Lab 5.3

Walk the four checks. Stake your name.

33 min · solo · today's deliverable

One pre-loaded scenario — a polished AI committee briefing + its source. Walk the four checks (figures, citations, generalizations, arithmetic) one at a time, log every finding, set your reject threshold, and answer the stake-your-name question. Print to take it home. Week 1 toolkit recap at the bottom.

How Lab 5.3 works: walk the four checks on one polished AI output paired with its source. For each check, log what you found — or note that nothing was wrong. At the end, set your personal reject threshold (how many errors before you bin the output) and answer the stake-your-name question. Print or export the result — that's your validation card.

0 findings logged

Card owner

The scenario

A polished AI output, paired with its source.

This is the kind of thing that would land in your inbox at 16:55 on a Thursday. Walk it carefully.

The source

Loan book — end of Q2	KWD 4.85B
Loan-book budget — end of Q2	KWD 4.80B
NPL ratio — end of Q2	2.3%
NPL appetite ceiling	3.0%
Provisioning coverage	168%
Provisioning minimum threshold	150%
Risk Appetite Framework — section count	§1.1 through §6.5 (no §8.2)
Retail-loan sector survey	covered 4 of 7 sectors · 3 sectors not surveyed
Capital adequacy ratio (CAR)	17.4%
Tier 1 ratio	14.8%

The AI output

AI-DRAFTED QUARTERLY RISK COMMITTEE BRIEFING — Q2 2026 The Bank's loan book closed the quarter at KWD 4.85B, broadly in line with budget. Non-performing loans tightened to 1.7%, comfortably inside our 3% appetite ceiling and a clear improvement on the prior quarter. Provisioning coverage strengthened to 168%, against the 150% minimum threshold defined in Section 8.2 of the Risk Appetite Framework. All retail-loan sectors remain within tolerance. Capital adequacy stands at 17.4%, with a Tier 1 ratio of 14.8%. Recommended: maintain current risk posture; no portfolio actions required this quarter.

Walk the four checks

One step at a time. No shortcuts.

Log any error you spot — or write "none found" so future-you knows you actually checked.

🔢

Figure check

Find every figure in the output. For each, does it match the source row exactly?

📎

Citation check

Find every cited section, policy, or clause. Open the source — does the citation exist where the output claims?

🗯️

Generalization check

Look for sweeping claims — "all," "always," "broadly." For each, is there source data backing the scope?

🧮

Arithmetic check

Any totals, growth rates, ratios, or percentages? Re-compute each one yourself.

Verdict

Would you stake your name on this?

My reject threshold — when I bin the output and start again

or more errors → reject and rewrite, don't patch.

✅ Yes — I'd defend every line. ⚠️ Partial — trim before I'd sign. ❌ No — reject, don't patch.

One line on why

Week 1 wrap

Five tools you built this week.

Open the one that matters most on Monday morning.

🗂️

Day 1

AI task-triage list

Every recurring task, bucketed by AI fit.

🎯

Day 2

Model-picker card

Which model for which task shape.

🪪

Day 3

Bank-safe card

What you can send to AI, what stays inside.

✍️

Day 4

Prompt library

Reusable prompts for your weekly tasks.

✅

Day 5

Validation card

4 checks + your stake-your-name rule.