the second hundred · metaphor 109
When you misjudge someone — trust the wrong person, doubt the right one — is it because you genuinely can't tell, or because of where you set your bar? These are two different failures, and one dial does not fix the other.
Jealousy, hiring, medical tests, airport security run the same machine. Each takes a noisy reading — a look that lingered, a résumé, a blood marker, a shape on the scanner — and forces a verdict: flag it or clear it. Two independent things decide how often that verdict is wrong. One is how far the evidence actually pulls the real cases apart from the innocent ones — how well you can tell. The other is where you plant your threshold — how much proof you demand before you act.
Confusing them wrecks the conversation. "You're being paranoid" is a complaint about the bar; "you can't read people" is a complaint about the eye — and the fix for one does nothing for the other. Move your threshold and you only trade one error for its opposite: fewer false accusations bought with more things missed. The single move that shrinks both at once is not a better bar. It is better evidence.
Two crowds, one readinggap between peaks = d′
ROC · every bar you could pickAUC —
Two dials, not one
d′ is the distance between the two humps, measured in units of their own blur. At d′ = 0 the honest and the lying draw identical readings — the two crowds are the same picture, and no bar can separate them. Pull the humps apart and the overlap — the confusable middle, where a reading could plausibly come from either crowd — shrinks. That overlap is the entire source of error; everything else is bookkeeping.
The criterion is the vertical line where you switch from "clear" to "flag." It slides freely and it changes nothing about the humps. Two people with the same eye and different bars will disagree about nearly everyone, and each is right about their own error budget. So when a judgment goes wrong, there are two entirely separate post-mortems: the eye that couldn't tell, and the bar you chose. "You're too paranoid" indicts the bar. "You're a bad judge of character" indicts the eye. People argue past each other for years because they never say which.
What to try
Move the criterion. Watch the two error cards. Every step that shrinks false alarms grows misses by almost exactly as much, and vice versa. On the right panel the dot only slides along the curve — you are choosing a point on a fixed menu, never leaving it. This is the tradeoff you cannot argue your way out of: at a fixed eye, less of one error is more of the other, always.
Now raise d′. Both error cards fall at once, and the whole ROC curve bows up toward the corner — the menu itself gets better. That is the only move that beats the tradeoff instead of paying it. In life it is the slow, expensive one: gather more evidence, wait for another data point, learn the tells, get a second opinion. Fiddling with the bar is free and instant, which is why people reach for it first — and why they keep being surprised that the total error won't come down.
In the wild
A partner reads a late reply, a new name. The jealous complaint "you never believe me" is about a bar set too low; "you can't see what's obvious" is about a poor eye. Lowering the bar to catch every betrayal guarantees you accuse the faithful too.
A résumé is a noisy reading of a future hire. Raise the bar and you miss brilliant odd candidates; lower it and you wave through plausible frauds. Better interviews raise d′ — the same bar now sorts far more cleanly.
Sensitivity and specificity are hit rate and correct-rejection rate. Screening rare disease pushes the bar high — a rare signal makes most positives false. The panic over "too many false positives" is often a base-rate story, not a bad test.
The honest model
The model is two Gaussians of equal spread; d′ is the gap between their means measured in that spread. Everything on the instrument is those two integrals and nothing else. The hit rate and false-alarm rate are the tail areas past your line; the ROC is those two areas plotted against each other as the line sweeps from far right to far left; the AUC = Φ(d′/√2) is exactly the chance that a random real case reads higher than a random innocent one.
Where should the bar sit? Not a matter of skill — it is set by two things the eye knows nothing about: how rare the real thing is, and how much each error costs you. Move the base-rate and cost sliders and the ring on both panels shifts to mark the bar that minimizes expected cost — the likelihood-ratio rule, k* = ln(β)/d′. A rare signal or cheap false alarms push the optimal bar high; catastrophic misses push it low. That is why the right level of suspicion for a smoke alarm and for a murder trial differ even when the eye is identical — and why "just be more skeptical" is advice about the wrong dial.
The mapping
| Signal detection | Judging under uncertainty |
|---|---|
| d′ (separation) | How well you can actually tell the honest from the lying — your read, not your rule. The one thing evidence can improve. |
| the criterion / bar | How much proof you demand before you act on the reading. A free choice; it moves errors around but never removes them. |
| a false alarm | Trusting — no, accusing — wrongly: flagging someone who did nothing, sounding the alarm on the innocent. |
| a miss | Failing to catch the real thing: the liar believed, the disease cleared, the threat waved through. |
| the ROC curve | Every threshold you could pick with the eye you have — the full menu of tradeoffs, and the proof you can't be off it. |
| moving the bar | Trading one error for the other. Never a way to reduce both — that takes a better eye. |
| base rate & cost | Where the bar should sit — a fact about the world and your values, not about your skill at reading it. |
Where the metaphor tears
Real evidence rarely stacks into two tidy Gaussians of matching spread. Let the liars vary more than the honest — often they do — and the ROC goes lopsided, a single d′ stops summarizing the eye, and the "unbiased" bar is no longer the crossover. The clean symmetry that makes the picture legible is the first thing reality breaks.
The model crushes a whole person to a single scalar reading on one axis. Actual trust integrates dozens of channels over time, context-dependent, some of them constructed by the observer's own mood. Cases the model calls "confusable" a fuller look would often separate — the overlap is partly an artifact of refusing to look at more than one dimension.
Φ and the likelihood ratio hand you the cost-minimizing bar only after you have priced a miss against a false alarm — a betrayal survived against an innocent wrongly accused. That price is an ethical choice no amount of Gaussian tail can make for you. The mathematics proves the tradeoff is unavoidable; it stays silent on which error you should be willing to make.