 65-question paper closed-book, in one sitting, without pausing
• I recorded an answer for every question before checking any answer
• I computed per-domain totals, not just the headline score
• I read the walkthrough for every question — including the ones I got right
• I identified, for each miss, which of the five distractor patterns caught me
• I listed the specific chapters my misses point at, by number
• I can state what compensatory scoring means and what it implies for triage
• I can name the four question types and the trap in each
• I know that a select-two question scores zero for a single selection

After the Session

Work the chapters your misses point at — not the whole course, and not this paper again. The walkthroughs name a chapter for every question precisely so that revision can be specific.

If your per-domain totals are weak in Domain 3, start there regardless of the others: it carries 28% of the exam, more than Domains 4 and 5 together, and it is a single connected body of knowledge rather than two.

Confirm the exam's duration and the current guide version on the official exam page before you book. This course was authored against v1.1 (2026-04-30) and the lifecycle was verified 2026-08-07; if you are reading this materially later than that, re-check before relying on the domain weights, because they are what this paper's sampling is built on.