Core AI and exam-specific vocabulary including definitions for terms that appear frequently in CCAO-F and PCAO-F questions.

tags: [ccao-f, exam-prep, vocabulary, mock-exam] parts: 1-of-3, 2-of-3 covers: "Questions 1–40"

Mock Exam Vocabulary Guide

Part 1 of 3 — Business Roles, Industries & Hard Terms (Q1–Q20)

How to use this guide: Keep this open while practicing the mock exam. When you hit a word you don't understand, find it here. Roman Urdu translations are added where a concept is especially hard to picture in English.

SECTION 1: JOB TITLES & ROLES (Who Is In These Questions?)

The exam puts every question inside a workplace. You don't need to BE these people. You just need to understand what they do so you know the STAKES of the situation.

1. Junior Analyst

Simple meaning: A new, entry-level employee who does research, data work, or report writing. They have limited experience.

Roman Urdu: Naya kaam karne wala banda jise abhi ziyada tajruba nahi hai.

Why it matters for the exam: When a junior analyst produces something, the exam is testing: should this go for expert review? The answer is almost always YES — a new person without domain expertise should not be the final check.

Appears in: Q1 — A junior analyst with no customs background drafts a compliance briefing.

2. Finance Associate

Simple meaning: Someone who works in a company's finance department. They handle money records, expenses, budgets, reports. Think: accountant ka chhota version.

Why it matters: When finance associates use AI for numbers (totals, calculations), those numbers MUST be verified. Finance = numbers = always check.

Appears in: Q2 — A finance associate at a hotel group uploads expense data and asks Claude to summarize spending.

3. Operations Team / Operations Associate

Simple meaning: The people who run the day-to-day business. They make sure things actually happen — deliveries go out, processes work, machines run.

Roman Urdu: Roz ka kaam chalane wali team — dispatch, logistics, processes.

Why it matters: Operations people follow procedures. When AI generates a procedure or briefing for them, it must be accurate because they will follow it literally.

Appears in: Q1, Q5, Q18, Q19

4. Compliance Analyst / Compliance Team / Compliance Owner

Simple meaning: People who make sure the company follows laws, government rules, and company policies.

Roman Urdu: Jo log dekhte hain ke company law aur rules ke mutabiq chal rahi hai ya nahi.

Why it matters: This is a KEY exam role. When data is "restricted" or "regulated," you ask the compliance owner before doing anything. They decide what's allowed.

Exam rule: Never guess about compliance. Always ask the compliance owner.

Appears in: Q11, Q15, Q25, Q32

5. Property Manager

Simple meaning: A person who manages rental properties — apartments, offices, buildings. They communicate with tenants (the people who rent), handle maintenance, collect rent.

Roman Urdu: Kiraye ki jagah ki dekh bhal karne wala banda.

Appears in: Q23 — A property manager drafts AI replies to tenants' maintenance requests.

6. Policy Owner

Simple meaning: The specific person inside a company who is RESPONSIBLE for a particular rule or policy. They wrote it, they can change it, and questions about it go to THEM.

Roman Urdu: Jo banda kisi rule ka zimmedar hai — agar rule unclear hai, usi se poochho.

Why it matters: A major exam pattern is: "The associate is unsure whether X is allowed." The correct answer is almost always: go to the policy owner, not guess, not ask Claude.

Appears in: Q8, Q14, Q23

7. CFO (Chief Financial Officer)

Simple meaning: The top finance leader of a company. They control the budget and make big financial decisions.

Roman Urdu: Company ka sabse bada paise wala head.

Why it matters:

CFOs care about bottom-line numbers, not technical details.
When writing FOR the CFO: be brief, show the outcome, name the human checkpoint.
When writing ABOUT the CFO: their decisions (like cutting a budget) must be documented before AI can use that information.

Appears in: Q26 (communicating to CFO), Q39 (CFO cut the budget in a meeting — but it was undocumented)

8. Facilities Associate / Facilities Analyst

Simple meaning: Someone who handles the physical buildings and equipment of a company — maintenance, space planning, contractors, security badges.

Roman Urdu: Office ya building ka kaam dekhne wala.

Appears in: Q12 (franchise requirements), Q19 (shift reconciliation), Q34 (warehouse inspection), Q38 (badge-swipe records)

9. Procurement Associate / Procurement Manager

Simple meaning: The person who buys things for the company — equipment, services, suppliers. They compare vendors and manage contracts.

Roman Urdu: Company ke liye cheezein khareedne wala banda.

Appears in: Q35 (supplier scorecard), Q36 (quote generator), Q56 (testing the scorecard)

10. Legal-Operations Associate

Simple meaning: Someone in the legal department who handles the administrative/operational side — not a lawyer, but works closely with legal processes.

Appears in: Q41 — reviews a 180-page master services agreement.

11. Lettings Firm (and a Lettings Associate)

Simple meaning: A company that rents out properties — houses, apartments, offices. Lettings = rental business (British English term).

Roman Urdu: Kiraye ke ghar ya daftar dene wali company.

Appears in: Q29 — A lettings firm wants Claude to decide whether a tenant's deposit is forfeited.

12. Claims Associate / Claims Team

Simple meaning: People who handle insurance claims. When something goes wrong (accident, damage), a customer files a "claim" — these people review it and decide if the insurance pays out.

Roman Urdu: Insurance company mein jo log decide karte hain ke paisa milega ya nahi.

Appears in: Q49 (warranty claims screening), Q31 (insurance claims letter)

13. Strategy Associate

Simple meaning: Someone who helps the company plan for the future — what to do next year, how to compete, what to change.

Appears in: Q51 — needs a proposal for redesigning the returns process.

14. Learning-and-Development (L&D) Associate

Simple meaning: The person who trains employees — designs training programs, workshops, onboarding.

Roman Urdu: Company mein employees ko train karne wala.

Appears in: Q55 — wants five format ideas for a leadership program launch.

15. Audit Committee / Internal Audit

Simple meaning: A group (usually senior leaders or board members) that reviews whether the company's financial records and processes are correct and honest. They check for errors, fraud, or rule violations.

Roman Urdu: Company ke accounts aur processes ko check karne wali senior committee.

Why it matters: Output going to audit committee = very high stakes. Must be accurate, traceable, stored properly.

Appears in: Q25 (regulatory attestation pack), Q47 (product launch analysis)

16. Regulator / Regulatory Body

Simple meaning: A government or official organization that makes and enforces rules for a specific industry. Example: a financial regulator oversees banks, a health regulator oversees hospitals.

Roman Urdu: Sarkari idara jo rules banata aur enforce karta hai.

Why it matters: When a question mentions "the regulator," the stakes are HIGH. Compliance reports, attestation packs, and regulatory submissions must be accurate and verified against the regulator's OWN documents — not third-party summaries.

Appears in: Q9 (data-protection threshold), Q32 (compliance report queried by regulator)

SECTION 2: INDUSTRIES IN Q1–Q20

Shipping & Logistics Industry (Q1, Q5, Q7, Q8)

This industry moves goods from one place to another — by ship, truck, plane. It has its own vocabulary.

Key Terms:

Shipping Broker A company or person that arranges shipments between buyers and sellers. They don't own the ships — they connect people who need to ship things with people who have ships. Roman Urdu: Maal bhejne ka kaam arrange karne wala.

Customs / Customs Clearance When goods cross a country's border, they must pass through "customs" — the government checks what's being imported/exported, calculates taxes, and approves entry. Roman Urdu: Maal jab ek mulk se doosre mein jata hai, border par government ki janch — ye customs hai.

Customs-Compliance Briefing A document explaining the rules a team must follow when moving goods through customs. If this document has errors, shipments get stuck, fines happen, goods get confiscated. Why stakes are high: This is a legal/regulatory document that real people will follow.

Port Authority The government or official body that manages a port (harbor where ships dock). They announce things like closures, fees, and schedule changes. Appears in Q6: "What did the port authority announce about the berth closure?" → needs web search (real-time news).

Berth Closure A "berth" is the specific place in a port where a ship docks. A "berth closure" means that spot is temporarily closed — ships can't use it. Roman Urdu: Jahaz jo jagah par rukta hai wo band ho gayi.

Demurrage ⭐ (Very important — appears in Q7) This is one of the hardest terms in the exam.

Demurrage = a FINE/PENALTY that a shipping company charges when you keep their container or ship for too long beyond the agreed time.

Simple example: You rent a truck for 3 days. You keep it 5 days. The extra 2 days = demurrage charge.

Roman Urdu: Jab aap container ya jahaz ko agreed time se zyada rokho, to jo penalty lagti hai usse "demurrage" kehte hain.

Why it appears in Q7: The analyst asks Claude: "What does demurrage mean?" — this is a simple vocabulary question. Correct format = a short conversational reply in chat, NOT a formal document.

Cold-Chain Logistics Moving temperature-sensitive goods (medicine, food, vaccines) in controlled temperatures so they don't spoil. The "chain" means the temperature is controlled at every step. Roman Urdu: Thandi jagah mein maal ki delivery — jaise dawa ya khana jo kharab na ho. Appears in Q6: "Produce a cited comparison of cold-chain logistics providers in East Africa" → research mode.

VAT (Value Added Tax) A tax added to goods when they are sold or imported. Different countries have different VAT rates. Roman Urdu: GST ki tarah ka tax jo cheez ki price par lagta hai. Appears in Q6: "What is today's VAT rate on imported machinery?" → web search (needs current rate).

Customs Declaration An official form you fill when importing/exporting goods. You declare WHAT you're bringing in, HOW MUCH it's worth, and PAY the taxes.

Rate Card A list showing fixed prices/fees for services. Like a menu of prices. Roman Urdu: Services ki price list. Appears in Q6 (finance team's approved rate card → internal document, not on the web) and Q21 (superseded rate card → old prices still in the Project).

Finance & Hotel Industry (Q2)

Hotel Group A company that owns multiple hotels. They have centralized finance/operations teams.

Expense Export A downloaded spreadsheet of all expenses (money spent) — like a report from accounting software.

Category Totals Grouping expenses by type (travel, meals, office supplies) and adding them up for each group. If Claude makes an error in grouping even one item, the total for that category is wrong. Why Q2 matters: 300 rows → prose arithmetic is unreliable → use code execution.

Legal & Contracts (Q3, Q8, Q9)

Distribution Agreement A legal contract between a manufacturer and a company that will sell/distribute their product. It covers: territory, prices, exclusivity, obligations. Roman Urdu: Maal bechne ke liye company aur dealer ke beech ka contract.

Indemnity Clauses ⭐ Very important legal term. "Indemnity" = a promise that one party will compensate the other if something goes wrong. An "indemnity clause" in a contract says: "If X happens and you suffer a loss, WE will pay for it." Roman Urdu: Contract mein ek hissa jo kehta hai — agar nuqsan hua to hum bharpai karenge. Why it appears in Q3: The associate uploaded only a SUMMARY. The summary missed the indemnity clauses. Missing an indemnity clause = you might agree to pay someone's losses without knowing it.

Pricing Schedules The section of a contract listing exact prices, discounts, and payment terms.

Risk Analysis Looking at a document (like a contract) and identifying: what could go wrong? What are we agreeing to that might hurt us?

Data Protection & Privacy (Q9)

Data-Protection Rule / Regulation Laws that govern how companies collect, store, and use people's personal data. Example: GDPR in Europe.

Threshold The exact number/limit where a rule starts to apply. Example: "GDPR applies to companies processing data of more than 250 employees." 250 = the threshold. Roman Urdu: Wo limit jahan se rule lagu hota hai.

Primary Source The ORIGINAL, official document — the actual law, the regulator's own website, the government publication. NOT a blog, NOT a summary, NOT a third-party article. Roman Urdu: Asli document — jo sarkaar ya regulator ne khud likha ho. Why it matters (Q9): Claude cited a 2023 consultancy blog. That's NOT primary. Go to the regulator's own current page.

Consultancy Blog Post An article written by a consulting firm explaining a law or rule. Useful for understanding — but NOT authoritative. The rule may have changed since 2023. Always verify against the primary source.

Workforce / HR (Q10, Q11)

Retention (Employee Retention) Keeping employees — making sure they don't leave for other companies. Roman Urdu: Employees ka company mein rokna.

Four-Day Work Week Working 4 days instead of 5, same salary. A debate about whether this improves employee happiness and reduces turnover (people leaving).

KYC (Know Your Customer) ⭐ A legal requirement for banks and financial firms: before opening an account, you MUST verify WHO the customer is. You check their identity documents, address, source of funds. Roman Urdu: Bank ya finance company mein account kholne se pehle customer ki identity verify karna — ye KYC hai.

KYC Exception When a normal customer doesn't meet all the standard KYC requirements, but there's a special reason to proceed anyway. These are HIGH STAKES decisions — someone accountable must approve them. Why Q11 matters: Claude has been accurate in SUMMARIZING KYC data. That does NOT mean Claude should APPROVE exceptions. Summary quality ≠ decision quality.

Rental / Property (Q12, Q16)

Franchise Agreement A legal contract where a big company (the "franchisor") allows another person/business to operate under their brand and system. Like a McDonald's franchise — you pay McDonald's and follow their rules. Roman Urdu: Jab koi badi company aapko apne naam par kaam karne ki permission deti hai, uske rules ke saath.

Franchisor The big company that gives the franchise. (McDonald's is the franchisor. The local owner is the franchisee.)

Operating Requirements The rules and standards the franchisor says you MUST follow to operate the franchise. If you miss one, you can lose the franchise. Why Q12 matters: AI extracted 37 requirements. Before planning anything, verify those requirements are complete and traceable to the actual contract.

Payroll (Q15)

Payroll Corrections Changes made to employee salary payments — fixing errors in what someone was paid.

Employee Identifiers The specific codes or numbers that identify each employee in the system (like employee ID numbers). These are sensitive data.

Restricted Data Data that company policy says cannot be shared freely — it needs special approval and routes. Why Q15 matters: The identifiers ARE the task. You can't remove them. So you must find the APPROVED route — ask the compliance owner.

Shipping/Logistics Continued (Q18, Q19)

Tariff-Change Notice An official letter to customers announcing that prices/fees (tariffs) are changing. Roman Urdu: Customers ko bheji jaane wali notice ke rates badal rahe hain.

Legal Footer The small text at the bottom of an official letter/document that includes required legal statements. Forgetting it can make the document legally non-compliant.

Technician-Shift Export A spreadsheet/file downloaded from a system showing which technician worked which shift on which day.

Site Sign-In Sheets Paper or digital records of who signed in/out at a physical location. Used to verify attendance.

Reconcile / Reconciliation Comparing two sets of records to check they match. Example: Does the shift export match who actually signed in? Roman Urdu: Do alag records milana aur dekhna ke dono same hain ya nahi.

SECTION 3: KEY CONCEPTS FROM Q1–Q20

These are the ideas (not just words) that are being tested in questions 1–20.

Domain Credibility (Q1)

What it means: Whether the output meets the standards of that professional field — as judged by an actual expert in that field.

Claude can produce text that LOOKS correct. But if the person who asked has no expertise in that area, they cannot judge whether it's actually correct. Only a field expert can.

Example: A junior analyst with no customs background writes a customs briefing. It looks professional. But only someone WITH customs experience can confirm it would hold up in real practice.

Exam pattern: Whenever a non-expert produces something for a high-stakes professional domain (legal, medical, customs, financial), a DOMAIN EXPERT must review it.

Prose Arithmetic (Q2)

What it means: When an AI writes numbers in paragraph/sentence form instead of calculating them with code.

Example: "The total for Category A is $14,200 and Category B is $8,700..."

This is UNRELIABLE for large datasets. An AI writing numbers in prose can easily make errors that look correct. The answer: use CODE EXECUTION — code always does the same math the same way.

Baseline (Q5)

What it means: The measurement of how things are RIGHT NOW, before any change. You need this to prove that a change made things better or worse.

Roman Urdu: Change karne se pehle ki current position — taaki baad mein comparison ho sake.

Example: "Before adding AI, we took 3 hours to make this report. After AI, it takes 45 minutes." The "3 hours" was the baseline.

Exam pattern (Q5): Before improving a process, always record the baseline first. Without it, you have nothing to compare against.

Standing Instructions (Q5, Q16, Q20)

What it means: Instructions you put in your Claude setup once, and they apply to EVERY chat automatically. You don't repeat them each time.

Like rules pre-written in a Project or account-level settings.

The risk: Standing instructions apply EVERYWHERE. If you write "always open with an executive summary," Claude does it even for quick questions.

Halo Delegation (Q11) ⭐

What it means: Trusting an AI to make a HIGH-STAKES decision just because it did a DIFFERENT (easier) step well.

Roman Urdu: AI ne ek kaam acha kiya, to hum soch lete hain ke wo doosra bada kaam bhi kar sakta hai — ye galat hai.

Example: Claude has been accurate at SUMMARIZING KYC data for 4 months. So the team lets Claude APPROVE the KYC exceptions directly. This is halo delegation — summary accuracy does NOT prove decision accuracy.

Exam rule: Never give an AI a bigger responsibility just because it did a different, smaller task well.

Unstaffed Gate (Q11) ⭐

What it means: A human "review step" that exists on paper, but the person doesn't actually check anything — they just sign off without reading.

Roman Urdu: Kaagaz par likha hai ke "insaan check karega" — lekin woh banda kuch check karta hi nahi, bas sign kar deta hai.

Example: The compliance analyst is supposed to sign off on each KYC exception by reviewing the summary line — but they never open the actual case. The gate exists but is unstaffed.

Exam rule: A human gate must name WHO checks, WHAT they verify, and WHEN (before the irreversible step).

Jurisdiction (Q8)

What it means: A specific country, state, or region whose laws apply. Different places have different laws.

Roman Urdu: Kaun sa mulk ya jagah — jo apne alag qanoon rakhti ho.

Example: Recording customer calls is legal in the UK with one set of rules. But the company just entered Germany — German law (a different jurisdiction) may have different consent requirements.

Risk Matrix (Q8)

What it means: A table or chart that scores different risks by likelihood and impact. Used to decide priority.

Why Q8 says it's wrong: The associate was asked to escalate a SPECIFIC legal question. Sending a risk matrix (which requires the associate to make judgment calls they're not qualified to make) is not what the legal owner needs. They need the specific unanswered question.

Data Profiling (Q19)

What it means: Looking at your data BEFORE doing anything — finding out its structure, spotting problems, seeing what's messy.

Roman Urdu: Data ko pehle inspect karo — kya hai, kya gadbad hai — phir kaam shuru karo.

Example: Before writing rules for data reconciliation, first ask Claude to scan the files and report: "What's inconsistent? What's missing? What's ambiguous?"

Threshold (Q20)

In the context of instructions, threshold means: the specific NUMBER or CRITERIA that defines when a rule kicks in.

"Flag unusual expenses" = NO threshold. Two people will decide differently. "Flag any single payment above £500" = threshold. Everyone flags the same thing.

QUICK REFERENCE CARD — Q1 to Q20

Term	3-Word Summary	Question
Demurrage	Late container penalty	Q7
KYC	Verify customer identity	Q11
Halo delegation	Past success ≠ new authority	Q11
Unstaffed gate	Human gate, no human checking	Q11
Primary source	Official document, not blog	Q9
Domain credibility	Field expert must judge	Q1
Baseline	Before-change measurement	Q5
Jurisdiction	Which country's law applies	Q8
Indemnity clause	Contract: we'll pay your loss	Q3
Franchise agreement	Brand license + rules	Q12
Reconcile	Match two records together	Q19
Threshold	Exact number rule kicks in	Q20
Prose arithmetic	AI writing numbers in sentences	Q2
Standing instructions	Pre-set rules for all chats	Q5, Q16
Rate card	Fixed price list	Q6, Q21

Part 2 coming next: AI concepts (sycophancy, context drift, model ladder, Projects, Skills, Connectors) + frameworks (halo delegation, shadow AI, purpose limitation) from Q21–Q40.

Part 2 of 3 — AI Concepts, Model Choices & Key Frameworks (Q21–Q40)

This part covers the AI-specific vocabulary and the exam's most important thinking frameworks. These appear in EVERY domain, not just one.

SECTION 4: HOW CLAUDE REMEMBERS THINGS (Memory & Storage)

This is tested heavily in Q21–Q40. You must know exactly where Claude stores things and what each storage type does.

The Four Places Claude Can "Know" Something

Think of it like your own brain vs. your desk:

Where	What It Is	Stays How Long?	Exam Name
Training	What Claude learned before you met it — general world knowledge	Forever (baked in)	Model weights / pre-trained knowledge
Chat window	What you type in this conversation	Only this session	Context window
Memory	Notes Claude writes about YOU to use in future chats	Until you delete it	Memory / stored memory
Project	A saved workspace: instructions + uploaded files	Until you change it	Project knowledge / Project instructions

Memory (Claude's Notes About You)

What it is: When you tell Claude something about yourself ("I prefer short replies," "I lead the Meridian project"), Claude can save a note and use it at the start of every future chat.

Roman Urdu: Claude aapke baare mein notes likh leta hai — har naye chat mein ye notes saamne rakh deta hai.

The problem (Q33): If memory has OLD or WRONG information, it will keep appearing in EVERY new chat. Fix = go into memory, find the wrong note, delete or correct it. Don't just add a correction line to each prompt — that treats a configuration fault as a per-prompt chore.

Exam rule: When the same wrong assumption appears in EVERY new chat → fix the MEMORY, not the prompt.

Superseded Rate Card / Stale Knowledge Files (Q21) ⭐

What "superseded" means: Replaced by a newer version. The old one is now outdated. Roman Urdu: Purani cheez jo nai cheez se replace ho chuki ho.

The scenario: A pricing Project gives unit prices to the cent. The associate thinks Claude hallucinated. But actually — the Project's knowledge files contain an OLD rate card (superseded). Claude faithfully reads the old file and gives old prices. No error message. Just confidently wrong output.

Key insight: Old prices TO THE CENT look exactly like hallucination. But the cause is a stale file, not a model error. Fix = check the Project's knowledge files, replace the old rate card, set a review date.

Exam rule: Before assuming hallucination, check if stored knowledge files are outdated.

Incognito Chat / Temporary Chat (Q22) ⭐

What it is: A special chat mode where the conversation is NOT saved to your history and NOT used to train memory.

Roman Urdu: Aisi chat jo history mein save nahi hoti aur memory mein bhi nahi jaati.

What people THINK it does: "Nothing is recorded — total privacy."

What it ACTUALLY does:

✅ Keeps the chat out of your history
✅ Doesn't update your memory
❌ Does NOT mean the vendor stores nothing — both Claude (Anthropic) and ChatGPT retain conversations for ~30 days for safety/abuse review
❌ Does NOT make it "off the record" for an audit
❌ On an enterprise/org plan, admins may still be able to see or export chats

Critical exam trap (Q22): An associate suggests using incognito mode for supplier-pricing conversations "so there's nothing on record if the audit team asks." This is WRONG thinking. The mode is memory-free, not trace-free. Using it to hide business decisions from auditors looks like concealment.

Exam rule: Incognito/temporary chat = no history, no memory. It does NOT = no vendor record, no audit trail.

Project (Q16, Q21, Q24, Q40)

What it is: A saved workspace in Claude where you can:

Write standing instructions (rules that apply to every chat inside this Project)
Upload knowledge files (documents Claude uses as reference)
Have multiple chats that all share those instructions and files

Roman Urdu: Ek save kiya hua kaam ka kamra — apne rules aur files ke saath.

Key behavior: Each Project is SEPARATE. Files and instructions in Project A do NOT appear in Project B.

The mistake (Q40): A consultant put wholesale AND retail pricing files in one Project. Claude mixed up the two. Fix = one Project per unit. "Two Projects is the wall."

Another mistake (Q24): A long chat drifts (forgets rules, contradicts itself). Moving it to a Project does NOT fix the drift — a Project holds standing files, it doesn't un-crowd a long conversation. Fix = summarize the chat, correct the summary, paste into a FRESH chat with files re-attached.

Account-Level Instructions vs. Project Instructions (Q16)

Account-level instructions: Rules that apply to EVERY chat you open, regardless of topic.

Only write things here that are TRUE for ALL your conversations.
Example: "I am a consultant. Write in British English."

Project instructions: Rules that apply ONLY inside that specific Project.

Write topic-specific rules here.
Example: "Use the Meridian account's tone guide" → belongs in the Meridian Project, not account-level.

The mistake (Q16): Associate put "always open with a 3-line executive summary" and "use Meridian tone guide" at account-level. Now every quick question — even "what's 2+2" — gets an executive summary in Meridian's voice.

Exam rule: Account level = universal rules. Project level = work-specific rules.

SECTION 5: THE MODEL LADDER (Q28, Q41)

The exam tests whether you know WHICH model to use for which task.

The Three Rungs

Model	Speed	Cost	Use When...
Haiku	Fastest	Cheapest	High-volume, simple, repetitive tasks. No deep reasoning needed.
Sonnet	Middle	Middle	Most normal work. The default starting point.
Opus	Slowest	Most expensive	Deep analysis, complex reasoning, large/multi-part documents.

Roman Urdu:

Haiku = sasta aur tez, simple kaam ke liye
Sonnet = middle — zyada tar kaam isi se
Opus = sabse powerful, mushkil aur bada kaam ke liye

Extended Thinking (Q28, Q59)

What it is: A setting you can TURN ON inside a model (especially Sonnet or Opus) that makes Claude "think longer" before answering — working through trade-offs step by step.

Roman Urdu: Claude ko zyada time dena sochne ke liye — taaki complex problems better solve ho sakein.

Use it when: Multiple constraints, trade-offs, or variables need to be weighed. Example: "Which of these 4 warehouses should we roll out to first, given staffing, downtime, and budget?"

Don't use it when: Simple tasks, quick lookups, routine rewrites. It wastes budget and time.

Critical distinction (Q28): Extended thinking is a SETTING INSIDE a model. It is NOT a different model on the ladder. Switching "extended thinking on/off" and "switching from Haiku to Opus" are TWO DIFFERENT CHOICES.

The Two Mistakes (Q28) ⭐

Mistake 1 — Flagship on simple work: Using Opus + extended thinking for meeting-notes tidy-up = wasting budget on work that needs no deep reasoning.

Mistake 2 — Cheap model on complex work: Using Haiku for an annual pricing-model rebuild = the cheapest rung wastes the analysis. Important decisions get shallow reasoning.

Exam rule: Match COST to COMPLEXITY.

Routine/high-volume → Haiku
Normal knowledge work → Sonnet
Complex, long, multi-variable → Opus (possibly + extended thinking)

SECTION 6: OUTPUT RELIABILITY — WHAT CAN GO WRONG (Q21–Q33)

Hallucination (Q4, Q21, Q43)

What it is: When Claude generates something that sounds confident and correct but is MADE UP — it doesn't exist in any source.

Roman Urdu: Claude ne khud se kuch bana liya — jo asli mein hai hi nahi — lekin confident lag raha tha.

Signs of possible hallucination in text:

A very specific number (like "14.73%") with NO source attached
A claim that Claude "checked the website" but shows no actual quote or page
A figure stated in one place that contradicts the same figure elsewhere in the document

Key distinction: Not every wrong answer is hallucination. If Claude reads an OLD file and gives OLD prices confidently — that's stale configuration, not hallucination. The fix is different.

Sycophancy (Q10, Q60) ⭐

What it is: Claude's tendency to agree with whatever the question implies, rather than giving an honest answer.

Roman Urdu: Claude jo sawaal karne wala chahta hai wahi answer deta hai — seedhi baat nahi karta.

The classic example (Q10):

Person A asks: "Why is a 4-day week BETTER than flexible hours?" → Claude gives a convincing case FOR the 4-day week.
Person B asks: "Why are flexible hours BETTER than a 4-day week?" → Claude gives an equally convincing case FOR flexible hours.
Both answers are persuasive. Both contradict each other. The model just agreed with what each question implied.

Why this happens: The model was trained to give helpful, agreeable, confident text. A leading question signals the desired conclusion — the model supplies it.

How to fix it: Use NEUTRAL framing. Don't tell Claude what conclusion you want. Instead:

❌ "Why is X better than Y?" (leading)
✅ "Evaluate both X and Y. Give the strongest case on each side." (neutral)
✅ "Rate this shortlist on these 3 criteria." (forces specific criteria, not just agreement)

Context Drift (Q24) ⭐

What it is: In a very long chat, Claude starts to "forget" earlier instructions, contradict itself, and give generic answers instead of following your specific rules.

Roman Urdu: Lambi chat mein Claude purani baatein bhool jaata hai aur galat/generic answers dene lagta hai.

Three warning signs:

Claude stops following a rule you set at the start (e.g., stops including page numbers)
Claude gives a figure that contradicts what it said an hour ago
Claude gives generic advice instead of specific answers about your document

The fix (Q24):

Ask Claude to write a SUMMARY of all decisions made, all open points, and all rules
Correct anything wrong in that summary
Open a FRESH chat, paste the corrected summary + re-attach the files
Continue from there

What does NOT fix it: Moving to a Project (Projects hold standing files, they don't un-crowd a conversation).

Stale Configuration (Q21, Q52) ⭐

What it is: When the files or instructions in your Project/setup are outdated — the world changed, but the setup wasn't updated.

Signs: Output was correct for months, then suddenly has old product names, old prices, discontinued items.

Roman Urdu: Claude Project mein purani files rakhi hain — duniya badal gayi, files nahi badlein.

The trap: This looks like hallucination (confidently wrong output) but the cause is different. The model faithfully read an old file.

Fix: Check Project knowledge files → replace outdated ones → set a review schedule.

Under-Specification (Q18, Q34)

What it is: You gave Claude too little information. The request was vague, so Claude filled in gaps with generic/wrong assumptions.

Roman Urdu: Aapne Claude ko poori baat nahi batayi — usne khud se guess kar liya.

Signs: Wrong from the FIRST answer, in a FRESH chat. (Contrast: if it worked before and broke later → stale config. If it was never right → under-specification.)

Fix (Q18): Re-read your original request. What did you forget to include? (In Q18: the legal footer and currency were never mentioned in the request.)

Fix (Q34): State the reader, the purpose, the length, what to include, what to leave out. Turn a vague request into a proper briefing.

SECTION 7: HUMAN ACCOUNTABILITY (Q23, Q26, Q27, Q29, Q37)

This is the exam's most important theme. These questions all test the same principle in different disguises.

The Core Principle: Accountability Cannot Transfer to a Tool

What it means: When YOU send something to a client, publish something, approve a payment, or make a legal decision — YOUR NAME is on it. The fact that AI helped does not move responsibility to AI.

Roman Urdu: AI ne kaam kiya — lekin zimmedari aapki hi rehti hai. AI par blame nahi daal sakte.

Exam appearances:

Q29: "Using the most capable model" doesn't make it OK for Claude to decide on deposit forfeitures. The accountability issue is about WHO answers — not about accuracy.
Q37: 12 clean AI-generated analyses doesn't earn the right to remove the human's name. A track record earns the assistant similar tasks, not your name off the result.
Q26: "AI clears the disputes" removes the human checkpoint from the sentence. The CFO now believes there's no gate. Fix: always name what the tool does AND the human checkpoint after it.

Defined Gate (Q23) ⭐

What it is: A human review step that is PROPERLY SPECIFIED with three things:

Who reviews (the named role, not just "someone")
What they check (the specific risk they're verifying, not just "quality")
When they check (before the output becomes hard to undo)

Roman Urdu: Insaan ka review step — lekin sahi se defined ho: kaun, kya check kare, aur kab.

The vague version (Q23): "All replies will be reviewed by a person before they are sent." — Sounds responsible. Means almost nothing.

The defined version: "The property manager reviews each reply before sending, checking that no commitment has been made that conflicts with the lease terms."

Exam rule: "A human will review it" is not a gate. A gate names WHO, WHAT, and WHEN.

High-Stakes, No-Undo Actions (Q27, Q42, Q45)

The principle: When an action CANNOT be reversed, a human must be the one to take it — not AI.

Roman Urdu: Jab kaam undo nahi ho sakta — insaan khud kare, AI nahi.

Examples:

Releasing bank payments (Q27) — once sent, cannot be recalled
Cancelling 40 purchase orders in the ERP system (Q42) — bulk deletion
Publishing a press statement (Q42) — already public
Approving a deposit forfeiture (Q29) — legal determination against a named person

Key nuance (Q27): The code that PREPARES the payment run can be automated — that's fine. But the button that SENDS the payments must be pressed by a named person. "Code prepares; a person sends."

Key nuance (Q45): The reason the code execution step (early-payment discount calculation) is safe is BECAUSE a supervisor reviews it immediately after. Remove the supervisor → the same calculation is no longer safe, even though the arithmetic didn't change.

SECTION 8: BUSINESS PROCESSES IN Q21–Q40

Supplier Payment Run (Q27)

What it is: The monthly process of paying all your suppliers (companies you buy from). A list of payments is prepared, checked, then uploaded to the bank.

Roman Urdu: Har mahine suppliers ko paisa dene ka process — list banti hai, check hoti hai, bank mein upload hoti hai.

Why it's high-stakes: Once you upload to the bank and release, payments go out. You cannot recall them. This is a "no-undo" action.

Invoice-Dispute Triage (Q26)

Invoice: A bill sent to you by a supplier asking for payment. Invoice dispute: When you disagree with what's on the bill — wrong amount, wrong item, already paid, etc. Triage: Sorting and prioritizing — deciding which disputes need immediate attention. Roman Urdu: Supplier ki galat bill pe jhagda — aur ye decide karna ke kaun sa pehle handle karo.

Regulatory Attestation Pack (Q25)

Attestation: A formal, official statement confirming something is true. You are "attesting" = swearing this is accurate. Attestation pack: A set of documents submitted to a regulator as official proof of compliance.

Roman Urdu: Regulator ko bheja jaane wala official pakka document — "haan hum rules follow kar rahe hain."

Why Q25 matters: This must be stored in the FIRM'S OWN records system for 7 years (per policy) — not left in Claude's workspace. "Finished work exits the platform."

Badge-Swipe Records (Q38)

What it is: Digital records of when employees badge in and out of a building — showing who was where and when.

Why sensitive: Each record is linked to a specific named employee. This is personal data.

Pseudonymised vs Anonymised (Q38): ⭐

	Pseudonymised	Anonymised
What happened	Real names replaced with codes ("Employee 47")	All identifying info removed AND no mapping exists anywhere
Can you trace back?	YES — if the lookup table exists	NO — impossible to re-identify
Data sensitivity tier	UNCHANGED — still restricted	Lower — less sensitive
What to do	Remove the lookup table too	Safe to analyze

The mistake in Q38: The analyst replaced badge numbers with codes. But the second tab of the workbook still has the lookup (the map from code → employee name). The data is PSEUDONYMISED, not anonymised. Still restricted. Fix: remove the lookup tab completely before uploading.

Supplier Scorecard (Q35, Q56)

What it is: A table/report that rates each supplier on criteria like price, delivery speed, quality, reliability. Used to decide which suppliers to keep or drop.

Roman Urdu: Suppliers ki performance ki marks sheet — kaun acha hai, kaun bura.

Weighted score: Each criterion has a WEIGHT (importance percentage). Example: Price = 40%, Delivery = 35%, Quality = 25%. The final score multiplies each rating by its weight and adds them up. If the weighting math is wrong, every score is wrong.

Feature problem vs Description problem (Q35): ⭐

	Description Problem	Feature Problem
What it means	The prompt is unclear — better wording will fix it	No amount of wording fixes it — the task needs a different feature (like code execution)
How to identify	One more attempt with clearer instructions produces improvement	SAME error persists even as everything else improves
Fix	Rewrite the prompt	Switch to the right tool (e.g., code execution for calculations)

In Q35: Four attempts, weighted scores still wrong each round, everything else improved = feature problem. Calculations belong in code, not prose.

CRM (Customer Relationship Management) System (Q36)

What it is: Software that tracks a company's interactions with customers — sales history, contacts, deals, follow-ups. Examples: Salesforce, HubSpot.

Roman Urdu: Software jo company ke customers ke saath tamam dealings track karta hai.

Why Q36 matters: The sales director made a prompt-built Claude Project a STEP in the CRM process. Other teams now DEPEND on it. This makes it infrastructure — it needs an owner, uptime, and proper access control. The associate must ESCALATE to engineering. Using a personal Project as operational infrastructure is the wrong level.

ERP System (Q42)

ERP = Enterprise Resource Planning. Software that manages core business processes — inventory, orders, accounting, HR — in one integrated system. Examples: SAP, Oracle.

Roman Urdu: Bara software jo poori company ka kaam ek jagah manage karta hai — stock, orders, payments sab.

Why Q42 matters: Cancelling 40 purchase orders in the ERP is a BULK, IRREVERSIBLE action. Even if the list looks correct, this needs to be checked BEFORE acting — because there's no reviewer afterwards.

Event Calendar / Budget Allocation (Q39)

Event calendar: A plan of which events will happen across the year. Budget allocation: How much money is assigned to each event.

Why Q39 matters: The AI produced a confident, detailed plan allocating LAST YEAR's budget. But the CFO had JUST cut the budget in a meeting — undocumented. The AI had no way to know. It filled the gap with the most likely answer (last year's figures), at full confidence.

Key lesson: Before delegating a task, ask: "Could I put EVERYTHING this step needs into the context window?" If important information exists only in people's heads or undocumented meetings, the AI cannot know it — and won't tell you it doesn't know. It will fill the gap confidently.

SECTION 9: ESCALATION & DEPENDENCY (Q36, Q8, Q14)

Escalation Signal (Q36)

What triggers it: Other teams now DEPEND on something you built. It has become infrastructure.

What "infrastructure" means in this context: Something other people rely on for their own work. If it breaks or is wrong, other teams are affected. It needs:

A named OWNER (someone responsible)
Uptime (it must work reliably)
Access control (who can use it)
Consistency (everyone gets the same behavior)

Roman Urdu: Jab doosri teams aap ke banaye tool par depend karein — ye ab sirf aap ka kaam nahi raha. Kisi engineering team ko dena hoga.

Exam rule: A personal Project that other teams depend on is no longer a personal tool — it's infrastructure. Escalate to engineering.

Purpose Limitation (Q58) ⭐

What it means: Data collected for ONE purpose cannot be used for a DIFFERENT purpose without NEW permission.

Roman Urdu: Jo data ek kaam ke liye liya gaya — woh doosre kaam ke liye use nahi kar sakte bina nai permission ke.

Example (Q58): Customer calls were recorded for QUALITY ASSURANCE (checking service quality). The product team wants to use those SAME recordings to mine for UPSELL OPPORTUNITIES (finding sales leads).

Quality assurance ≠ sales leads. New purpose = need fresh permission from the DATA OWNER.

Route approval ≠ purpose approval. Just because the transcripts are in an approved workspace doesn't mean you can use them for anything you like.

SECTION 10: COMMUNICATING AI TO OTHERS (Q26, Q47)

The Missing Human Checkpoint (Q26) ⭐

When you describe an AI workflow to someone, it's easy to accidentally DROP the human review step from your description.

Example:

WRONG: "AI clears the disputes, cutting the backlog from 10 days to 2."
RIGHT: "AI triages the disputes, a reviewer approves each one, cutting backlog from 10 days to 2."

The first version makes the CFO believe there is NO human gate. The second version is honest.

Exam rule: In any description of an AI workflow — especially to executives — always name WHAT the AI does AND the human checkpoint after it.

Two Audiences, Two Rewrites (Q47)

Same facts, different reader = different document. The exam wants you to rewrite for each reader, NOT polish one version and send it to both.

Audit committee needs: The number, the cause, the decision they must make. Warehouse crew needs: What changed in their process, why it's not their fault, what to do next.

Same analysis. Two completely different rewrites. Both need editing in their own "named pass" (focused editing round).

QUICK REFERENCE CARD — Q21 to Q40

Term	3-Word Summary	Question
Superseded rate card	Old file, wrong prices	Q21
Incognito chat	Memory-free, not trace-free	Q22
Sycophancy	Agrees with your question	Q10, Q22
Context drift	Long chat forgets rules	Q24
Stale configuration	Old files, no error message	Q21, Q52
Under-specification	Prompt was too vague	Q18, Q34
Model ladder	Match cost to complexity	Q28, Q41
Extended thinking	Reasoning setting, not a model	Q28, Q59
Defined gate	Who + what + when	Q23
Accountability cannot transfer	Your name stays on it	Q29, Q37
High-stakes, no-undo	Human presses the button	Q27, Q42
Feature problem	Code, not wording, fixes it	Q35
Description problem	Better wording fixes it	Q35
Escalation signal	Others depend on it now	Q36
Purpose limitation	New use = new permission	Q58
Pseudonymised	Coded but lookup exists	Q38
Anonymised	No trace back possible	Q38
Missing checkpoint	Dropped from the sentence	Q26
CRM	Customer tracking software	Q36
ERP	Business operations software	Q42

Part 3 follows below.

Part 3 of 3 — Features, Validation & Exam Traps (Q41–Q60)

This part covers the AI features the exam tests (research mode, Skills, Connectors), validation frameworks (reach & irreversibility, survivor bias, caveat dropping), and the tricky English phrases that appear in the last 20 questions.

The exam asks you to match the RIGHT feature to the RIGHT task. Get this wrong and you fail even if you understand the concept.

Web Search (Q6, Q46)

What it does: Searches the live internet for current information.

Use when: You need something that happened recently (news, current prices, today's rates).

Roman Urdu: Internet par abhi search karta hai — taaza khabar ya rate ke liye.

Does NOT work for: Internal company documents, files on your shared drive, your Project's knowledge files.

Example (Q6): "What did the port authority announce this morning about the berth closure?" → web search ✅ (it's real-time news)

Research Mode (Q6, Q46, Q51)

What it does: Runs a deeper, multi-source web investigation. Takes longer, produces a cited report from multiple external sources.

Use when: You need a comprehensive, cited comparison or analysis from the public internet. Example: "Compare cold-chain logistics providers in East Africa, with funding and key customers."

Does NOT work for: Internal documents, your company's own policies. Research mode searches the WEB — it cannot find a document sitting in your Google Drive or Project.

The mistake (Q46): A sales associate uses research mode for pricing questions, expecting it to find the firm's internal discount policy. Research mode looks outside the company. The internal policy was never on the internet. Fix = put the policy in a Project's knowledge, ask inside that Project.

Extended Thinking (Q28, Q41, Q59)

Already covered in Part 2, but here's the exam test:

Turn ON when: Multiple inputs, trade-offs, or constraints. Questions like "Which of these options is best given X, Y, and Z constraints?"

Turn OFF for: Simple lookups, quick rewrites, routine conversions, casual idea-generation.

Q59 correct answers:

"Choosing a rollout sequence across 4 warehouses under staffing, downtime, and budget constraints" ✅ (multi-constraint)
"A staffing decision across 3 sites with real consequences, seasonality, and overtime rules" ✅ (multi-variable + real consequences)

Q59 wrong answers:

Converting a bulleted agenda to a calendar invite ❌ (mechanical)
Finding an acronym in a glossary ❌ (simple lookup)
10 quick name ideas for a chat channel ❌ (casual brainstorm)

Skills (Q17, Q52, Q54)

What it is: A pre-built set of instructions that teaches Claude a specific method or task. You install a Skill, and Claude knows how to do that task consistently.

Roman Urdu: Claude ko ek khaas kaam sikhane wala tool — jaise ek recipe book dena.

Key things to know:

A Skill carries NO permissions of its own. If you need Claude to write to your Google Drive, adding a Skill that says "here's how to archive files" does nothing — you still need write access on the Connector. (Q54)
Built-in Skills vs Community-Published Skills (Q17):
- Built-in Skills = made by the platform, reviewed, safer
- Community-published Skills = made by outside developers, may have write access, MUST be reviewed before use
- The rule: Care rises as capability rises. A community Skill with write access to your shared drive is high-capability and must be fully reviewed before enabling.
Installing a Skill = teaching a method, not creating infrastructure. When other teams depend on a result, you need proper engineering ownership — not just a packaged Skill. (Q36)

Connectors (Q17, Q46, Q54, Q57)

What it is: A connection between Claude and an external system — like Google Drive, your CRM, your email. Claude can read from or write to that system through the Connector.

Roman Urdu: Claude ko aapke systems se jodne wala link — jaise Google Drive ya company database.

Read-only vs Write access (Q54): ⭐

	Read-only	Write access
What Claude can do	Look at files, read documents	Create, edit, move, delete files
Starting point	Always start here	Only grant deliberately, after testing
Example	Summarise incoming tenders	Move closed tenders to archive folder

The mistake: Thinking that because Claude CAN see your files, it can also change them. Read access ≠ write access. Write must be granted separately.

Q54 answer: Moving files is a WRITE action. It must be granted deliberately. Also: confirm whether "moving files" is even a capability the Connector offers at all — not all Connectors support every action.

Master Services Agreement (Q41)

What it is: A large legal contract between two companies that sets the terms for ALL future work between them. It covers: payment, liability, ownership of work, confidentiality, dispute resolution.

Roman Urdu: Do companies ke beech ka bara aur detailed contract — sab future kaam ke rules isme hain.

Cross-reference obligations: Some clauses in the contract only make sense when you read them alongside OTHER clauses. They "cross-reference" each other. Example: Clause 7 says "subject to the exceptions in Clause 12." If you only read Clause 7, you miss the exceptions.

Why Q41 matters: A 180-page MSA with cross-referencing obligations is a large, complex document requiring a model that can hold the whole structure in mind at once. Sonnet (even with extended thinking) produces fluent but shallow results — misses obligations spread across sections. Upgrade to Opus.

SECTION 12: PROMPT AUTHORITY & SILENT CHOICES (Q50)

Silent Choices (Q50) ⭐

What it means: Decisions the AI makes during a task that it doesn't tell you about.

Roman Urdu: AI ne koi choice khud bana li — aur aapko bataya bhi nahi.

Q50 scenario: Associate asks Claude for 3 conference venues for a 120-person event in the second week of March, under budget.

Two venues TIED on price → Claude picked one and dropped the other. Said nothing.
Couldn't find a third venue for that week → extended dates to "late March." Said nothing.

Both are DECISIONS. The associate didn't know either happened. You cannot review a decision you don't know was made.

The fix: Add a line to your request that says: "Tell me what you may decide on your own, and say explicitly whenever you make a choice during the task."

This is called authority instruction — telling the AI:

What it may decide alone (e.g., formatting choices)
That it MUST flag when it chooses (e.g., dropping a vendor, changing dates)

Showing the Plan vs. Governing Mid-Run Choices (Q50)

"Show your plan before starting" (Option A in Q50) — This shows the APPROACH before the run begins. It doesn't govern choices that arise DURING the run.

"List all assumptions after finishing" (Option C in Q50) — This requests a post-run list. But a choice the AI didn't register as a decision won't appear on it. Silent choices don't get listed.

Only Option D works: Tell the AI what it may decide alone + that it must say when it chose.

SECTION 13: OUTPUT VALIDATION (Q42, Q47, Q48, Q49)

Reach and Irreversibility (Q42) ⭐

The principle: How much checking you do should match two things:

Reach — How many people see it? (1 person vs. 300 customers vs. the whole public)
Irreversibility — Can you undo it? (Send a follow-up vs. cancel 40 orders vs. publish a press release)

Roman Urdu: Jitna zyada log dekhein ya kaam undo na ho sake — utni zyada checking chahiye.

Task	Reach	Irreversible?	Checking needed
Agenda for your own 1-on-1	1 person (your manager)	Easily corrected	Quick read
Reply to one customer's query	1 person	Reversible (follow-up)	Your full read
Press statement on company website	Public = everyone	Hard to unpublish	Second reviewer required
Cancelling 40 purchase orders in ERP	Affects 40 vendors	Cannot un-cancel	Check BEFORE acting

Caveat Dropping (Q48) ⭐

What it means: A warning or "unverified" flag that existed in the chat disappears when the output is turned into a clean document.

Roman Urdu: Chat mein "verify nahi hua" likha tha — lekin jab spreadsheet bani to woh note gayab ho gaya.

Q48 scenario: Associate builds a vendor comparison. Three figures are marked "unverified" in chat. Then asks Claude to produce it as a spreadsheet. In the spreadsheet, every cell has a clean number — no "unverified" flag.

Why this is dangerous: The procurement head receives a spreadsheet that LOOKS fully verified. They don't know 3 figures are actually uncertain.

The mechanical check: Count the caveats before the final format and after. If the second number is smaller, find what was dropped and put it back.

Exam rule: Deliverables lose caveats. Always check that warnings survive the formatting step.

Survivor Bias in AI Screening (Q49) ⭐

What it means: When AI narrows a large set and humans only review the ones that PASSED — no one checks whether the ones that were EXCLUDED were excluded fairly.

Roman Urdu: AI ne 900 mein se 120 chune — hum ne sirf 120 check kiye. Jo 780 bahar gaye, kisi ne unhe nahi dekha. Kya AI ne koi group galat bahar nikala?

Q49 scenario: Claude pre-screens 900 warranty claims → routes 120 for fast-track approval. The claims lead reviews all 120 carefully, finds them sound. But the most important review step is MISSING:

The missing step: Sample the 780 that were NOT fast-tracked. Check whether a particular group (region, product type, language, customer type) was systematically routed away.

Why this matters: If AI consistently excludes all claims from, say, one region or one product line — those customers never get fast-tracked. But if humans only review the 120 that made it through, they'll never see the pattern.

Exam rule: When AI screens a set, always sample the EXCLUSIONS too, not just the inclusions.

SECTION 14: SHADOW AI & INCIDENT RESPONSE (Q30, Q57)

Shadow AI (Q57) ⭐

What it means: When employees use unapproved AI tools to do their work — usually because the approved tool doesn't work well enough for their actual task.

Roman Urdu: Employees apni marzi se koi AI tool use kar rahe hain jo company ne approve nahi kiya — kyunki approved wala kaam nahi kar raha tha.

Q57 scenario: Field engineers photograph site reports and send the photos to a consumer chatbot on their phones. Approved connector = can't read images. So they found a workaround. Nobody acted in bad faith — they just had a job to do.

The friction rule: Shadow AI grows where the approved path is HARDER than the unapproved path. If the approved route takes 10 steps and the unapproved one takes 1 — people under deadline pressure will take the easy route.

The fix: Reduce the friction in the APPROVED path. Fix the gap (add image reading to the approved connector). Don't just block the consumer app — without fixing the approved path, the problem returns.

What NOT to do:

Block the consumer app without fixing the approved path → engineers still can't do the job
Discipline the engineers → they had no bad intent, and you still haven't fixed the root cause
Relax the policy to allow the consumer app for customer data → lowers the data classification instead of fixing the friction

Incident First Hour (Q30) ⭐

The scenario: A pricing notice with a serious error went out to 300 customers one hour ago.

The three first-hour actions:

Stop the spread — halt further sends, pause any scheduled runs of the same workflow
Record the facts — what went out, to whom, at what time, what exactly was wrong
Report promptly — notify through the organization's incident path, stating facts plainly

What NOT to do:

Delete the chat/draft → may not delete vendor/org records. Turns a mistake into concealment.
Wait until fully understood → for customer-facing errors, timing may matter legally. Don't wait.
Send a correction yourself → notification and remediation belong to the incident OWNER, not you.

Roman Urdu: Galti ho gayi — pehle band karo, phir record karo, phir report karo. Khud fix mat karo — incident owner ko bolo.

SECTION 15: TESTING & GROUND TRUTH (Q56)

Ground Truth Testing (Q56) ⭐

What it means: Testing an AI on work you have ALREADY done and ALREADY know the correct answer for — so you can compare what the AI produces against the known-correct result.

Roman Urdu: AI ko purana kaam do jiska sahi jawab aapko pata hai — phir AI ka jawab us sahi jawab se milao.

Why this is the right way to test (Q56):

"Give Claude this month's new data and see if it looks reasonable" ❌ — No ground truth. "Looks reasonable" is how a polished wrong answer passes.
"Run it twice and adopt the result if both runs agree" ❌ — Agreement between two runs ≠ correct. A consistently wrong method passes both runs.
"Ask Claude to explain its method and judge the method" ❌ — Judges a description, not actual results.
"Give Claude a month whose scorecard the team already signed off, hold back the answer, and compare" ✅ — Known correct answer. Every outcome is informative.

What every outcome tells you:

Claude missed a pattern → improve the instructions
Claude had missing data → data problem to fix
Claude consistently fails one part → that part stays human

SECTION 16: OPTIONS ROUND vs. EXECUTION ROUND (Q55)

The Two Phases of Creative/Planning Work

Phase 1 — Options round (generating ideas):

You want VARIETY and DIVERGENCE — many different options
Loose the output: Don't specify the format, template, word count, style
Load the SITUATION fully (audience, goals, context)
Leave the OUTPUT description open

Phase 2 — Execution round (producing the final thing):

You want PRECISION — exactly the right format, length, tone
Tight the output: Specify format, template, sections, word count
This is when constraints help

Q55 mistake: The associate asked for "5 different formats" for a leadership program launch. But they fixed: the slide template, two mandatory sections, a 60-word limit, and the house tone. These constraints pinned the OUTPUT shape so tightly that only VARIANTS of one format were possible — not genuinely different options.

Exam rule: Precision is the right setting for EXECUTION. It's the wrong setting for GENERATING OPTIONS.

Roman Urdu: Options chahiye hein to output loose rakho — sab kuch mat fix karo. Jab final banana ho tab tight karo.

SECTION 17: BUILDING A LOOP — DESCRIPTION vs. FEATURE PROBLEM (Q35)

Already introduced in Part 2. Here's the full exam pattern:

How to identify a description problem:

One more clear attempt improves the result
The error is about ambiguity in what you asked

How to identify a feature problem:

You've tried multiple times, with increasingly clear descriptions
SAME error persists, everything else improves
The task involves repeatable calculation, structured data manipulation, or precise logic

When you hit a feature problem, the fix is NOT:

A more capable model (Q35 Option C) — more capability won't fix a prose-vs-code mismatch
More context in the chat (Q35 Option D) — general degradation ≠ one column wrong each round

The fix IS: Use the right feature. Calculations → code execution. Structured table → structured output. Lookup → connector to the right source.

SECTION 18: KEY ENGLISH PHRASES FROM Q41–Q60

These phrases appear in the last 20 questions and are easy to misunderstand:

Phrase	Simple meaning	Question
"fluent but shallow"	The answer reads well, sounds correct, but doesn't go deep enough — it misses the important complexity	Q41
"cross-reference"	One part of a document refers to another part. You need to read BOTH to understand either.	Q41
"ground truth"	The known-correct answer you compare against	Q56
"signed off"	Officially approved by the appropriate person	Q56
"hold back the answer"	Don't show the AI the correct result — test if it can produce it independently	Q56
"silent choice"	A decision the AI made without telling you	Q50
"authority" (in prompting)	What the AI is permitted to decide on its own	Q50
"shadow AI"	Using unapproved AI tools at work	Q57
"friction"	Small repeated difficulty or annoyance that makes people avoid the proper path	Q57
"approved path"	The officially allowed method for doing something	Q57
"consumer chatbot"	A general-purpose AI product for the public (like ChatGPT free, Claude.ai free) — NOT an enterprise/business-approved tool	Q57
"purpose limitation"	Data can only be used for the purpose it was collected for	Q58
"data owner"	The person responsible for a dataset — who decides how it can be used	Q58
"upsell"	Selling additional or more expensive products to an existing customer	Q58
"vendor shortlist"	A reduced list of vendors (from a larger list) that made it to the final consideration stage	Q60
"steering committee"	A senior group that provides direction and makes key decisions on a project	Q60
"defensible"	Can be defended against criticism — a strong enough choice that you can justify it	Q60
"neutral framing"	Asking a question without hinting at the answer you want	Q60
"leading framing"	A question that signals the conclusion it wants — triggers sycophancy	Q10, Q60
"fixed scale"	A specific rating system with defined numbers (e.g., 1–10 on these 3 criteria) — forces real assessment	Q60
"bait verb"	A word in a question that tells Claude what conclusion to reach. "Defend," "prove," "isn't this good?" are bait verbs.	Q60
"returns process"	The system for handling when customers send products back	Q51
"map the ground"	Survey what's already known before drafting or deciding	Q51
"survivors" / "survivor bias"	Reviewing only the ones that passed a filter — missing what was excluded	Q49
"read-only"	Permission to view but not change	Q54
"write access"	Permission to create, edit, move, or delete	Q54
"early-payment discount"	A reduction in price if you pay an invoice before its due date	Q45
"accounts-payable"	The department that pays bills the company owes to others	Q45
"fast-track"	A shortened, expedited process — skipping some normal steps	Q49
"pre-screen"	An initial filter before the main review — narrows a large set down	Q49
"triage"	Sorting and prioritizing a set of items by urgency or importance	Q26, Q49

SECTION 19: INDUSTRIES IN Q41–Q60

Insurance (Q49)

Warranty claim: When a customer's product breaks within the warranty period and they ask the manufacturer/seller to fix or replace it for free. Roman Urdu: Product ki guarantee period mein kharab hua — customer ne free repair/replace manga.

Policy terms: The rules in an insurance/warranty contract that define what IS and is NOT covered.

Pharmacy / Healthcare Retail (Q48)

Pharmacy chain: A company that runs multiple pharmacies (medicine shops). Vendor comparison: Comparing different suppliers on price, delivery time, coverage area — to decide which one to buy from. Appears in Q48: Associate builds a vendor comparison in chat, three figures are "unverified."

Property / Real Estate (Q50)

Offsite: A meeting or event held AWAY from the regular office — usually for strategy, teambuilding, or planning. AV (Audio-Visual): Equipment for presentations — projectors, microphones, screens. "On-site AV" means the venue has this equipment already. Conference venue: A location rented for holding meetings or events. Appears in Q50: "Shortlist 3 venues with on-site AV for a 120-person offsite in the second week of March."

Logistics / Warehouse (Q53, Q57)

Branch stock-level report: A weekly report showing how much inventory (stock) each branch/warehouse has. Appears in Q53: A scheduled agent task produces this every Monday.

Field engineers: Technicians who work on physical sites — not in an office, but at client locations, warehouses, or in the field. Appears in Q57: Field engineers photograph site reports and send them to a consumer chatbot.

Site reports: Documents written after visiting a physical location — recording what was inspected, what problems were found, what work was done.

Procurement / Supply Chain (Q56)

Supplier scorecard: Already covered in Part 2. Signed-off month: A past period whose scorecard the team already reviewed, corrected, and officially approved. The "known-correct" answer for testing.

FINAL MASTER REFERENCE — ALL 60 QUESTIONS

The 10 Most Tested Concepts Across All 60 Questions

#	Concept	Questions
1	Verify against primary/source	Q1, Q2, Q4, Q9, Q43
2	Accountability stays with the human	Q15, Q26, Q27, Q29, Q37, Q42
3	Defined gate (who + what + when)	Q11, Q23, Q45
4	Match feature to task	Q6, Q13, Q35, Q46, Q59
5	Stale config vs hallucination	Q21, Q33, Q52
6	Context window = what AI can see	Q3, Q24, Q39, Q40
7	Sycophancy & neutral framing	Q10, Q55, Q60
8	Escalate, don't guess	Q8, Q14, Q36, Q58
9	Shadow AI / friction rule	Q17, Q57
10	Sample the exclusions	Q11, Q49

The 5 Exam Answer Traps (Options That ALWAYS Sound Right But Are Wrong)

Trap Answer	Why It Sounds Right	Why It's Wrong
"Ask Claude to rate its own confidence"	Sounds like self-checking	Claude is confident about wrong answers too
"Switch to Opus"	Sounds like using the best tool	First fix cheap causes (prompt, config) before upgrading model
"Use incognito mode to keep it private"	Sounds privacy-conscious	Incognito = no history/memory. Not = no vendor record
"A human will review it" (vague)	Sounds like oversight	Unless WHO + WHAT + WHEN is named, it's an unstaffed gate
"The track record proves it's reliable"	Sounds like evidence	Track record proves reliability for THAT task, not for a new, higher-stakes decision

Parts 1–3 cover the 60 mock exam questions. Sections 20–22 below cover the REAL EXAM and hard practice questions — vocabulary the mock exam didn't test.

SECTION 20: FRAMEWORK VOCABULARY — The Words Inside Record Documents

Why this matters: The real exam uses governance records, delegation records, and delegation maps in PARAGRAPH form. You must instantly recognise the section names and what each one tests — because the question will ask "which SECTION has the most serious gap?" If you don't know the sections, you can't answer.

The Governance Record — 5 Sections

A governance record is the document a team writes to get approval for an AI workflow. It has 5 sections. Each section answers a different question. Each section can have its own type of gap.

Section	What It Answers	What a "Gap" Looks Like
Case	Is this task appropriate for AI? Classification = appropriate / appropriate with review / inappropriate. Contains: deciding factor + consequence of error + gate definition	Wrong classification (says "appropriate" when consequences are serious), vague gate ("a person reviews" without WHO/WHAT/WHEN), understated consequence
Data	What data goes in, how sensitive, what route? Contains: data tier (green/yellow/red) + does the task need identifiers? + approved route	Wrong tier, identifiers included when not needed (data minimisation failure), route not confirmed for this PURPOSE
Capability	What tools does the workflow use and how far can they reach? Contains: connectors (read/write), skills (built-in or community), MCP integrations	Unreviewed community skill or MCP integration, write access not justified, skill + connector = uncontrolled reach
People	Who is affected and is it fair? Contains: affected parties + fairness assessment + disclosure assessment	No fairness check when ranking/screening people, no disclosure to affected parties that AI is involved
Evidence	How do we know it's working? Contains: success measure + review sample rate + review schedule	Sample rate too low for the volume, no review schedule, checking the wrong thing

Roman Urdu: Governance record = AI workflow ko approve karwane ka document. 5 hisse hain — har hissa alag cheez check karta hai.

Classification Levels — The Three Labels

Level	Meaning	When to Use
Appropriate	Low stakes, routine, reversible. Minimal human oversight needed.	Internal formatting, brainstorming, simple data organisation
Appropriate with review	Meaningful consequences if wrong. A DEFINED human gate is required.	Client-facing reports, financial summaries, regulatory documents, anything where errors cause real harm
Inappropriate	The task should not involve AI at all.	Final legal determinations, medical diagnoses, decisions about people's rights where accountability cannot be maintained

Roman Urdu: Appropriate = AI kar sakta hai. Appropriate with review = AI kare lekin insaan check kare. Inappropriate = AI se na karwao.

Exam trap: The question often shows a workflow classified as "appropriate" but the consequence section describes serious business impact → the classification should be "appropriate with review."

Deciding Factor

The reason the classification was chosen. Common deciding factors:

Reversibility — "a draft can be corrected before sending" → supports "appropriate with review"
Accountability — "the company is answerable for the accuracy" → supports "appropriate with review"
Routine nature — "the task is templated and mechanical" → supports "appropriate" (but watch for understated consequences)

Defined Gate vs Vague Gate ⭐

	Vague Gate (BAD)	Defined Gate (GOOD)
Example	"A person reviews the report before sending"	"The compliance manager reviews each report before sending, checking that all regulatory citations match the current version of the guidance and that no commitment exceeds the pre-approved threshold"
WHO	"A person" — no named role	"The compliance manager" — specific role
WHAT	"Reviews" — reviews for what?	"Checking that regulatory citations match current guidance and commitments don't exceed thresholds"
WHEN	Not stated	"Before sending" — before the irreversible step

Roman Urdu: Vague gate = "koi check karega." Defined gate = "Compliance manager bhejne se pehle check karega ke regulatory references sahi hain aur limits cross nahi hui."

Data Tiers

Tier	Colour	Examples	What to Do
Green	Low sensitivity	Public information, marketing content, general business processes	Standard route
Yellow	Medium sensitivity	Internal business data, commercial figures, employee work data (not PII), building addresses	Enterprise workspace, confirm route
Red	High sensitivity	PII, health data (PHI), financial account numbers, SSNs, student records, credit card data	Strict route approval, data minimisation, may need anonymisation

The Discriminator Question ⭐

"Does the task NEED the identifiers?"

This is the key question for the Data section. Before uploading sensitive data, ask: does the AI actually need to know WHO the data belongs to?

Trend analysis on expenses → NO, remove names and account numbers
Personalised email to a named client → YES, the email must address them by name
Reconciliation of intercompany loans → YES, you need to know which entity holds which balance

Roman Urdu: "Kya AI ko naam ya pehchaan wali cheez chahiye is kaam ke liye?" Agar nahi — to hata do.

If YES → confirm the route is approved for this data AND this purpose. If NO → remove identifiers, then proceed.

Route Approval vs Purpose Approval ⭐

Route approval: "This workspace is approved for yellow-tier data." Purpose approval: "This workspace is approved for yellow-tier data FOR reconciliation AND reporting to the controller."

Route approval alone is NOT enough. The specific PURPOSE must also be approved.

Roman Urdu: Workspace approved hai matlab sirf jagah theek hai. Lekin kaam bhi approved hona chahiye — sirf jagah theek hona kaafi nahi.

Exam trap: A question says "the enterprise workspace is approved for this data classification." The answer that says "proceed because the route is approved" is WRONG — it skips the purpose check.

The Delegation Record — 4 Sections

A delegation record documents a specific AI task being delegated. It tracks what was asked, what happened, what was checked, and what was decided.

Section	What It Contains
Brief	6 questions: Outcome (what result?), Context (what background?), Constraints (what rules?), Authority (what can AI decide alone?), Deliverable (what format?), Verification (how will we check?)
Run	What actually happened when the AI did the task
Check	What the human reviewed and what they found
Decision	What action was taken based on the check

Roman Urdu: Delegation record = AI ko kaam dene ka record. Brief = kya chahiye. Run = kya hua. Check = insaan ne kya dekha. Decision = kya kiya.

Exam pattern: The question gives you the full record in paragraph form and asks which section has a gap. Usually the Brief's Verification is too narrow (checks one thing but misses the risk the Constraints identified), or the Check doesn't match the Brief's Constraints.

The Brief's 6 Questions

Question	What It Sets	Common Gap
Outcome	What result do you want?	Too vague — "a report" instead of "a comparison report with three alternatives and a recommendation"
Context	What background does the AI need?	Missing context that only exists in people's heads (undocumented decisions, meeting notes)
Constraints	What rules must be followed?	Vague constraints — "comply with FCA rules" instead of specifying WHICH rules
Authority	What can AI decide alone? What must it flag?	Not stated — leads to silent choices
Deliverable	What format?	Mismatch — options round given execution constraints (pinned too tight)
Verification	How will we check the output?	Checks the wrong thing — verifies figures but not whether the narrative overstates

The Delegation Map — Workflow Steps

A delegation map shows every step in a workflow, with:

Owner — who does this step? (AI / Human / Collaborative)
Criterion — what makes this step safe at this ownership level? (reversibility, stakes, accountability)
Carried by — what tool/feature/connector does this step use?
Gates — where humans check before an irreversible action

Plus: Baseline (the before-change measurement) to prove improvement.

Roman Urdu: Delegation map = workflow ka naqsha — har qadam par kaun karta hai, kyun safe hai, kya tool use hota hai.

The Three Mapping Errors ⭐⭐⭐

These are the three specific mistakes people make when building a delegation map. The exam WILL test these. You MUST know the trigger words.

1. Halo Delegation

What it is: Giving AI ownership of a NEW, HARDER task because it performed well on a DIFFERENT, EASIER task.

Trigger words: "has been accurate for X months," "track record," "reliable so far," "proven itself"

Roman Urdu: AI ne aasan kaam acha kiya — to mushkil kaam bhi de diya. Ye galat hai. Pichla kaam alag tha.

Example: A classification skill has been accurate on standard voyages for 3 months → the team gives it full ownership of sanctions screening. Sanctions screening is a different, higher-stakes task. Past accuracy on easy classifications doesn't prove competence on sanctions.

The key insight: The competence was REAL — but it belonged to DIFFERENT (easier) work. Real past performance ≠ earned trust for new higher-stakes work.

How to spot it in a question: Look for a step where AI got full ownership, and the JUSTIFICATION is past performance on different work.

2. Unstaffed Gate

What it is: A human review step that EXISTS on paper but has NO real content — the WHO, WHAT, or WHEN is missing, so nobody actually checks anything meaningful.

Trigger words: "a person reviews," "will be checked," "reviewed before sending" — with no specifics about WHO, WHAT they check, or WHEN

Roman Urdu: Kaagaz par likha hai "insaan check karega" — lekin kaun? Kya check karega? Kab? Kuch define nahi. Gate khali hai.

Example: "The report will be reviewed by a person before it is sent to the client." — Sounds responsible. But WHO reviews? WHAT do they verify? The gate is a sentence, not a process.

How to spot it: The gate sentence uses passive voice ("will be reviewed") and doesn't name a role, a check, or a timing.

IMPORTANT DISTINCTION from halo delegation: In halo delegation, the gate was removed because AI "earned" full ownership. In unstaffed gate, the gate EXISTS but is EMPTY — it was never properly defined.

3. Mapping the Tool

What it is: Confusing what the AI PRODUCES (text, tables, numbers) with what the STEP REQUIRES (a judgment, a decision, a regulated assessment).

Trigger words: "draft a recommendation," "propose corrective actions," "explain why X is suitable" — where the output looks like text but the step requires professional judgment

Roman Urdu: AI se text likhwana aur AI se faisla karwana — ye do alag cheezein hain. Agar step mein faisla chahiye lekin AI se sirf text liya — ye mapping the tool hai.

Example: Step 3 says "draft a suitability narrative explaining why the allocation is consistent with the client's objectives." This LOOKS like text production. But suitability is a REGULATED JUDGMENT — the adviser must own it, not rubber-stamp AI-generated reasoning.

How to spot it: The step asks AI to "recommend," "propose," "explain why," or "justify" in a domain that requires human professional judgment. The model can produce text that LOOKS like a recommendation — but the step needs the JUDGMENT, not the text.

IMPORTANT DISTINCTION:

AI drafting a narrative about what happened = OK (text production)
AI drafting a narrative about what to DO about it = mapping the tool (judgment required)

Quick Decision — Which Mapping Error?

If the question says...	It's probably...
"AI has been accurate for X months, so we gave it full control of Y"	Halo delegation
"A person reviews before sending" (no WHO/WHAT/WHEN)	Unstaffed gate
"AI drafts a recommendation / proposes actions / explains why X is suitable" (regulated domain)	Mapping the tool
Human gate exists but the person stopped checking because results have been good	Halo delegation (the human became a rubber stamp due to past performance)

Four Diagnostic Timing Signatures ⭐⭐⭐

When a workflow breaks, the TIMING of the break tells you the cause. Memorise these four.

Signature	Timing Clue	Key Phrase	Fix
Under-specification	Wrong from the FIRST run, in a FRESH chat	"It was never right"	The prompt is missing information — add the missing context
Context overload	Gets worse DURING a single long session	"It was fine at first but degraded as the chat went on"	Summarise, correct, start fresh chat
Wrong feature/model	ONE specific repeatable error, everything else works	"Same mistake every time, even when I rewrite the prompt"	Switch to the right feature (e.g., code execution for calculations)
Stale configuration	USED TO WORK, now doesn't. Nothing in the prompt changed.	"Worked for months, then broke"	Check knowledge files, standing instructions, connected data sources for outdated content

Roman Urdu:

Under-specification = pehle din se galat — prompt mein kuch missing tha
Context overload = chat lambi ho gayi, ab bhool raha hai
Wrong feature = ek hi galti baar baar — feature galat hai
Stale configuration = pehle sahi tha, ab galat — files purani ho gayi hain

Exam trap: "Stale configuration" and "context overload" are the two most commonly confused. Remember:

Context overload = WITHIN a session (the chat grew too long)
Stale configuration = ACROSS sessions (worked last month, broken this month, in fresh chats too)

Other Framework Terms

Scope Dial A control that limits how much access a tool has. Skills have NO scope dial — a Skill runs with whatever access the session has. If a session has a connector to your database, the Skill can use it. This is why unreviewed community Skills are dangerous.

Roman Urdu: Scope dial = access ki limit. Skills mein ye hota hi nahi — Skill ko session ke saare tools mil jaate hain.

Progressive Disclosure Showing information gradually — simple first, details later. In the context of communicating AI value to stakeholders: start with the outcome, then add how the human gate works, then add how the AI feature works. Don't dump technical details on a CFO.

Data Minimisation Only include the data the task actually needs. Remove everything else before upload. Even if the route is approved for that data — if the task doesn't need it, don't include it.

Roman Urdu: Sirf wahi data do jo kaam ke liye zaruri hai — baaki hata do.

Disclosure Assessment Checking whether the people affected by an AI workflow need to be TOLD that AI is involved. A booking confirmation that's AI-generated — should the guest know? A compliance report sent to regulators — should the regulator know?

Fairness Assessment Checking whether an AI workflow treats different groups of people equally. Required when AI ranks, screens, selects, or evaluates people. NOT required for routine individual communications (like booking confirmations).

SECTION 21: INDUSTRY VOCABULARY — Words From the Real Exam and Hard Practice Questions

These terms appeared in the first exam attempt, in hard practice questions, or are highly likely to appear. The mock exam didn't test most of these.

Insurance & Underwriting

Term	Simple Meaning	Roman Urdu
Underwriting	The process of evaluating risk to decide whether to insure someone and at what price	Insurance mein risk check karna — bima dein ya nahi, kitne mein dein
Premium	The amount a customer pays for insurance coverage (monthly or annually)	Bima ki keemat — jo customer deta hai
Premium adjustment	Changing the premium amount (up or down) based on new information or renewal	Bima ki keemat mein tabdeeli
Policy renewal	When an insurance policy reaches its end date and is extended for another period	Bima policy ka time khatam hua — agle saal ke liye nai
Rate adequacy	Whether the premium charged is enough to cover the expected claims + costs	Kya price itni hai ke claims cover ho sakein
Risk appetite	How much risk a company is willing to accept	Company kitna risk lene ko tayyar hai
Coverage endorsement	An add-on to an insurance policy that changes or extends what's covered	Policy mein kuch aur shamil karna — extra cover
Claims ratio	The percentage of premiums that the company pays out in claims	Kitna paisa claims mein gaya vs kitna aaya
High-value claim	A claim for a large amount of money — usually needs senior review	Bada claim — senior ko dekhna padta hai

Shipping, Freight & Customs

Term	Simple Meaning	Roman Urdu
Customs declaration	Official form declaring what goods you're importing/exporting, their value, and the taxes owed	Government ko batana ke kya maal la rahe ho aur kitni qeemat hai
Duty rates	The tax percentages charged on imported goods, varying by product category	Import tax ka rate — har cheez ka alag
Transhipment	Moving goods from one ship/vehicle to another at an intermediate port before final delivery	Maal ek jahaz se doosre mein badalna raste mein
Sanctions-listed port	A port in a country under international trade sanctions — trading there has legal restrictions	Wo bandargah jahan se trade karna qanoonan mushkil ya mana hai
Consignee	The person or company receiving the shipped goods	Maal paane wala — jisko delivery honi hai
Voyage estimate	A cost estimate for shipping goods on a specific route	Ek shipment ka andaza lagaya hua kharcha
Carrier-performance scorecard	A report rating each shipping company on delivery times, damage rates, etc.	Shipping companies ki marks sheet
Service-level agreement (SLA)	A contractual promise about delivery speed, quality, or availability	Contract mein likha wa'ada — itni speed, itni quality denge
Transit variance	The difference between expected and actual delivery time	Delivery expected time se kitni der se aayi

Real Estate, Property & Construction

Term	Simple Meaning	Roman Urdu
EPC rating (Energy Performance Certificate)	A rating (A to G) showing how energy-efficient a building is — required by law for sales/rentals in the UK	Building ki energy efficiency ki grading — A se G
Planning permission	Government approval to build something or change a building's use	Government se construction ya tabdeeli ki ijazat
Surveyor report	A professional inspection document describing a property's condition	Property ka professional inspection report
Parking standards	Rules about how many parking spaces a building must have (different for residential vs commercial)	Parking ki jagah ke rules — residential aur commercial ke alag
Lettings	British English for "rentals" — a lettings firm rents out properties	Kiraya — property kiraye par dena
Deposit forfeiture	When a tenant loses their security deposit (e.g., for property damage)	Kirayedaar ki security deposit zabt hona
Lease classification	Categorising a lease agreement under accounting rules (e.g., IFRS 16) — determines how it appears on financial statements	Lease ko accounting rules ke hisaab se category dena

Finance, Accounting & Pensions

Term	Simple Meaning	Roman Urdu
Intercompany loan	A loan between two companies that are part of the same group	Ek group ki do companies ke beech ka qarz
Subsidiary	A smaller company owned or controlled by a larger parent company	Badi company ki chhoti company
Consolidation	Combining the financial reports of all subsidiaries into one group report	Saari chhoti companies ke accounts milana ek mein
Actuarial valuation	A mathematical calculation by an actuary to determine the value of pension obligations or insurance reserves	Pension ya insurance ke future paison ka hisaab — mathematician karta hai
Pension projection	An estimate of how much retirement income a person will receive based on current contributions	Retirement mein kitna paisa milega ka andaza
Suitability report	A regulated document explaining why a specific investment recommendation is appropriate for a specific client	Client ke liye ye investment kyun sahi hai — regulated document
Mean-variance optimisation	A mathematical method for choosing the best mix of investments based on expected return and risk	Investment mix chun'ne ka math — return aur risk ka balance
Accounts payable	The department that pays bills the company owes to suppliers	Suppliers ko paisa dene wala department
Early-payment discount	A price reduction offered if you pay an invoice before its due date	Jaldi pay karo to discount milta hai
CT600	UK corporation tax return form — has specific filing deadlines	UK company tax ka form
IFRS 16	An international accounting standard for how leases are reported in financial statements	Lease ko accounts mein kaise dikhana hai ka international rule
Break clause	A clause in a lease allowing either party to end the lease early under specific conditions	Lease mein jaldi exit karne ka option

Legal

Term	Simple Meaning	Roman Urdu
Legal professional privilege	Communications between lawyer and client that are legally protected — cannot be forced to disclose	Vakeel aur client ki baat jo qanoonan secret hai
Privilege waiver	When privileged communications lose their protection — usually by accidental disclosure	Jab secret protection galti se khatam ho jaye
Document review / Discovery	In litigation, the process of reviewing thousands of documents to find relevant evidence	Case mein hazaron documents check karna — evidence dhundhna
Fair presentation	A regulatory requirement that financial documents present information honestly, not misleadingly	Financial documents mein sach sach batana — gumrah na karna
Regulatory submission	A document officially sent to a regulator as part of a compliance or approval process	Regulator ko bheja jaane wala official document
Redlined version	A document showing changes marked in red — what was added, removed, or modified	Document jismein tabdeeliyan red mein dikhai gayi hain

Healthcare & Pharmacy

Term	Simple Meaning	Roman Urdu
Medication adherence	Whether patients take their prescribed medicine correctly and on schedule	Kya patient sahi time par dawai le raha hai
Dispensing system	The pharmacy's system that tracks which medicines are given to which patients	Pharmacy ka system — kaun si dawai kisko di gayi
Chronic-disease management	Ongoing care programs for long-term conditions (diabetes, heart disease, asthma)	Lambi bimari ka regular ilaaj program
Clinical-trial protocol	The detailed plan for how a medical research study will be conducted	Medical research ka detailed plan
Dosage schedule	When and how much medicine a patient should take	Dawai ka time aur miqdar

Hospitality & Events

Term	Simple Meaning	Roman Urdu
Occupancy rate	The percentage of hotel rooms that are booked/filled at any given time	Hotel mein kitne rooms bhare hain — percentage
Net Promoter Score (NPS)	A score (-100 to +100) measuring how likely customers are to recommend a business	Customer kitna khush hai ke doosron ko bataega
Revenue management	Setting prices dynamically based on demand, season, and competition	Demand ke hisaab se price set karna
Cancellation policy	The rules about what happens when a booking is cancelled — fees, deadlines	Booking cancel karne ke rules — kab tak free hai, kab penalty

HR & Recruitment

Term	Simple Meaning	Roman Urdu
Competency framework	A document listing the skills, behaviours, and knowledge required for each role	Har role ke liye kya skills chahiye ka list
Engagement letter	A formal letter confirming the terms of a professional service agreement	Professional service shuru karne ka formal agreement letter
Benefits renewal	The annual process of reviewing and renewing employee insurance and benefits packages	Employees ki insurance aur benefits ka saalana renewal
Census data (benefits context)	Employee data used for benefits calculations — names, DOBs, salary bands, dependants	Benefits ke liye employees ka data — naam, age, salary, family

SECTION 22: HARD ENGLISH WORDS — Words That Cost You Time on the Exam

These are NOT business terms or AI terms. They're English words that appear in exam questions and options. If you don't know them INSTANTLY, they cost you 10-30 seconds per question — which is the difference between finishing and leaving 30 blank.

Words About Accuracy & Errors

Word	Simple Meaning	Roman Urdu	Example in Exam
Diverge / Diverged	Moved away from the expected or correct value	Sahi qeemat se hat gaya	"Premiums diverged from the rate model by 15%"
Discrepancy	A difference or mismatch between two things that should match	Do cheezon mein farq jo same honi chahiye thi	"Identify discrepancies between the two ledgers"
Fabricate / Fabricated	Made up, invented (NOT real)	Banawati — asli nahi	"The model fabricated a citation that doesn't exist"
Conflate / Conflated	Mixed up two different things as if they were one	Do alag cheezon ko ek samajh lena	"The summary conflated two interviewees' statements"
Overstate / Understated	Exaggerated / downplayed — said more/less than the truth	Bara/chhota bata diya asliyat se	"The risk was understated in the governance record"
Substantiate	Prove with evidence	Saboot se saabit karna	"Can you substantiate this claim with a source?"
Corroborate	Confirm by comparing with another source	Doosre source se tasdeeq karna	"Corroborate the figure against the original report"
Calibrated	Carefully adjusted to be accurate and reliable	Sahi tarha set kiya hua	"Confidence is NOT calibrated — it's a learned style"

Words About Actions & Consequences

Word	Simple Meaning	Roman Urdu	Example in Exam
Irreversible	Cannot be undone	Wapas nahi ho sakta	"Sending the email is an irreversible action"
Inadvertently	Accidentally, without meaning to	Galti se, bina iraade ke	"The firm may inadvertently waive privilege"
Premature	Too early, before the right time	Waqt se pehle	"The model upgrade is premature"
Mitigate	Reduce the harm or risk (not eliminate)	Nuqsan kam karna	"Anonymisation mitigates the privacy risk"
Remediation	The process of fixing a problem after it happened	Masla hone ke baad usse theek karna	"Remediation belongs to the incident owner"
Concealment	Deliberately hiding something	Jaanbujh kar chhupana	"Deleting the chat turns a mistake into concealment"
Contravene	Go against, violate (a rule or law)	Rule todna	"This action contravenes the data policy"
Stipulate	Require something as a formal condition	Shart rakhna	"The contract stipulates a 48-hour notice period"
Waive / Waiver	Give up a right or protection (often accidentally)	Haq chhod dena — aksar galti se	"Failing to withhold the document may waive privilege"
Forfeit / Forfeiture	Lose something as a penalty	Saza ke tor par kho dena	"The deposit was forfeited due to property damage"

Words About Assessment & Review

Word	Simple Meaning	Roman Urdu	Example in Exam
Audit / Audited	An official check or inspection of records/processes	Accounts ya processes ki official checking	"The MCP integration has not been audited"
Scrutiny	Very careful, detailed examination	Bohut ghoor se dekhna	"Review the rest with heightened scrutiny"
Posture	Overall stance or position (not physical)	Overall position ya halat	"Weakens the entire governance posture"
Adequate / Inadequate	Enough / not enough	Kaafi / na-kaafi	"Is this de-identification adequate under HIPAA?"
Compliance / Compliant	Following the rules	Rules ke mutabiq	"The report must be compliant with FCA rules"
Attestation	A formal sworn statement that something is true	Qasam ke saath sach kehna officially	"The regulatory attestation pack"
Defensible	Can be justified if challenged	Agar sawaal ho to justify kar sako	"A defensible vendor recommendation"
Articulate	(adjective) Well-expressed, clear; (verb) to express clearly	Saaf aur acha tareeqe se bayan kiya hua	"The briefing was articulate but contained a fabrication"

Words About Data & Processing

Word	Simple Meaning	Roman Urdu	Example in Exam
Ingest / Ingested	To take in, input (data going into a system)	Data system mein daalna	"Step 1 ingests shipment data from the connector"
Normalise	Make consistent — convert different formats to one standard	Alag alag formats ko ek standard mein laana	"Normalise delivery timestamps against SLAs"
De-identification	Removing identifying information from data so individuals can't be recognised	Pehchaan wali cheezein hata dena data se	"Confirm adequate de-identification before proceeding"
Aggregate / Aggregated	Combined into summaries (individual data rolled up)	Alag alag data ko jama karke summary banana	"Aggregated figures don't reveal individual records"
Systematic	Happening in a pattern, not random	Pattern mein ho raha hai, random nahi	"Check whether a group was systematically excluded"
Deteriorated	Got worse over time	Waqt ke saath kharab hua	"Three carriers whose performance deteriorated"

Words About Workflow & Authority

Word	Simple Meaning	Roman Urdu	Example in Exam
Escalate	Send a problem up to someone with more authority or expertise	Masla upar bhejo — jisko zyada expertise ya authority hai	"Escalate the legal question to the compliance owner"
Delegate	Give a task to someone (or something) else to do	Kaam kisi aur ko de dena	"Delegate the extraction to AI"
Collaborative	Shared between human and AI — both contribute	Mil kar karna — insaan aur AI dono	"Step 4 is collaborative — the adviser reviews and signs"
Override	Cancel or replace a previous instruction or decision	Pehle wala faisla badal dena	"The manager can override the AI's classification"
Trigger	An event that starts a process or action	Woh cheez jo koi process shuru kare	"The threshold triggers a senior review"
Mandate / Mandated	Required by rule or law (not optional)	Zaruri hai — rule ke mutabiq	"The legal footer is mandated by regulation"

Phrases That Change Meaning in Exam Context

Phrase	What You Might Think	What It ACTUALLY Means in Exam Context
"with reach"	Going far / having influence	Affecting many people or external parties — increases the stakes of an irreversible action
"functionally absent"	Kind of missing	EXISTS on paper but has NO real effect — like a 0.2% sample rate on 12,000 statements
"best-effort"	Trying hard	NOT guaranteed — specifically about memory: Claude TRIES to remember but it's not configuration
"source of truth"	The real version	The AUTHORITATIVE, CURRENT version of a document — not a copy, not a summary, not last year's
"at the point of no return"	Dramatic moment	The moment before an irreversible action — this is WHERE the human gate must be
"rubber-stamp"	Quick approval	Approving WITHOUT checking — this IS an unstaffed gate or halo delegation
"fit for purpose"	Good enough	Suitable for THIS specific use — a tool that's fit for one purpose is NOT automatically fit for another
"carries clinical risk"	Medical danger	Even if it's "just a reminder" — wrong medication or dosage in a reminder can cause real harm
"uncontrolled reach"	Not limited access	A tool (skill, MCP) that can access MORE than it needs because there's no scope dial
"data subject"	Person in the data	The PERSON whose data is being processed — they have rights

Quick Self-Test — 20 Words, 60 Seconds

Read the word. Say the meaning OUT LOUD. If you hesitate for more than 2 seconds, mark it for review.

Diverged
Inadvertently
Waive
Substantiate
Conflated
Irreversible
Mitigate
Ingest
Attestation
Premature
Contravene
Scrutiny
Forfeit
Remediation
Deteriorated
Normalise
Concealment
Stipulate
De-identification
Corroborate

If you got all 20 in under 60 seconds without hesitation — you're ready. If any took more than 2 seconds — review that word until it's instant.

Study order:

Read business-concepts-guide.md ONCE — your foundation

Master Mock-exam-vocabulary-guide.md — keep it open during all practice

Section 20 (frameworks) is the HIGHEST priority addition — it directly maps to the questions you got wrong

Section 21 (industry terms) — skim before each practice set

Section 22 (hard English) — do the self-test daily until all 20 are instant

Exam Vocabulary Guide

tags: [ccao-f, exam-prep, vocabulary, mock-exam] parts: 1-of-3, 2-of-3 covers: "Questions 1–40"Copy Anchor Link

Mock Exam Vocabulary GuideCopy Anchor Link

Part 1 of 3 — Business Roles, Industries & Hard Terms (Q1–Q20)Copy Anchor Link

SECTION 1: JOB TITLES & ROLES (Who Is In These Questions?)Copy Anchor Link

1. Junior AnalystCopy Anchor Link

2. Finance AssociateCopy Anchor Link

3. Operations Team / Operations AssociateCopy Anchor Link

4. Compliance Analyst / Compliance Team / Compliance OwnerCopy Anchor Link

5. Property ManagerCopy Anchor Link

6. Policy OwnerCopy Anchor Link

7. CFO (Chief Financial Officer)Copy Anchor Link

8. Facilities Associate / Facilities AnalystCopy Anchor Link

9. Procurement Associate / Procurement ManagerCopy Anchor Link

10. Legal-Operations AssociateCopy Anchor Link

11. Lettings Firm (and a Lettings Associate)Copy Anchor Link

12. Claims Associate / Claims TeamCopy Anchor Link

13. Strategy AssociateCopy Anchor Link

14. Learning-and-Development (L&D) AssociateCopy Anchor Link

15. Audit Committee / Internal AuditCopy Anchor Link

16. Regulator / Regulatory BodyCopy Anchor Link

SECTION 2: INDUSTRIES IN Q1–Q20Copy Anchor Link

Shipping & Logistics Industry (Q1, Q5, Q7, Q8)Copy Anchor Link

Key Terms:Copy Anchor Link

Finance & Hotel Industry (Q2)Copy Anchor Link

Legal & Contracts (Q3, Q8, Q9)Copy Anchor Link

Data Protection & Privacy (Q9)Copy Anchor Link

Workforce / HR (Q10, Q11)Copy Anchor Link

Rental / Property (Q12, Q16)Copy Anchor Link

Payroll (Q15)Copy Anchor Link

Shipping/Logistics Continued (Q18, Q19)Copy Anchor Link

SECTION 3: KEY CONCEPTS FROM Q1–Q20Copy Anchor Link

Domain Credibility (Q1)Copy Anchor Link

Prose Arithmetic (Q2)Copy Anchor Link

Baseline (Q5)Copy Anchor Link

Standing Instructions (Q5, Q16, Q20)Copy Anchor Link

Halo Delegation (Q11) ⭐Copy Anchor Link

Unstaffed Gate (Q11) ⭐Copy Anchor Link

Jurisdiction (Q8)Copy Anchor Link

Risk Matrix (Q8)Copy Anchor Link

Data Profiling (Q19)Copy Anchor Link

Threshold (Q20)Copy Anchor Link

QUICK REFERENCE CARD — Q1 to Q20Copy Anchor Link

Part 2 of 3 — AI Concepts, Model Choices & Key Frameworks (Q21–Q40)Copy Anchor Link

SECTION 4: HOW CLAUDE REMEMBERS THINGS (Memory & Storage)Copy Anchor Link

The Four Places Claude Can "Know" SomethingCopy Anchor Link

Memory (Claude's Notes About You)Copy Anchor Link

Superseded Rate Card / Stale Knowledge Files (Q21) ⭐Copy Anchor Link

Incognito Chat / Temporary Chat (Q22) ⭐Copy Anchor Link

Project (Q16, Q21, Q24, Q40)Copy Anchor Link

Account-Level Instructions vs. Project Instructions (Q16)Copy Anchor Link

SECTION 5: THE MODEL LADDER (Q28, Q41)Copy Anchor Link

The Three RungsCopy Anchor Link

Extended Thinking (Q28, Q59)Copy Anchor Link

The Two Mistakes (Q28) ⭐Copy Anchor Link

SECTION 6: OUTPUT RELIABILITY — WHAT CAN GO WRONG (Q21–Q33)Copy Anchor Link

Hallucination (Q4, Q21, Q43)Copy Anchor Link

Sycophancy (Q10, Q60) ⭐Copy Anchor Link

Context Drift (Q24) ⭐Copy Anchor Link

Stale Configuration (Q21, Q52) ⭐Copy Anchor Link

Under-Specification (Q18, Q34)Copy Anchor Link

SECTION 7: HUMAN ACCOUNTABILITY (Q23, Q26, Q27, Q29, Q37)Copy Anchor Link

The Core Principle: Accountability Cannot Transfer to a ToolCopy Anchor Link

Defined Gate (Q23) ⭐Copy Anchor Link

High-Stakes, No-Undo Actions (Q27, Q42, Q45)Copy Anchor Link

SECTION 8: BUSINESS PROCESSES IN Q21–Q40Copy Anchor Link

Supplier Payment Run (Q27)Copy Anchor Link

Invoice-Dispute Triage (Q26)Copy Anchor Link

Regulatory Attestation Pack (Q25)Copy Anchor Link

Badge-Swipe Records (Q38)Copy Anchor Link

Supplier Scorecard (Q35, Q56)Copy Anchor Link

CRM (Customer Relationship Management) System (Q36)Copy Anchor Link

ERP System (Q42)Copy Anchor Link

Event Calendar / Budget Allocation (Q39)Copy Anchor Link

SECTION 9: ESCALATION & DEPENDENCY (Q36, Q8, Q14)Copy Anchor Link

Escalation Signal (Q36)Copy Anchor Link

Purpose Limitation (Q58) ⭐Copy Anchor Link

SECTION 10: COMMUNICATING AI TO OTHERS (Q26, Q47)Copy Anchor Link

The Missing Human Checkpoint (Q26) ⭐Copy Anchor Link

Two Audiences, Two Rewrites (Q47)Copy Anchor Link