Published on: June 28, 2026
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Send Strategic Nudge (30 seconds)Green in-lane · amber a little out · red drift. Every panel is a real commit, byte-identical on recompute. Tap any panel to open its shareable receipt.
We believe the hardest problem in an AI information landscape is not capability — it is attribution. Take it at its limit, because the limit shows the shape most clearly: in a near future where an autonomous agent, or a person directing one, could cause catastrophic harm for very little money, the decisive question is not can it act but can you tell one intent from another in time to do anything about it. When you cannot — when the same observable result could be a mistake, a shrug, a bad week, or a deliberate act — you cannot attribute it, and what you cannot attribute, you cannot learn from. That is the real danger: not the harm, but that the harm arrives untrainable, leaving no signal you can act on. The way out is not to get better at reading intent, which is hopeless. It is to read where the work landed — a decidable coordinate that makes the invisible legible without accusing anyone of anything.
You do not need to know why an output went where it went — undecidable, and often unknowable — to know where it went relative to where it was supposed to go. The first is a verdict on a person's heart. The second is a measurement. Only the second is learnable, and only the second is fair.
Start where you have actually lived it, because the limit case is just this — smaller. The work that was technically delivered but somehow off. The contribution that was fine on paper and useless in practice. The room that got a little colder after a particular person walked into it. You could not point to one thing — no missed deadline, no broken rule — and that is exactly what made it impossible to name. The injury was real and the evidence was nowhere, because the injury lived in the gap between what was promised and where the work actually came to rest, and nobody was measuring that gap. This is the grip the problem has on you: a harm you can feel but cannot locate is a harm you cannot address — and the same grip, unchanged, is what waits at the top of the stakes.
Here is what the shift hands you to give. Instead of a feeling you cannot defend, you get to put down a located fact: this is where the work came to rest, and that is not where it was contracted to rest. You are not accusing anyone of bad intent — you could not prove that and should not try. You are placing a coordinate on the table that anyone can check for themselves. That is a gift you can give a teammate without poisoning the room, a gift you can give a vendor without a fight, a gift you can give your own future self instead of a vague unease you will have forgotten by next week. The thing you contribute is not a verdict on a person. It is a measurement of an outcome — and a measurement is something other people can pick up and use.
The capability you gain is a single, repeatable move: stop trying to litigate why, and locate where. It is one skill, and it is the same skill whether the stakes are a soured afternoon or an irreversible act — only your tolerance for the loss changes, never the method. Learn it on the small things and it is already the tool you need on the large ones: you do not have to become a better mind-reader as the danger grows, which is a race no one wins. You get to keep doing the one thing that works at every size — turning "something is off" into "here is the coordinate it landed on." Each time you do it, you add to a record you can actually reason over, instead of a pile of impressions you cannot.
Be honest about why the obvious responses fail, because that is what makes the move necessary rather than nice. You cannot draw the line by reading intent: souring a mood is not sabotage, poor service is not by itself sabotage, and an employee who simply does a bad job might be indifferent, incapable, or aimed — you cannot tell, and the disagreement runs even deeper, because people do not agree on whether the actor had a choice at all. And you cannot route around it by complaining, because of a trap built into the situation: the moment you raise something you cannot prove, you are filed as a complainer, and a complainer's reports are treated as unfalsifiable and waved off — so the very subtlety that let the harm work also makes the objection self-defeating. Naming it without evidence does not raise the alarm; it discredits the alarm-raiser. Meanwhile stability lulls everyone: a system that runs smoothly is assumed to be doing what it claims, even though "it works at the level it works" and "it does what it says" are different facts. You can be losing steadily inside something that looks, from outside, perfectly fine.
The trap in one line: the harms subtle enough to deny attribution are exactly the harms too subtle to complain about credibly. No provable signal, no standing — so nothing is learned, and the pattern runs free. That is what "untrainable" means here, and why feeling it is not enough.
So anchor on the one thing that does not move. Whether an act was meant — wise, good, aimed at you — is undecidable, and asserting it is defamatory. Where an output landed — which region of a fixed map it occupies, against the region it was supposed to occupy — is a placement of finite things on a finite grid: no mind-reading, computed the same way by anyone, anywhere, every time. That is the part you can sign and the part someone else can re-run and get the same answer. It does not say you did this on purpose. It says this is where it came to rest, and that is not where it belonged — a sentence that is at once true, checkable, and survivable to say out loud. The certainty is small and that is the point: it is small enough to actually be certain, which is exactly what nothing else in the gray zone could offer you.
With this in hand you stop being the person who senses something is wrong and cannot say it, and become the person who puts a checkable fact on the table — which means you can finally act on a deck that was stacked precisely because the harm stayed invisible. You are not a complainer; you are someone holding a measurement, and that is a different kind of person to be in a room. And there is a larger thing you become: the operator who closed the small gray zone honestly, which is the only way anyone earns the right — and the instrument — to close the large one in time. The civilizational version of this problem will not be solved by people who could never solve the everyday version. You become someone who can, by learning to locate instead of litigate while the stakes are still survivable.
Why there is no conventional alternative. The problem of unobservable effort is old and formalized: the principal-agent literature — Holmström's 1979 work on moral hazard — shows that when you can see outcomes but not the action or intent behind them, you cannot contract on the thing you actually care about; you are forced to a proxy, and the proxy is gameable. And reading intent directly is not merely hard — for any sufficiently expressive system it is formally undecidable, the same Rice's-theorem wall that defeats "is the AI aligned." Both roads dead-end at the same place: you cannot adjudicate the why. The only move left that has an answer is to measure the where.
Why it is not overreach. Because it refuses to claim what it cannot know. The instrument never says an act was malicious, negligent, or chosen; it reports a location and a distance from the contracted location, and stops. No motive, no mind-reading, no naming a saboteur — which is exactly what keeps it defensible instead of paranoid. The fence is the feature: a tool that claimed to detect intent would be the thing to distrust.
Why nobody else has it. Because the whole conversation is stuck on the undecidable side — arguing intent, quality, good faith, sentiment — none of which can be settled, all of which invite the complainer trap. Measuring where an output lands, signed and recomputable, sidesteps the argument no one can win and produces the artifact the situation always lacked: an attributable, learnable fact that names a coordinate instead of a culprit.
The next time something is off and you cannot name it, notice the urge to build a case about why — and set it down, because that road ends in the complainer trap every time. Reach for the located fact instead: where did the work land, and where was it supposed to land. Claim your own coordinate at /pixel — the decidable spot where your work, in your lane, is yours to stand on and to point to. Read the slice — the decidable slice of alignment lays out the where-not-whether boundary in full, and Black-Scholes didn't touch the stock shows how a located fact becomes something you can even price. The deck only stays stacked while the harm stays invisible. Make it a coordinate, and you can finally learn from it — which is the one thing the gray zone was built to prevent.
