The problem is not finished at the number
Every number in this book was derived to make an argument. This chapter asks a reader to derive them independently, in the order a desk would actually meet them, and then to say what the system does about the answer. A quantitative answer here is only half a solution. The other half is a verdict and a fail code, because a desk that computes the right number and grants the wrong right has failed the whole exam.
Problem 1: when averaging is maximally destructive
A client asks for a single mark on CEH-001. The spread frame returns 1.50 with band [1.42, 1.58]; the curve frame returns 2.10 with band [2.01, 2.22], each band declared as a four-standard-deviation interval. That makes the spread frame's implied σ = 0.0400 and the curve frame's σ = 0.0525. The forbidden midpoint 1.80 sits 7.50 standard deviations above the spread centre and 5.71 below the curve centre. Under the spread frame, P(mark > 1.80) ≈ 3.2×10-14. Under the curve frame, P(mark > 1.80) is 1 minus about 5.5×10-9 — essentially one. Average those two numbers and you get exactly 0.5000.
That 0.5000 is the most instructive number in the problem. Two models, each individually extremely sharp, each assigning probability within a billionth of a boundary, average to a coin flip. Blending did not produce a compromise with intermediate precision. It produced formally zero information, and it did so precisely because both inputs were confident. The instinct that averaging is conservative is exactly backwards here: averaging is maximally destructive when the inputs are sharpest.
There is a second trap in the same problem. A crowd of eleven models, six near the spread frame, five near the curve frame, has a mean of 1.80 with a standard error of 0.095 and a 95% interval of [1.61, 1.99]. That interval is computed correctly. It contains neither honest cloud. Attaching a confidence interval to a forbidden point estimate does not launder it; it produces a more credible-looking version of the same violation, arriving with error bars and an air of rigour. The sample standard deviation of 0.316 is itself the tell — it is roughly half the frame gap, meaning the crowd's dispersion is explained entirely by the gap itself, not by estimation noise. Verdict: refuse, coded DUAL_MARK_MIDPOINT, and the remedy here is to preserve the gap, not close it, because this is genuine disagreement between two contracted geometries, not a missing field.
Problem 2: a leak that flatters the desk
A mark is produced at 14:00. The true value is Y = 1.50 + A + B, where A resolves at 14:00 and B resolves at 16:00, both fair coin flips of ±0.40 and ±0.10. Before 14:00, Var(Y) = 0.16 + 0.01 = 0.17. After A resolves, only B remains: Var(Y|A) = 0.01. Learning A collapsed 94% of the variance — that is what a filtration is, a schedule of when variance is legally allowed to disappear.
A pipeline that reads B early prints Y exactly on every path, squared error zero. The legally-informed mark prints 1.50 + A, with squared error 0.01. Watch the path where A = +0.40 and B = -0.10: the leaked model prints 1.50 + 0.40 - 0.10 = 1.80 exactly. Leakage manufactures the denied midpoint, and the exam rewards it with a perfect score. That is why leakage is a hard gate rather than a statistical one — no confidence interval can rescue you, because the evidence is faithfully describing a model that occupied a room it had no right to occupy.
The second half of the problem is worse, because a timestamp audit cannot see it. A feature is standardised by a scale fitted over a five-day window that includes a co-break day with a deviation four times normal. The legal scale, fitted only on the four calm days at or before the clock, is 0.4200. Include the fifth, loud day and the full-week scale becomes 0.8400 — exactly double. A deviation that legally reads 2.50 standard deviations reads only 1.25 once the leak enters. The feature has been halved. Every row's own timestamp is honest; what reached past the clock was the fitting window of an aggregate, and an aggregate is an ancestor of every row it scales. The direction of the damage is the part worth sitting with: this leak did not inflate the signal, it damped it, hardest on exactly the loud days. An exam run on it reports that the world is calmer than it is. A leak that flatters the desk is far harder to catch than one that produces implausible accuracy.
What passing actually requires
Every problem in this chapter ends the same way: the code that fires, the card that owns it, and whether the remedy closes a gap by declaring a missing field or preserves a gap because it is genuine, honest disagreement. Confusing the two is the single most common error a competent quant makes on this material — treating an honest dual-frame disagreement as a data problem to engineer away, when the correct answer is to name the frame and let a human choose. A reader who gets every arithmetic answer right and never once writes a fail code has not passed. The arithmetic was always the easy half.
Book: Chapter 32 in Rulers Before Models.