Right superior oblique palsy · measured 2026-08-01
Roll the head and watch the two eyes counter-roll. The left one manages it. The right one cannot intort, so the further the head tilts onto the right shoulder, the further apart the two retinal images rotate — until fusion is arithmetically impossible. Then switch the correction on.
Work in progress — and the numbers will move
This is an open notebook, not a result. One person, measured by that person. Several numbers on this page moved on 2026-08-02, and one of them moved because the instrument was wrong rather than because the eye changed.
A hand on a dial stops near wherever it started: the four left-tilt points finished 0.20, 1.09, 2.81 and 5.90° from their opening offset, ordered by how long the subject spent rather than by head roll. So the app now does the moving and he only says when the image is single. Against that instrument, resting rotation reads ~4° rather than 5.5, and at the right shoulder nothing fuses at any setting — the 9–11.5° plotted there is where he stopped dialling, not where the images came together.
The vertical is in a worse state than this page previously implied. The model term predicting it was fitted from the headset’s vertical vergence, and that channel was later shown to be synthesised from head pose rather than measured; the fitted value was never reset, and it disagrees in sign with every vertical the subject has dialled. Whether applying the correction helps is still untested — and now known to have been untestable, because every attempt so far applied that vertical.
The examination itself, driven by this subject’s measurements. Each step halves the field of candidate muscles; after three, one is left.
Expected worse in left gaze — the right eye adducted, where the superior oblique does most of its lifting work. This rig measures torsion across gaze but has never measured the vertical across gaze, so the page states this step rather than demonstrating it.
Result Higher right eye, worse in left gaze, worse tilting right — right superior oblique. The three steps only ever point at one muscle, which is why the test survives with no equipment at all.
This demonstrates the test on one person’s measurements. It does not perform one, and it cannot tell you anything about your own eyes.
Each iris carries a marker — faint where it should sit, solid where it does. The right eye also rides up, and drifts a little toward the nose: same failing muscle, second and third consequences. Both are drawn life size — a few pixels, because that is what a few degrees is on an eye this big. The dashed references are what make them readable; the figures beside the eye are the measurement.
One line per eye. Fused when they lie on top of each other; doubled when they scissor apart.
The simplest possible scene: a level horizon. Each eye receives its own copy, rotated by that eye’s own torsion. Superimposed, the right eye’s horizon drops away to the left.
The green band is this subject's measured cyclofusional range, ±3.5°. Inside it the images can be pulled together; outside, they cannot.
Three separate things are being applied to the right eye’s image, and they are not equally well known. Bars are drawn in common angular units so they can be compared — 1 Δ = 0.573°.
Measured Across the whole roll range, 15 hand-dialled nulls over four sessions. Grows from about 3° at the left shoulder to 11.6° at the right. No spectacle lens can deliver this component at all.
Unresolved Hand-dialled at head level this reads 9.05, 8.40 and 6.63 Δ across three runs, against a prescription of record of ~11. But the newer instrument returns values of the opposite sign (+10.51, +5.34, +2.57, −0.00), and the model term that predicts it was fitted from the headset’s vertical vergence — a channel later shown to be synthesised from head pose rather than measured, and never reset afterwards. Applying it during a rotation measurement halved how often anything fused. Neither the size nor the sign is settled, and how it changes with head roll has still never been measured.
Barely measured Around 1 Δ base-out — the eye drifts toward the nose, because the muscle covering for the palsy also adducts. And inconsistent: 0.61, 2.48 and 0.00 across the same three runs. Small enough that it may be nothing at all.
A prism can already carry the lower two. The top one is the reason this needs a headset — and at every head position it is the largest of the three.
The curve is fitted; the ticks are corrections this subject dialled by hand across four sessions. Nothing here comes from the headset's eye tracker, which was found to be synthesising its per-eye vergence.
Alfred Bielschowsky described it in 1935, and it is still how a fourth nerve palsy is identified in a clinic — usually as the third step of a three-step examination. It needs no equipment at all.
Step one: with the head straight, which eye sits higher? Step two: does that difference get worse looking left, or looking right? Step three — the head tilt: does it get worse with the head tipped toward one shoulder or the other? Each step halves the field of candidate muscles, and after three steps one muscle is left.
The third step is the clever one, and it is the step this page is about. Tilting the head demands that the eyes counter-roll — and intorting the right eye is a job shared by the superior oblique and the superior rectus. When the superior oblique is weak, the superior rectus has to supply the missing twist, and because it is also an elevator the eye rises as a side effect. So the eye that was already higher goes higher still.
A right superior oblique palsy therefore reads: right eye higher, worse looking left, worse tilting right. That last part is the positive Bielschowsky sign, and it is why a person with this palsy holds their head toward the other shoulder without ever being taught to.
What the page adds is resolution. In a clinic the third step is three head positions and a judgement of “worse”. Here it is a continuous sweep, and instead of a judgement the subject dials the image until it goes single — so “worse” becomes a number of degrees, at every angle in between.
When the head rolls, the otoliths command ocular counter-roll — the eyes rotate against the head to keep the world upright. It is involuntary, and it only ever recovers a fraction of the roll, about 15%.
Rolling toward the right shoulder asks the right eye to intort. Intorsion comes from two muscles: the superior oblique — the palsied one — and the superior rectus. The right eye is asked for a rotation it cannot make, so it falls short, and the two retinal images rotate apart.
Rolling toward the left shoulder asks the right eye to extort instead, which the inferior oblique and inferior rectus supply perfectly well. The failing muscle is simply not called on, and the mismatch shrinks — from 11.5° at the right shoulder to about 3° at the left.
That is why the head tilts left. It is not a habit. It is the one posture that moves the mismatch back inside the fusion range, found without instruments, and paid for in neck.
No tracker made or sold measures the sign of cyclotorsion — it needs iris-pattern tracking, which no headset does. So the direction the right eye’s horizon falls is taken from what you actually see, and every panel here follows it. The magnitude is measured; the sign is testimony. That is worth stating plainly, because it is the one number on this page no instrument could have supplied.
To make up the missing intorsion the brain recruits the superior rectus — which is also an elevator. So the right eye lifts as it tries to twist, and a vertical deviation grows alongside the torsional one. That is the mechanism behind Bielschowsky’s head-tilt test, and it is drawn in the first panel: the right eye’s pupil rides up inside a globe that stays put, which is what a hypertropia actually looks like. It slides a little toward the nose at the same time, because the same muscle also adducts. Both are drawn at true scale: there was a 2× magnification here to make the rise easier to see, and it has been removed, because a drawing that magnifies one axis is no longer a drawing of the measurement. A few degrees on a 27-pixel eye is a few pixels, so the panel prints the figures in prism dioptres — up and in — beside the arrow.
That half is a prediction. The 8 Δ at head level is measured — 9.05, 8.40 and 6.63 across three runs, dialled once the rotation was already right. How it grows with head roll has never been measured, because the channel meant to measure it was synthesising its output. The animation assumes it scales with the torsion, both coming from the same failing muscle. That is a reasonable guess and nothing more, which is why it is drawn hatched wherever it appears.
There is a sideways part too, and it goes the way most people guess wrongly: inward. The superior rectus is not only an elevator — its secondary action is adduction — so the recruitment that lifts the eye also pulls it toward the nose, and the correction of record carries 1 Δ base-out, which is what an inward deviation is given.
It is now drawn, at the same true scale as the rise so the two stay in proportion — and the result is that you can barely see it. That is the honest picture. The horizontal came out around 1 Δ and inconsistently: 0.61, 2.48, 0.00 across the same three runs, against 8 Δ of vertical. An eighth of the rise is about half a pixel of travel, so the number is printed beside the eye rather than left to the drawing. Exaggerating it until it showed would have made the smallest of the three components look like one of the big ones, and it may yet be nothing at all.
A prism bends light up, down or sideways. No optic rotates a retinal image. That is why the torsional part of a fourth-nerve palsy is normally left uncorrected and carried by head posture — and why 27 prism dioptres, dialled at 37° of right roll, achieved nothing at all. It was the wrong control for the problem.
Two of the three components could be carried by ordinary spectacles today. The vertical — about 8 Δ, dialled at head level once the rotation was already right — is exactly what a prism does, and the prescription of record already calls for roughly 11. The horizontal, around 1 Δ, may be nothing at all. It is the rotation that has nowhere else to go, and at every head position it is the largest of the three.
A headset has two independent displays and knows its own roll against gravity, ninety times a second. So it can pre-rotate what the right eye is shown by exactly the amount that eye is going to fall short — about 5.5° at level, rising toward 11° at the right shoulder. Turn the correction on above and watch the two images come back together.
One person. Fifteen corrections dialled by hand across four sessions in one day, then five more sessions with an instrument that moves the image itself so the subject only has to say when it is single — because a hand on a dial tends to stop near wherever it started, and the four left-tilt points above finished 0.20, 1.09, 2.81 and 5.90° from their opening offset, ordered by how long he spent. The sign of the tilt is still his own report: no tracker measures cyclotorsion.