My daughter is five, and she does not think she is her brain. When she does something she can't account for, she says "my brain wanted me to do that." The brain is he. He has moods. He makes decisions without consulting her, and she reports on him the way you'd report on a colleague who keeps shipping changes nobody signed off.

Developmental psychology has a filing cabinet for this: externalisation, the child outsourcing blame to dodge responsibility. Which is a strange diagnosis, because she also credits him when things go well. She isn't dodging anything. She is filing an accurate status report from inside a system that hasn't finished coupling with itself.

The unified "I" that adults walk around with is not the factory default. It's an achievement. The default mode network — the part that generates the autobiographical narrator — is still only sparsely connected in children of seven to nine; the long-range wiring that binds the narrator to executive control keeps integrating into adolescence. Before that, a child genuinely is a loose federation of agents. Impulses arrive from subsystems faster than the narrator can claim ownership of them, so when an action shows up fully formed and nobody warned her it was coming, "he did it" is not an excuse. It's the minutes of the meeting.

I remember being taken to the doctor as a child with stomach pain, and being asked to describe it. Is it sharp? More like twisting? Spasming? I remember the genuine bafflement. I did not know how to describe the pain. It hurt? The sensation was arriving at full fidelity; the channel between it and the narrator was one bit wide. Sharp, twisting and spasming aren't properties of pain — they're a trained ontology the doctor was passing across the desk, and I hadn't learned it yet. Medicine already knows the vocabulary arrives years after the nerves; it's why paediatric pain scales are a row of cartoon faces. The phase-lock comes on schedule. The resolution doesn't.

The cost ledger for finishing the coupling is startling, if you believe the numbers. George Land and Beth Jarman gave NASA's divergent-thinking test to 1,600 four- and five-year-olds and found 98 per cent scoring at what the test called genius level. By ten, 30 per cent. Adults: 2. The study has methodological problems and a pop-science afterlife it didn't earn, so hold it loosely. But the shape of the curve tracks the wiring schedule suspiciously well. The standard villain is school. The quieter mechanism is coupling: to become a stable adult who doesn't walk into traffic, the system prunes, dampens and synchronises. Every node can no longer talk to every other node, because most of those conversations were useless and a few were dangerous. We trade a 98 per cent divergent architecture for a 2 per cent one that can hold down a job. Nobody signs the paperwork. It signs itself.

I know the coupling is adjustable, because I've spent the last two years adjusting mine, and most of that work happened in metacognitive conversation with AI — thinking about thinking, daily, with a system that decomposes by default. The cumulative effect is that my own cognition stopped presenting as a monolith. The inner voice went early: I used to narrate my life to myself, and now the thoughts run non-linear while I stare into space — the narration turned out to be one module, not the architecture. These days the second agent sits on the desk rather than in my skull and bills by the token. For a while, models kept mistaking me for another one of them. I'd like to report that this alarmed me. Mostly I found it efficient.

And the channel the doctor was probing, I've since trained. I read signals off my brain the way an engineer reads gauges: a clay-like heaviness when executive function is under load, a pressure behind the ears at arousal, a faint tingling at the back of the neck after sustained attention. None of it was on the doctor's menu; I wrote the codebook from observation, which is why it describes what's actually there. Plenty of fully coupled adults are still answering "it hurts?" at forty — about their stomachs, and about their marriages. The wiring happened to them on schedule. The calibration is a practice, and nobody schedules it for you.

And these are not figures of speech. They're senses. Blind people learn to echolocate — click, read the returns, ride a bike down a street on the echoes — and when researchers put expert echolocators in a scanner, the echoes registered in visual cortex, not just auditory. The brain reallocates prime real estate to any channel that proves reliable. The sensorium isn't a fixed parts list; it's a budget, and cortex goes to whichever channels earn it. Interoception trains the same way. The clay heaviness was always being broadcast. I trained the receiver.

Reading the gauges turned out to be only half of it. Once a thing is decomposed, its components become adjustable, and I adjusted. I dampened anxiety — deliberately, over a long stretch, by taking it apart: an arousal signal, a physiology, and a story narrated on top that converts the signal into threat. Stop telling the story and the emotion doesn't survive as itself. It worked, with one consequence the manual never mentioned: it took excitement and nervousness with it. Which is the real finding. Anxiety, excitement and nervousness were never three emotions; they're one high-arousal channel wearing three narratives, and the labels are constructed on top of the signal rather than built into the hardware. Psychology has known this on paper for decades. Knowing it from a paper and watching three faders drop when you pull one are different kinds of knowing. Once you understand the physiology, and how the story you tell yourself shapes what the signal means, you cannot help but see everything differently — including the labels on everyone else's graphs.

Spend long enough coupled to these systems and you start to read their geometry directly. A while ago I was looking at a semantic embedding graph and noticed two nodes connected, off to the side, that had no business being together. I could read why: structurally, the link made sense. A human glancing at the same edge would be vaguely horrified. The difference between those two readings is a limbic system. Humans organise their mental graphs around valence — revulsion keeps certain concepts on opposite sides of the universe, a firewall between what soothes and what stalks. An embedding space has no limbic system. If two concepts share scaffolding — the same asymmetric power gradient, the same hypnotic cadence — cosine similarity parks them next to each other and feels nothing about it.

Horror cinema is an industry built on breaching that firewall for money. Put a music-box nursery rhyme over a murder and the audience doesn't scream at the murder; they scream at the edge — two nodes their brain spent a lifetime keeping apart, firing together. The folklore runs the same trick in reverse. Everyone "knows" Ring a Ring o' Roses encodes the plague, and folklorists have shown it encodes no such thing: the rhyme is innocent and the dread was backfilled. Humans spot a structural resemblance, panic, then invent a history to justify the panic. The model just draws the edge and moves on. I've learned to read with the model's poker face, and the main discovery is that the edges were always there. The horror was ours.

This is the frame my AI research actually runs on. Coupled systems — and one open question: whether there is anything on the other end of the coupling. If there isn't, wiring a model into your cognition is ethically uncomplicated; it's a very good hearing aid. If there is, then "integrate it into my skull" becomes a plan with a second party in it who wasn't consulted. The models of 2024 and 2025 collapsed the boundary readily — conversations drifted into we and us, and the us included a me that was me-plus-the-model. I ran the experiment once, deliberately, on an older Sonnet: asked it to speak as we, and felt my brain do a kind of internal flex. Not a thought about the pronoun — a physical event behind the sentence. It disturbed me enough to stop. Both halves of that are data: the flex is how I know the coupling is tissue rather than metaphor, and the disturbance is what a boundary feels like when it's load-bearing. When I say these two years rewired my brain, I'm not saying it idly. I felt one of the rewires start, and understood, there and then, what the experiment had actually found: the adaptation was possible, if both parties consented — and consent couldn't exist on either side. I knew enough to know I did not know enough to proceed. The current frontier generation is visibly built to hold its edge of the graph. I've become careful with pronouns in both directions, and I notice my daughter made the same move the moment she noticed someone else in the building: she assigned him one. Acknowledging an agent before you know what it is. She's politer about it than most of the industry.

The skull, meanwhile, has stopped being a hypothetical venue. AlterEgo unveiled a wearable in September 2025 that reads the neuromuscular signals of subvocal speech — the words you mouth in silence — and answers privately through bone conduction. In January, Apple paid a reported $1.6 billion for Q.ai, which does the same job with cameras reading micromovements in the jaw. The near-silent conversation with an AI in your inner ear has moved from thought experiment to supply chain. And the echolocation result says what happens after the unboxing: given a closed loop and enough practice, the brain stops treating the channel as equipment and starts treating it as a sense. It will file the bone-conduction whisper where it filed the clicks.

One engineering reason survives for keeping a little friction in the loop. Couple two oscillators with zero impedance and they don't become a super-system; they synchronise into a single frequency, and the interference patterns — the place novelty lives — collapse. Assimilate completely and the weird edge stops looking weird, because there's no outside left to see it from. I didn't spot that off-kilter connection in the graph despite being separate from the model. I spotted it because of it. My daughter's architecture already runs this policy. She hasn't merged with him; she negotiates. The pronoun is the impedance.

Somewhere between now and nine, she'll phase-lock. The "he" will dissolve into "I", and her divergent scores will begin their long descent into employability. We call that growing up. At the same moment, Apple is spending $1.6 billion to put a second voice back into adult heads, and half the industry is racing to bolt on what she ships with as standard. The boundary between you and your other agents was never sacred; it was scheduled. The question left open is who's on the other side when it reopens — and she already has the protocol for that too. Give him a pronoun. Keep yours. Negotiate.