1.2 Scoring, K-levels & Chapter Weights

How the 44-point paper is built

ISTQB publishes the Version 2.0 blueprint in its Exam Structures and Rules tables (candidates often see table pack v1.18 or a later pack such as v1.19; the CT-AI Version 2.0 numbers below are the ones this chapter teaches). The live exam is 40 multiple-choice questions totaling 44 points.

The extra four points are not bonus credit. They exist because cognitive levels are weighted:

• 36 questions at K2 (Understand), each worth 1 point
• 4 questions at K3 (Apply), each worth 2 points

You still need 29 points to pass. Missing one K3 item costs twice as much as missing one K2 item. That is why a candidate can answer 28 questions correctly and still fail, or miss several K2 distinctions and still pass if the application items are solid.

What K1, K2, and K3 mean in practice

ISTQB learning objectives use a short cognitive scale. K1 Remember is recall: name the term, pick the definition. K2 Understand is explanation and comparison: why locked models differ from adaptive ones, which quality characteristic a scenario illustrates. K3 Apply is doing the technique on a given scenario or data set. The syllabus appendix also defines K4 Analyze, but the Version 2.0 CT-AI exam composition you must plan for is the 36 K2 / 4 K3 split above.

Translate that into study habits:

• K2 rewards clean distinctions. If two terms sound similar—overfitting versus underfitting, locked versus adaptive, training versus validation versus test data—write a one-sentence contrast and a tiny example. The exam will often ask which statement is true, which risk is illustrated, or which technique belongs to which test level.
• K3 rewards procedure. You should be able to compute a metric from a confusion matrix, design a red-team probe for an LLM, apply a dataset constraint to a sample, or derive a metamorphic follow-up test when there is no perfect oracle.
• K1 still matters for keywords. Terms listed under syllabus chapter headings, defined in the ISTQB glossary or in syllabus Chapter 9 (AI-specific terms), are K1 even if no numbered learning objective repeats them. If the heading lists narrow AI, metamorphic testing, or data pipeline, you must remember the official name and meaning, not a nearby synonym you invented in a blog comment.

A worked K2 picture: a deployed fraud model is locked, so production data does not retrain it overnight. An e-commerce recommender that keeps fitting new click streams is adaptive. A K2 item will ask you to compare testability, not to write Python.

A worked K3 picture: you are given a 2x2 confusion matrix for a cancer-screening classifier. True positives 80, false negatives 20, false positives 10, true negatives 90. Precision, recall, and similar ML functional-performance metrics are calculable from those four cells. That style of item is why Chapter 3 carries eight points from seven questions.

Official chapter weights

The 40 questions and 44 points are distributed across the seven examinable syllabus chapters as follows:

Check the arithmetic once so you trust the table in a mock exam. Questions: 6+3+7+7+6+9+2 = 40. Points: 6+3+8+8+7+10+2 = 44. Each K3 chapter has one more point than it has questions (7/8, 7/8, 6/7, 9/10). That is the fingerprint of a two-point application item sitting in that chapter.

Where the four K3 items live

Memorize the four application homes with a story, not a slogan:

1. Chapter 3 — ML functional-performance metrics. A product owner says 'the model is 95% accurate, ship it.' You still need to calculate precision, recall, or a related metric from a confusion matrix and interpret what the number hides. A screening test that almost never predicts disease can look 'accurate' while missing the patients who matter.
2. Chapter 4 — red teaming GenAI / LLMs. You are not only checking that the bot answers the return-policy FAQ. You try jailbreaks, prompt injection, and other hostile or curious probes to see whether the model leaks data, ignores policy, or invents commitments the company cannot keep.
3. Chapter 5 — dataset constraint testing. Training rules say applicants must be aged 18-65 and income must be a positive number. K3 asks you to apply those constraints: which rows violate them, which checks belong in the pipeline, and what happens if production data silently drops the constraint.
4. Chapter 6 — metamorphic testing. An X-ray classifier has no perfect oracle for every image. You rotate or slightly brighten a film that the model already called 'fracture present.' The metamorphic relation says the diagnosis should not flip for that harmless change. If it does, you have a test case without needing a radiologist on every pixel.

Those four techniques are also the ones that feel slow in a 90-second budget. Practice them with pencil and a timer. K2 reading will not substitute for one worked confusion matrix, one red-team charter, one constraint table, and one metamorphic follow-up.

What is examinable — and what is not

The Version 2.0 syllabus is explicit. All syllabus sections are examinable except:

• the Introduction (the front matter that explains purpose, business outcomes, and how the document is organized)
• Hands-On Objectives
• References
• Appendices

Everything else in the seven teaching chapters can appear on the paper. An item may also combine more than one section. A Chapter 6 metamorphic question can still expect you to remember a Chapter 4 oracle problem. Do not study chapters as sealed boxes.

Standards and books cited in the syllabus are background. Their extra pages are not examinable beyond what the syllabus itself summarizes. You are not sitting an ISO/IEC 25059 reading test. You are sitting a test of the quality-characteristic ideas the syllabus actually teaches, including those it draws from that standard.

Hands-on exercises are not a second exam

Version 2.0 puts real exercises in the accredited course: create a small ML model, prepare data, score metrics, implement a perceptron, explore an LLM, test input data, apply metamorphic testing. Hands-on objectives are classified H0 (demo or video), H1 (guided), and H2 (exercise with hints). They are not separately examined. There is no lab practical with a separate pass mark.

They still earn their place. Each exercise is rehearsal for a K3 application item. If you only read about metamorphic relations and never derive one, the two-point Chapter 6 question is the item you will reread three times while the clock dies. Self-study candidates should still do a tiny version of each exercise: a spreadsheet confusion matrix, a one-page red-team charter against a public chatbot, a constraint checklist on a toy table, a metamorphic pair of images or sentences.

Accredited-course minimum timings are another published hint, and they are not identical to exam points. Chapter 3 is 375 minutes of instruction but 8 exam points. Chapter 6 is 225 minutes but 10 exam points. Chapter 7 is 30 minutes and 2 points. Use training minutes to respect conceptual density. Use exam points to decide mock-exam time. The 19.5-hour accredited minimum is 1,170 minutes; it is a course-design floor, not a promise about how many items each hour produces.

Study implications for a 29-point target

Spend most of your time on chapters 3, 4, 5, and 6. Together they are 7+7+6+9 = 29 questions and 8+8+7+10 = 33 points. That is already more than the passing score if you were perfect there—and nobody is. Those four chapters also hide all four K3 items.

A practical weekly split for an independent candidate who already holds CTFL:

• Chapter 6 (9 questions, 10 points): model risks, documentation review, probabilistic performance testing, adversarial examples, metamorphic testing, drift, over/underfitting, A/B testing, back-to-back testing. This is the densest exam chapter. Give it the first and last revision slots.
• Chapter 3 (7 questions, 8 points): ML forms, workflow, data roles, metrics, neural-network structure. Drill the confusion-matrix arithmetic until it is boring.
• Chapter 4 (7 questions, 8 points): locked versus adaptive, statistical testing, oracles, GenAI/LLM testing, red teaming, ML test levels, risk-based testing. Write two red-team charters, not twenty pages of notes.
• Chapter 5 (6 questions, 7 points): input-data risks, bias testing, data pipelines, representativeness, dataset constraints, label correctness. Build one constraint table and one bias-check list you can reuse.
• Chapter 1 (6 questions, 6 points): enough K2 coverage that you cannot skip it, but do not let hardware-and-hosting reading crowd out Chapter 6.
• Chapter 2 (3 questions, 3 points): quality characteristics and acceptance criteria. Small on the paper, easy to under-drill until a fairness or transparency stem surprises you.
• Chapter 7 (2 questions, 2 points): development and deployment testing. Learn it, then stop. Two points will not rescue a weak Chapter 6.

Keyword drill is cheap and required

Because keywords are K1 even without a numbered learning objective, keep a running glossary. When a chapter heading lists terms, write the official wording once. Do not substitute a nearby phrase from a vendor white paper. On a timed paper, K1 keyword items are the fastest points you can still drop by being sloppy.

Independent practice, official rules

This OpenExamPrep chapter is independent teaching about the published Version 2.0 structure. It does not replace ISTQB's sample exam, and it does not grade a live sitting. Use the official sample questions to feel K2 versus K3 difficulty, then come back here when you need the blueprint explained with examples. If you remember only three planning sentences, make them these: 29 of 44 points, four two-point K3 homes, and chapters 3-6 first.