Blog··7 min read

Behind the Honorable Mention: the Design and Engineering of dark-moon.org

A three.js moon on NASA maps, the product's orbital map live on the page, a no-removal policy for weak devices, and a gate for every visual change.

DarkMoon research series·E. Website & recognition·Article 22 of 23·All 23 articles

On 28 October 2026, dark-moon.org received an Awwwards Honorable Mention. Our submission described the site as « a shader-built moon you glide over on scroll, a live 3D attack-path map in a flat black dashboard, liquid-glass cards, a spatial menu ». This article explains how those pieces are built, and the rules we kept to while building them.

Every figure below comes from the site's source code or our engineering notes. Where we have no measurement, we say so.

The stack

The site is a Next.js (App Router) static export. All 3D runs on three.js and WebGL. The heavy modules are loaded on demand: the orbital map imports three.js, so it only loads through next/dynamic as its own async chunk.

A moon you glide over

The background is a single three.js sphere (128 × 128 segments) lit by one directional light placed behind it. The overall staging follows a reference site we studied, 21hrs.space, but our code and assets are our own. We did not reuse third-party shader code. An early full-screen ray-sphere attempt was dropped as too slow, and we also avoided code from GPL-licensed planet shaders.

The textures come from NASA. The albedo, normal and height maps are derived from NASA's CGI Moon Kit (Scientific Visualization Studio): the LRO LROC WAC colour mosaic and the LOLA elevation data. A script in the repository converts them into 4096 px and 8192 px greyscale albedo, normal maps at two resolutions, and a 2048 px height map. The design rule is strict: original NASA maps only, no invented texture.

The scroll drives a camera descent. The validated hero framing is keyframe 0. As you scroll, the page moves the camera continuously (with damping) through further keyframes, closer to the moon's limb section after section. Each framing leaves one band of detailed ground and one calm black area for the text. The material never changes during the descent. Up close, a finer, rotated sample of the same texture fades in, so there is no texture swap and no popping.

Brand colours are painted exactly. Tone mapping is off. ACES tone mapping shifted every brand colour towards grey, so the scene now paints the official hex values as they are.

Glass that does not shake on iOS. The « liquid glass » cards first used a separate glass library. It was later replaced everywhere by glass that the moon renderer draws itself, in a lens pass that is then copied into each card. On iOS this was the only approach we found that did not make the cards shake while scrolling.

One bug worth sharing. On an iPhone (Apple GPU), the starfield collapsed into a single squashed dot. The star rotation angle was computed in the shader from wall-clock time, about 2.7 × 10⁷ radians. At that magnitude the GPU's sin/cos returned values close to zero, so the rotation matrix flattened every star onto one point. We now compute the angle on the CPU, modulo 2π. We also reproduced the bug on the device with a standalone diagnostic page before shipping the fix.

The product's own map, live on the page

The « live 3D attack-path map » is not an illustration. The site runs the orbital map engine of DarkMoon's demo application, copied as it is. Only the glue changes: React hooks instead of Angular, keyboard input on the map instead of the whole document, no mouse-wheel zoom, and no rendering while the map is off-screen. The engine is fed with the demo campaign the product ships: OWASP Juice Shop, 17 assets and 57 findings.

On the site, the guided tour plays by itself and a click pauses it. On phones, creating the WebGL context was a single 75 to 390 ms task that cannot be split. It landed in the first scroll, so we now create the context ahead of time, during page load, and hand it to the engine when the engine boots.

The orbital map's design is documented in the public research repository: research/graph-layouts/orbital-graph.md.

Weak devices: lower the quality, never remove a visual

This is a product-owner rule, and it was earned the hard way. Once, a safeguard of ours kept the VR headset scene from being built until the GPU had proven itself. On a fast laptop where the visitor scrolled straight away, the headset never appeared, and we had to ship a hotfix. The rule since then: on a weak machine we lower resolution, effects and frame rate. The moon, the headset, the glass and the orbital map always stay. No visual depends on a performance measurement.

Here is how the code applies the rule:

  • A signal, not a switch-off. The background renderer watches frame times only after the page is revealed, plus a grace period, so loading hitches never decide anything. If frames stay slow, it sets a dm-lowgpu class on the page. Every layer keeps drawing. The orbital map, for example, is then built only as it nears the screen, in an idle slot.
  • Capped frame rates are not weak GPUs. A 30 fps cap from Low Power Mode or Energy Saver is not treated as a slow GPU. On Firefox, which compiles shaders on the main thread, compile stalls are not counted as frame times. Counting them once wrongly flagged an Intel iGPU that then held 50 to 60 fps.
  • Strong devices get the best. A phone draws the sky at its screen's own pixel density, up to 3×. It steps down only if it measurably misses frames, and never below 1.3×. Desktop is capped at 1.5×.
  • Test the immediate scroll. Every low-end test includes a scroll right after load, not only after the page has settled.

A gate in front of every visual change

Any change that ships to the browser goes through scripts/qa/gate.mjs before it can be committed. Git hooks refuse a visual commit whose exact file tree has no PASS receipt. The gate builds the current production version and the candidate side by side, then checks four things:

  1. Visual. Playwright captures at 1440, 1024, 390 and 320 px, at three scroll positions. Each page is captured twice, and the difference between the two captures measures the WebGL noise floor. Only changes above that floor are flagged.
  2. SEO. Parity of titles, meta tags, hreflang, JSON-LD, headings and links on every exported page, plus the text, buttons and links of the rendered page.
  3. Accessibility. axe-core against WCAG 2.2 A/AA. Any new serious or critical violation fails.
  4. Performance. Shipped JavaScript and CSS per page, Lighthouse mobile medians over alternating runs, and, when available, a benchmark in native Windows Chrome on a real GPU.

The gate has limits, and we keep them in writing. Its captures use software WebGL (SwiftShader), so real-GPU rendering is still judged by eye. Time-driven text, such as the map's captions, is excluded from the text check.

Iterating with a jury in mind

The week of the nomination, we reviewed the site against the four criteria Awwwards weighs (Design 40 %, Usability 30 %, Creativity 20 %, Content 10 %) and fixed what we found. Examples: one idea per section on the home page, the orbital map moved out from behind a text link, logos given equal optical weight, and a real 404 page in place of the host's generic one. Those reviews were internal. They are not Awwwards scores, and we quote no score here.

Research, source code & reproducibility

  • Award record: Awwwards: Darkmoon, AI Pentest Platform.
  • Visualization research (public): ASCIT31/darkmoon-visualization-research. It covers the orbital graph, the graph layout families, the low-end rule as it applies to the 3D map, and the full list of what was not measured.
  • Attack-path engine research (public): ASCIT31/darkmoon-attack-paths-research.
  • Moon data: NASA SVS CGI Moon Kit (LROC WAC mosaic, LOLA elevation). Credits ship with the site in public/textures/moon/CREDITS.txt.
  • What is not claimed: we publish no current Lighthouse or frame-rate figure in this article. The gate measures each change against production, but those are relative checks, not a benchmark of the site. The research repository's benchmarks page lists which measures exist and which do not.

Related reading: Our Website dark-moon.org Receives an Awwwards Honorable Mention · From an Award-Winning Website to the Product: How dark-moon.org's Visual Research Shapes the DarkMoon Dashboard and Terminal Explorer · Visualizing Cyber Attack Paths: From Node-Link Diagrams to Orbital Graphs

Next
Run it against your own lab

Darkmoon is open source (AGPL-3.0) and self hosted. Clone it, point it at a target you own, and read every line.