I have spent years building walls into machines. Into a company: the optimizer never holds the pen on its own objective. Into a loop: meter the truth on its own clock, and pre-register the floor before you have a stake in where it lands. Good walls, all of them, and all aimed at some subsystem inside the product. Every one quietly assumed the same thing. The assumption is a lie.

They assumed I could build a trustworthy wall. Feynman's first principle says the wall-builder is compromised: you must not fool yourself, and you are the easiest person to fool. So the walls kept guarding the machine and never once guarded the one embedded agent I can't step outside of, can't airgap, can't audit from any clean vantage. The designer. This essay is about turning the wall on him. On me.

You cannot out-discipline yourself

Here is the shape every self-fooling failure shares. Once you see it you can't unsee it: the faculty doing the grading is downstream of the faculty doing the wanting. It's the same organ. So the instinct (build a stronger check, a stricter threshold, a smarter reviewer, try harder to be honest) is doomed, because a stronger wall built by the same hands that want through it just raises the cleverness of the breach. You end up with a better-defended delusion, which is worse than a naked one, because now it has credentials.

That's why "try harder to be objective" is not advice. It's the disease describing itself. The fix isn't a stronger wall. It's a different wall.

Strength was never the variable. Location was. You cannot build a wall tall enough to keep you from fooling yourself, for the plain reason that you're the one building it. You can only build it somewhere your hands aren't.

Which forces a move that feels almost undignified, and is the whole point of this piece. Every other essay on this site let me be the engineer. The one holding the instrument, deciding what the readout means. This one demands I climb onto the slide. Stop being the thing that measures. Become the thing that's measured. You are not the scientist here. You are the specimen.

The machine I already built

I know this wall works because I once sold it. Years ago I built a behavior-change company on continuous glucose monitors. A CGM is nothing but the anti-self-fooling machine rendered in hardware. Think about where dieting actually fails. It fails in two places, and neither is willpower. It fails in when you measure. You step on the scale after a good week, never a bad one. And in how you interpret. "That oatmeal was basically healthy," a sentence that has never once been checked against a blood draw. Both leaks are discretion. Both are the wanting, dressed as judgment.

The CGM removes the discretion. It meters a shape. Your glucose curve, 288 times a day, every five minutes, whether you want the news or not. It does not care how you feel about breakfast. And the thing I watched, over and over, was not that people got more disciplined. It's that people who were certain their diet was fine got metered, saw a curve they could not argue with, and changed behavior that a decade of resolve had never touched. The empathy in it surprised me: the monitor doesn't shame you. It just shows you the shape. The shape does the work shame never could. Because the verdict finally lived somewhere outside them.

I built that for other people's blood sugar and then spent years forgetting to build it for my own thinking. This essay is me remembering.

Three ways to leave the room

The protocol is three moves, and they are one idea in three positions, each of them a way of moving custody of the verdict outside the self.

Measure the generated shape. Self-deception always strikes at the moment of interpretation. You get a result, then you decide what it means. The deciding is where the wanting seeps in. So measure the artifact that exists before you touched its meaning. Feynman's own version was exactly this: don't report "the data supported the theory." Report the raw scatter and let the shape testify, because the summary is where you cheat and the scatter is where you can't. The glucose curve, not "I ate clean." Whether money moved, not "the launch felt strong." A shape, not a story.

Meter it. This is the move people skip. It's the sharpest. A wall you check when you feel like it is a wall you check when you expect to pass. The sampling is corrupted before the measurement even begins. A meter runs on its own clock, continuously, indifferent to whether the news is good. That is the entire difference between a bathroom scale after a good week and a monitor that reads you 288 times a day and doesn't ask permission. Metering strips out your discretion over the timing of the truth. Timing is the room self-deception lives in.

Pre-register the floor. The last hiding place is deciding, after the fact, what would have counted as success. Un-pre-registered, you move the goalposts by inches you never feel. The result comes in soft. You discover you "always thought" the bar was lower. So plant a Schelling fence before you have a stake in where it lands, because your pre-stake self is the only disinterested broker you will ever get. Write the floor down while you still don't care, make it append-only, and let the motivated future you be judged by it instead of allowed to rewrite it. That is you, from outside, going near the invariant. The one arrangement where the optimizer doesn't hold the pen.

It's all the same wall

Here is what unsettled me when I lined them up. These three moves are one architecture. A wall between what happened and what you get to say about it, rebuilt for a human instead of a machine. The measured shape is the ground truth that exists before interpretation. The meter is the append-only record, a clock the wanting can't pause. The pre-registered floor is the invariant you may read but never edit. Feynman's first principle, the embedded-agent's airgap, Amazon's controllable-input meter, my old glucose monitor, and the append-only ledger in my own code are not analogies for each other. They are the same wall, several floors apart, all answering one question:

How does a mind that is part of the world it's judging keep from grading its own paper? The answer is never "grade harder." It's "put the grade somewhere your wanting can't reach."

Register the registrations

And even this fails. I'd be fooling myself to pretend otherwise, which would be a fittingly stupid way to end an essay about not fooling yourself. A meter can be gamed if you get to choose which shape to meter. A pre-registered floor can be gamed if you get to choose which experiment to register. Run ten, quietly, and report the one that cleared. Selection is the final room. It has no window.

So you meter the selection too. Keep an append-only record of every floor you set and every meter you started, not just the ones that survived. Register the registrations. That is the bottom turtle. It's the same turtle the whole way down: the only defense against a mind that fools itself is a record it wrote before it had a reason to lie, kept somewhere it can no longer reach.

The general statute under that turtle: leave a trace, not a memory. A trace is a shape that outlives the session. Record what happened durably enough that a different mind (including your own, next week, or another agent picking up where this one failed) can resume from it without you. Memory dies with the worker. A committed trace is the shape, kept.

This is the becoming the other essays were circling without naming. Grit, curiosity, simplicity. Real, but downstream. The founder's deepest virtue isn't any of them. It's the willingness to become the specimen: to build the one wall you most don't want, in the one place you can't see, and hand the pen to a version of yourself who didn't yet care how the story turned out. You are the easiest person to fool. So stop trying to be honest. You can't, not reliably, not from the inside. And build so that honesty doesn't depend on you being honest. That is the only wall that holds. It's the only one your hands aren't on.

Part of an ongoing series on durable systems.