Two pupils,
and one is averaged away.
Deep in the brainstem sits the locus coeruleus — the brain’s main noradrenaline source, and the sole supplier to the cortex. You cannot see it fire, but it drives the sympathetic tone that widens the pupil, so the black disc becomes a needle you can read across a room, and there are two of these needles, which don’t always read the same.
A pupillometry conjecture · DR McCulloch
Most instruments measure one pupil, or average the two into a single diameter. But averaging is a low-pass filter: the mean survives, the left-right difference is discarded. If the arousal asymmetry is what matters, as my own eyes seem to show under load, then the instrument erases it before anyone looks. I built an instrument to make that erasure visible, and set it beside the eyes it happened to: my own, under load.
WATCH THE ERASURETwo systems, one aperture
The pupil is not a muscle but a hole, the gap left by two opposing muscles in the iris, each wired to a different half of the autonomic nervous system. Nothing about a pupil reading is ever a single quantity. It is always a subtraction: the parasympathetic ring pulling it shut against the sympathetic mesh hauling it open. And the opening limb runs mostly ipsilateral, so the two eyes can widen by different amounts.
Sphincter pupillae
Two things close it. The light reflex, in which more light on the retina brings more sphincter tone and a smaller pupil, is a fast subcortical loop, but the sphincter also fires in the near triad: focus on something close and the pupil constricts along with convergence and accommodation, and that response does run through the cortex. So even the constrictor is not one mechanism.
Dilator pupillae
The arousal limb, and it is largely ipsilateral: each side’s sympathetic chain drives its own pupil. Central arousal, noradrenaline from the locus coeruleus, recruits this limb under load, and because the wiring is ipsilateral it can widen one eye more than the other.
The light reflex is just that, a short loop that never touches the cortex, whereas the arousal limb is the one that reaches it. It may carry, however faintly, a lateralised trace of how hard each side is being driven, though, as I set out in the honest limits, that pupil-to-hemisphere mapping is an assumption I lean on, not a fact.
The puzzle that makes the load
Escalating mental arithmetic is what actually drives the dilation. It comes from my own session, “High IQ with low left locus coeruleus activation”, 12 February 2026, a piece of mental arithmetic built to load one thing and one thing only. The base task: 347 → subtract 58 → ×3 → add 167 → ÷7, which runs 347 → 289 → 867 → 1034 → 147.71.
Every single operation is trivial given time, which means arithmetic ability is deliberately not the variable under test. The variable is maintenance: how long you can hold a running number while the next operation runs on top of it.
Load builds across the steps until a division by a non-factor (7 into a four-digit number) forces you to hold the accumulated intermediate result while running a fresh sequential sub-calculation, the long division. That is the step where the held number goes.
And it doesn’t fade so much as go: the sustained-firing representation holding that number collapses rather than decays, there one moment and gone the next.
Pick a tier and step through the chain, watching the right pupil climb until it goes.
The pupil here is the readout, not the target. Escalating maintenance load drives task-evoked dilation through the LC-NE system; on my own reported asymmetry the right pupil tracks the climb and peaks at the collapse step. What an instrument does with two pupils that read differently →
The measurement problem
Here both pupils are driven live. LIGHT works the reflex and raising AROUSAL widens the two of them together, while COGNITIVE LOAD is the dial that pulls them apart: it models the right hemisphere recruiting and dilating the right pupil past the left, the same asymmetry my own eyes show under load. At rest they read the same, so choose what the instrument is, whether the true bilateral state, a binocular device that averages or a device that samples one eye, and watch |L−R|, the gap between the two pupils, vanish the moment you stop measuring both.
These pupils move on a model of ipsilateral recruitment, not on fitted data, but at the real scale: at full load the gap between them reaches about 0.7 mm, a touch beyond the widest asymmetry I measured in the cohort (~0.58 mm), and still under a millimetre. A difference that small is trivially erased by averaging the two eyes, and it is exactly where the group difference lives.
My own eyes, under load
A sequence of close-ups of my own eyes across a piece of mental arithmetic, annotated in my handwriting, blue dots for resting bilateral tone turning red as asymmetric demand climbs, thin rules drawn between the two pupils to hold the gap still. Read it top to bottom: at baseline the pupils are symmetric, but by step 3 the right pupil is clearly the larger, and it widens further until the task collapses, “lost it”.
- baselinereading the task, eyes level, convergent, pupils symmetric. both LC nuclei at resting tonic output. blue dots bilateral; no asymmetric demand, the system looks resourced.
- step 1–2gaze shifts upper-right. left dlPFC engaging for verbal computation. pupils still roughly symmetric. single operation, low maintenance demand. says “789” aloud. done.
- step 3strabismus more pronounced. dots redder. right pupil clearly dilated. gaze drifting upward-left as right-hemisphere visuospatial strategy recruits the left verbal…
- step 4tries to hold it with a gaze upwards right, right pupil visibly larger
- step 5 → lost itthe asymmetry peaks, then the task collapses.
This is one person in one session on one task, with a real strabismus muddying the very divergence it points to, so some of what you see is my eyes rather than my arousal, and it settles nothing on its own. It’s where the conjecture came from, not evidence that the conjecture is right.
Tonic, and phasic
Adaptive gain theory splits locus-coeruleus firing into two registers, and the pupil carries both. The slow tonic baseline is the resting diameter. The sharp phasic transient is the dilation that rides on top when something task-relevant lands: a decision, a target, the moment the sum comes.
At a low tonic baseline the phasic bursts are crisp and well-timed: focused engagement, exploiting the task in front of you. Push the tonic baseline high and the pupil sits wide and jittery while the phasic responses flatten: restless, scanning. Both my pupils carry this in time, and they carry it differentlyfrom each other.
The artefact in the cohort
Consider two arousal signatures with the same averaged pupil and opposite laterality. A binocular device logs one number for each, and the two numbers are almost identical. Everything that distinguishes them lives in |L−R|, the exact quantity the average throws away.
This is a broadly raised, roughly symmetric arousal signature whose two pupils agree to within a tenth of a millimetre, so their mean carries it faithfully and a binocular device sees it fine.
Here the right pupil out-dilates the left by half a millimetre, and the two eyes look almost identical, which is exactly the problem. The mean is the same 4.35 mm as the symmetric case, so averaging cannot tell them apart. What told them apart lived in |L−R|, and |L−R| is what averaging deletes.
There is a prediction here, and it has already been observed. If autism’s LC-NE signature lives partly in lateral asymmetry, an averaging instrument should track ADHD (the symmetric, global-arousal story) more strongly than core autism. That is precisely what Kim et al. (2022) found: their LC-NE pupil measures correlated with ADHD symptoms more than with autism-specific symptoms. The usual reading is “autism has a weaker LC signature.” I read it the other way:the measurement cannot see the axis autism’s signature sits on. It’s still a conjecture rather than a result, though a testable one: stop averaging and look at |L−R|.
From the data
I re-analysed the Cilia binocular eye-tracking recordings (27 autistic and 29 typically-developing children) from the raw per-eye traces, keeping the left and right pupils separate and never averaging them. Every child watched the same videos on the same clock, so at each moment I subtracted the cohort's own mean pupil diameter, computed from the other children at that video and that timepoint. That removes the part of the response common to everyone, the light reflex, and leaves each child's departure from the group, in millimetres.
In the raw recordings the two groups' within-trial pupil variability is the same: right-eye within-trial standard deviation 0.215 mm in the autistic group and 0.209 in the typically-developing group (Cohen's d = 0.06, p = 0.83). After the shared response is removed they separate. The autistic children's residual standard deviation is higher, 0.222 versus 0.171 mm (d = 0.94, p = 0.001), and so is their phasic amplitude (d = 0.93, p = 0.001). Removing the common response lowers the typically-developing children's variability, because their pupils track the shared stimulus; the autistic children's stays high. The shared response was inflating the variance within each group and hiding the difference between them.
The difference is present in both eyes (left eye d = 0.77). It gets larger, not smaller, when I restrict the comparison to boys to remove the sex imbalance in the sample (d = 1.21, p = 0.0004). It holds when the reference is built only from typically-developing children, so it is not an artefact of comparing a child against itself. It does not correlate with autism severity (CARS: all |ρ| < 0.2).
The left-minus-right difference needs no cleaning: light constricts both pupils together, so the reflex cancels in R−L: the reference's own left–right difference is 0.06 mm against a 0.15 mm effect. Its upper tail is heavier in the autistic group. The 99th percentile of |R−L| is 0.87 versus 0.67 mm (d = 0.78, p = 0.007), with about twice as many left–right divergences held for 150 ms or longer. Inter-eye correlation falls during those divergences in both groups.
Twenty-seven and twenty-nine children is a small sample, and the autistic children contributed roughly a third as much usable data. These are passive video recordings, not a titrated cognitive load, so this is a trait-level difference in how these pupils move, not a measured response to being pushed. Nothing scaled with severity, and the direction of the asymmetry is inconsistent (near-symmetric on average), so this concerns the size of the left–right divergence, not a fixed weaker side.
Averaging the two eyes, or pooling across conditions, removes exactly the between-child variance where the autistic difference sits. It is undetectable until you stop averaging and take the shared response out.
The median, and what it hides
The difference I found only appears because I stopped taking the median. A median, or any mean, is a decision to keep one number and throw the rest away: the spread, the moments, the two sides. Often that is the right decision, since a summary has to leave things out to be a summary. But a summary is also a claim about what matters, and the claim is usually made before the answer is known. Collapsed to a mean, the two groups were identical. The difference was in the variance, in how much each child’s pupil moved once the shared response was taken out, and the median is built to remove exactly that.
It is the same move whenever a person, or a system, is reduced to one measure: an IQ, a severity score, a diagnosis, a mean pupil size. Sometimes the measure serves a purpose. Sometimes it only marks how little we have understood, a placeholder for the thing we could not yet see. Either way the collapse is not free. It keeps the part of the person the measure was built to hold and drops the part it was not, and the part it drops is often where the person is.
So I kept the eyes apart and worked the difference and the tail, not the average. The average had already decided the two groups were the same.
A developmental conjecture
Why would autistic children's two eyes diverge further, a heavier tail of left–right difference, than typically-developing children's? One explanation, a hypothesis to test, not a result: it reads autism partly as a lateral difference, an asymmetry in activation between the two hemispheres, plausibly rooted in neural-crest laterality, the same neural-crest cascade the sibling Pilearytixmodel is parameterised on.
A left-right difference set that early does not stay local. In development it would bias connection across the hemispheres, and reduced interhemispheric connectivity is an established finding in autism (Anderson et al. 2011). That is the mechanism the conjecture leans on: not a missing region, but two halves that coordinate less tightly than they should.
The autistic group’s heavier tail of left–right difference (and, more weakly, a lower inter-eye correlation that does not reach significance here) would be a peripheral, bodily readout of reduced interhemispheric coordination: two hemispheres that coordinate less tightly, read out at the eyes as two pupils that diverge further and track each other less.
The Cilia-cohort children were not doing this effortful arithmetic ladder, they were young children simply viewing images and video. So the difference in that data is task-independent, near-resting. That is arguably cleaner evidence than my own effort-driven dilation: the heavier asymmetry appears with no imposed executive load at all. The developmental under-connectivity hypothesis predicts exactly this, a resting difference that does not need a hard task to appear, and it therefore does not inherit the general-effort confound that weakens the arithmetic probe.
How loose the needle really is
The instrument is seductive, and the seduction outruns the evidence. So I have to make the case against my own argument. Joshi and colleagues (2016) recorded the locus coeruleus directly and found pupil size does track its firing moment to moment. But the more careful work qualifies it hard: Megemont, McBurney-Lin and Yang (2022) titled their paper, flatly, “pupil diameter is not an accurate real-time readout of locus coeruleus activity.”
And my conjecture is exactly that, a conjecture. It rests on n = 1: my own eyes, one task, one sitting, with a strabismus that muddies the very asymmetry it points to. The pupil sums the light reflex and the arousal limb into one number, lags by hundreds of milliseconds and reflects several neuromodulators at once. It is a needle on a very soft spring, two of them, and I am claiming the difference between them matters.
Two mechanisms, one curve
The puzzle cannot separate the left-LC-deficit story from ordinary working-memory span. Held intermediate results decay past a load threshold in everyone. “It disappeared” is the standard signature of phonological-loop and central-executive overrun, and does not, on its own, require alpha-2A-to-alpha-1 noradrenergic spillover or a lateralised locus coeruleus lesion.
The dilation is a weaker probe still. Task-evoked pupil response indexes central effort in general, not left-hemisphere engagement, and not LC integrity.
My own earlier result already forced this. The right pupil dilated for verbal, visuospatial and planning tasks alike, which killed the contralateral-content reading and left “right prefrontal executive effort” as the referent. It is the same right-pupil pattern the EYES sequence shows (Act IV), and I have to be honest that it points to general right-PFC effort, not left-LC specificity.
So the instrument demonstrates load-linked dilation and a maintenance ceiling, but it does not localise that ceiling to the left locus coeruleus or to noradrenaline, because two mechanisms fit the one curve equally well.
The arithmetic-task dilation (my own eyes, the puzzle above) is the weak, effort-confounded probe. The stronger evidence needs no task at all: the cohort re-analysis (Act VII), where autistic children show higher within-trial pupil variability once the shared response is removed, and a heavier tail of left–right difference, both in passive viewing with no imposed load. The honest position keeps the two apart, and rests the claim on that re-analysis, not on the pupil that happened to widen while I did the sum.
“Lateral stimulation shouldn’t just average… that isn’t how brains work.”design note · the sibling piece Pilearytix
Pilearytix is the model of this: a whole arousal engine built hemisphere by hemisphere, right-dominant, its two connectomes never collapsed into one. This page is the body that goes with it, two eyes kept apart so you can watch the number that erases them do it in front of you.