Bertrand's chord, for a reason that isn't a math trick
Bertrand's chord problem is the cleanest demonstration in classical probability that a question can look complete while being underspecified. Draw a chord "at random" in a unit circle; what is the probability it is longer than the inscribed equilateral triangle's side? Three defensible constructions of "at random" give three different answers: 1/3, 1/2, 1/4. Every derivation is correct. What differs is the measure on the space of chords, and the phrase "at random" never specified one. No error surfaces anywhere. Three teams would ship three floats, each with a clean proof, and a fourth team would average them to 0.3611 and call it resolved.
The classical lesson is that geometric probability needs a declared invariance. The operational lesson, the one this book actually cares about, is stronger: the question was never a question, and it still produced numbers.
The desk version is worse
The desk version isn't abstract. A risk officer asks a direct compliance question about CEH-001: what is the probability the mark exceeds 1.80? Two carded frames each supply a mark and a conformal band. The spread frame's 90% band is [1.42, 1.58] around 1.50, implying a residual standard deviation of 0.0486. The curve frame's band is [2.01, 2.22] around 2.10, implying 0.0638.
Under the spread frame: z = (1.80 - 1.50)/0.0486 = 6.17, giving P(mark > 1.80) = 3.4×10-10. Under the curve frame: z = (1.80 - 2.10)/0.0638 = -4.70, giving P(mark > 1.80) = 0.999999. The ratio between the two answers is roughly three billion to one. One frame says the event is a one-in-three-billion impossibility. The other says it is a near-certainty. Both are computed correctly, from carded artifacts, by the same script.
Two more answers, both wrong for reasons that feel responsible
A third reading is the one a well-meaning integration actually ships. Take the convex hull of both bands, [1.42, 2.22], and place a uniform law on it, on the grounds that the desk doesn't know which frame is right and uniformity is the honest expression of that ignorance: (2.22-1.80)/(2.22-1.42) = 0.5250. A fourth reading blends the two frames with equal weight, which is what a consensus service does by default: 0.5×(3.4×10-10) + 0.5×0.999999 = 0.500000.
| Reading | Answer | Feels like |
|---|---|---|
| Spread frame | 3.4×10-10 | Certainty it won't happen |
| Uniform over convex hull | 0.5250 | Responsible ignorance |
| 50/50 blend of frames | 0.500000 | Consensus |
| Curve frame | 0.999999 | Certainty it will happen |
Four honest answers to one question, spanning the entire unit interval from near-zero to near-one. None of the four is a mistake. Each follows correctly from a legitimate choice about how to treat the frame disagreement. That range is the whole argument: without a declared frame, "the probability" is not a well-posed object, and a system that returns any single number, however defensible, has silently made the choice for you.
Even "maximum ignorance" hides a coordinate choice
The standard rescue from Bertrand-style ambiguity is a principle: choose the distribution of maximum entropy subject to the constraints you actually have. It is a good principle, and it is not a rescue, because entropy is defined against a measure, and the measure is itself a coordinate choice. Put a nine-point grid uniformly across the hull of the two frames, 1.40 to 2.20, with no constraint beyond support. In the mark coordinate the mean comes out to exactly 1.8000 — the denied midpoint is not an operator's laziness, it is a theorem about the coordinate the grid happened to be laid down in. Lay the same grid uniformly in the log coordinate instead, the natural choice for a ratio, and the mean becomes 1.7738. In the reciprocal coordinate, natural for a rate, it's 1.7477. In the squared coordinate, natural for a variance-like target, it's 1.8259.
The spread across those four equally defensible coordinates is 0.0782 — 13% of the 0.60 frame gap, and, the comparison that actually matters operationally, 98% of the conformal half-width the spread frame's own 90% band carries. The ambiguity introduced purely by which coordinate you chose to be uniformly ignorant in is almost the same size as the entire prediction band built from two hundred held-out points. None of this makes maximum entropy wrong. It makes it conditional on a lock: the right instrument only after a frame card names the coordinate, the support, and the constraints, never before.
Why this is the deep chapter, not the repeated one
Earlier chapters introduce the ruler as a claim to accept: a question without a frame has several correct answers, a crowd of near-optimal models is a set rather than a sample, conformal coverage and calibration are different promises. This chapter pays for those claims with arithmetic instead of asking for trust. A desk that believes "frames matter" because someone said so behaves like a desk that heard a slogan. A desk that can produce a three-billion-to-one ratio on demand, from its own carded numbers, behaves like a desk that has a gate. The disagreement-abort rule that follows from this arithmetic is simple to state and expensive to skip: a query that names a threshold but not a frame gets refused, not answered, because every answer available to it is honest and none of them is the one number the requester actually wants.
Book: Chapter 28 in Rulers Before Models.