{"path":"research/mobile-argument-ux.md","content":"# Mobile-First Argument Interaction Design\n\n**Research date**: March 28, 2026\n**Context**: Deliberus collaborative deliberation platform — exploring how to make structured argumentation work as a first-class mobile experience, not a degraded desktop experience.\n\n---\n\n## The Problem Statement\n\nNo successful mobile implementation of argument graph visualization exists. This is documented across the field: Kialo degrades to a nested list on mobile. DebateGraph is not optimized for mobile at all. Every failed argumentation platform was effectively desktop-only. This is not a coincidence — it reflects a category error in how these tools have been designed.\n\nThe category error: treating the graph as the product. If the argument graph is the interface, mobile is always going to lose. A 500-node DAG on a 390px screen is inherently unworkable.\n\nThe reframe: **the argument is the product, the graph is a map.** Maps are useful for orientation but you don't navigate a city by constantly staring at the map. You walk the streets, and consult the map when you need to reorient. Mobile argument UX should work the same way — the core interaction is moving through arguments one at a time, with the graph available for orientation when needed.\n\nThis research explores what that reframed mobile experience could look like.\n\n---\n\n## 1. Mobile Interaction Paradigms That Could Apply\n\n### 1.1 Card-Based UX\n\n**Pattern**: One argument per screen, full-bleed card. User swipes or taps to act.\n\n**Proven precedents**:\n- **Tinder** (swipe to accept/reject): Proven at hundreds of millions of users that binary swipe gestures are learnable and satisfying. The key insight from Tinder is that full-screen focus eliminates comparison anxiety — you evaluate this thing, now, on its own terms.\n- **Duolingo** (lesson cards with immediate feedback): Shows that bite-sized units with clear progress create habit-forming engagement loops. Each lesson card is self-contained and completable in under 2 minutes.\n- **Flashcard apps (Anki, Quizlet)**: Spaced repetition on mobile proves that structured review of specific items is deeply compatible with mobile sessions. The card metaphor maps directly to reviewing individual arguments.\n\n**Application to Deliberus**:\n\n```\n┌─────────────────────────────────┐\n│  CONTENTION                     │\n│  \"Carbon taxes reduce emissions  │\n│   more reliably than cap-and-   │\n│   trade systems\"                │\n│                                 │\n│  [↓ 3 supporting premises]      │\n│                                 │\n│  Quality: ████████░░ 78%        │\n│  Community: 342 evaluations     │\n│                                 │\n│  ────────────────────────────── │\n│                                 │\n│  [WELL-ARGUED]  [AGREE]         │\n│  [FLAG ISSUE]   [COUNTER]       │\n│                                 │\n│  ← prev        next →           │\n└─────────────────────────────────┘\n```\n\nKey decisions for card UX:\n- Swipe direction semantics need to feel natural. \"Agree/disagree\" as left/right could work, but may feel too binary for structured argumentation. Tap-based controls may be better — swipe for navigation (move to next/prev argument), tap for evaluation.\n- The card should show argument quality scores immediately — this is a key differentiator from social media cards, which show engagement metrics. Deliberus cards show epistemic quality metrics.\n- Pull-down or tap a disclosure trigger to reveal the full premise chain without leaving the card.\n\n**Stack metaphor**: A \"deck\" of arguments on a topic, ordered by controversy or quality. Work through the deck. This creates a natural mobile session: \"I have 10 minutes, let me evaluate some arguments on climate policy.\" The deck UI makes progress tangible.\n\n### 1.2 Vertical Scroll (Feed-Based)\n\n**Pattern**: Arguments as rich cards in a vertically scrolling feed, similar to Twitter/X or a Reddit thread but with structured metadata visible.\n\n**Proven precedents**:\n- **Twitter/X**: Demonstrated that text content with structured metadata (who, when, how many responses) works at scale on mobile.\n- **Reddit mobile**: Adapted threaded discussions to mobile by collapsing depth — top-level comments are full-width, replies indent and compress. This preserves the hierarchy while keeping the primary content readable.\n- **Hacker News on mobile** (via third-party apps like Apollo): Shows that technical/intellectual content works on mobile when the UX respects the content rather than trying to gamify it.\n\n**Application to Deliberus**:\n\nThe feed approach works well for browsing a debate landscape — see the top arguments on a topic, scroll through, pull to expand any argument's premise chain. Each feed item is:\n\n```\n┌─────────────────────────────────┐\n│ PRO  \"Carbon taxes are easier   │\n│       to administer than        │\n│       cap-and-trade\"            │\n│                                 │\n│ Quality: 82% · 12 premises      │\n│ [▼ expand] [💬 34] [↑ 128 ↓ 12]│\n└─────────────────────────────────┘\n\n┌─────────────────────────────────┐\n│ CON  \"Carbon taxes are          │\n│       politically unpopular     │\n│       and get repealed\"         │\n│                                 │\n│ Quality: 71% · 8 premises       │\n│ [▼ expand] [💬 21] [↑ 89 ↓ 44] │\n└─────────────────────────────────┘\n```\n\nThe feed approach has a failure mode: infinite scrolling becomes passive. Users consume without evaluating. Design must create friction at the right points — when a user has scrolled past 5 arguments without evaluating any, the interface should gently prompt: \"You've been browsing — want to weigh in on an argument?\"\n\n**Embedded mini-visualizations**: Each feed card could include a small inline visualization showing the argument's position in the broader debate graph — a 50x50 thumbnail showing this node highlighted in context. Tapping the thumbnail zooms into the full graph view. This keeps the spatial understanding available without making it the primary interface.\n\n### 1.3 Conversational UI\n\n**Pattern**: AI presents arguments conversationally, user responds naturally, AI structures the response into the argument graph.\n\n**Proven precedents**:\n- **Deliberation.io's Socratic mode**: Validated at city scale (Washington D.C., July 2025) that AI-mediated 1-on-1 reflection improves deliberation quality. The pushiness levels (1-3) and \"one question per turn\" constraint are well-engineered.\n- **Claude mobile app**: Shows that high-quality intellectual conversation works on mobile without any special UX — the conversation IS the interface.\n- **Duolingo's AI conversations** (2024+): Language learning through natural conversation proves that structured learning can be disguised as casual chat.\n\n**Application to Deliberus**:\n\nA conversational mode where an AI presents a contention and asks the user to respond. The AI then extracts the user's argument structure from the natural language response:\n\n```\nAI: \"The debate is: 'Should AI-generated content require\n     disclosure?' Here's the strongest argument for\n     mandatory disclosure: [argument text]. Do you\n     find this convincing? Why or why not?\"\n\nUser: \"I think it's a valid point but it ignores how\n       hard disclosure would be to enforce on social\n       media platforms. There's no way to verify...\"\n\nAI: \"That's a 'practical enforceability' objection.\n     I've noted that as a counterargument. The\n     proponents respond: 'Platform-level detection\n     tools like watermarking make this more\n     feasible than enforcement of prior disclosure\n     laws.' Does that change your view?\"\n```\n\nBehind the scenes, the conversation is building structured argument nodes. The user never sees the graph during this interaction — they experience it as a debate with a knowledgeable interlocutor. The graph is built from the conversation and available for later analysis.\n\nThis mode is ideal for first-time users and for mobile. It removes the cognitive overhead of argument mapping entirely — the user just talks. The structure emerges from the conversation.\n\n**Short-authored-input refinement (Apr 8, 2026):** A short question or topic in the landing textbox is a first-class version of this conversational mode, not a failed source document. The first move should be a truth-graph/background-knowledge response matched to the user's brevity, lingo, and apparent level; the second move should be provisional structure that the user can confirm before anything becomes durable graph content. In mobile terms: capture the raw utterance immediately, publish deliberately after a one-tap or light-edit confirmation. See [truth-graph-evidence-system.md](truth-graph-evidence-system.md).\n\n**Key design challenge**: The AI must be transparent that it is extracting structure. A confirmation UI after each exchange: \"I've recorded your objection as: [structured form]. Does this capture what you meant? [Edit] [Looks right]\"\n\n### 1.4 Map/Explore Mode\n\n**Pattern**: Simplified graph view with semantic zoom, optimized for touch navigation.\n\n**Proven precedents**:\n- **Google Maps**: Pinch-to-zoom with semantic zoom (street-level detail appears at high zoom, city-level labels at low zoom). Pan gestures feel natural. Tap to select, then get detail in a bottom sheet.\n- **Apple Maps**: The \"bottom sheet\" pattern — primary view is the map, tapping a location brings up a detail sheet from the bottom that covers part of the screen without replacing it.\n- **Notion's graph view / Obsidian mobile**: Knowledge graph browsing on mobile. Obsidian's mobile graph view is actually reasonably usable for small-to-medium graphs (under ~200 nodes).\n\n**Application to Deliberus**:\n\nThis mode works for orientation and navigation, not for deep reading. Design it as an entry point, not a primary experience:\n\n```\n┌─────────────────────────────────┐\n│  [Galaxy view: topic clusters]  │\n│                                 │\n│   ○ Climate      ○ Economy      │\n│      ↕               ↕          │\n│   ○ Energy   ── ○ Policy        │\n│                                 │\n│  Pinch to zoom into a cluster   │\n└─────────────────────────────────┘\n            ↓ zoom in\n┌─────────────────────────────────┐\n│  [Constellation: key claims]    │\n│                                 │\n│  ● Carbon tax [68% supported]   │\n│  ↕                              │\n│  ● Cap-and-trade [54%]          │\n│  ↕                              │\n│  ● Carbon removal [47%]         │\n│                                 │\n│  Tap any node to focus          │\n└─────────────────────────────────┘\n            ↓ tap node\n┌─────────────────────────────────┐\n│  [Bottom sheet: argument card]  │\n│  \"Carbon tax is more effective\" │\n│  [Read full] [Evaluate] [Map]   │\n└─────────────────────────────────┘\n```\n\nThe semantic zoom levels from the graph-visualization.md research map cleanly onto this:\n- **Galaxy** (zoomed out): Topic clusters\n- **Constellation**: Key contentions with pro/con balance indicators\n- **Star system**: Individual arguments\n- **Planet** (zoomed in): Full premise chains and evidence\n\nOn mobile, users would typically navigate from Galaxy → Constellation → tap to bottom sheet → read in card view. The full Star system and Planet views are better suited to tablet or desktop.\n\n---\n\n## 2. Microinteractions for Mobile Argumentation\n\nThe atomic interactions are the make-or-break factor for mobile. Each action must take under 2 seconds and require fewer than 3 taps from any state.\n\n### 2.1 Voting on Argument Quality (Not Just Agree/Disagree)\n\nThe critical differentiation from social media: Deliberus separates **epistemic quality** from **personal agreement**. A user can find an argument well-structured and evidenced (high quality) while personally disagreeing with its conclusion.\n\nThree-axis evaluation, designed for mobile:\n\n```\n┌─────────────────────────────────┐\n│  Rate this argument:            │\n│                                 │\n│  LOGIC:   ○ ○ ● ○ ○   Strong   │\n│  EVIDENCE: ○ ○ ○ ● ○   Good     │\n│  STANCE:  [AGREE] [NEUTRAL] [DISAGREE] │\n│                                 │\n│  [SUBMIT]  [SKIP]               │\n└─────────────────────────────────┘\n```\n\nAlternatively — and perhaps better for mobile — a single gesture-based rating: slide a thumb from \"weak argument\" to \"strong argument\" on a continuous scale, then separately tap agree/neutral/disagree. Two interactions total. This maps cleanly to the Bayesian model Deliberus's academic foundations suggest: quality (prior) + personal stance (likelihood update).\n\n**Anti-pattern to avoid**: The Polis agree/disagree/pass model is mobile-friendly but epistemically impoverished. It can't distinguish \"I disagree with the conclusion but respect the reasoning\" from \"this argument is weak and I disagree.\" Both are important signals in structured argumentation.\n\n### 2.2 Expanding Premise Chain (Progressive Disclosure)\n\nTap to expand, never navigate away:\n\n```\n[ARGUMENT TEXT]\n  ▼ Show 3 supporting premises\n\n→ tap expands inline:\n[ARGUMENT TEXT]\n  ∧ Hide premises\n  ├─ P1: \"Studies show...\"  [Q: 85%]\n  ├─ P2: \"The mechanism...\"  [Q: 72%]\n  └─ P3: \"Historical data...\" [Q: 68%]\n  [+ ADD PREMISE]\n```\n\nEach premise is itself tappable to expand to its own card. Breadcrumbs at the top track depth:\n\n```\nClimate Policy > Carbon Tax Effectiveness > [current]\n```\n\nTap any breadcrumb to jump up the chain.\n\n### 2.3 Flagging a Logical Issue\n\nOne-tap, pre-categorized issues rather than free text (free text is terrible on mobile):\n\n```\n┌─────────────────────────────────┐\n│  Flag logical issue:            │\n│                                 │\n│  ○ Unsupported premise          │\n│  ○ Non sequitur                 │\n│  ○ Strawman                     │\n│  ○ False dichotomy              │\n│  ○ Missing evidence             │\n│  ○ Scope too broad              │\n│  ○ Other (add note)             │\n│                                 │\n│  [FLAG]  [CANCEL]               │\n└─────────────────────────────────┘\n```\n\nThis categorized approach does double duty: it guides users to think clearly about what type of issue they've spotted, and it provides structured data for Deliberus's logical analysis layer. Free text flags are noise; categorized flags are signal.\n\nWalton's argument schemes (from docs/academic-foundations.md) provide the right ontology here — the categories should map to recognized fallacy types from argumentation theory, not pop-psychology fallacy lists.\n\n### 2.4 Adding Evidence (Attach URL, Photo, Quote)\n\nThe mobile advantage: camera and share sheet. Adding evidence on mobile should leverage what phones do better than desktops:\n\n- **Share from browser**: User reads an article on mobile, taps share → Deliberus → selects which argument to add the URL as evidence to. The citation is auto-extracted (title, author, publication, date).\n- **Photo evidence**: Photograph a physical document, chart, or physical phenomenon. OCR extracts text, user selects the relevant quote.\n- **Quote selection**: Long-press in the browser to select text, share to Deliberus, system prompts \"which argument does this support/contradict?\"\n\nThe share-sheet integration is the highest-value mobile-specific evidence feature. Desktop doesn't have this. It makes evidence addition a natural part of reading rather than a separate task.\n\n### 2.5 Creating a Counter-Argument (Voice → Structure)\n\nThis is where mobile has a genuine advantage over desktop. Voice input for argument creation:\n\n1. User taps microphone icon on an argument card\n2. Speaks their counter-argument naturally (30-90 seconds)\n3. SuperWhisper/Whisper transcribes\n4. LLM extracts argument structure: conclusion + premises\n5. UI shows the extracted structure for user confirmation before posting\n\n```\n┌─────────────────────────────────┐\n│  Your counter-argument:         │\n│  [transcription...]             │\n│                                 │\n│  Extracted structure:           │\n│  CONCLUSION: \"Cap-and-trade     │\n│    enables innovation while     │\n│    carbon taxes don't\"          │\n│  PREMISES:                      │\n│  ├─ P1: \"Price signals alone...\" │\n│  └─ P2: \"The EU ETS shows...\"   │\n│                                 │\n│  [EDIT] [POST] [DISCARD]        │\n└─────────────────────────────────┘\n```\n\nThis is the voice-note equivalent for argumentation. It lowers the contribution barrier dramatically. Structuring an argument in writing requires cognitive effort that many people don't invest on mobile. Speaking it, then reviewing the extracted structure, is much lower friction.\n\nThe LLM extraction step is consistent with Deliberus's broader architecture (Claimify-style extraction) — this is just applying it to real-time voice input rather than corpus text.\n\n### 2.6 Sharing an Argument Chain (Deep Link)\n\nShare a specific argument with its full premise chain context:\n\n```\n📎 \"Carbon taxes are more effective than cap-and-trade\"\n   with 3 supporting premises — on Deliberus\n\n[deliberus.com/argument/xyz?depth=2]\n```\n\nThe deep link opens to that specific argument in card view, with breadcrumbs showing where it sits in the larger debate. This is a crucial viral loop — arguments are shared as standalone epistemic units, pulling people into the platform.\n\n**Twitter card support**: When a Deliberus argument link is shared to Twitter/X or iMessage, the preview card should show the argument quality score prominently (\"Logic: 84%, Evidence: 71%\") alongside the claim text. This signals that Deliberus is different from opinion-sharing — it's evidence-sharing.\n\n### 2.7 Bookmarking for Later Deep Reading\n\nSimple one-tap bookmark with optional context note:\n\n```\n[★ Bookmark]  →  \"Save for later\" with optional tag:\n                  ○ Read closely later\n                  ○ Need to fact-check\n                  ○ Potential evidence for my argument\n                  ○ [custom tag]\n```\n\nOffline sync: bookmarked argument subgraphs (argument + its premises to depth 2) are cached locally. Users can read saved arguments on the subway with no connectivity. Contributions queue for sync when back online.\n\n---\n\n## 3. Voice Input as Mobile Differentiator\n\nVoice as a first-class contribution mechanism, not an afterthought.\n\n**The pipeline** (already proven in Fredrik's stack):\n1. Wake phrase or tap microphone — Whisper (or SuperWhisper) handles transcription\n2. LLM extracts argument structure from natural speech\n3. User reviews and confirms the extracted structure\n4. Structured argument is added to the knowledge graph\n\n**Why this matters**: The activation energy for structured argument creation on desktop is high. Users must: click \"new argument,\" fill in a conclusion field, add premises one by one, link evidence. On mobile, this friction is multiplied. Voice removes the friction almost entirely — speaking is natural, transcription is fast, LLM extraction handles the structuring.\n\n**Existing tools** for reference:\n- **Otter.ai**: Real-time transcription with speaker identification. Shows that voice→structured text is mature technology.\n- **Fireflies.ai**: Meeting notes with action item extraction. Demonstrates LLM extraction of structured content from conversational speech.\n- **SuperWhisper** (Fredrik's current setup): High-quality voice dictation already integrated into the development environment. The same infrastructure is deployable on mobile.\n\n**Critical design principle**: The user must always be shown the extracted structure before it is posted. LLMs make extraction errors. The confirmation step is not optional — it is the trust layer. Without it, users will quickly learn not to trust that their voice contributions accurately represent their intent.\n\n**Privacy consideration**: Voice input creates a more sensitive data stream than text. Audio should be processed on-device where possible (Whisper.cpp runs on iPhone), or the audio should be deleted immediately after transcription. The stored record should be the text transcript, not the audio.\n\n---\n\n## 4. Notification Design for Argument Platforms\n\nMost notification systems are engagement-maximizing by default. Deliberus should be truth-seeking by default. This requires explicit notification design choices.\n\n### 4.1 When to Notify\n\n**High-value notifications** (send immediately):\n- \"Your argument has a new counter-argument\" — someone engaged with your specific reasoning, not just agreed/disagreed with your conclusion. This is intellectually meaningful.\n- \"New evidence found for your bookmarked argument\" — when a new study or source is added as evidence to an argument you're tracking.\n- \"A premise you support has been flagged\" — logical issues identified by the community in reasoning you've endorsed.\n- \"Your argument was evaluated as well-argued by [N] people\" — positive epistemic feedback (quality, not just agreement).\n\n**Digest notifications** (send daily/weekly):\n- \"The climate debate evolved: 3 new top-rated arguments this week\"\n- \"Your contributed argument now has 45 evaluations — here's the quality breakdown\"\n- \"Debate you're following: new consensus emerging on premise quality\"\n\n**Do not notify** (engagement traps):\n- \"X people agreed with you\" — agreement notification drives validation-seeking, not quality-seeking\n- \"Trending in your network\" — FOMO-based notification, disconnected from argument quality\n- \"New activity in [debate]\" — too broad, signals nothing meaningful\n- Every new comment on a thread you participated in (Reddit-style) — overwhelming and unfocused\n\n### 4.2 Notification Framing\n\nThe language matters enormously. Two notification designs for the same event:\n\n**Engagement-optimizing** (avoid): \"💬 5 people replied to your post!\"\n**Truth-seeking** (use): \"Your premise 'Carbon taxes are regressive' has a well-argued counter — 3 logical objections raised.\"\n\nThe second notification is longer but the user who reads it will arrive at the app prepared to engage with a specific epistemic challenge, not to see a like count.\n\n### 4.3 Belief Calibration Updates\n\nAn ambitious but powerful notification type: when the aggregate quality assessment of an argument changes significantly based on new evidence or evaluations:\n\n\"The argument you find 'well-evidenced' has been downgraded by the community: a key premise was found to misrepresent a cited study. Review?\"\n\nThis is essentially a Bayesian update notification — the community's epistemic state has changed and you should update yours. No platform does this today. It requires the quality-scoring infrastructure that Deliberus is designed to have.\n\n---\n\n## 5. Offline-First and Progressive Web App\n\n### 5.1 PWA Capabilities\n\nA PWA (Progressive Web App) for Deliberus has significant advantages:\n- Install from the browser without an App Store — lower acquisition friction\n- Same codebase as the web app — no platform split to maintain\n- Background sync API (Chrome/Android, not Safari/iOS — see caveat below)\n- Push notifications (Chrome/Android, limited on iOS even post-iOS 16.4)\n- Service worker caching for offline argument browsing\n\n**iOS caveat** (March 2026): Apple's PWA support on iOS remains limited. Background sync is unreliable. Push notifications require the user to \"Add to Home Screen\" first. For a broad audience, a native iOS app is eventually necessary. But a PWA is the right starting point — it validates the mobile UX before committing to native development.\n\n### 5.2 Offline Cache Strategy\n\nWhat to cache aggressively:\n- **Bookmarked argument subgraphs**: The argument + its full premise chain to depth 3. This is the primary offline use case: user saves arguments to read carefully later (commute, offline time).\n- **Debates the user is actively participating in**: Cache the full debate structure for debates where the user has contributed arguments.\n- **Draft contributions**: If a user starts writing/speaking an argument and loses connectivity, the draft is queued for upload. Never lose user-generated content.\n\nWhat not to cache:\n- The full argument database — too large, and stale graphs are worse than no graph\n- Real-time quality scores — these change as new evaluations come in and stale scores mislead\n\n### 5.3 Contribution Queuing\n\nWhen offline, allow users to:\n- Evaluate arguments (queue vote locally)\n- Draft counter-arguments (save locally, upload when connected)\n- Bookmark arguments (local immediately, sync when connected)\n\nShow a clear \"pending sync\" state so users know their contributions aren't yet visible to others. When sync happens, show a brief confirmation.\n\n**Conflict resolution**: If an argument was deleted while the user was offline and they submitted an evaluation of it, silently discard the evaluation. Don't show an error — it's not the user's fault and there's nothing to do. If an argument text was edited while offline and the user submitted a quality score, the score applies to the edited version — flag this in the audit log.\n\n---\n\n## 6. Successful Mobile Knowledge Tools for Inspiration\n\n### 6.1 Duolingo — Bite-Sized Mastery Loops\n\n**What it does**: Breaks language learning into 5-10 minute lessons with immediate feedback, streak tracking, and gamified progression. Over 575 million registered users.\n\n**Mobile design lessons for Deliberus**:\n- **Session completion** must be tangible. \"You evaluated 5 arguments\" is a better session end state than scrolling until you close the app.\n- **Streaks work**, but only when the daily task is well-defined. \"Evaluate at least 3 arguments\" is a good daily commitment. \"Participate in deliberation\" is too vague.\n- **Immediate feedback** after every evaluation: show the community's quality score compared to yours. This is genuinely educational — users learn what makes arguments strong by comparing their judgments to the collective assessment.\n- **Celebrate quality, not quantity**: Duolingo awards XP for completing lessons. Deliberus should award recognition for well-received contributions, not just for volume of interaction.\n\n### 6.2 Pol.is — Mobile-Native Opinion Voting\n\n**What it does**: Shows statements one at a time, collects agree/disagree/pass. PCA + k-means finds opinion clusters and consensus.\n\n**Why it works on mobile**: The interaction model — one statement, three buttons — is inherently mobile-friendly. No complex UI. Users can participate meaningfully in 2 minutes.\n\n**Lessons for Deliberus**: The Pol.is interaction model is a good pattern for one specific use case: initial topic framing. Before a full argument evaluation session, show users 10-15 key claims from a debate in Pol.is style, quickly. This orients them to the debate landscape without the cognitive overhead of full argument evaluation. Use it as an onboarding mode.\n\n**The Pol.is limitation** to not replicate: the agree/disagree model can't distinguish epistemic quality from personal stance. Deliberus must avoid this collapse. The mobile interaction can be simple (two swipes or three buttons) but it must preserve the logic/agreement distinction.\n\n### 6.3 Reddit Mobile — Threaded Discussion at Scale\n\n**What Reddit mobile does right**:\n- Collapse comment threads to reduce visual noise — users control depth\n- Distinguish top-level arguments from replies visually and spatially\n- Upvote/downvote is visible on every item without navigation\n- Share, save, report as swipe actions (on third-party apps)\n\n**What Reddit mobile does wrong** (Deliberus should avoid):\n- Upvotes conflate quality with agreement with popularity\n- No structure to threads — any reply to any comment, any depth\n- No logical relationship metadata — \"counterargument\" looks the same as \"tangentially related observation\"\n\nDeliberus can take Reddit's navigation and collapse patterns but must add the logical structure that Reddit threads lack.\n\n### 6.4 Blinkist/Headway — Expert-Curated Card Stacks\n\n**What they do**: Book summaries as a structured card stack. Each card is a key insight. Users swipe through. Completion is satisfying.\n\n**Lesson**: The \"curated deck\" model works. For Deliberus, an editorially or algorithmically curated \"debate digest\" — the 8 most important arguments on a topic this week — would be a compelling mobile product. Not the full argument graph, not infinite scroll: a curated, completable deck.\n\n### 6.5 Twitter/X — Microblogging on Mobile\n\n**The key Twitter insight**: Short-form content with rich metadata (who, when, engagement) is highly scannable on mobile. The metadata is as important as the content.\n\n**For Deliberus**: Argument cards need visible metadata at a glance — logic score, evidence score, stance distribution, number of premises, evaluation count. This metadata should be visible without tapping. It lets users decide whether to engage deeply or skip.\n\n---\n\n## 7. The Dual-Mode Strategy\n\nDesign mobile and desktop as **different renderings of the same epistemic data**, optimized for different interaction patterns.\n\n### 7.1 Mobile: Consume, Evaluate, Quick-Contribute\n\nMobile users are:\n- Browsing in short sessions (5-15 minutes)\n- Reading and evaluating existing arguments\n- Making quick contributions via voice or short text\n- Saving content for later desktop engagement\n- Sharing argument chains to bring in collaborators\n\nMobile UX goals:\n- Every action completable in under 2 minutes\n- Clear session completion state (\"You've evaluated 8 arguments today\")\n- Voice-first contribution\n- Offline-capable for bookmarked content\n- Deep-link sharing for viral distribution\n\n### 7.2 Desktop: Analyze, Construct, Navigate the Graph\n\nDesktop users are:\n- Engaging in longer, focused sessions (30+ minutes)\n- Constructing detailed argument chains with precise premise structure\n- Navigating the full argument graph\n- Doing comparative analysis (which arguments are strongest? where are the logical gaps?)\n- Writing extended evaluations or rebuttals\n\nDesktop UX goals:\n- Full graph visualization with semantic zoom (see graph-visualization.md)\n- Multi-pane layout (graph + argument detail + evaluation panel)\n- Rich text editor for argument construction\n- DAG layout with hierarchical navigation\n\n### 7.3 Making Mobile Users Feel First-Class\n\nThe risk of a dual-mode strategy is that mobile becomes second-class — the place where you browse but can't do \"real\" work. Three design principles to avoid this:\n\n**Principle 1: Voice contribution is mobile-exclusive prestige**. Frame voice argument creation as the premium interaction mode, not a compromise. \"Deliberate by voice — speak your argument, we'll structure it.\" This makes the mobile-only feature a differentiator rather than a consolation.\n\n**Principle 2: Mobile evaluation is the primary data source**. The quality scores that desktop users rely on for analysis come from mobile users' evaluations. Mobile users are the crowd that makes the platform work. Make this explicit: \"Your evaluation of this argument has been counted by 340 people using it for analysis.\"\n\n**Principle 3: Seamless handoff**. Bookmarked arguments, drafted contributions, and browsing history sync instantly to desktop. \"Continue on desktop\" is a one-click action from any mobile state. The mobile session feeds the desktop session, not the other way around.\n\n---\n\n## 8. Text-Based Wireframe: The Core Mobile Loop\n\nA concrete walkthrough of the core mobile session from launch to completion:\n\n```\nLAUNCH → \"Today's deliberation digest\"\n┌─────────────────────────────────┐\n│ 🔥 Hot debate: AI Governance   │\n│ 5 new top-rated arguments       │\n│                    [START DECK] │\n├─────────────────────────────────┤\n│ 📌 Your bookmarks (3 unread)   │\n│                          [VIEW] │\n├─────────────────────────────────┤\n│ 🔔 Your argument challenged     │\n│ \"Carbon tax\" has 2 counters     │\n│                   [RESPOND] 🔴  │\n└─────────────────────────────────┘\n\n→ tap [START DECK]\n\nARGUMENT CARD (1 of 5):\n┌─────────────────────────────────┐\n│ AI Governance                   │\n│  1 of 5 ───────────────────     │\n│                                 │\n│ \"AI systems above a capability  │\n│  threshold should require       │\n│  mandatory third-party safety   │\n│  audits before deployment\"      │\n│                                 │\n│ Logic:    ████████░░  79%       │\n│ Evidence: ███████░░░  71%       │\n│ 189 evaluations                 │\n│                                 │\n│         [▼ 4 premises]          │\n│                                 │\n├─────────────────────────────────┤\n│ STANCE:                         │\n│ [AGREE]  [NEUTRAL]  [DISAGREE]  │\n│                                 │\n│ QUALITY:                        │\n│ [WEAK]  ●────────●  [STRONG]    │\n│         (drag)                  │\n│                                 │\n│ [🏳 FLAG ISSUE]  [🎤 COUNTER]  │\n└─────────────────────────────────┘\n\n→ user drags quality to 65%, taps DISAGREE\n\n→ INSTANT FEEDBACK (500ms overlay):\n┌─────────────────────────────────┐\n│ Community: Agree 61% / Dis 31%  │\n│ Your quality rating: ↑ avg       │\n│ You tend to rate this type of   │\n│ regulatory argument conserv...  │\n│                  [NEXT →]       │\n└─────────────────────────────────┘\n\n→ user taps [🎤 COUNTER]\n\nVOICE COUNTER:\n┌─────────────────────────────────┐\n│ 🎤 Recording... (tap to stop)  │\n│                                 │\n│ \"I think this assumes that      │\n│  safety audits can actually     │\n│  detect capability levels       │\n│  reliably, which hasn't been    │\n│  demonstrated...\"               │\n│                                 │\n│ [■ STOP]                        │\n└─────────────────────────────────┘\n\n→ EXTRACTED STRUCTURE:\n┌─────────────────────────────────┐\n│ Your counter-argument:          │\n│                                 │\n│ CONCLUSION:                     │\n│ \"Safety audit feasibility is    │\n│  unproven at scale\"             │\n│                                 │\n│ PREMISE:                        │\n│ \"No demonstrated method exists  │\n│  to reliably assess AI          │\n│  capability thresholds\"         │\n│                                 │\n│ [✏ EDIT]  [✓ POST]  [✗ DROP]  │\n└─────────────────────────────────┘\n\n→ user taps [✓ POST]\n→ argument added to graph, continue deck\n\nSESSION COMPLETE (5/5 cards):\n┌─────────────────────────────────┐\n│ ✓ Deliberation session done     │\n│                                 │\n│ You evaluated 5 arguments       │\n│ Your ratings: ~avg community    │\n│ You contributed 1 counter       │\n│ 3-day streak 🔥                 │\n│                                 │\n│ [EXPLORE MORE]  [DONE]          │\n└─────────────────────────────────┘\n```\n\n---\n\n## 9. Key Design Principles Summary\n\n**1. Argument-first, graph-second.** The graph is always available but never required. Users can participate fully in structured argumentation without ever seeing the graph visualization.\n\n**2. Separate quality from stance.** Every evaluation must distinguish \"is this well-argued?\" from \"do I agree with the conclusion?\" These collapse in social media and must not collapse here.\n\n**3. Voice as a first-class contribution method.** LLM-based argument structure extraction from natural speech is the mobile-native contribution pathway. It must be reliable and always subject to user confirmation.\n\n**4. Pre-categorized microinteractions.** Free text on mobile is friction. Flag categories, evaluation sliders, tap-to-expand premises — every common interaction should require zero typing.\n\n**5. Session completion is tangible.** Mobile sessions should have a clear beginning, middle, and end. The \"deck\" metaphor provides this. Infinite scroll does not.\n\n**6. Notifications are epistemic, not engagement-maximizing.** Notify when someone challenged your reasoning, not when someone agreed with it. Never notify for raw engagement metrics.\n\n**7. Offline-first for consumption, queued sync for contribution.** Reading arguments offline is essential. Contributing arguments offline with queued sync is achievable and important for developing-world adoption.\n\n**8. Mobile users are the primary data source.** The quality scores that make Deliberus valuable come from mobile evaluations. Design the evaluation UX for clarity and speed without sacrificing epistemic richness.\n\n---\n\n## 10. Open Questions for Design Work\n\n- **Swipe vs. tap for evaluation**: Swiping (Tinder-style) is more satisfying but limits the number of discrete evaluation states. Tapping button rows is less fluid but more expressive. Prototype both.\n- **Quality slider granularity**: Should users rate on a 5-point scale, a 100-point continuous slider, or a 3-point categorical? More granularity produces better data but more cognitive load per evaluation.\n- **Conversational AI mode as default vs. optional**: Should new users always start in conversational mode and opt into card/feed mode? Or should both be accessible from launch? Likely A/B test candidate.\n- **Voice argument quality**: LLM extraction from voice will occasionally miss the user's intent. What is the threshold for acceptable extraction quality? Below what accuracy does voice contribution become trust-destroying?\n- **Deck curation algorithm**: How should the \"daily digest\" deck be curated? Highest controversy? Most contested premises? Arguments where the user's past positions suggest they'd have useful perspective? This is a significant algorithmic design question.\n- **PWA vs native**: At what user scale does committing to native iOS/Android become necessary? The PWA limitation on iOS (background sync, push notifications) may drive this decision earlier than expected.\n\n---\n\n## Sources and Cross-References\n\n**Internal research**:\n- [graph-visualization.md](graph-visualization.md) — §8 Mobile section established the baseline problem statement; §3 Semantic Zoom provides the multi-level model adapted for mobile\n- [polis-deep-dive.md](polis-deep-dive.md) — Agree/disagree/pass interaction model; bridging statement computation relevant to mobile evaluation design\n- [deliberation-io.md](deliberation-io.md) — Socratic dialogue pushiness levels; validated conversational deliberation model\n- [adoption-problem.md](adoption-problem.md) — Contribution barrier is the core adoption problem; mobile UX is one of the primary levers\n- [academic-foundations.md](../academic-foundations.md) — Walton's argument schemes as ontology for flagging categories\n\n**External references**:\n- Horak et al. (2021), \"Responsive Visualization Design for Mobile Devices\" — 76 strategies for adapting visualizations to small screens\n- \"The Structure of Bridging\" (Blair et al., 2025, Toronto/Harvard) — group-aware consensus metrics relevant to mobile evaluation aggregation\n- Duolingo mobile UX analysis (internal, 2025) — gamified learning loop patterns\n- Polis source code (compdemocracy/polis) — agree/disagree/pass UX and underlying algorithm\n- SuperWhisper / Whisper.cpp — voice transcription infrastructure already proven in Fredrik's stack\n"}