↩ ExploRational · A Catalogue of Divergent Rationality
explorational · a divergent move
The wisdom of crowds averages away independent errors — but you are one person. Can you extract part of that gain from a single head by making a second estimate from a different starting point and averaging it with your first?
Herzog and Hertwig gave the trick a name in 2009: dialectical bootstrapping. Make your estimate. Then assume that first estimate is wrong, and ask why — what did you overweight, what did you ignore, which anchor were you standing on? Re-derive a second estimate from that different ground, and average the two. In their experiments the averaged answer beat the first guess reliably, as if a lone person had convened a tiny committee of one and taken its consensus. A crowd, borrowed from inside a single skull.
But the whole gain rides on one word: independence. The wisdom of crowds works because independent people make errors that point in different directions, so averaging cancels them. A second guess that merely echoes the first — same anchor, same reasoning, a cosmetic nudge — carries the same error, and averaging two copies of the same error removes nothing. This is token opposition: the ritual of a second opinion with none of its substance. The instrument below makes the gain a computed function of exactly how independent the second guess really is, so you can watch it appear and vanish.
the signature instrument
Each trial has a true value; you make a first estimate and a second, each carrying bias and noise. The error correlation ρ sets how alike the two errors are — ρ = 1 means the second guess is the first in disguise, ρ = 0 means fully independent, ρ < 0 means deliberately over-correcting. The displayed values are computed from thousands of live Monte-Carlo trials. The theoretical curve σ²(1+ρ)/2 is overlaid so you can see the simulation land on it.
Two guesses, one head
RMSE of one guess vs. their average, over live trials
—
—
a crowd of one
The classic wisdom of crowds is an averaging machine. Ask a thousand people to guess the weight of an ox and the guesses scatter — some high, some low — but because those errors are roughly independent, they cancel when you average, and the mean lands astonishingly close. The signal survives; the noise, being unaligned, washes out. The catch is that it takes a crowd, and you are one person with one head, one anchor, one first impression that tends to colonise everything that follows.
Dialectical bootstrapping is the move that opens the hypotheses dimension by multiplying — from a single estimate to several — while working alone. Instead of consulting other people, you consult a version of yourself that started from somewhere else. The discipline is what forces the difference: don't just re-guess, which returns you to the same anchor. Assume the first answer is wrong, name the reason it might be, and let that reason move you to different ground. Then average. If the two estimates truly stand on different footings, their errors are less aligned, and the average inherits a little of the crowd's cancellation. You have manufactured a second, cheaper witness — and the second witness is worth exactly as much as it is independent.
what to try
With bias off, drag ρ from +1 down to 0. At ρ = 1 the average's error sits exactly on the single-guess line — no gain. As ρ falls, the simulated dots and the theory curve descend together along σ²(1+ρ)/2; at ρ = 0 the error drops by about 29%. Keep going into negative ρ and the error keeps falling — over-correction cancels more than independence alone.
Now raise systematic bias with the second guess still set to echoes. Slide ρ all the way to −1: the noise cancels perfectly, yet the average's error refuses to reach zero — it floors at the bias. Averaging kills noise; it cannot touch a lean both guesses share.
Leave the bias high and drag the second guess from echoes toward attacks bias. Watch the floor fall: a guess re-derived to lean the other way cancels the shared bias as well as the noise. This is the difference between a token second look and a real one.
independence is the whole gain
Split each estimate's error into two parts. There is noise — the wobble that would land differently if you guessed again — and there is bias, the systematic lean that lands the same way every time. Averaging treats them completely differently.
Take two estimates whose noise has variance σ² and correlation ρ. The variance of their average is σ²(1+ρ)/2 — a small, exact, and merciless formula. At ρ = 1 it returns σ²: averaging two identical errors buys you nothing. At ρ = 0 it halves the variance, cutting the error by about 29%. At ρ = −1 it hits zero: two errors that are mirror images annihilate. The only lever in that formula is ρ. Not how many guesses, not how confident you feel — how correlated the errors are. Token opposition fails precisely because it leaves ρ near 1 while dressing up as diversity.
Bias obeys no such formula, because there is nothing to cancel. If both guesses lean high by the same amount, their average leans high by that same amount — ρ can go to −1 and the noise can vanish, but the bias floor stays put. That is why a genuine second estimate has to do more than add fresh noise from a fresh angle: it has to attack the bias, standing on ground that leans the other way. A crowd of independent people escapes shared bias only when their biases genuinely differ; a crowd of one escapes it only when the second guess is re-derived hard enough to lean against the first. Add noise and you shave the wobble; attack the bias and you move the floor.
model
No faked numbers. On each trial the first estimate's error is e₁ = S + n₁ and the second's is e₂ = S·m + n₂, where S is the systematic bias, m is the second guess's alignment (+1 echoes the bias, 0 ignores it, −1 attacks it), and n₁, n₂ are Gaussian noise with standard deviation σ. The two noises are drawn correlated: n₂ = σ·(ρ·z₁ + √(1−ρ²)·z₂) with independent standard normals z₁, z₂, so that each noise has variance σ² and their correlation is exactly ρ. A seeded generator makes every run reproducible; reseed draws a fresh world.
From the trials the instrument computes root-mean-square error directly: RMSE = √(mean of error²), once for e₁ alone and once for the average (e₁+e₂)/2. The overlaid curve is the closed form it should match — √(bias² + σ²(1+ρ)/2), where the averaged bias is S·(1+m)/2. The single-guess reference line sits at √(S²+σ²); the crowd ceiling at σ/√10, the error of ten independent, unbiased estimators. The simulated dots wear a little sampling scatter — that scatter is honest, and it shrinks as you raise the trial count. Read the instrument as a measuring device, not a slogan: it will happily show you the settings where the second guess earns nothing.
the move ↔ the model
| The move | The model |
|---|---|
| the first estimate | your immediate guess — e₁ = S + n₁, carrying whatever anchor you reached for first. |
| the second estimate | a guess re-derived from different ground — e₂ = S·m + n₂, worth only as much as it truly differs. |
| correlation ρ | how independent the second guess really is; the single lever that sets how much error the average removes. |
| the average | the cheap internal ensemble — (e₁+e₂)/2, the consensus of your committee of one. |
| shared bias (m near +1) | the error averaging cannot remove: a lean both guesses hold, surviving even perfect noise cancellation. |
| token opposition | a second guess that only pretends to differ — ρ near 1, m near +1 — buying confidence at no accuracy. |
how this opening fails
risk · token opposition
The catalogue's named risk for this move. The ritual of a second estimate is easy to perform and easy to fake: re-guess without re-deriving, nudge the first number by a token amount, and you have two estimates whose errors are nearly identical — ρ near 1. The instrument shows the payoff of that theatre is essentially nothing, yet it feels like diligence, and the feeling is worse than useless because it manufactures confidence where no new information exists. A second look is only worth taking if it can genuinely land somewhere the first would not. If nothing about your reasoning changed, you did not consult a second witness; you cross-examined the first one twice.
risk · the bias averaging cannot remove
Averaging is a variance machine, not a truth machine. When both of your estimates share a systematic error — a bad conversion, a wrong base rate, an assumption you cannot see well enough to argue against — no amount of internal re-derivation escapes it, because the same blind spot sits under both guesses. The dialectical trick can attack a bias you can name and lean against; it is helpless against one you cannot. Past a point, the move is not a cleverer second guess but a genuinely external check: a measurement, a reference class, a person whose errors are not made of the same material as yours. Know which problem you have before you reach for the average.
Average with a second guess from a different starting point, and one head borrows a little of the crowd's wisdom.
can you use it?
RECOGNITION — Which is the crowd within? A: you look at your timeline estimate again and pad it 10% to be safe. B: you assume the first estimate is wrong, name its anchor, re-derive from a base rate, and average the two. C: you collect three colleagues' independent estimates and average.
B. A is token opposition — same anchor, ρ near 1, nothing cancels. C is the real crowd: excellent, but it takes other people.
THE NEAREST NEIGHBOR — What distinguishes dialectical bootstrapping from double-checking your answer?
The second estimate must stand on different ground — assume the first is wrong, name why, re-derive — so the errors decorrelate. Checking re-walks the same path and inherits the same error.
PRODUCTION — You must forecast beta signups; your gut says 2,000. Make the second estimate before looking.
Assume 2,000 is wrong — it anchors on the launch list's best week. Re-derive from conversion base rates of similar betas: about 800. Average: 1,400. Yours works if the second estimate names the first's likely error and reasons from a different footing — a nudge to the same number earns nothing.