Tesseract

Interactive audiovisual piece
Self-initiated experiment · 2026 · Built using Claude Opus 5.5 · three.js (WebGPU, TSL), Web Audio

A tiny 4D universe, experienced on a flat screen.

The question
What makes an interactive experience feel immersive on a 2D screen?
How it was made
Built using Claude Opus 5.5 in two days, from research done first, and judged against it at every pass.
What it shows
The research decides how well the model builds, and judgment closes the loop.

See it live

tesseract.postminimal.agency

A moment after a tap, the light around the signal fans out.

Tesseract is a small universe on the surface of a four-dimensional sphere, unfolded, turned and heard with your hands. A fourth dimension can't be shown, only felt through interaction, which makes it a pure test of the question. The subject was a pick, and it turned out to be a good one.

PhaseYouIt answers
1ThresholdPress and holdA line, a square, a cube, a tesseract, each with a voice and a line of narration
2BloomWatchThe tesseract presses onto a 4D sphere, and its light falls into linked circles
3WorldArriveOne hint at a time, each gone once it's done
DragIt turns through the fourth dimension with the hand, then coasts
ScrollIt turns toward you, so you seem to move forward, and the light warms and cools
TapA bell, its echo from the far side, then the homecoming
HoldTime slows into long exposures, and the sound sinks like a record
4Outside timePress TThe visit as one shape, a knot for each signal, and a playhead to play it back
5ReturnCome backYour last signals are still arriving, and hints already learned are skipped
The interaction model. In each phase of a visit, what you do and how the piece answers, in light, sound and words.

The gesture is the tutorial.

Browsers won't play sound without a gesture, so instead of a start button the piece asks you to press and hold. A point, moved, draws a line, then a square, a cube, a tesseract. Each dimension latches, so letting go early costs nothing. Only then does the next lesson appear. Drag, and it turns through the fourth dimension, which shows as colour, warm for anaHinton's names for the two directions along the fourth axis, ana and kata. and cool for kata.

Press and holdor hold the space bar
Press and hold to unfold it, then drag to turn it. The lesson appears when it's needed and leaves once it's done.

Every lesson after that works the same way. Each hint is keyed to what you've done, not to a clock. It waits its turn, is skipped if you already did the thing, and disappears when you do. Two explanations interrupt on purpose, at the moment they become true, when your first signal reaches the far side and when it comes home.

A first visit. Each hint appears when it's needed and leaves once it's done.

You travel the fourth dimension by hand.

The universe is the surface of a 4D sphere, so moving through it means turning it. Every movement is a rotationIn 4D, things turn in a plane, not around an axis. W is the fourth axis.. Scroll, and it feels like going forward, but you stay put and the universe turns toward you, like a globe turning under your feet. Dragging turns it through the fourth dimension. Hold Shift for an ordinary 3D turn, and on a phone, tilt to lean it through.

Its hint, then the travel. You scroll, and the universe turns toward you.
How four dimensions become pixels

The universe is computed in four dimensions. It's turned there, projected from one of its poles down to three dimensions, then filmed by an ordinary camera. The projection keeps circles as circles, so the linked rings of the Hopf fibrationThe 3-sphere filled with circles, each one linked with every other. stay linked on screen.

  1. 01
    On the 3-sphere4D

    Every particle is a point on the surface of a 4D ball, on a Hopf fiber or a tesseract edge.

    • Hopf fibers
    • Tesseract edges
  2. 02
    Turned by the handSO(4)

    Plane rotations, composed into one matrix. Drag turns XW and YW, scroll turns ZW.

    • XW
    • YW
    • ZW
  3. 03
    Projected to 3DStereographic

    From one of the poles. Circles stay circles, so linked rings stay linked.

    • W kept as colour
  4. 04
    Seen through a lens3D to 2D

    An ordinary perspective camera with depth of field, so points out of focus become soft discs.

From four dimensions to the screen, top to bottom. W, lost in the projection, comes back as colour.

Every particle starts as a unit 4-vector. On a Hopf fiber, it's

p(η,δ,φ)=(cos⁡(φ+δ)sin⁡ηsin⁡(φ+δ)sin⁡ηcos⁡φcos⁡ηsin⁡φcos⁡η)p(\eta, \delta, \varphi) = \begin{pmatrix} \cos(\varphi + \delta)\sin\eta \\ \sin(\varphi + \delta)\sin\eta \\ \cos\varphi\cos\eta \\ \sin\varphi\cos\eta \end{pmatrix}

η\eta picks the torus, δ\delta the fiber on it and φ\varphi the point along it. Advancing φ\varphi with time moves light along every fiber at once.

The hand turns it with plane rotations, each one turning uu toward vv by θ\theta and leaving the other two directions alone. They compose into one matrix RR in SO(4)All the rotations of four-dimensional space..

Ruv(θ)=I+sin⁡θ (vu⊤−uv⊤)+(cos⁡θ−1)(uu⊤+vv⊤)\begin{aligned} R_{uv}(\theta) = I &+ \sin\theta\,(v u^{\top} - u v^{\top}) \\ &+ (\cos\theta - 1)(u u^{\top} + v v^{\top}) \end{aligned}

The projection to three dimensions is taken from a point on the W axis.

q=λλ−w′ (x′,y′,z′)q = \frac{\lambda}{\lambda - w'}\,(x', y', z')

With λ=1\lambda = 1 it's stereographic projectionProjecting a sphere onto flat space from one of its poles. from the pole. With λ=1.8\lambda = 1.8 it's a perspective view from a 4D eye, which is how the tesseract is drawn, and the bloom slides λ\lambda from one to the other. Near the pole the denominator nears zero, so it's clamped, and those points fade out with distance rather than running off to infinity.

The projection packs the far side of the sphere into the centre and stretches the near side. Each particle's brightness is scaled by that local stretch, (1−w′)−1.1(1 - w')^{-1.1}, clamped, so the dense centre stays readable. The coordinate the projection throws away, w′w', comes back as colour, warm toward ana and cool toward kata.

A tap goes the other way. The ray through the pixel meets the unit sphere, and inverse stereographic projection lifts that point onto the 3-sphere.

p=(2x, 2y, 2z, r2−1)r2+1r2=x2+y2+z2\begin{gathered} p = \frac{(2x,\ 2y,\ 2z,\ r^2 - 1)}{r^2 + 1} \\[4pt] r^2 = x^2 + y^2 + z^2 \end{gathered}

The inverse rotation R⊤R^{\top} then puts it in the universe's own frame, so the signal turns with the universe.

/**
 * Stereographic projection from the pole (0, 0, 0, 1)
 * of the unit 3-sphere.
 */
function stereographic(
    q: Vec4Like,
    out: { x: number; y: number; z: number }
): number {
    const k = 1 / Math.max(1 - q[3]!, 0.03)
    out.x = q[0]! * k
    out.y = q[1]! * k
    out.z = q[2]! * k
    return k
}

/**
 * Inverse stereographic projection: any point of ℝ³
 * back onto the unit 3-sphere.
 */
function inverseStereographic(
    x: number, y: number, z: number, out: Vec4
): Vec4 {
    const r2 = x * x + y * y + z * z
    const d = 1 / (r2 + 1)
    out[0] = 2 * x * d
    out[1] = 2 * y * d
    out[2] = 2 * z * d
    out[3] = (r2 - 1) * d
    return out
}
Both directions, from the source. The CPU copy matches the GPU's, so taps and sound line up with the pixels.

Direct, then physical.

While your hand is on it, the universe follows the hand exactly, with no easing in between. Let go, and it coasts on the momentum you gave it, then slows down. Anything that answers you directly is critically damped, as fast as possible without overshooting, a rule borrowed from control engineering.

Move the ring. Eased motion trails the hand, underdamped motion overshoots, and critically damped motion arrives fastest without overshooting.
Why it feels the same at any frame rate

Frame rates run anywhere from 30 to 120 a second, depending on the screen and the load. A spring integratedAdvanced in small time steps from its velocity, rather than solved outright. the usual way, one step per frame, feels different at each, and a stiff one can blow up when frames drop.

This spring isn't integrated. Its equation has an exact solution, and each frame evaluates it from the current offset x0x_0 and velocity v0v_0.

x¨+2ωx˙+ω2x=0x(t)=(x0+(v0+ωx0) t) e−ωt\begin{gathered} \ddot{x} + 2\omega\dot{x} + \omega^2 x = 0 \\[4pt] x(t) = \big(x_0 + (v_0 + \omega x_0)\,t\big)\,e^{-\omega t} \end{gathered}

ω\omega sets how fast it settles. No error builds up between frames, so it behaves the same at any frame rate. The camera, the cursor and the clock all use it. The springs allowed to wobble, like the latch settling, are stepped in fixed slices of 1/2401/240 s instead, small enough to stay stable when frames drop.

/**
 * Critically damped: the fastest settle with no overshoot.
 * Exact for any dt.
 */
step(dt: number): number {
    const x = this.value - this.target
    const e = Math.exp(-this.omega * dt)
    const t = (this.velocity + this.omega * x) * dt
    this.velocity = (this.velocity - this.omega * t) * e
    this.value = this.target + (x + t) * e
    return this.value
}
The solution above, evaluated at each frame's dt.

Every tap is heard, and comes back.

A tap rings a glass bell whose pitch is its colour. The wave spreads through the 3-sphere, meets itself on the opposite point and comes home, a real property of the sphere, and each meeting rings again. Sound and light read one function, so every echo lands on the frame its light converges. Every sound is synthesized live.

A signal reaches the far side of the universe and comes back.
Tap the ring to send a signalsound on
The same function, one dimension down. Tap the ring, and the pulse meets itself on the far side and comes home.
Why the echo is always in sync

On the web, sound is usually scheduled ahead on the audio hardware's clock. That doesn't work here, because holding can slow the universe's time at any moment, so an echo's arrival time isn't known in advance.

Instead, light and sound read the same clock, the universe's, once per frame. The GPU gets each signal's wavefront radius from it, and the score counts how many times the front has closed to a point.

n=⌊age3.4 s⌋n = \left\lfloor \frac{\text{age}}{3.4\ \text{s}} \right\rfloor

An odd nn rings the echo from the far side, an even one the homecoming. Because there is only one clock, light and sound can't drift apart, and slowing time slows both.

const age = now - s.born
const crossings = Math.floor(age / this.halfPeriod)
while (s.crossings < crossings) {
    s.crossings++
    const level = s.amplitude * Math.exp(
        -(s.crossings * this.halfPeriod) / this.lifetime
    )
    listener({
        kind: s.crossings % 2 === 1 ? 'focus' : 'return',
        signal: s, level, crossing: s.crossings
    })
}
The score, counting crossings on the same clock. 'focus' is the far side, 'return' is home.

Your visit, laid out as a timeline.

Hold, and time slows. Light smears into long exposures and the sound sinks like a slowing record. Press T and the light becomes your visit, your hand's path with time as its length and a knot for every signal. From the present it's a scribble, from the side a shape, and a playhead plays it back. Come back later and your last signals are still arriving.

Press T. The whole visit, seen from outside time, played back.

Judgment closes the loop.

The piece was built using Claude Opus 5.5, in two days. The research came first, so the model had something solid to build from. Judgment did the correcting, and what it found went back into the thesis before the next prompt.

  1. ResearchThe thesis, what it should be and why
  2. BuildClaude Opus 5.5 writes the code
  3. JudgeCompare the result with the thesis
  4. RefineImprove the thesis, not just the output
prompt again
Research sets the thesis. Each pass is judged against it, and the thesis improves before the next.
The public sources it stands on (8)
  1. Page Laubheimer, Onboarding Tutorials vs. Contextual Help, NN/g, 2023hints that arrive when they're needed
  2. Apple, Designing Fluid Interfaces, WWDC 2018following the hand, then momentum
  3. Karl J. Åström, Richard M. Murray, Feedback Systems, 2021critical damping (free book)
  4. Jos Leys, Étienne Ghys, Aurélien Alvarez, Dimensions, 2008stereographic projection and the Hopf fibration (free film)
  5. C. H. Hinton, A New Era of Thought, 1888the tesseract, ana and kata
  6. John Chowning, The Synthesis of Complex Audio Spectra by Means of Frequency Modulation, 1973FM synthesis, the glass bells
  7. Charles Howard Hinton, What is the Fourth Dimension?, 1880motion as a cross-section, the timeline
  8. W3C, WCAG 2.2, Animation from Interactionsmotion that can be turned off

An experiment by Post Minimal, Montréal, 2026. Built using Claude Opus 5.5, with research and direction by Bahae Ben. three.js (WebGPU, TSL) and the Web Audio API. Every gesture has a key, reduced motion is respected, and sound is never required.