What the ET actually looks like¶
The exam structure below is officially announced by the course team — not our inference from sessions or papers. Around it, everything else comes from four real papers (T1-2026 AN + FN, T3-2025 AN + FN) and three live revision sessions (ET-1, ET-2, ET-3).
The official structure — 80 marks, 90 minutes¶
| Fact | Detail |
|---|---|
| Total marks | 80 |
| Duration | 90 minutes (1.5 hours) |
| Negative marking | None — same as every paper before |
| Section 1 | 30 MCQ/MSQ questions · 39 marks — graded automatically |
| Section 2 | 9 short-answer questions · 41 marks — "Applied AI Judgment", manually graded (not auto-graded) |
Marks per question are not fixed — within each section some questions carry 1 mark, some 2, and Section 2's questions are the heavy ones (4–5 marks each). One written answer can outweigh three objective questions.
Must remember
There is no negative marking. Never leave a question blank. A wrong guess costs the same as a skipped question — nothing. If 30 seconds are left, mark something on every unanswered question — and write something on every short answer too: an LLM grader can only reward what's on the page.
What "Applied AI Judgment" means¶
The official name for Section 2 sounds abstract, but it's simple in practice:
You're given a real situation involving an AI or data system — and your answer is the judgment you make about it.
The situations look like: a flawed API design, a suspicious pattern in a dataset, a claim someone made about a model's performance, a buggy script, a client's vague request. Your answer shows four things:
- What's really going on — the mechanism, not the surface symptom
- What the evidence supports — and what it doesn't
- What else it could be — the alternative you tested before concluding
- What to do next — a safe, reversible action
No code, no syntax — "we are least concerned about syntax" (ET-3). The rubric rewards reasoning over conclusions: a well-argued alternative answer beats an asserted "correct" one. If you practiced the P2-style thinking of evidence → alternatives → safe action, that is exactly this skill.
One confirmed question format inside this section: agent-prompting — given a software problem, write the prompt you'd give a coding agent (like Codex) to solve it. Role, context, task, constraints. Method: the grading guide.
How the 90 minutes should be spent¶
With 39 questions on the paper, pace matters more than ever:
Section 1 — 30 questions, 39 marks ~40-45 min ≈ 80-90 seconds per MCQ
Section 2 — 9 questions, 41 marks ~45-50 min ≈ 5 minutes per written answer
Must remember
Section 2 is worth more than half the paper (41 of 80). The classic mistake is burning 60 minutes perfecting MCQs and sprinting the written answers. When your Section-1 time budget is up, move — a missed MCQ costs 1-2 marks; a rushed or blank written answer costs 4-5.
The six official topics — the study map¶
The official study guide names six topics: topics 1–5 feed Section 1, topic 6 is Section 2. Every week of this course has a home in them — nothing you studied is wasted:
| # | Official topic | What it covers | Learn it | Revise it |
|---|---|---|---|---|
| 1 | Observability & Monitoring | percentiles vs averages, rates vs counts, health vs readiness, AI cost tracking | Week 6 | Data & ML |
| 2 | Data Pipeline Integrity | stable identity, change detection, partial runs, idempotent retries, reproducibility, provenance | Week 6 | Data & ML |
| 3 | CI/CD & Release Security | secret isolation, supply chain, risky infra review, progressive rollouts | Week 7 | Git & security |
| 4 | Reliable AI/LLM Systems | verify output, structured output, grounding, authz in code | Week 3 | LLMs & prompting |
| 5 | Web/API/Infra Fundamentals | statelessness, error design, CORS/authN/authZ, identity vs delegation, git history, containers | Week 2 | Web & APIs · Git & security |
| 6 | Applied AI Judgment | robust prompts, evidence choice, high-leverage questions, rubric design, probability × impact, minimal fixes, valid vs invalid claims | The grading guide | same |
Some Section-1 questions will be scenario-stemmed — a short 2-3 sentence situation followed by one or two objective questions. That's the classic PYQ scenario-block style (the FastAPI-proxy story, the StreamFlix gateway), just compacted. The full scenarios live in Section 2.
How marks were laid out in previous papers (history)¶
Past terms were 40-mark papers with a difficulty ramp — useful for how questions are built, not for this term's totals:
Part 1 — 1-mark questions (~8-9 questions) warm-up, single facts
Part 2 — 2-mark questions (~5 questions) reasoning, "why does this happen"
Part 3 — 3-mark questions (~2 questions) diagnosis, multi-step thinking
Part 4 — Scenario blocks (~3 blocks) one story, 5-6 questions each
Part 5 — Short answers (~2 questions) write 200 words, graded by an LLM
This term keeps the kinds of thinking (warm-up recall → reasoning → diagnosis → scenario judgment) but reorganizes them into the two official sections, with the written half grown from ~4 marks to 41.
What the questions are actually about¶
Not syntax. Not memorized definitions. The papers test one thing: do you understand why a system behaves the way it does?
The recurring shapes:
| Shape | Example from a real paper |
|---|---|
| Why did this break? | UI on :3000 fetches API on :8000 → browser throws CORS error. Why? |
| What happens if...? | You run sqlite3 new.db and the file doesn't exist. What happens? |
| Pick the professional fix | API returns 500 under load. What do you check? |
| Spot the wrong belief | "COUNT(column) and COUNT(*) are the same." True or false? |
| Choose the right tool | Storing large numeric arrays: XML, JSON, CSV or Parquet? |
| Judge the situation (Section 2) | A dashboard shows a suspicious spike. What is it, and what do you do? |
Must remember
Almost every question is a scenario + one concept. If you know the concept, the question takes 20 seconds. If you don't, the options can't save you. That's why this site is organized by concept, not by paper.
Strategy the professor gave away¶
These came straight from the revision sessions:
Study concepts, not answers
The professor's ET-3 advice on PYQs: use them to predict the core concepts, not the exact questions — this term's question shapes are new. The last three papers are enough; don't deep-dive older ones.
The options are designed to guide you (Section 1)
Read all four before answering. Many questions are built so that wrong options are plausible-sounding nonsense — eliminate those and the answer is what's left.
These option tricks cover Section 1 only
Elimination and option-reading work where there are options. Section 2 (41 marks) has none — for that, the LLM-grading guide is the strategy.
What is NOT in this term's exam¶
OpenRefine is excluded
The professor said it explicitly in ET-2: not in this term's syllabus. If you see
OpenRefine questions in old papers (text clustering, value.toTitlecase()), skip
them — you will not be asked.
Geospatial is almost certainly out
Folium, GeoPandas and CRS appeared in T3-2025 and even T1-2026, but the current term's course content contains zero geospatial material. Learn only the 30-second version: distance functions expect (latitude, longitude) — swapped coordinates point somewhere else entirely. That concept still transfers.
Spreadsheets and Seaborn are gone too
Old-syllabus leftovers. Don't study them for this exam.
Next steps: Read How to write answers an LLM grades well — Section 2 is 41 marks and this method is how you collect them. Then take Mock-1 under a 90-minute timer.