Published on: June 26, 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 most important word in AI safety — "aligned" — has no fixed meaning. You can never satisfy it because there is nothing fixed there to satisfy. You're told to make the model good, helpful, safe. Try to pin any of those down at the speed and scale of an autonomous agent and it slips through your fingers, so you run another eval, and the anxiety never resolves. But there is a part of the question that does have a fixed, checkable meaning: did the agent stay in the lane it was hired for? That part is decidable. And here is the claim that reorders the field: the decidable part is not the small thing you settled for when the big thing was too hard — it is the only part that means anything. Decidability is meaning. Are you out of your pixel? is a question with an answer. "Is the AI aligned?" mostly isn't.
"Decidable" sounds like a compromise — the narrow slice you accepted because the whole was intractable. It is the reverse. The undecidable part has no stable meaning to hold an agent to; the decidable part is the part you can actually hold it to. We didn't settle for the small thing. We found the part that means something.
Most people hear "undecidable" as "hard," a tooling gap we'll close with more compute. It is not. Rice's theorem (1953) says every non-trivial semantic property of a program — any property about its behaviour rather than its syntax — is undecidable. There is no general procedure that decides, for an arbitrary program, whether it is "correct," "safe," or "good." And that is not merely a statement about what we can compute; it is a statement about meaning. A property with no decision procedure has no computable extension — no stable line you could actually draw, by machine, between the cases that satisfy it and the cases that don't. Maybe the line exists somewhere in principle; the point is you cannot compute it, and for an autonomous agent acting at machine speed, a meaning you cannot compute is the same as a meaning that isn't there. Ask ten careful people where exactly "helpful" ends and "obsequious" begins and you'll get ten lines — not because language is empty, but because there is no fact of the matter a program can decide against. The undecidable term isn't a meaning you're failing to measure. At autonomous speed and scale, it has no checkable meaning at all.
This is why "more evals" never lands the plane. You cannot converge on a target that has no fixed location. The endlessness of alignment-as-quality is not a sign you need a better grader — it is the signature of a question with no decidable answer.
Now turn it around. Computable measure theory gives a clean equivalence: a property is decidable exactly when its characteristic function is computable — when there is an actual procedure that returns yes or no for every case. And a measurement is nothing more than the evaluation of such a function. So decidability is not a weaker, sadder kind of meaning. It is the precondition for any meaning you can verify. If you cannot decide a property, you cannot measure it; if you cannot measure it, you are not holding the agent to a meaning at all — you are holding it to a feeling, and dressing the feeling in a number. The whole game, then, is to find the property of an AI's work that is both load-bearing and decidable — the one with a computable characteristic function — and price that, instead of forever litigating a word that was never going to resolve.
Meaning you can check = a decidable property = a computable characteristic function = a measurement. That chain is the same chain. Break it anywhere and you are back to vibes. The question is not "can we measure alignment?" but "which part of alignment is even a measurable thing?"
So which part is decidable? Not "is the work good" — that stays the model's problem and yours to judge. The part with a fixed, computable meaning is where the work landed: did the agent stay in the domain it was hired for, or fire in one it never declared? That is a question about placement, and placement is decidable. Two finite texts — what the work declared and what it did — are placed on a fixed 144-node lattice; the distance between them is read as a king move (the Chebyshev metric, where a diagonal shift weighs the same as a lateral one, so the simultaneous small-domain, small-role drift can't hide); and the answer is a coordinate. The walk that does the placing is real and physical — the recursive definer walk on a chip, about 14 milliseconds, no model in the loop — and that physicality is the why: the meaning isn't asserted, it's computed, signed, and recomputable by anyone offline. The pixel is that coordinate. Out of your pixel is the decidable event: the agent left the lane, and we can tell, byte for byte.
The definer walk is not a metaphor for the meaning — it is the meaning, made physical. A coordinate you can recompute on your own machine is a meaning no one has to trust. That is what it looks like when semantics runs on silicon instead of living in a debate.
Here is the part that's in it for you, whichever seat you hold. If you run an enterprise, your board wants autonomous AI deployed and you are the bottleneck, because one agent touching the wrong system is an extinction-level compliance event — and you cannot get comfortable on an undecidable promise that it'll behave. Stop trying. Hold it instead to the decidable thing: a cryptographically sealed, recomputable receipt that it stayed in the lane you assigned. You stop being the bottleneck and become the enabler. If you write insurance, the reason you exclude agentic AI is that you can't price a feeling of safety — but a decidable event is a countable event, and a countable event is a frequency, and a frequency is a premium. The decidable slice is the loss-data input you've been missing. Run on real agents, the in-lane breach rate lands near 10 percent today, with a tight interval — a number, not a fear. Same primitive, two readouts: permission to deploy, and a line you can finally write.
You don't need a smarter eval. You need a decidable boundary — and the receipt that proves the agent stayed inside it, recomputable on your own machine, your IP never leaving it. That is the "go" button, and the priced line, and they are the same object read two ways.
There is a deeper turn here, and it is worth saying plainly because it dissolves a problem the field has carried for forty years. The symbol-grounding problem asked how a symbol gets its meaning — and the answer everyone kept reaching for was understanding, some inner comprehension that ties the word to the world. But understanding is undecidable, so the problem never closed; it just generated more philosophy. Flip it through what we've said: the part of meaning that can actually be grounded is the decidable part — placement, not worth. You don't need the machine to understand "tax law" to decide whether its output landed in the tax-law lane or the marketing lane; you need a fixed map and a recomputable walk. Language has more in it than placement — of course it does — but the more is precisely the undecidable part, the part you cannot hold an agent to. The part you can build on, insure, and check is placement: the public, computable fact of where a thing sits relative to everything else. So this is not "meaning is nothing but coordinates." It is narrower and stronger: the slice of meaning that is groundable — that a stranger can recompute without trusting you — is placement, and decidability is the floor it stands on. We didn't solve understanding. We found the part of meaning that doesn't need it, and that part is enough to insure.
The undecidable term tends to have no meaning, full stop. The decidable placement has a meaning anyone can recompute. So the "narrow" slice was never narrow — it was the load-bearing floor, and everyone walked past it looking for the ceiling.
Three honest questions before you act on a claim this large. Why is there no alternative? Because the alternative is to keep grading the undecidable, and Rice's theorem says that target has no fixed location — no amount of compute converges on a thing that isn't there. The only move is to change the question to a decidable one. Why is this not too good to be true? Because it is strictly bounded and says so: it prices placement, not quality; it names its open inch (a reworded sentence reads coarsely, on purpose, because surface paraphrase is the model's job, not the fence's); and the numbers are measured and recomputable, not asserted — a blind cross-domain separation of 0.90, off-domain work rejected 10 times out of 10, a scrambled-null around 4.5 sigma. Why has nobody else done it? Because the whole field was looking inside the box for understanding — the undecidable problem — instead of placing a rigid geometry around the output, where the question becomes decidable and the answer becomes a coordinate. The mechanics of the decidable slice are in The Decidable Slice of Alignment; what it means for the people holding the bag is in The Exclusion Is the Liability.
Stop chasing the word that won't resolve. Find your pixel — the coordinate where your work is supposed to land — and decide, for once, whether it stayed there.
Decidability is meaning. Find the coordinate, recompute the receipt: thetadriven.com/pixel. The part of alignment you can actually check was never the narrow part — it was the only part that ever meant anything.
![tolerance panel for commit 1910ba5 — chore(blog): attach commit tolerance panel as OG image [panel-attached]](/blog/post-panels/2026-06-26-decidability-is-meaning/1910ba5de48b98a26ec2e2a50f2a5ae4c56dbe46.png)
