A Target That Moves
Description
The missing move, named at the end of the last episode, is a definition of merit that does not stand still. This episode states the mechanism at full strength. Keep a short, fixed, public menu of merits — judgment, honesty, numeracy, foresight, perspective-taking. Each time a post is filled, draw a few items from the menu at random; only those count, this time, and nobody knows in advance which. Why it helps: gaming needs a fixed target, and Frankel and Kartik showed the deeper version — the more a single signal decides, the more it reveals about gaming skill and the less about the real trait; spreading the decision across a randomly drawn menu lowers the stakes on any one signal and lets each go back to measuring honestly. Security has a name for the same idea: moving-target defense, where randomizing a system's configuration denies an attacker a stable exploit. A faction can capture a fixed exam by owning the preparation, but it cannot build a multi-generation pipeline for a target redrawn each time it's used. The part most easily missed: if the menu items are real capacities people already build in ordinary life, then no single path maximizes a person's odds of influence, yet getting genuinely better at anything on the menu raises them — so the smartest way to 'game' the system becomes actually becoming good at something. The mechanism keeps the competence floor pure sortition discarded, via a Sandel-style threshold (a bar to pass, not a rank to win). It does not beat Gibbard; nothing does. It does not make capture impossible — it makes capture expensive, and pays for the attempt in real skill.
- Muddled Information — Frankel & Kartikhttps://navinkartik.com/Papers/FK-MudInf.pdf
- A Survey on Moving Target Defense — MDPI Applied Scienceshttps://www.mdpi.com/2076-3417/13/9/5367
- Draw lots for Harvard entry, says Michael Sandel — Times Higher Educationhttps://www.timeshighereducation.com/news/draw-lots-harvard-entry-says-michael-sandel
Script
Cold open
Take everything we've said merit should reward — judgment, honesty, numeracy, foresight, the ability to hold someone else's point of view — and put them on a list. A short, fixed, public list. Now, each time you fill a post, don't use the whole list. Draw a few items from it at random, and only those count, this time. Nobody knows in advance which ones.
Frame
We ended last time on a single missing property: a definition of merit that doesn't stand still. This is the mechanism that supplies it. It has an awkward name — randomizing the criterion — and it does exactly one new thing. It makes the question 'what counts as merit, here' have a different answer each time, drawn fresh from a fixed menu. Everything strange about it follows from that one move.
Why would randomizing which merit counts help at all?
Why would that help? Because gaming needs a target that holds still — that's the wall from episode one. Frankel and Kartik gave the sharper version: the more a single signal decides, the more it ends up revealing about your skill at gaming it, and the less about the real trait underneath. So spread the decision across a menu, drawn at random, and the stakes resting on any one signal fall. And as they fall, each signal drifts back to measuring the thing it was supposed to measure all along.
What stops a faction from just capturing it anyway?
What stops a faction from simply capturing it anyway? The security world has done the thinking here. They call it moving-target defense: keep changing the system's configuration at random, so an attacker can never build a stable exploit against a fixed surface. A faction can capture an exam the way it captured the keju — by owning the preparation, for generations. But you cannot build a multi-generation pipeline aimed at a target that gets redrawn every single time it's used. The capture has nothing left to stand on.
What does that do to the person trying to get ahead?
Now the part that's easy to miss — and it's the whole point. If the items on the menu are well chosen — real capacities people already build in ordinary life — then no single path gives you the best odds of influence, but getting genuinely better at anything on the list raises them. Which means the smartest way to game this system is to actually become good at something. The effort you'd spend cheating and the effort you'd spend improving collapse into the same effort. The gap between gaming and merit closes — not because we blocked gaming, but because the best available game is the real one.
Doesn't randomness throw away competence, like the lottery did?
Doesn't randomness just throw the competence away again, like the bare lottery did? No — because the floor stays. You still have to clear a threshold. That's Sandel's move: merit as a bar to pass, not a rank to win. So the post never lands on someone who can't do it. You simply don't always clear the bar on the same axis. Qualify on whatever got drawn this time — and then, among everyone who cleared, draw by lot.
Turn
So this is the move the whole series has been walking toward. Capture and Goodhart both need a target that holds still. Every fix before this changed what we measure and left it standing there to be studied. This one leaves the menu fixed and public — and randomizes which part of it decides, each time. It does not beat Gibbard. Nothing beats Gibbard; there is still a way to game it, and the theorem promises there always will be. What changes is the price. To beat a target that moves across a good menu, you'd have to be broadly, genuinely capable across the whole thing — which is exactly what we were trying to find. It doesn't make capture impossible. It makes capture expensive, and it charges the attempt in real skill.
Closer
That's the idea, stated at full strength. A fixed menu. A random draw. A threshold. A lottery. It sounds almost too tidy — and the right response to an idea that sounds too tidy is not to admire it. It's to try to break it. Because there are at least three good ways this could fail, and if it's worth anything, it has to survive all three.