Design System · Component Documentation

Atlas Signal UI

Component reference for the Atlas Signal brief viewer. IBM Plex type system on a warm editorial paper palette. Each section shows live renders with CSS class annotations.

Component · 01

Snapshot Strip

The at-a-glance panel placed directly below the pull quote in a brief view. A .snap-plain prose block sits above a dashed-rule divider; a two-column .snap-grid holds evidence strength (left) and readiness level (right).

Plain-language summary

Researchers interviewed 20 everyday users of the Merlin bird-identification app to find out what they actually want from AI explanations. The surprise: users care less about technical model internals and more about practical guidance — how to trust the AI's answer, how to take better photos so the model works better, and how to give developers useful feedback. They strongly preferred explanations that pointed to recognisable parts, the way a human birder would explain.

Evidence strength
Moderate
Rigorous mixed-methods design, but qualitative findings from n=20 in one app context.
Readiness level
Concept
Lab
Replicated
Pilot
Deployed
Findings align with prior human-centred XAI work and are directly actionable for product design — a quantitative cross-domain replication is still owed.
.snapshot.snap-plain + .snap-grid (1fr · 1.6fr) separated by a dashed border-top
Evidence Badge

Three tiers of evidence strength. All color, fill, and border are set by the modifier class — no inline styles. Always includes a .dot span as first child.

Strong
.ev-strong
#2d6a4f · #d8eadd
Moderate
.ev-moderate
#8a5a14 · #f3e3c4
Emerging
.ev-emerging
#8e3527 · #f3d8d2
.evidence-badge.ev-{strong|moderate|emerging} · 11px IBM Plex Mono · pill border-radius
Readiness Track

Five-step linear progress indicator. Steps before the active stage take .past (solid ink dot); the active step takes .here (accent dot with glow ring). The connector line is a pseudo-element on .rs-track::before.

Concept
Lab
Replicated
Pilot
Deployed
Stage 0 — Concept  ·  .here on index 0

Concept
Lab
Replicated
Pilot
Deployed
Stage 2 — Replicated  ·  .past on 0–1 · .here on 2

Concept
Lab
Replicated
Pilot
Deployed
Stage 3 — Pilot

Concept
Lab
Replicated
Pilot
Deployed
Stage 4 — Deployed  ·  all .past · final .here
.rs-step · .rs-step.past · .rs-step.here  —  .rs-dot background inherits from modifier · label weight set via CSS
Component · 02

Section Blocks

Expandable content sections for the full paper brief. The section number badge (§ 01) and heading sit in .bsec-head; the + toggle rotates 45° via CSS when .open is present. Three modes: collapsed, open, and skim (all visible, no toggle controls).

Collapsed — default state

Body is hidden (display: none). Click any header to open it.

§ 01

Research problem

+

What do real end-users of a deployed AI app actually want from explainability — and how does that map onto the four dominant XAI approaches the research community has been building?

§ 02

Methodology

+

Mixed-methods study with 20 end-users of Merlin, the Cornell bird-identification app. Each participant completed a one-hour interview combining a survey adapted from the XAI Question Bank and an interactive feedback session with four XAI approaches.

.bsec (default) · body hidden · first sibling strips top border via :first-of-type
Open state

Add .open to reveal the body. The + toggle rotates to × via transform: rotate(45deg).

§ 03

Key findings

+
  • Users wanted practically useful information that improves their collaboration with the AI, more than technical system details.
  • They intended to use XAI for four purposes beyond understanding: trust calibration, learning the task themselves, supplying better inputs, and giving developers feedback.
  • Part-based explanations were strongly preferred over heatmaps because they resemble human reasoning.
  • The "help me help the AI" need was unanimous across all participant backgrounds.
.bsec.open · body visible · toggle rotated 45° · click to collapse
Skim mode — all sections open, no toggle

Applied at the brief level, not per-section. No .toggle element is rendered. All bodies are always visible via .bsec.skim .bsec-body { display: block }.

§ 04

Limitations

  • Single application (bird identification) — findings may not transfer to high-stakes domains like medicine or law.
  • n=20, qualitative — effect sizes and quantitative generalisation are out of scope.
  • Developers and deployers of Merlin were not interviewed; only end-users.
§ 05

Questions to validate

  • Do part-based explanations beat heatmaps in higher-stakes domains like radiology or fraud review?
  • Does adding "how to supply better inputs" guidance measurably improve real-world AI accuracy and adoption?
  • Can a single XAI pattern library transfer cleanly across consumer, enterprise, and clinical AI apps?
.bsec.skim · no .toggle element rendered · body always display: block · non-interactive
Tokens · 04

Color Palette

All surfaces, text, and borders are drawn from a single warm-paper palette. The accent cascade (--accent--accent-soft--accent-deep) covers interactive, tinted, and deep-ink contexts. Tier colors are defined directly — not via variables — so each tier can cascade --sig-color and --sig-bg to child elements.

Paper
--paper
#f6f4ef
--paper-2
#efece4
--paper-3
#e7e2d5
Ink
--ink
#1a1a1a
--ink-2
#2b2a26
--ink-3
#4a4843
--ink-4
#76736a
--ink-5
#a7a39a
Rules
--rule
#d9d4c5
--rule-strong
#b9b3a1
Accent
--accent
#c8553d
--accent-soft
#edcab9
--accent-deep
#8e3527
--highlight
#f0d97a
Tier colors — foreground · background
Critical · 85+
sig-color
#2d6a4f
sig-bg
#d8eadd
Strong · 70–84
sig-color
#5a7a2f
sig-bg
#e2eccb
Notable · 55–69
sig-color
#a06a2f
sig-bg
#f3e3c4
Emerging · <55
sig-color
#c8553d
sig-bg
#f3d8d2
Tokens · 05

Typography

Three IBM Plex families, each with a specific editorial role. Sans carries all UI and body text. Mono labels, metadata, code, and scores. Serif appears only in the pull-quote drop cap and decorative quotation mark — a contrast accent, not a body face.

IBM Plex Sans
--font-sans
UI · Body · Headings
Display 600
Heading 500 — Section title
Body 400 — A short feed of studies worth your attention, distilled and scored.
UI label 500 — 13px medium
IBM Plex Mono
--font-mono
Labels · Meta · Scores
94
Nature Neuroscience · Vol. 28 · 2025
Brief · Neurotech / BCI · NeurIPS '17
Section label — 9px uppercase tracking
IBM Plex Serif
--font-serif
Drop cap · Pull quote mark
"
A
Used sparingly — decorative quotation mark & drop cap only.
Component · 06

Signal Pill

The primary score display in the paper list — a stacked column of number, tier dot + label, meter bar, and index. Color cascades from the tier class on the parent wrapper via --sig-color. Distinct from the sidebar signal display: larger score, explicit tier row, 60px meter, and row index.

94
Critical
№ 01
.sig-critical · 85+
76
Strong
№ 02
.sig-strong · 70–84
62
Notable
№ 03
.sig-notable · 55–69
44
Emerging
№ 04
.sig-emerging · <55
.signal-pill.sig-{tier}.signal · .tier-row.tier-dot + .tier-label · .signal-meter > i · .signal-idx
Component · 07

Buttons

Four variants. Default sits on paper with a ruled border. Primary inverts to ink fill. Ghost drops the border entirely for chrome contexts (topbar). Accent uses the terracotta as a call-to-action fill.

.btn
.btn.primary
.btn.ghost
.btn.accent
Base: 7px 12px padding · 6px radius · 13px 500 weight · display: inline-flex · hover shifts border to --ink
Component · 08

Badges & Labels

Inline status markers. Tier chip is the sidebar-weight sibling of the signal pill tier label — styled as a pill with dot. Peer reviewed is a green check badge rendered inline with author metadata. Saved marker is a bare accent-colored glyph. KBD renders keyboard shortcuts with a bordered mono box.

Tier Chip — all four tiers
Critical
.tier-chip.sig-critical
Strong
.tier-chip.sig-strong
Notable
.tier-chip.sig-notable
Emerging
.tier-chip.sig-emerging
Tier chip colors are driven by --sig-color / --sig-bg on the tier class — the same variables used by the signal pill and sidebar display.
Status badges
Peer reviewed
.peer-reviewed-badge
◆ SAVED
.saved-marker
KBD — keyboard shortcut
Toggle sidebar \
Open brief
.kbd · 10px mono · paper-2 bg · rule border · 3px radius
Component · 09

Tag Chip

Filterable topic tags used in the sidebar tag cloud and brief sidebar. Default is a ruled pill on a transparent background. Active state (.on) inverts to solid ink fill. Used both as filter buttons (sidebar) and as read-only labels (brief metadata).

.tag-chip — default
.tag-chip.on — active filter
.tag-cloud wrapper (flex-wrap: wrap; gap: 4px) · .tag-chip 10px mono · pill · .on = ink bg + paper text
Component · 10

Topbar

Sticky chrome bar above the main content column. Left: sidebar toggle (.side-toggle) with an SVG panel icon. Center: flex spacer. Right: a primary CTA button and a mono back link. Renders on paper background with a bottom rule.

.topbar · .side-toggle (32×32 icon button) · .topbar-spacer (flex: 1) · .btn.primary · .topbar-back (mono 10px uppercase)
Component · 12

Masthead

The library page header. A .masthead-meta strip carries mono publication metadata separated by a flex-1 rule line. The display heading and deck follow. Below, .masthead-tabs provides a tab row that sits flush on the section border.

Atlas Signal Vol. 03 · No. 41 Friday, May 16, 2026
48 papers · 214 application threads

This week's signal from the literature.

A short feed of studies worth your attention — distilled, scored, and mapped to the industries where they might actually change something.

.masthead-meta .rule (flex:1 horizontal spacer) · 42px 600 heading · 16px sub · .masthead-tabs .tab.active (2px bottom border flush)
List Toolbar

Sits between the masthead tabs and the paper list. Signal legend on the left, view toggle on the right.

Signal Emerging Notable 55+ Strong 70+ Critical 85+
.signal-legend · tier class on .lg-item drives --sig-color into the .lg-dot background · .view-toggle button.on = ink fill
Component · 13

Paper Row

The primary list-view item. A 3-column grid: 100px signal pill column, 1fr body (title + finding + metadata), 220px aside (field, tags, top application scores). Hover shifts background to paper-2.

94
Critical
№ 01

A Unified Approach to Interpreting Model Predictions

SHAP unifies six existing XAI methods under a single framework grounded in cooperative game theory, producing consistent feature attributions that satisfy three desirable axioms — local accuracy, missingness, and consistency.

NeurIPS '17 · 2017 S. M. Lundberg, S.-I. Lee Peer reviewed ◆ SAVED 1d ago
Machine Learning / XAI
SHAP · Shapley values · SVM
Top applications
95 · Financial services 91 · Healthcare AI 87 · Insurance
76
Strong
№ 02

Emergent 3D Spatial Reasoning in Vision-Language Models

A 7B-parameter VLM fine-tuned on 18M robot manipulation trajectories generalises to unseen objects and rooms, achieving 81% success on a novel-object reach-and-place benchmark.

arXiv preprint · 2025 Y. Chen, N. Patil, A. Ferrari, +6 4d ago
AI / Robotics
VLM · embodied AI · robotics
Top applications
84 · Warehouse logistics 79 · Lab automation 71 · Home robotics
.paper-row · grid 100px 1fr 220px · 24px gap · hover paper-2 · .body h3 22px 500 · .meta 11px mono flex-wrap · .aside 11px mono right col
Component · 14

Pull Quote

The headline finding, displayed as an editorial pull quote. A 2px top border anchors it; the oversized Serif quotation mark (::before) sits at left: -8px in accent color. Text and attribution indent 28px to clear the quote mark.

74-word/minute decoding of silently-imagined speech with 6% word error rate — the first real-time speech BCI to feel conversational rather than dictation-paced.

The headline finding
.pull-quote · 2px solid ink top border · ::before 72px Serif quote mark in accent · text 21px 400 · attrib 10px mono uppercase
Drop Cap

Applied via .drop-cap on a paragraph. The ::first-letter pseudo-element uses IBM Plex Serif at 56px, floated left, with accent color. Used on the first paragraph of the Research Problem section.

What do real end-users of a deployed AI app actually want from explainability — and how does that map onto the four dominant XAI approaches the research community has been building: heatmap, example, concept, and prototype-based explanations?

.drop-cap::first-letter · 56px IBM Plex Serif 500 · accent color · float: left · 6px 8px 0 0 padding · line-height: 0.88
Component · 15

Application Engine

Full-bleed dark panel below the brief. Ink background inverts all text to paper tones. An .engine-switch pill tab row switches between four presentation modes. All four treatments are shown statically below.

Panel header + switch
Application Intelligence

Where does this research create leverage?

Four ways to explore the commercial and institutional implications of this paper's core finding.

.engine · ink bg · .engine-kicker accent-soft mono · .engine-switch button.on = paper bg ink text · others rgba(255,255,255,0.08) bg
Treatment 1 — Industries grid
Financial services
Credit scoring, fraud detection — auditable per-decision attribution
Fit95
Healthcare AI
Clinical risk model explanation and fairness auditing
Fit91
Insurance
Underwriting model transparency for regulators
Fit87
Legal / compliance
Model fairness and audit documentation
Fit84
Manufacturing
SVM quality-control model explanation
Fit77
HR & recruitment
Candidate scoring model auditing
Fit72
.ind-grid · 3-col CSS grid · 1px gap in rgba(255,255,255,0.1) · .ind-cell 2-col (name+note / score) · score 22px mono accent color
Treatment 2 — Idea cards
IDEA · 01

Per-prediction explanation panel for SVM dashboards

Product teams using SVM classifiers
Wrap LIME around an existing model in a day — the explanation bar chart becomes the interface between the algorithm and the end user, no model changes required.
IDEA · 02

Model behaviour audit report

Enterprise AI teams & compliance
Use SP-LIME's representative set to produce a one-page summary of how the SVM behaves across the data space — boardroom-ready, regulator-friendly.
IDEA · 03

XAI onboarding screen for clinical decision tools

MedTech and clinical AI vendors
Clinicians reject opaque models. A LIME explanation at the moment of recommendation converts "I don't trust this" into "I understand and agree."
.ideas · auto-fit minmax(240px, 1fr) grid · .idea-card · hover shifts border to accent · .idx accent-soft · .who mono uppercase with bottom border
Treatment 3 — Graph / cluster map
LIME — local XAI
Core concept
Input types
Tabular Text Images SVM features
Output formats
Feature weights Highlighted text Superpixels Rules
Verticals
Healthcare Finance Legal HR
.graph flex · .graph-center (180px fixed, circle core in accent) · .graph-clusters 3-col grid · .node pill hover → accent bg
Treatment 4 — Guided Q&A
Q1
What changes if we believe this result?

Any SVM output becomes explainable without changing the model. The explanation layer sits on top — making existing classifiers auditable overnight.

Q2
Who feels the disruption first?

Organisations using black-box SVMs for consequential decisions — credit, hiring, clinical triage. Their "we can't explain the model" defence is gone.

Q3
What's the smallest viable product?

A one-day integration: wrap your existing SVM with LimeTabularExplainer and render the output as a feature-importance bar in your existing UI.

Q4
What kills it?

Explanation instability. If running LIME twice gives different answers, users lose trust in the explanations themselves — not just the model.

See methodology, caveats, and validation questions →
.guided-step · 36px 1fr grid · .gidx accent-soft mono index · .gq 16px 500 paper · .ga 14px rgba body · .guided-cta paper-bg button
Component · 16

Mobile Frame

A CSS-only phone mockup for previewing the mobile reading experience. 320×640px content area with a dynamic island notch, status bar, scrollable inner body, and a sticky 4-tab navigation. Used as an overlay demo triggered by a FAB.

9:41 ● ◐ ▮
VOL. 03 · NO. 41 · MAY 16

This week's signal.

● CRITICAL 94 · XAI · 1d ago

A Unified Approach to Interpreting Model Predictions

SHAP unifies six XAI methods under a game-theoretically grounded framework, producing consistent feature attributions…

● STRONG 76 · AI / Robotics · 4d ago

Emergent 3D Spatial Reasoning in Vision-Language Models

A 7B-parameter VLM fine-tuned on 18M robot manipulation trajectories generalises to unseen objects…

● NOTABLE 62 · Behavioral · 5w ago

12-Minute Mindfulness Protocol Improves Joint Value Creation by 23%

In 312 dyadic negotiations, a brief guided mindfulness exercise before bargaining raised total joint surplus…

Signal
Search
Compare
Library
Anatomy
.notch — dynamic island bar, absolute top-center
.status — time + signal indicators, space-between
.mob-inner — scrollable content area
.mob-head — publication header + search input
.mob-row.sig-{tier} — paper card, tier color on signal meta
.mob-tab — absolute bottom, 4-col grid, .ti.on = ink
320×640 content · 40px border-radius · 8px solid #0a0a0a border · box-shadow 32px 80px depth