If you’ve ever fed a compiled binary into a decompiler, you know it’s a small miracle that it works at all. The names are gone, the comments are gone, the elegant structure the original programmer had in mind is flattened into a soup of instructions — and yet the tool hands you back something legible. Loops become loops again. Function boundaries reappear. Types get reconstructed. You can read the thing.

Why does that work? Not because decompilers are magic, but because the binary was authored. Somewhere upstream there was a shared source language: a type system, calling conventions, a standard library, a compiler that mapped source to machine code the same way every time. The decompiler doesn’t recover the original; it recovers a program in the shared language that the binary provably speaks. The legibility was baked in before the compilation ever happened.

Now try to decompile a brain — mine, say, while I’m looking at a ripe tomato — and read off what my red is like.

Here the miracle fails, and it fails for a specific, non-mystical reason: there was never a shared source language. Nobody wrote my visual system in a common tongue and compiled it down. The mapping from my neural states to my felt experience wasn’t authored at all — it was grown, sub-symbolically, over decades, by one particular developmental trajectory, one sensory history, one connectome that exists in exactly one copy. There are no calling conventions. There is no type system. There is no standard library of qualia against which you could resolve my “red” the way a decompiler resolves a call to printf. To read my red off my tissue, you would have to reverse-invent a private interpretive language whose vocabulary has exactly one speaker — and your only Rosetta stone is the external world we happen to share.

That, I think, is the real shape of ineffability. Not a magic seal. A missing shared source.

The puzzle, stated carefully

Let me be precise, because two different problems get smuggled under one word.

There is privacy — the question of who can have a given mental state. Can you literally host my experience, or only your own? And there is ineffability — the question of what can be conveyed. Even granting that we each have our own experiences, why can’t I hand you mine in words? This post is only about the second. Privacy is a separate essay.

So: suppose you and I agree on every relational fact about my color experience. Red is more like orange than like blue. It’s the color of the tomato, of arterial blood, of the stop signal. It contrasts with green and shouts louder than pastel. We agree on all of it — and still, I cannot convey to you what it is like for me. Something survives the agreement. What is it, and why won’t it move?

Two standard answers, both of which I reject

The first is brute pessimism: experience is just ineffable, full stop, and there’s nothing more to say. I find this both lazy and, more importantly, false — because it predicts nothing, and where it does implicitly predict, reality contradicts it. If my inner color space were a sealed private theatre with no shareable content, researchers should not be able to line up the structure of color experience across different people. But they can. Using representational similarity analysis, hyperalignment, and — most strikingly — unsupervised Gromov–Wasserstein alignment of color structure (Kawakita and colleagues), you can match one person’s color relations onto another’s with no shared labels at all, using nothing but the internal geometry. Something real and shared is clearly getting out. Brute pessimism can’t explain why.

The second answer is deflationary illusionism, in Keith Frankish’s sense: there is no determinate “what it’s like,” the felt residue is a kind of introspective user-illusion, and once you see that, the puzzle dissolves. I respect the discipline of this move, but it buys tidiness at too high a price. Nearly every human being who reflects on their own experience reports the same thing — that there is a specific this-ness to their red that they cannot pass on. To deflate the puzzle you have to treat that near-universal report as systematic error. I would rather take the report seriously and explain it than declare the witnesses collectively confused.

Where my theory enters

The Four-Model Theory says, in one honest paragraph: consciousness is an ongoing self-simulation. There are two axes of models — what they’re about (the world, or the self) and how they’re built (implicitly learned, or explicitly generated) — which gives a 2×2 of model kinds. A system becomes conscious when its explicit self-model closes the loop on the very system that is running its explicit world-model, and does so at criticality — the edge-of-chaos regime where a system is maximally sensitive and expressive without tipping into noise. Qualia, on this view, are level-constitutive properties of that running simulation: perfectly real at the computational level, and incoherent as claims about the substrate. The value in a spreadsheet cell is real — you can compute with it, it has consequences — but “which transistor is the number 42?” is not a question with an answer. Asking which neuron is my red makes the same category mistake.

That last point is what lets me split ineffability cleanly in two.

The clean split

Relational structure — red-is-more-like-orange-than-blue — is a substrate-independent invariant. It’s a pattern of relations, and patterns of relations are exactly the kind of thing that can be lifted off one substrate and matched onto another. This is why the alignment work keeps succeeding. Structure is the decompilable part. It travels.

Absolute encoding — the specific state that is my red, this-particular-realization — is the grown private code. It has no meaning outside the one system that grew it, because there is no shared language it was ever written in. It is refractory to description. You cannot get it in words, not because it’s hiding, but because words route through the shared world, and the absolute term never passed through the shared world on its way into me. The only way to have it is to instantiate it. It can be run; it cannot be recited.

Mary, one more time

Frank Jackson’s Mary knows every physical fact about color from inside her black-and-white room, then steps out and sees red for the first time. Does she learn something? The endless fight is over yes-or-no. My theory says: of course — but be precise about what.

Everything Mary could be told is the relational geometry — the entire structure, complete and exact. What she gains only by seeing is an absolute term, now realized in her own substrate for the first time. The residue she picks up on stepping outside is not a new fact about the world; it’s the instantiation of the one thing that was never conveyable in the first place. Mary’s Room isn’t a knock-down argument for or against physicalism. It’s a clean demonstration of exactly which half of experience refuses to travel in words.

The payoff — and it’s testable

Once you split it this way, ineffability stops being a mystical absolute and becomes a graded, directional asymmetry: structure flows freely between minds, absolute encoding does not, and the gap runs along a seam you can point to.

And here is the part I care about most, because it exposes the theory to being wrong. On the very same data where you can align two people’s color structure, a decoder that tries to read one brain’s absolute “red” off the other should fail — and, crucially, should keep failing even after the structures are perfectly aligned. Structural alignment should buy you no traction on absolute cross-brain decoding. I know of no rival framing that predicts that exact pairing: success on structure, persistent failure on the absolute term, on one dataset, with the failure surviving alignment. Brute pessimism predicts failure everywhere; illusionism predicts there’s nothing there to decode. Only the split predicts the asymmetry.

Where this stops

I want to be honest about the borders, because a theory that claims everything explains nothing.

This explains the ineffability of report — why felt experience can’t be handed over in words. It does not explain the existence of feeling. Why there is something it is like to run this simulation at all remains the hard part, and I am not pretending otherwise. On the Hard Problem, the Four-Model Theory’s standing is exactly what it was before this argument: partial. Nothing here moves that needle.

And “can’t be conveyed” is not “can’t be transferred.” I’m making a claim about description, not about metaphysical sealing. If you could re-instantiate my absolute encoding in another substrate — actually grow or copy the private code — the experience could in principle come along. What can’t happen is compressing it into language and decompressing it in your head. No shared source code, no compilation.

Not a rival — a complement

So I don’t see this as a competitor to the structuralist theories of consciousness. They own the shareable half, and they describe it beautifully — the geometry, the invariants, the alignments. What I’m adding is the account of why that half is shareable in the first place, why the other half isn’t, and exactly where the seam between them runs.

The structure decompiles because, at the level of relations, we do share a source language. The absolute term doesn’t, because there was never a standard library of qualia — only a world we both look at, and two private compilers that were each grown exactly once.