Capstone Exam Simulation (AIF-C01)

Fifteen lessons gave you the content. This one gives you the container: how the AIF-C01 is actually built, how it is scored, how the clock behaves, and how AWS-style distractors are engineered. Everything before this lesson was domain knowledge; everything here is test craft, and on a compensatory 100-1,000 scale with 700 to pass, test craft is worth as much as another hour of revision.

By the end of this lesson you will be able to:

• state every published logistics fact (question count, time, scale, cost, validity, retake rules) with its source;
• explain the four official question types and what each one demands of your time budget;
• run a 90-minute clock plan that never leaves an item blank;
• decode the six distractor families AWS uses on this exam;
• allocate 40 study hours across the five domains by their published weights;
• grade Set A honestly, fix the weakest domain first, and defend the comparative verdict on domain weighting;
• say what the exam does not publish, so you never assert an invented number.

---

1. Exam logistics: what AWS publishes

1.1 The facts, each with its source

Nothing in this table is folklore. Each row comes from the current exam guide (v1.1, published 30 April 2026), the AIF-C01 certification page, or the AWS certification policies and FAQ pages as retrieved in October 2026.

• 📚 Did you know? The 15 unscored items look exactly like the 50 scored ones. AWS does not tag them, colour them or otherwise separate them; which means the only rational strategy is to treat every item as scored. A pretest item you answer badly costs you nothing, and a scored item you skip costs you the full marks, so the incentive never points toward "leaving it for later and never coming back".

1.2 The four mechanics that decide your score

Read these four sentences as the operating system of the exam:

1. Compensatory scoring means a brilliant Domain 3 cannot rescue a weak Domain 4. You do not need to be great everywhere; you must not be bad anywhere.
2. No guessing penalty means a blank is strictly worse than a guess: guessing is never a risk, it is free expected value.
3. Hidden pretest items mean your raw count of "obviously right" answers maps to a scaled score you cannot see, so never compute a pass from "X of 50".
4. Only domain weights are published: per-objective weights do not exist, so any study plan built on "this topic is 8 %" is built on a guess.

1.3 Booking, delivery, retake and results

The administrative half of the exam is examinable in the sense that it changes how you plan your attempt, so treat it as a small checklist rather than trivia:

---

2. Question formats and what each one costs you

2.1 The official types

All 15 items in Set A are multiple choice, which is the format you can drill fastest. The live exam adds the other three in unpublished counts (see section 12), so practice them as skills, elimination, sequencing, pair-checking, rather than as a predictable quota.

2.2 The arithmetic of the clock

$$\text{seconds per item} = \frac{90 \times 60}{65} \approx 83\ \text{s}$$

Against scored items only, the budget looks like $\frac{5{,}400}{50} = 108$ s, but you cannot tell which 50 are scored, so 83 s is the number you plan with. A first pass at 60 s per item leaves 25 minutes for flags, ordering and matching, which is where the slow formats and the second-guesses live.

• 📚 Did you know? Two stalls of three minutes each erase six minutes; that is seven multiple-choice items you will have to answer in the last third of the exam at double speed. The exam does not reward depth on one item; it rewards coverage of all 65, because coverage is the only way to guarantee you never convert a knowable item into a blank.

2.3 Attacking the three formats Set A does not contain

Set A trains the single-best-answer muscle. The live exam adds three formats whose failure modes are different, so practise the procedure even though you cannot predict the quota:

A useful habit: in matching, the longest right-hand side is often the most specific and therefore the least likely to be a distractor, but verify it against the domain knowledge rather than guessing from length alone. In ordering, if two adjacent steps could swap without changing meaning, the stem is usually telling you the criterion (by time, by dependency, by cost): find it and apply it consistently.

---

3. Exam-day strategy: the 90-minute clock

3.1 The phase plan

3.2 Worked example 1: where the time goes

5,400 seconds across 65 items is about 83 seconds each. Plan 60 seconds for the first pass (65 minutes), which leaves 25 minutes: 13 for flagged items, 9 for ordering and matching, 3 for the final sweep. If two items consume three minutes each, you have spent six minutes of your review budget, so the correct move is always flag, guess, move, never "one more careful read".

3.3 Worked example 2: the guess protocol

There is never a reason to leave an item blank. The protocol that keeps guessing disciplined:

1. Predict the answer from the stem before you read the options; it stops the distractors from anchoring you.
2. Eliminate two options and write a one-line reason for each (wrong layer, wrong rung, wrong family, over-scoped, absolute wording).
3. If two remain, choose the one that is most specific, most AWS-native and least absolute.
4. Move within 60 seconds. Four minutes on one item is four items lost elsewhere.

---

4. Decoding the distractors

AWS does not write random wrong answers. It writes plausible near-misses that punish a specific confusion, and the same six families repeat across all five domains.

4.1 Worked example 3: elimination that survives contact with the options

A fraud model's false negatives cost 50x a false alarm. Which metric do you prioritize?

• Accuracy dies: majority-class inflation makes it meaningless under heavy imbalance.
• Precision dies: it penalizes the cheaper error, so it optimizes the wrong side of the ledger.
• R-squared dies: it is a regression metric and this is a detection problem.
• Recall survives: $\text{Recall} = \frac{TP}{TP + FN}$, and minimizing false negatives is exactly what the stem asks for.

Test every option against the stem's pain, not against general goodness: each option must answer the specific thing that hurts.

4.2 Worked example 4: what "700" actually means

Exam forms are equated, so AWS never publishes how many of the 50 scored items equal 700: any "35 of 50 = pass" claim is unsourced (section 12). What you can act on: compensatory scoring means a soft Domain 4 still sinks you, so target >= 80 % on mixed practice with zero blanks, then use the report's section classifications to pick what to revise.

4.3 Three questions for every option

Slow candidates read options; fast candidates interrogate them. Ask the same three questions of each choice, in this order:

1. What does this measure or do?: name the mechanism in one clause. If you cannot, the option is a distractor you have not yet recognized.
2. Does that mechanism answer the stem's pain?: the stem, not the topic. An option can be true and still irrelevant.
3. Is it scoped, hedged and AWS-native?: the narrowest true answer beats the broadest true answer, and a hedged statement beats an absolute one.

4.4 The trap map, domain by domain

---

5. Domain weighting and the simulation blueprint

5.1 The only weights AWS publishes

• 📚 Did you know? Domain 4 is worth 14 % of the exam from just two task statements, while Domain 1 spreads 20 % across a much longer list of objectives. That means each individual Domain 4 objective carries far more expected score than the average Domain 1 objective, and it is why a candidate who "knows ML well but never read the NIST AI RMF" can lose a passing margin in only a handful of items.

---

6. The customization ladder, and the drills that decide Set A

6.1 The ladder you must never skip

The single most repeated decision on this exam: given a problem with a foundation model, which lever comes next?

RAG changes what the model knows; fine-tuning changes how it behaves. Answer every "what should we do next?" question by asking which of those two is broken.

6.2 Worked example 5: the two-hour fix vs the two-week fix

"Answers must cite current internal policy documents the model has never seen." A better prompt will not help, because the model has no access to those documents. The correct next step is RAG (Bedrock Knowledge Bases: retrieved context plus citations). Only if the style of the answers is still wrong after retrieval do you consider fine-tuning, and then you need Provisioned Throughput to invoke the customized model at all. Candidates who jump straight to fine-tuning answer a question nobody asked.

6.3 Worked example 6: weight-driven hour allocation

Multiply weight by total hours: $0.20 \times 40 = 8.0$ h (D1), $0.24 \times 40 = 9.6$ h (D2), $0.28 \times 40 = 11.2$ h (D3), $0.14 \times 40 = 5.6$ h (D4), $0.14 \times 40 = 5.6$ h (D5). Split D4's 5.6 hours into about six dense sessions (NIST, ISO, EU AI Act, fairness, transparency, oversight) and D3's into prompting 3 / RAG 3 / agents 3 / evaluation 2.

6.4 Stem keywords and the first move they demand

The fastest way to lose Set A is to start reading options before you have decided what the stem is asking. Train this lookup table until the first move is automatic:

---

6.5 Trap Patterns & How to Beat Them

Sections 4 and 6 gave you the six distractor families and the stem lookup table. This section compresses the rest of the strategy material into three artefacts you can rehearse in the final week: a twelve-row trap map, a pacing plan keyed to the item ranges you can actually see on screen, and a short drill that punishes the reflexes most candidates bring into the exam hall. The map is deliberately wider than section 4's six families because the live exam recombines them: one item can bait you with a pricing unit and an absolute claim at the same time, and only one of the four options will fail on both counts.

6.5.1 The twelve trap families and the reflex that beats each one

Two rows do most of the damage to a first attempt: T2 (you know the service but not what it bills for) and T5 (you know both levers but not which one this stem needs). If you only rehearse two rows of this table before exam day, rehearse those two.

6.5.2 The pacing plan, keyed to item ranges

Section 3.1 splits the clock into phases; the plan below splits the same 90 minutes into item ranges, which is what you can actually count on screen. The two are compatible: ranges S1 to S4 are your first pass with ordering and matching handled where they appear, S5 is the flagged review, and both plans reserve the final three minutes exclusively for the no-blank check.

Derived check-points (arithmetic, not an AWS figure): the ceiling is 83 s per item; more than three minutes on one item costs you two items elsewhere; a sustainable flag budget is about one flag in every six items (roughly ten in total); and if S2 runs five minutes over, you cut time from S5: never from S6's blank check.

• 📚 Did you know? Because the average budget is 83 seconds, three minutes spent re-reading one stem is not three minutes lost; it is two whole items pushed into your last ten minutes. Working backwards from that ceiling, the derived flag allowance is about one flagged item in six, roughly ten flags across the exam. Flag twenty and you have not identified the hard items; you have quietly scheduled a third of the paper for a review window that cannot hold it.

6.5.3 Trap mini-drill: four items that punish the usual reflexes

These four are deliberately short, because the trap usually resolves in the first five seconds: the moment you notice which unit, which direction or which window is in play. Answer all four before reading the explanations.

6.5.4 Where each trap already lives in this course

You do not need new material for these traps; you need to reconnect them to the lesson that owns them. Read this as your revision routing table: if a row surprises you, that lesson is your next stop.

6.5.5 2025 to 2026 Updates Quiz

The trap map has one more family worth its own check: material that changed while you were studying. Guide v1.1 (30 April 2026) added objectives, added services to the in-scope list and removed one, and several services moved into maintenance or closed to new customers during 2025 to 2026. Both items below come from verified 2025 to 2026 changes: answer them before you read the explanations.

• 📚 Did you know? Seven objectives were added when the guide moved from v1.0 (26 March 2026) to v1.1 (30 April 2026): traditional ML vs foundation models, token-based pricing, context engineering, agentic AI, prompt versioning, business-alignment metrics, and hallucination detection with grounding. AWS also states that guide updates reach the live exam about one month after publication, so an item that feels "too recent to be examined" is precisely the item v1.1 was written to test.

---

7. Practice Questions, Set A: 15 single-best-answer items

Set A mirrors the real exam's dominant format: one correct response, three distractors. It is deliberately weighted like the published domains: 3 items for D1, 4 for D2, 4 for D3, 2 for D4 and 2 for D5. Set a 15-minute timer, answer every item, and never leave one blank. The answer key, organized by domain, follows in section 8.

7.1 How to sit Set A

---

8. Answer key and explanations, organized by domain

8.1 Domain 1: Fundamentals of AI and ML (Set A items 1-3)

8.2 Domain 2: Fundamentals of Generative AI (Set A items 4-7)

8.3 Domain 3: Applications of Foundation Models (Set A items 8-11)

8.4 Domain 4: Guidelines for Responsible AI (Set A items 12-13)

8.5 Domain 5: Security, Compliance and Governance (Set A items 14-15)

8.6 Rapid key for the rest of the 40-item bank (Sets B and C)

Set A is your graded simulation; these 25 items are your ungraded reinforcement set. Cover the answer column, recall the key, then check.

8.7 Reading your Set A score by domain

A single wrong answer is information; a pattern of wrong answers in one domain is a curriculum problem, and the rubric in section 9 exists to tell those two situations apart.

---

9. Scoring rubric: turning a raw score into a decision

9.1 How to grade yourself honestly

1. Score Set A raw first: one point per item, no partial credit, exactly like the live exam's multiple choice.
2. Convert by domain, not by total. A 13/15 that hides two Domain 4 misses is more dangerous than an 11/15 spread evenly, because compensatory scoring punishes the dip, not the average.
3. Categorize every miss as knowledge gap (you did not know) or reading error (you knew it and picked the distractor). Knowledge gaps go to the study plan in section 10; reading errors go to the distractor table in section 4.
4. Re-sit Set A in 72 hours. Score inflation on an immediate re-sit is memory, not mastery; a 48-72 hour gap is the shortest honest interval.
5. Convert confidence, not luck. If you guessed between two options and got it right, you are at risk: re-run that item until the reason is correct, not just the letter.

$$\text{Domain accuracy} = \frac{\text{items correct in the domain}}{\text{items attempted in the domain}} \times 100\,\%$$

---

10. Study plan: 40 hours, driven by the published weights

10.1 Hours per domain

10.2 The six-week sequence

• 📚 Did you know? Because the exam guide updates reach the live exam roughly one month after publication (guide v1.1 was published 30 April 2026), a candidate testing in the same month as a revision may face items drawn from either version. The safe behaviour is to re-read the revision page for your course about a month before your date, and to refuse to build any answer on a service that only appears in older guides (section 12).

10.3 The last 48 hours, and the morning itself

The goal of the last 48 hours is not new knowledge; it is retrieval practice on what you already have, so that the four decision rules (guess protocol, ladder, least privilege, evidence-over-assertion) fire automatically at item 1.

---

11. Synthesis of lessons 01-15

11.1 The whole course on one page

11.2 The five threads that run through every domain

1. Choose the smallest sufficient tool. Prompt before RAG, RAG before fine-tuning, managed before DIY, scoped ARN before wildcard: the same instinct appears in D1 (prebuilt vs custom), D3 (prompt vs train), D4 (documented control vs aspiration) and D5 (least privilege).
2. Know the billing unit. D2 bills in tokens, batch is ~50 % below on-demand, Provisioned Throughput is hourly, and prompt caching can cut repeated-prefix cost dramatically. Cost questions are usually just unit questions.
3. Evidence beats assertion. Compliance evidence lives in AWS Artifact, model evidence in Model Cards, bias evidence in Clarify, configuration evidence in Config, action evidence in CloudTrail. An option that promises a result with no artifact behind it is usually wrong.
4. Humans stay accountable. Human-in-the-loop oversight, human evaluation, approval gates in MLOps, and the shared responsibility split all say the same thing: automation proposes, accountability does not transfer.
5. Language discipline. "HIPAA-eligible", not "HIPAA certified"; "reduces hallucination risk", not "eliminates hallucinations"; "in scope", not "compliant". The least absolute, most specific option wins.

11.3 Final integration drill

Take any stem from Set A and name the lesson that owns it. Items 1-3 belong to lessons 01-05, items 4-7 to lessons 08-09 and 15, items 8-11 to lessons 10-11 and 13, items 12-13 to lesson 12 (with the NIST framing from lesson 01), items 14-15 to lesson 14. If you cannot place an item, you do not yet have a revision target; you have a recognition gap, which is the more dangerous kind, because it feels like knowledge.

---

12. What this lesson does not assert

To keep an unverified claim from becoming a wrong answer, these points are deliberately flagged rather than taught:

1. Case-study items. The current guide lists four question types; an older AWS document also mentions case study. Whether case studies appear live is unconfirmed.
2. Language counts and ESL wording. The product page lists 12 languages while parts of the FAQ describe 8 plus +30 minutes for ESL accommodation. The two pages are not reconciled for AIF-C01.
3. On-screen pass/fail. The FAQ says most AWS exams do not display a result on screen; AIF-C01's specific behaviour is unstated and community reports conflict.
4. Raw-to-scaled conversion. AWS never publishes how many of the 50 scored items equal 700. Any "35/50 = pass" claim is unsourced.
5. Per-topic weights. Only domain percentages exist; objectives carry no published weight.
6. Live-form composition. The number of multiple-response, ordering and matching items is unpublished; Set A is 15 single-best-answer items by design.
7. Items per domain. The "scored items" column in section 5 is arithmetic from weights, not an AWS figure.
8. Legacy services as keys. Services absent from the current in-scope list are avoided as correct answers in this lesson.
9. Actual result timing. Five business days is the official ceiling; faster anecdotal releases are unverified.
10. Time-bound promotions. Any voucher or discount offer is not re-checked here and must be confirmed on the official page before you rely on it.
11. Official practice-exam contents. Skill Builder practice exams sit behind sign-in; their counts, formats and difficulty were not retrieved for this lesson.

---