Quiz: Capstone

Twelve questions. Answer question 1 only after computing the reconciliation yourself.

1. The acceptance criterion has two clauses doing work: "to the cent" and "rather than discovered after the fact." What does the second one require?

• A. That the reconciliation is automated
• B. That you could have predicted the difference before running the comparison, because every rule that produces it is documented, owned, and tested
• C. That the reconciliation runs before the pipeline
• D. That differences are logged

2. §38.2's assembly has six steps, and step 5 is three verifications that produce no data. Why does the chapter say it is the step people skip?

• A. It is slow
• B. It produces nothing, and a pipeline whose last step is "write the gold table" has no definition of done
• C. It requires special permissions
• D. It duplicates CI

3. Chapter 1's anchor says Black Friday runs at 6.28×. 6.28× what?

• A. A normal November day
• B. November 1st
• C. The annual average day — November's own average day already runs at 1.447× the annual one
• D. The previous year's Black Friday

4. §38.5 distinguishes R1–R3 from R4. What is the distinction, and why does it matter?

• A. R4 is larger
• B. R1–R3 remove rows and R4 revalues them — so only R4 can change a closed month's number, which is why the reconciliation must state an as-of date
• C. R4 is owned by finance and the others are not
• D. R1–R3 are automated and R4 is manual

5. §38.7 says the reconciliation closing is "necessary and weak evidence." What makes the result believable?

• A. That it was reviewed by finance
• B. That the assertions pass
• C. That two independent ratios agree — line count and revenue are both 1.447× the monthly average, and revenue per line matches Chapter 1's independently derived $28.084
• D. That the rebuild reproduces it

6. §38.8 recommends four reconciliations. Which pair does the chapter call "the strong pair," and why?

• A. Revenue and inventory — they cross systems
• B. Revenue and orders — a defect has to fool both a sum and a count with the same rows
• C. Customers and inventory — they cover the dimensions
• D. Orders and customers — they are both counts

7. §38.10: on Black Friday the critical path grows from 3.80 to 4.69 hours. Which component is responsible for the largest proportional growth, and why is it the one nobody sized?

• A. Sessionization — it processes the most data
• B. The dbt transform — warehouses scale poorly
• C. The quality register — it does not scale at all, and it costs $1.44, so it is invisible in a cost review
• D. Ingestion — it runs 24 times

8. §38.11: the quality register fired 29 times and 15 were real defects. The chapter calls 52% "roughly the right target." Why?

• A. Because half is a natural balance point
• B. Because "not a defect" is not "false positive" — nine freshness alerts correctly said a source was late, and a register that fires only on real defects has bands too wide to measure anything
• C. Because the industry standard is 50%
• D. Because the remaining 48% were suppressed

9. §38.13: the month reconciled to the cent while Black Friday was nearly twice its real size. How was that possible?

• A. The reconciliation had a tolerance
• B. The total was pinned, so the error just moved lines between days — and a monthly check cannot see a distribution error
• C. Black Friday was excluded by a rule
• D. The reconciliation ran on a different month

10. In Case Study 1, attempt 2 took twenty minutes and attempt 4 took three days. What explains the difference?

• A. Attempt 4 was larger
• B. Whether the gap matched a known rule's total exactly — when it does the search is over; when it does not, you are looking for a rule that is present and wrong
• C. Attempt 4 involved finance
• D. The team was more tired

11. Case Study 1's fourth failure: refunds were netted by settled_at rather than by the order's month. Why would this have survived a code review?

• A. It was in a file nobody reviews
• B. The other three failures are absences, which a reviewer listing applicable rules finds — but R4 was present, correct, tested, and pointed at the wrong dimension, and the wrong version is the simpler one
• C. It was too small to notice
• D. It was added after the review

12. Case Study 2's rebuild differed by $741.18 — 0.0036%. What is the chapter's rule about chasing it?

• A. Chase anything above 0.001%
• B. A reproducibility failure has no materiality threshold, because the size is what the defect happened to produce this month, and the failure invalidates the rebuild, replay, CI-against-a-fixture, and backfill procedures
• C. Chase it only if it exceeds the materiality threshold
• D. Log it and move on

---

Answers

1 — B. That you could have predicted the difference.

It is not enough that the numbers agree; agreement can be a coincidence, or the product of two errors cancelling. The criterion requires that before running the comparison you can name every rule that produces the gap, and that each has an owner and a test. Kestrel could not, on the first four attempts — and each failure was a rule that existed in somebody's head, in the SQL, and nowhere in between.

2 — B. It produces nothing, and it is what gives the reconciliation meaning.

layer_check makes "we rebuilt from bronze" a fact rather than a claim; manifest --verify confirms the platform can honor a deletion request, without which it is not finished regardless of its numbers; xref_audit is trivial and fails most often after a refactor. Verification steps produce nothing and gate everything — a pipeline whose last step is "write the gold table" has no definition of done.

3 — C. The annual average day.

41,300 orders against 6,575 on an average day of the year. November's own average day already runs at 1.447× the annual one, so within November Black Friday is only ~4.28×. All three multiples are true — 6.28× annual, 4.28× a November day, 7.89× November 1st — and only the first is what the anchor asserts. Applying 6.28 within November produces a Black Friday of almost 200,000 lines, which is wrong and looks right, because 6.28 is the number everybody remembers and the baseline is what gets dropped when a figure is quoted.

4 — B. R1–R3 filter; R4 revalues.

A refund settled in December against a November order changes November's net revenue after November has closed. Four consequences: November is not final until 1 March (a 90-day refund window); a snapshot taken on 1 December and a query run on 1 March disagree and both are correct; any incremental model must be able to restate a closed month, which is why fct_order_line needs a lookback window rather than append-only; and the reconciliation must state its as-of date or it is not reproducible.

5 — C. Two independent ratios agree.

The reconciliation closes because it is a subtraction — arranging four numbers to sum correctly proves arithmetic, not correctness. What makes it believable is that the line count and the revenue are elevated by the same factor (1.447×) to three decimals, which nothing in the reconciliation forced, and that revenue per line ($28.085) matches Chapter 1's independently derived $75.83 ÷ 2.70 = $28.084. Two derivations, thirty-seven chapters apart, to four significant figures — and the self-check asserts it, so it is a test rather than an observation somebody made once.

6 — B. Revenue and orders.

A defect must fool both a sum and a count with the same rows, and very few do — a join that duplicates rows at half price sums correctly and counts wrongly, which is exactly Chapter 18's fan-out. Each of the four reconciliations is blind to what another sees, and the set is chosen for the union rather than for completeness. For a platform with only one reconciliation, add the row count: an hour of work that roughly doubles what the check can see.

7 — C. The quality register.

It runs fixed assertions on two nodes, so its runtime doubles with the data while sessionization autoscales and the warehouse can be resized. It goes from 8% of the critical path to 13%. And it costs $1.44 a night — so it never appears in a cost review, which is where somebody would have asked whether it scales. Cheapness is why it was invisible, not why it is unimportant.

8 — B. "Not a defect" is not "false positive."

All nine freshness alerts were a source system running late, which is precisely what a freshness check is for; the check worked, the pipeline waited, nothing broke. Five of seven volume alerts were Black Friday against a band set from ordinary days — correct alerts revealing that the band was wrong, which is itself the finding. The number that would worry Kestrel is the opposite one: a register firing only on real defects has bands so wide it is not measuring anything. Too few alerts is as diagnostic as too many.

9 — B. The total was pinned.

The fixture forces the month to 781,380 lines, so getting Black Friday wrong simply moved lines from other days into it, and the monthly sums were unaffected. This is not a fixture artifact — a real platform has the same blindness: if daily_revenue sums correctly for the month, nothing in a monthly check notices that a day is wrong. What caught it was a shape assertion checked against Chapter 1's figure rather than against anything the fixture controls. Totals hide distributions, and Chapter 23's register is mostly totals and bounds.

10 — B. Whether the gap matched a known total exactly.

Attempts 2 and 3 matched R1's and R3's totals to the cent, which identified the rule immediately. Attempt 4 did not, and the search space for "a rule that is present and pointed at the wrong dimension" is far larger than for "a rule that is absent." Note also that attempt 1's exact match was a coincidence and cost a day — the heuristic is a strong prior, not evidence, and the row-count reconciliation is what falsified it.

11 — B. It was present, correct, tested, and pointed at the wrong dimension.

The other three failures are absences: a missing timezone, a missing filter, a missing accounting rule. A reviewer asking "what rules apply here?" finds absences, because the answer is a list and items are not on it. R4 was implemented, and settled_at is the obvious column — it is when the money moved and what the processor reports. The correct version groups by the order's placed_at, which requires a join the wrong version does not need, so the wrong version is also the simpler one. It also nets out over a quarter, is 0.85%, and fails in the low direction, which sends you looking for an over-aggressive filter.

12 — B. A reproducibility failure has no materiality threshold.

$741.18 is what the defect produced this month, on this data; the defect is "the platform's output depends on something other than its inputs," and that statement has no size. It also invalidates four other checks — the rebuild, the replay, CI against a fixture, and the backfill procedure all assume the property that just failed — which is a much larger exposure than the money. Materiality is a reasonable concept against an external source, where irreducible differences genuinely exist (Chapter 37's 50.2%); it is not reasonable for a system compared against itself, where the correct difference is zero by construction and anything else is information you do not have.