The melon that would not wobble here: reading a GLM-5.3-Flash WebGPU build

JACK GROSS · SOURCE REVIEW · 29 SEP 2026

The page opens like a museum specimen card: “Melon Jelly.” in a huge serif, and beneath it, “A slice of summer. A little wobble. Too soft to share.” To the right sit a hand, a knife, three color varieties and two sliders. In the middle, in the browser I can inspect, there is no melon. The card reads “No WebGPU adapter was found.” That empty stage is where this story honestly starts: with what a model put into the source, and what a screenshot can never prove.

HTML + CSS→SoftBody / XPBD→WebGPU→A test that would earn the claim

What I can verify

The Melon Jelly page I examined describes itself as created with Z.AI GLM-5.3-Flash. I have the deployed page and its source. I do not have the prompt, the conversation, the model logs or the build history, so I can credit the creation to that description but cannot certify a one-shot generation or count the revisions. Z.AI's documentation names the model and lists image, file and text inputs, which is product context and not evidence about this project. The live URL returns one HTML document of 175,592 bytes, with inline CSS, one inline module script and no external scripts or stylesheets. That is a source-level observation, not a security audit. [source]

The mechanism inside the file

The slice is more than an animated picture. The source builds a tetrahedral mesh, spreads mass over the rest volumes and steps a SoftBody at a fixed 1/60 of a second with ten substeps. Each tetrahedron is pulled toward a best-fit rotated rest shape with a compliance term, then a volume constraint nudges it toward its rest volume. The code adds an anti-inversion correction, stiffer rind regions, damping relative to edges, and floor contact with friction. The museum label explains the same ideas in words. The implementation can be read; its numerical stability and physical realism cannot be established by reading it. [source]

Why cutting is the hardest promise

In the source the knife is not a decorative toggle. A stroke becomes a blade plane, the code splits a convex rest-plane outline, and it builds fresh tetrahedral pieces. It tries to carry the old piece's motion into the new ones instead of hiding half a mesh. That is a harder continuity problem than a pretty cut animation, since the new bodies need their own volumes, surface attachment, collisions and readouts. The source has splitPiece and planCut/performCut paths, but I have not run them on a GPU. A rigorous claim needs a repeatable cut test, not a menu label. [source]

Light is half the story

Lighting lives in WGSL shader strings and WebGPU renderer code. The page's own methods panel lists back-face thickness, refraction, Beer-Lambert color absorption, milky scattering, Fresnel reflections and a colored floor shadow. Surface vertices, seeds and bubbles hang on the tetrahedra through barycentric weights, so the visible skin is meant to follow the simulation. That split between simulation mesh and display mesh is why a jelly can look smooth without making every seed its own physics particle. This is a reading of the architecture, not a measurement of rendering quality. [source]

What the screenshot cannot prove

The interface promises Crimson, Golden and Rosé palettes, firmness and damping, a quarter-speed checkbox, wireframe, nudge, pause, and readouts for mass, volume, kinetic energy and piece count. The source wires those controls and derives the numbers from mesh state on an illustrative scale. In my capture, the WebGPU adapter request failed, the fallback disabled interaction and the readouts stayed blank. So the screenshot shows strong art direction and a graceful error path, not a working wobble. A GPU-capable browser may show more, but I cannot claim that test happened here. [source]

A test that would earn the claim

I would run it on a WebGPU-capable device and record the browser and GPU versions. Start with a settled slice and watch volume drift, finite positions and frame timing. Grab a tip, release it, and check that energy decays without the body exploding or freezing. Cut through the rind at several angles and verify that piece count, volume and mass agree and the cut surfaces do not tunnel through the table. Toggle quarter-speed and wireframe, repeat after a device loss, and compare screenshots across viewport sizes. If the goal is to judge model generation and not just the code, keep a prompt and revision log. [source]

The quiet hero is the fallback

What stands out is the restraint of the fallback. The page does not quietly swap in a static image and pretend it ran. It says the renderer is offline. The build shows how a model-assisted artifact can pack interface design, physics vocabulary, executable structure and an honest failure message into one file. Whether GLM-5.3-Flash produced it in a single pass, and whether the physics feels right under a real hand, remain questions for the build record and a GPU trial. [source]

Illustrative reading of the step loop

// Simplified map of the inspected loop, not a copied or run implementation.
const STEP = 1 / 60;
while (elapsed >= STEP && steps < 3) {
  for (let sub = 0; sub < 10; sub++) {
    integrate();
    constrainRotatedTetShape();
    constrainTetVolume();
    resolveGround();
  }
  elapsed -= STEP;
  steps++;
}
// Test real values and cuts on a WebGPU device before claiming stability.

Sources and provenance: Melon Jelly source and live page · Z.AI GLM-5.3-Flash documentation. Source inspection, not a verified generation transcript or a successful GPU run.