Blog··8 min read

Our UI Experiments: What Worked, What Failed and Why

A record of the DarkMoon dashboard redesign: rejected directions, reverts, a 20-board study the owner overruled, and what actually held in production.

DarkMoon research series·B. Data visualization & UX·Article 13 of 23·All 23 articles

Introduction

Most design write-ups show the final screen. The DarkMoon visualization research repository treats rejected attempts as first-class results, and this article follows that rule. It covers the redesign of the DarkMoon Pro Angular dashboard between late September and 20 November 2026: what we tried, what the product owner rejected, what we reverted, and what remained in production.

Owner quotes below are translated from French and labelled OBSERVED (from review notes); reverts are VERIFIED git objects in the private Pro repository, cited by short hash.

Problem

The dashboard had to show pentest results to two audiences (operators and decision-makers), on screens from 1440 px down to 320 px, in light and dark themes, without ever presenting a hypothesis as a proof. Before November 2026 the team had no stable design system: successive reference products were used as style sources, and the global stylesheet list still records their names.

The volume of exploration was large. A branch census found about 70 UI branches and ~120 worktrees created between late October and mid-November 2026; the 6 fe/* exploratory branches were never merged (VERIFIED).

Methodology

There was no user testing with external participants. The method was:

  1. Owner review rounds on real captures of every screen (OBSERVED).
  2. An AI design-review skill that had to choose between open options and cite its rule, with a fixed precedence: owner's words > brand charter and kit > generic advice (OBSERVED).
  3. AI adversarial critics scoring captures. Their scores, such as 4.5/10, are internal LLM judgements, not usability measures (OBSERVED).
  4. Measure the reference. When a reference kit existed, computed styles were compared with Playwright until zero deviation, then per-component pixel diffs (OBSERVED).
  5. One structured study for the campaign header: 20 mock-up boards scored against four owner complaints (empty space inside the box, overlapping borders, wrapping rows, views not visible on phones).

Real results

What failed

Date (2026)AttemptOutcomeLesson recorded
26 SepRebuild the orbital graph « in its own way »« That is not it at all »; lab engine reusedimprove the artefact that was liked, do not rebuild it
29 Oct« Le Brief » hybrid dashboard (plain-language brief, very large figures, Guided/Expert switch)« very very ugly »; internal AI judges 4.5/10 twiceAI judges are not a substitute for the owner
29–30 OctInvented cards (activity chart, severity bars, « Top risky targets »)« I forbid you to invent cards, modernise what exists »; revertedadd information, never filler
30 OctFull-screen 3D dashboard and editorial variants« that's not a dashboard »3D is an accent (infra map, path focus)
30 Oct~15 fintech style layers« too banking, not cyber enough »domain tone matters
30 OctLiquid glass, severity emblems, ambient field4 revertseffects without a job are removed
30 OctNode-editor flow graph pixel-matched to a design shot« no smooth transitions »; never mergeda reference is a mood for an experience, not an image to trace
31 Oct~95 « night hunt » design lots« no look kept »volume of options did not converge
3–4 NovKit skins; invented sidebar, project tree, rails, matricesrejectedmeasure the reference, then reproduce
5 NovCard border, grain, glass and gap experiments17 reverts—
18 NovWording change on agent-declared link labelsreverted (gain not demonstrated, n = 1)do not change vocabulary without evidence
20 NovSpace Mono display face in the product UIretired site-wide; kept only in the PDF reportbrand fonts belong to documents, not to dense UI
Campaign overview before owner pass 2: uppercase monospace chips in Space Mono, dark theme.
Campaign overview before owner pass 2: uppercase monospace chips in Space Mono, dark theme.Original capture: ASCIT31/darkmoon-visualization-research
Campaign overview after owner pass 2: Space Mono retired, chips in Google Sans with icons and a status pill.
Campaign overview after owner pass 2: Space Mono retired, chips in Google Sans with icons and a status pill.Original capture: ASCIT31/darkmoon-visualization-research

What held

Date (2026)ChangeEvidenceWhy it held
26 SepFix the map data before the renderer: server-side finding attachmentlab test 57/57 findings attached (OBSERVED, one lab campaign)removed the « all findings on one node » defect that no restyle could fix
29 Oct → 6 Nov« Exposure path · not proven » wording; guided story only for proven pathscommits (VERIFIED)never presents inventory as an attack
1 → 9 NovReference kit copied with locked tokens, on real fields only33 per-lot branches merged (VERIFIED)measured copy instead of invention
3–4 NovELK layered path graph; operator-declared paths as a never-counted classcommits (VERIFIED)deterministic reading, clear provenance
17 NovReserved heights for asynchronous cardsdashboard CLS 0.1065 at 1440×900 → fixed (VERIFIED, Playwright layout-shift observer)measured layout stability
19–20 NovTreatment plan by asset × weakness class66 per-finding fixes → 14 actions (VERIFIED, captures)removes contradictory counts
20 NovBoard 13 campaign header; old rules deleted from 11 stylesheets instead of out-rankedcommits (VERIFIED)no void, no overlapping borders, every view visible on phones

The study that the owner overruled

For the campaign header, the study produced 20 mock-up boards rendered at 1440, 834, 390 and 320 px. Its ranking was Board 02 > 01 > 07. The owner chose Board 13, which the study had rated « heavy on desktop ». No user test arbitrated between them, and the repository says so (UXDR-16).

Campaign header before Board 13 on a 390 px phone: capsule pills where some views are hidden behind horizontal scrolling.
Campaign header before Board 13 on a 390 px phone: capsule pills where some views are hidden behind horizontal scrolling.Original capture: ASCIT31/darkmoon-visualization-research
Campaign header after Board 13 on a 390 px phone: one card with a date picker and a Review button above five equal cells (Map, Paths with badge 2, Findings with badge 68, Overview, More).
Campaign header after Board 13 on a 390 px phone: one card with a date picker and a Review button above five equal cells (Map, Paths with badge 2, Findings with badge 68, Overview, More).Original capture: ASCIT31/darkmoon-visualization-research

A technical detail mattered as much as the visual choice. The old header was styled by five stacked layers using !important and specificity chains up to 24 :not(#id) selectors. Board 13 removed those rules from 11 stylesheets rather than adding a sixth layer on top (OBSERVED in the study's design-system notes; VERIFIED commit body).

Diagram. Figure 1. How options were filtered. The strongest rejections concerned meaning, not style.
Figure 1. How options were filtered. The strongest rejections concerned meaning, not style.

The pattern behind the verdicts

Read together, the rejections share one property: the owner's strongest objections were epistemic, not aesthetic. Two figures for one notion, a hypothesis drawn like a proof, an invented card or narrative: these became rules. Style choices were accepted when they reproduced a measured reference and rejected when they were invented in between.

Limits

  • No user study. No task time, error rate or SUS score exists for any version (not measured). « What worked » means « accepted by the owner and kept in production », not « performed better for users ».
  • One decision-maker. Decisions reflect one product owner plus AI critics.
  • Board 13 was implemented under time pressure; the post-deploy owner review is not recorded (UNKNOWN).
  • Many rejected directions have no published capture, because those captures show lab addresses or unreviewed local data. The failures table rests on owner notes and git reverts, not on images.
  • Performance was not part of the comparison. Production initial bundle 1.23–1.26 MB raw in three build logs (VERIFIED); no Lighthouse run on the Angular client (not measured).
  • Rules were amended. « Never add a card » became « you may add cards if they are well made, no duplicates or filler » on 20 November.

Figures

  • Figure 1 (Mermaid): the filtering process.
  • Before/after pairs: screenshots/comparisons/ in the research repository (real UI, synthetic demo data).
  • The 20 boards: experiments/alternatives/campaign-bar-boards (mock-ups, not the product).
DarkMoon Pro Attack paths tab on a phone (390 px wide) on an anonymised real OWASP Juice Shop campaign: the compact campaign header with counters as badges, the note that no path could be proven, and the six Possible paths marked Inferred.
Figure 2. What shipped on phones: the Board 13 header at 390 px, five equal cells with counters as badges and Reports under More. Anonymised real OWASP Juice Shop campaign of 26 April 2026 (demo.dark-moon.org).Captured 21 November 2026 from demo.dark-moon.org (current interface).
DarkMoon Pro campaign map in the light theme on an anonymised real OWASP Juice Shop campaign: the campaign header card with equal-width icon-over-label tabs above 30 assets grouped into Internet, Edge, Application and Data zones, with labelled edges and per-asset finding counts.
Figure 3. What shipped on desktop, light theme: the same header card with equal-width icon-over-label tabs, the layout the study had rated « heavy on desktop ». Anonymised real OWASP Juice Shop campaign of 26 April 2026 (demo.dark-moon.org).Captured 21 November 2026 from demo.dark-moon.org (current interface).

Scientific sources (verified)

All URLs fetched on 2026-11-20. The campaign-bar study's wider source list (VEILLE.md) was not re-verified for this article.

Research, source code & reproducibility

Related reading: Progressive Disclosure in Vulnerability Management Interfaces · Redesigning a Cybersecurity Dashboard Around Attack Paths · Designing Explainable Security Graphs Without Misrepresenting Evidence · Building DarkMoon in the Open: Research, Benchmarks, Failures and Reproducibility


If you have run a structured comparison of navigation patterns for security consoles, we would be glad to read it; the research repository accepts issues and pull requests.

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.