{"path":"research/feed-algorithm-design.md","content":"# Feed Algorithm Design for Truth-Seeking: Anti-Engagement Principles and Epistemic Quality\n\n*Research compiled March 28, 2026 for the Deliberus project.*\n\n---\n\n## Overview\n\nDeliberus faces a design challenge that no major platform has fully solved: what should an argument feed optimize for when the goal is epistemic quality rather than user retention? The sketches show two poles — \"infinite scroll\" (engagement maximization, borrowed from social media) and \"organic redraw\" with a 60% quality threshold funnel (content-weighted surfacing). This document surveys the landscape of anti-engagement feed design, academic research on truth-seeking metrics, and synthesis principles for what a Deliberus feed algorithm might actually look like.\n\nThe core thesis of this research: **engagement metrics and epistemic metrics are not merely different — they are often directly opposed.** Designing for one actively degrades the other. The platforms that have gotten closest to epistemic quality did so by explicitly refusing engagement as a metric.\n\n**Implementation update (Apr 3, 2026):** The live `needs-help` feed has now begun to move in the direction argued for here. It no longer surfaces only unanswered critical questions; it also begins surfacing claims that still look bundled and likely want decomposition next. The important new detail is that this is no longer only a ranking signal: when decomposition is the clearest next intervention, the feed can now route directly into a focused decomposition path with backend-seeded guidance rather than an empty generic claim page. That work queue has also begun to descend into the CQ layer itself: bundled CQ polarity claims can now surface as decomposition targets when the next useful epistemic act is to unpack a yes/no CQ answer rather than only inspect the parent claim. This is still an early guidance layer, not yet the full clarification/deepening work queue envisioned below, but it marks a stronger convergence between the feed and the recursive-decomposition frontier.\n\n---\n\n## 1. Anti-Engagement Feed Design: What Already Exists\n\n### 1.1 Bluesky: Marketplace of Algorithms\n\nBluesky's most significant architectural contribution is not its chronological default, but its **algorithmic marketplace** — an open system where any developer can publish a custom feed algorithm, and users select which feeds they subscribe to ([Bluesky, July 2023](https://bsky.social/about/blog/7-27-2023-custom-feeds)). By mid-2025, 39,639 feeds had been created by 18,352 active users across a 37-million-user platform.\n\nThe marketplace model makes an important epistemological move: it **separates algorithm authorship from platform control**. No single company decides what is \"relevant.\" Users can subscribe to a chronological feed, a trending feed, a domain-specific feed built by a journalist cooperative, or a bridging-optimized feed built by a researcher. The platform provides the substrate; the epistemic choices are distributed.\n\n**Limitation for Deliberus**: Bluesky's feed marketplace is about content *selection* — which posts from the firehose to show. Deliberus's challenge is argument *evaluation* — how to surface arguments worth engaging with, weighted by epistemic properties. The marketplace model is an inspiration for governance (who controls ranking?) but not directly a ranking algorithm.\n\n**What Deliberus can borrow**: The governance insight. Deliberus should probably not have a single \"official\" feed ranking. Community-defined or user-modifiable ranking parameters let the platform remain epistemically neutral while enabling specialized use cases (education, policy deliberation, scientific communities).\n\n### 1.2 Mastodon: The Chronological Commitment\n\nMastodon's design philosophy is explicitly anti-algorithmic: users get a chronological feed of accounts they follow, with no algorithmic reranking, no promoted content, and no engagement optimization ([joinmastodon.org](https://joinmastodon.org/)). A 2025 CSCW study ([arxiv.org/abs/2504.18817](https://arxiv.org/abs/2504.18817)) found that most Mastodon users valued the chronological feed for its *transparency* — \"you know exactly what you missed while away\" — but were not opposed to *all* forms of curation, just to opaque algorithmic curation they couldn't inspect or control.\n\nThe key insight from Mastodon is not that chronological is optimal, but that **transparency about the ranking function is itself an epistemic value**. When users cannot see why something is surfaced, they cannot reason about whether to trust it or the system that surfaced it. A black-box algorithm is epistemically corrosive independently of whether its outputs are good.\n\n**What Deliberus can borrow**: Feed ranking should be auditable. Users should be able to see not just *what* the algorithm surfaced, but *why* — what properties of an argument caused it to appear. This is more achievable in a structured argument platform (where arguments have explicit properties: evidence quality, vote distributions, controversy score, recency) than in a social media platform (where content properties are opaque).\n\n### 1.3 Hacker News: Decay, Controversy Penalties, and Emergent Quality\n\nHacker News uses a deceptively simple formula: `Score = ((points - 1) / (hours + 2)^1.8) × penalties` ([righto.com](http://www.righto.com/2013/11/how-hacker-news-ranking-really-works.html)). Several properties are worth noting:\n\n- **Time decay is steep** (exponent 1.8): brand-new items with 2-3 votes can outrank old items with 30 votes. This creates natural content turnover and prevents accumulation effects.\n- **Controversy is penalized**: items that accrue many comments relative to votes (high controversy signal) are algorithmically downranked. This is the inverse of Twitter/X, which amplifies controversy because it drives engagement.\n- **Flagging and moderator intervention** can further penalize items by factors of 0.2–0.4. About 20% of front-page items receive some penalty.\n- **Subtlety**: the `-1` in the numerator means items with 1 point score zero — they don't reach the front page without community endorsement.\n\nThe social-protocols project ([github.com/social-protocols/quality-news](https://github.com/social-protocols/quality-news)) has gone further, proposing a \"quality news\" variant that attempts to detect and correct for **sample bias in upvoting** — early upvotes come from a non-representative subset of users (people who see items when new), so quality items that don't happen to get early votes are systematically underranked.\n\n**What Deliberus can borrow**: The controversy penalty is directly relevant — in a truth-seeking context, high controversy on an argument should increase inspection, not virality. Deliberus might invert this: high controversy arguments should surface to *more* scrutiny, not be amplified for their controversy alone. The time decay model is also relevant for controlling staleness — arguments about a topic that has been \"resolved\" in the community should decay in priority even if they were historically important.\n\n### 1.4 Wikipedia's Watchlist: Pull Over Push\n\nWikipedia's feed design is not an algorithm at all — it is a **pull-based watchlist** where editors subscribe to specific pages and see a chronological stream of changes to those pages ([Wikipedia Help:Watchlist](https://en.wikipedia.org/wiki/Help:Watchlist)). The 2025 Wikimedia annual plan explicitly aims to improve watchlist usability for editors with 100+ edits, focusing on \"helping editors find edits that relate to their interests.\"\n\nThis is the most radical departure from push-feed design: the user declares what they care about (by watching pages), and the feed is simply a log of changes to those things. There is no recommendation, no surfacing, no ranking — just a transparent event stream of changes to declared interests.\n\n**What Deliberus can borrow**: The \"watching\" primitive — users should be able to subscribe to specific argument threads, contentions, or topic areas and receive a feed of changes (new counter-arguments added, existing arguments updated, new evidence linked, vote distributions shifting). This is complementary to algorithmic surfacing, not a replacement. It gives users agency and reduces \"what did I miss?\" anxiety without relying on a black-box ranker.\n\n### 1.5 Substack: Author-Curated vs. Algorithm\n\nSubstack's recommendation system is hybrid — the algorithm surfaces newsletters, but editorial curation by the Substack team and by successful authors (who can recommend other newsletters to their subscribers) plays a significant role. The key design principle: **human curation by domain experts is algorithmically multiplied, not replaced**.\n\nFor Deliberus: argument threads that have been \"endorsed\" or highlighted by users with demonstrated track records of epistemic quality (calibrated belief updates, well-sourced arguments) might get additional surfacing weight. This is a reputation-weighted curation model, not pure algorithmic ranking.\n\n### 1.6 The Humane Technology Critique\n\nThe Center for Humane Technology (CHT), founded by Tristan Harris, has provided the most influential framework for understanding what is wrong with engagement-optimized feeds ([humanetech.com](https://www.humanetech.com/)). The core argument: engagement-maximizing design exploits psychological vulnerabilities (intermittent reinforcement, social comparison, outrage amplification) and produces what Harris calls \"human downgrading\" — weakening of attention, reasoning capacity, and democratic deliberation. The business model problem is structural: platforms that make money from attention cannot prioritize epistemic quality, because high-quality, nuanced content generates fewer clicks than outrageous, simple content.\n\nCHT has been criticized for overstating the strength of evidence connecting platform design to addiction-level harms. But the critique of the **engagement-as-proxy-for-value** assumption remains sound. Engagement is not quality; it is not even correlated with quality in reliable ways. The March 2025 Knight-Georgetown Institute report \"Better Feeds: Algorithms That Put People First\" ([kgi.georgetown.edu](https://kgi.georgetown.edu/research-and-commentary/better-feeds/)) formalized this critique into policy recommendations, arguing that feeds should rely on \"deliberative, onerous, clear, and stated preference signals\" rather than passive behavioral tracking (clicks, scroll time, hover).\n\n---\n\n## 2. What Should a Truth-Seeking Feed Optimize For?\n\n### 2.1 Epistemic Quality Metrics: The Research Landscape\n\nAcademic research on \"epistemic welfare\" in algorithmic systems is nascent but growing. A 2025 paper in *Communication Theory* ([academic.oup.com](https://academic.oup.com/ct/advance-article/doi/10.1093/ct/qtaf018/8240891)) frames the problem directly: \"the epistemic standard of reliability, when applied to algorithmic recommender systems, refers to how many false beliefs (errors) and justified beliefs (truths) the system disseminates.\" This is a useful first formulation, but operationalizing it is non-trivial — you need ground truth to measure false beliefs, and most interesting deliberative content concerns contested claims where ground truth is precisely what is in dispute.\n\nMore tractable epistemic quality proxies include:\n\n- **Calibration**: does exposure to this argument help users form beliefs that match the actual probability distribution of outcomes? Well-calibrated content presents uncertainty accurately rather than asserting false certainty. ([thesephist.com on epistemic calibration](https://thesephist.com/posts/epistemic-calibration/))\n- **Source quality**: does the argument link to high-quality evidence? Evidence quality can be assessed via citation analysis, fact-checking API integration, or community sourcing assessment.\n- **Argument completeness**: does the argument acknowledge its strongest counter-arguments? Arguments that ignore all counter-evidence are epistemically lower quality than those that engage with them.\n- **Belief-update potential**: does the argument contain information likely to update beliefs of people who currently disagree with its conclusion? (See §3.)\n\n### 2.2 Deliberative Polling: Structured Information Changes Minds\n\nJames Fishkin's Deliberative Polling® method ([deliberation.stanford.edu](https://deliberation.stanford.edu/what-deliberative-pollingr)) provides the most robust empirical evidence that structured information exposure changes beliefs. The method: random sample of citizens receives balanced briefing materials, deliberates in moderated small groups, and then re-measures opinions. Key findings:\n\n- **~70% of participants change their views** on at least one issue after structured deliberation with balanced information.\n- A 2024 \"America in One Room\" national poll on democratic reform showed an **18-point drop in overall dissatisfaction** with US democracy after deliberation, with significant depolarization across party lines on specific issues.\n- Changes are not random — they tend toward positions better supported by evidence, and away from positions held primarily due to partisan cue-following.\n\nThe key variable in Deliberative Polling is not who deliberates but what they deliberate *with* — the briefing materials are curated for balance and accuracy, not for engagement. Fishkin's empirical result is essentially a proof of concept for truth-seeking feeds: **structured, balanced information exposure moves the distribution of beliefs toward better-calibrated positions.**\n\nFor Deliberus: the feed should function as a personalized briefing document rather than a social stream. It should ensure that users who hold position X are exposed to the strongest arguments for not-X — not to \"balance\" as a value in itself, but because exposure to well-articulated counter-arguments is the mechanism by which beliefs update.\n\n### 2.3 Bayesian Surprise: Information-Theoretic Approach to Content Value\n\nBayesian surprise offers an information-theoretic approach to ranking content value: an argument is more valuable if it **maximally updates the beliefs of its intended audience**. Formally, this is the KL-divergence between posterior and prior beliefs after exposure to the argument.\n\nResearch on Topic-Level Bayesian Surprise for recommender systems ([arxiv.org/abs/2308.06368](https://arxiv.org/abs/2308.06368)) demonstrates that content with high Bayesian surprise — items that are surprising given a user's consumption history — tends to be rated more highly in post-hoc evaluation, even though users wouldn't have selected it in advance. This is the key asymmetry: **people don't know what will change their minds, so they can't filter for it voluntarily.**\n\nThis has direct implications for Deliberus: a purely preference-based ranking (show users arguments they already find compelling) will systematically underweight arguments that would most update their beliefs. The optimal epistemic feed must override stated preferences in favor of *estimated belief-update potential*.\n\nThe practical challenge: computing prior beliefs requires knowing what a user currently believes, which requires either explicit belief elicitation (asking users to rate their confidence on claims before and after exposure) or inference from voting behavior. Both are tractable in Deliberus's architecture — voting on arguments is the primary interaction primitive, and can double as prior belief registration.\n\n### 2.4 Controversy as Signal for Attention, Not Amplification\n\nIn engagement-optimized feeds, controversy is amplified — controversial content generates more interaction, so it rises. In an epistemic feed, controversy should function differently: **controversy is a signal that a question deserves more community attention, but not that it should be shown to everyone.**\n\nSpecifically, high controversy on an argument (approximately equal pro and con votes) indicates one of three things:\n\n1. The argument touches a **genuine empirical uncertainty** — the community doesn't know the answer because nobody does yet.\n2. The argument touches a **genuine value disagreement** — the community is divided because different people have different underlying values.\n3. The argument has **poor quality on both sides** — both supporting and opposing arguments are weak.\n\nCase 1 is an invitation for evidence-gathering. Case 2 is an invitation for value clarification. Case 3 is an invitation for argument improvement. All three warrant surfacing to *engaged contributors* who can help resolve the controversy — not to passive readers who will only pile on with low-quality takes.\n\nThe Deliberus feed should implement a **controversy routing mechanism**: highly controversial arguments get surfaced to users who have demonstrated ability to make high-quality contributions (evidence-linking, well-structured premises), not to users who primarily vote without engaging with argument structure.\n\n### 2.5 Evidence Gap Detection: Surfacing Claims That Need Work\n\nAn under-explored feed design primitive is **evidence gap detection** — surfacing claims that have been asserted but lack supporting evidence links. In a truth-seeking platform, the most valuable contribution isn't always adding a new argument — it's adding evidence to an existing, under-supported claim.\n\nResearch on automated fact-checking shows that argument mining systems can classify claims by their evidential support status (well-supported, partially supported, unsupported) ([hal.science paper on argumentation and fact-checking](https://hal.science/hal-05017906)). This classification, applied to Deliberus's argument graph, produces a natural feed priority signal: **under-evidenced claims that are load-bearing in important arguments should surface to users with relevant domain expertise.**\n\nThis inverts the typical social media feed logic (surface what is popular) in favor of a research/wiki model (surface what needs work). It also creates a constructive engagement invitation — \"this argument is widely cited but has no evidence linked, can you help?\" — rather than the adversarial frame of debate.\n\n---\n\n## 3. The \"Show You What Might Change Your Mind\" Algorithm\n\n### 3.1 Argument Convincingness vs. Argument Quality\n\nResearch on argument convincingness reveals a crucial distinction: **an argument can be rhetorically convincing without being logically strong, and logically strong without being convincing to its target audience.** These two dimensions are different and often decorrelated.\n\nA 2024 NLP study ([arxiv.org/html/2404.00750v1](https://arxiv.org/html/2404.00750v1)) showed that LLMs can predict which side of a debate participants find more convincing — and that combining predictions from multiple LLMs approaches human-level performance on this task. A 2025 follow-up explored multi-turn debate formats where models must update beliefs as counter-arguments arrive — precisely the dynamic that a Deliberus feed would generate.\n\nThe distinction matters for feed design: an algorithm optimizing for *argument quality* (logical validity, evidence strength) will surface different content than one optimizing for *argument convincingness to skeptics* (psychological resonance with people who currently disagree). For epistemic purposes, the latter is more important. An argument that is technically valid but completely unconvincing to the audience it needs to persuade has low epistemic utility.\n\n### 3.2 Polis's Bridging Statements as a Model\n\nPolis's bridging statement algorithm ([research/polis-deep-dive.md](polis-deep-dive.md)) identifies comments that achieve high agreement across participants who otherwise disagree. This is the closest existing implementation of a \"what might change minds across groups\" algorithm.\n\nA 2025 academic analysis by Blair et al. ([cs.toronto.edu/~nisarg/papers/bridging.pdf](https://www.cs.toronto.edu/~nisarg/papers/bridging.pdf)) proposes improvements using pairwise disagreement metrics and p-mean bridging — stable, interpretable metrics that work even with sparse vote data. The key insight: bridging is not about finding the \"middle ground\" position, but about finding *arguments* that people with opposing views can both endorse. These arguments often work by finding a shared premise beneath a surface disagreement.\n\nFor Deliberus: a bridging-weighted feed would prioritize arguments that achieve agreement across opinion clusters, not just aggregate agreement. An argument endorsed by 80% of users is less interesting than one endorsed by 70% of users who otherwise hold opposing positions — the latter actually bridges; the former might just reflect a consensus that already exists.\n\n### 3.3 Detecting Actual Belief Updates (Pre/Post Measurement)\n\nThe gold standard for \"did this argument change minds\" is pre/post belief measurement. If users register their confidence in a claim before and after reading an argument, the delta is a direct measure of the argument's epistemic impact.\n\nThis is standard practice in Deliberative Polling but rare in online platforms. The friction is high: users don't want to fill out surveys. But in Deliberus's architecture, voting on claims is already a belief-registration action. If users vote on a claim at time T₁ and then revote at time T₂ after engaging with new arguments, the vote delta is a measured belief update.\n\nThis creates a possible virtuous cycle: arguments that demonstrably caused vote changes on the parent claim get surfaced as high-epistemic-value; this brings more users to those arguments; more vote measurements become available; the signal gets stronger. The feed algorithm becomes **empirically calibrated to argument effectiveness** rather than relying on proxies.\n\nThe engineering challenge: tracking individual vote history over time and attributing belief changes to specific argument exposures. This requires more sophisticated data modeling than a simple vote count, but it is architecturally feasible.\n\n### 3.4 The Filter Bubble Complexity\n\nThe filter bubble hypothesis — that algorithmic personalization creates ideological isolation that drives polarization — has received significant empirical scrutiny, with mixed results. A large-scale study with ~9,000 participants found that exposure to algorithmically slanted content produced \"little to no change\" in political opinions ([Harvard Government Department, 2025](https://www.gov.harvard.edu/2025/05/08/short-term-exposure-to-filter-bubble-recommendation-systems-has-limited-polarization-effects/)). This challenges the naive version of the filter bubble story.\n\nHowever, the relevant question for Deliberus is not \"do filter bubbles cause polarization\" but \"does homogeneous information exposure prevent belief updating.\" These are different questions. Even if filter bubbles don't radicalize people, they may prevent the kind of belief updating that Fishkin's deliberative polling demonstrates is possible with balanced information exposure. A Deliberus feed designed for epistemic quality should err on the side of ensuring counter-argument exposure regardless of whether filter bubbles cause measurable polarization.\n\n---\n\n## 4. Feed Personalization Without Echo Chambers\n\n### 4.1 Exposure Diversity as a Design Principle\n\nHelberger et al.'s foundational work on \"Exposure diversity as a design principle for recommender systems\" ([ResearchGate](https://www.researchgate.net/publication/311957085_Exposure_diversity_as_a_design_principle_for_recommender_systems)) argues that diversity should be a first-class optimization objective alongside relevance, not a post-hoc correction. Diversity in this context means exposure to arguments from different perspectives, different evidential bases, and different underlying value assumptions — not diversity for its own sake.\n\nThe research distinguishes several types of diversity:\n- **Topical diversity**: exposure to different subtopics within a domain\n- **Perspective diversity**: exposure to arguments supporting different positions\n- **Source diversity**: exposure to evidence from different methodological traditions\n- **Temporal diversity**: exposure to both recent and historically settled arguments\n\nFor Deliberus, all four matter. An argument feed that shows only recent arguments misses historically important settled points. One that shows only pro-side arguments to pro-side users fails epistemically. One that shows only empirical evidence misses value-based arguments that may be the actual load-bearing points in a normative dispute.\n\n### 4.2 Serendipity Without Noise\n\nSerendipity in recommendation systems — surfacing content that is surprising but ultimately valued by the user — is distinct from random diversification. Research on Topic-Level Bayesian Surprise ([arxiv.org/abs/2308.06368](https://arxiv.org/abs/2308.06368)) shows that serendipitous recommendations (in the Bayesian surprise sense) are rated more highly in post-hoc evaluation than random diversification, even when users wouldn't have selected them in advance.\n\nFor Deliberus, serendipity might mean surfacing an argument from a domain the user doesn't usually engage with (e.g., surfacing an economic argument to a user who primarily engages with ethical arguments on the same topic), or surfacing a historical argument that pre-dates the current controversy and establishes the conceptual foundations differently. The criterion is not \"unexpected\" but \"useful-in-a-way-you-wouldn't-have-predicted.\"\n\n### 4.3 The Nutritional Metaphor\n\nThe \"information diet\" metaphor — treating information exposure the way we treat food nutrition — is a useful design heuristic. A nutritious diet is not one where you eat only what you crave; it is one that meets full nutritional requirements, including nutrients you don't notice you're deficient in. Similarly, an epistemically nutritious argument feed is not one that shows you only arguments you find compelling — it ensures full coverage of the argument landscape for a topic, including arguments that challenge your current views.\n\nOperationally, this suggests a **coverage objective** alongside a relevance objective: the feed should ensure that a user engaging with a topic is exposed to at least one strong argument from each major position cluster, not just the cluster they already agree with. This mirrors the Deliberative Polling briefing material design principle.\n\nThe \"organic redraw\" concept from Deliberus's Sketch 4 is consistent with this: the 60% quality threshold funnel doesn't just filter for quality — it redisplays the filtered result across the full argument space, ensuring the user sees the strongest representatives of each position, not just their preferred position.\n\n### 4.4 The Allostatic Regulation Model\n\nA 2025 paper on \"reducing echo chamber effects\" ([tandfonline.com, 2025](https://www.tandfonline.com/doi/full/10.1080/29974100.2025.2517191)) proposes an **allostatic regulator** for recommendation algorithms, based on opponent process theory. The idea: the feed actively monitors the user's exposure diversity and dynamically adjusts recommendations to counteract accumulated homogeneity. If a user has seen 10 consecutive pro-X arguments, the regulator surfaces pro-not-X arguments, not as a punishment but as a correction toward epistemic balance.\n\nThis is more sophisticated than simple diversity weighting because it tracks cumulative exposure rather than just current-item diversity. A user who just engaged with a series of one-sided arguments gets a counter-argument surfaced; a user who has already seen balanced coverage doesn't get redundant \"balance\" imposed.\n\n---\n\n## 5. Real-Time Feed Evolution\n\n### 5.1 The \"Organic Redraw\" Concept\n\nThe Deliberus sketch describes the feed as \"organic (redraw)\" with a 60% quality threshold — contrasting with infinite scroll's passive accumulation. The technical interpretation: rather than appending new content to the bottom of a list, the feed continuously recomputes salience and redistributes arguments in the displayed space.\n\nThis is architecturally closer to a **dashboard** than a timeline. Arguments don't \"fall off\" as time passes — they are continuously re-ranked based on:\n- New evidence linked\n- Changes in the vote distribution (shifting consensus)\n- New counter-arguments added\n- Changes in the user's own belief state (tracked via votes)\n- New community activity on the argument\n\nThe 60% quality threshold is a filter layer: only arguments meeting a minimum quality criterion (based on evidence links, vote distribution, structural completeness) appear in the primary feed at all. Below-threshold arguments exist in the database and can be found via search or exploration, but they don't compete for primary feed space.\n\n### 5.2 Continuous Reranking vs. Batch Updates\n\nContinuous reranking (update rankings on every new vote or argument addition) has two failure modes:\n- **Instability**: the feed changes too rapidly for users to develop a coherent picture of a debate\n- **Computational cost**: recomputing rankings across the full argument graph on every write is expensive at scale\n\nBatch reranking (recalculate every N minutes or hours) avoids both problems but introduces lag. The tradeoff depends on debate velocity — a fast-moving political debate needs more frequent reranking than a slow-moving philosophical one.\n\nPolis addresses this with its continuous async math worker ([polis-deep-dive.md §1.4](polis-deep-dive.md)): recomputation runs continuously in the background and results are pushed to clients, but the UI smooths out rapid changes through transition animations. The argument graph updates continuously; the user's *experience* of the update is batched and smoothed.\n\nFor Deliberus: the underlying argument graph should be updated continuously (every vote, every new argument), but the feed presentation should update on a slower cycle (every few minutes for active debates, daily for slow-moving ones). This prevents the feed from feeling unstable while ensuring that important new arguments surface promptly.\n\n### 5.3 Notification Design for Counter-Arguments\n\nThe most valuable notification a truth-seeking platform can send is: \"A new counter-argument has been added to an argument you endorsed.\" This directly invites belief revision — the user is shown an argument that challenges a position they have expressed agreement with.\n\nThis is the inverse of typical social media notification design, which notifies you when someone agrees with you (likes, shares) — positive reinforcement for already-held positions. A truth-seeking notification model preferentially notifies about challenges, not endorsements.\n\nImplementation: when a user votes positively on argument A supporting position P, they should receive a notification when:\n1. A new argument attacking argument A is added and receives meaningful positive votes from other users\n2. The vote distribution on argument A shifts significantly (e.g., from +60%/-40% to +45%/-55%), indicating the community has updated its assessment\n3. New evidence is added that substantially strengthens or weakens argument A's empirical premises\n\nNotification fatigue is a real risk. The solution is **threshold-based triggering**: notifications fire only when the epistemic change exceeds a minimum significance threshold (e.g., new counter-argument has > 50 votes, or vote distribution shifts > 10 percentage points).\n\n---\n\n## 6. Design Principles for Deliberus\n\nThe following principles synthesize the research above into concrete feed algorithm design guidance. These are principles, not decisions — consistent with the project's philosophy that the design space must be explored before choices are made.\n\n### Principle 1: Optimize for Belief Updates, Not Engagement\n\nEvery ranking signal should be evaluated by asking: \"Does surfacing this argument more increase the probability that users form better-calibrated beliefs?\" Engagement signals (click-through, time-on-page, re-sharing) should be treated as weak proxies, overridable by stronger epistemic signals. The Knight-Georgetown Institute recommendation ([kgi.georgetown.edu, March 2025](https://kgi.georgetown.edu/research-and-commentary/better-feeds/)): rely on \"deliberative, onerous, clear, and stated preference signals\" — in Deliberus, this means explicit votes and deliberate argument engagement, not passive scroll behavior.\n\n### Principle 2: Prioritize Coverage Over Relevance\n\nA user engaging with topic T should be guaranteed exposure to the strongest argument for each major position on T, even positions they disagree with. This is the Deliberative Polling briefing model translated to feed design. Relevance (show me what I'm interested in) is subordinate to coverage (ensure I've seen the full argument landscape). Fishkin's evidence: coverage drives belief updating; relevance-only feeds create the epistemic equivalent of reading only your own party's briefings.\n\n### Principle 3: Route Controversy to Contributors, Not Passive Readers\n\nHigh-controversy arguments (near-even vote split) should be routed to users who have demonstrated ability to make epistemic contributions (evidence-linking, well-structured premises), not surfaced to maximize raw audience size. Controversy is a signal of unresolved epistemic tension, not of entertainment value. This inverts the HN controversy penalty: HN penalizes controversy to reduce noise; Deliberus should *selectively* surface controversy to users who can *reduce* it.\n\n### Principle 4: Evidence Gaps Are Contribution Invitations\n\nUnder-evidenced claims that are structurally important (frequently cited, load-bearing in major argument chains) should surface to users with relevant domain expertise as explicit contribution invitations — \"this argument needs evidence.\" The feed should function partly as a research task queue, not just as a reading list.\n\n### Principle 5: Make Rankings Auditable\n\nEvery argument in the feed should show users why it was surfaced: \"Shown because you endorsed a related argument,\" \"Shown because this is a widely-endorsed counter-argument to a position you hold,\" \"Shown because this claim lacks evidence and you have expertise in this domain.\" Mastodon's transparency insight applied to algorithmic ranking. Users should be able to inspect the ranking function, not just its outputs.\n\n### Principle 6: Track Cumulative Exposure, Not Just Current Item Diversity\n\nThe allostatic regulation model: the feed tracks what a user has been exposed to over time and uses this to correct drift toward homogeneity. A user who has engaged exclusively with pro-X arguments gets counter-argument exposure actively increased, not as punishment but as correction. This requires per-user exposure history tracking — architecturally more complex than per-item ranking, but essential for epistemic balance at the individual level.\n\n### Principle 7: The Organic Redraw Is a Dashboard, Not a Timeline\n\nThe feed should present the current state of a debate as a dashboard — showing the strongest representatives of each position cluster, the most-contested arguments, the most evidence-hungry claims — not as a chronological stream. Content doesn't \"fall off\" the feed because it's old; it falls off because it's been resolved, superseded, or is no longer the most epistemically important instantiation of its position. This is the 60% quality threshold funnel from the sketch: continuous recomputation of what the feed should contain, not accumulation.\n\n### Principle 8: Prioritize \"Changed My Mind\" Evidence as the Strongest Ranking Signal\n\nIf the architecture tracks pre/post belief states (via vote changes after argument exposure), arguments that demonstrably caused belief updates in users who initially disagreed with them should receive the highest epistemic quality ranking. This is computationally the most demanding principle but the most principled: it creates an empirical feedback loop where the feed algorithm becomes calibrated to actual argument effectiveness, not proxy signals.\n\n---\n\n## Summary Table: Ranking Signal Taxonomy\n\n| Signal | Epistemic Direction | Priority | Notes |\n|--------|-------------------|----------|-------|\n| Caused measured belief update in prior disbelievers | Strongly positive | Highest | Requires pre/post vote tracking |\n| Endorsed by users across opinion clusters (bridging) | Strongly positive | Very high | Polis-style computation |\n| Evidence links to high-quality sources | Positive | High | Needs source quality API |\n| Acknowledges strongest counter-arguments | Positive | High | Argument completeness check |\n| High controversy + low quality coverage | Positive (route to contributors) | High (for contributor routing) | Not for general surfacing |\n| User hasn't been exposed to this position | Positive (coverage) | Medium-high | Exposure history tracking |\n| Under-evidenced + structurally important | Positive (for domain experts) | Medium (targeted) | Contribution invitation |\n| Recency | Weakly positive | Low | Time decay, not primary |\n| Raw vote count | Neutral | Low | Popularity ≠ quality |\n| High engagement (clicks, time-on-page) | Negative (weak) | Overridden by epistemic signals | Engagement proxy |\n| High controversy + large audience surfacing | Negative | Suppress | Route to contributors instead |\n| Pure positive vote count from already-agreeing users | Weakly negative (echo signal) | Downweight | Bridging matters more than raw agreement |\n\n---\n\n## Open Questions for Deliberus Design\n\n1. **Who computes \"argument quality\"?** The signal taxonomy above presupposes that argument quality can be assessed. Can it be computed from structural properties (evidence links, premise completeness), or does it require human editorial judgment, or LLM evaluation? What is the appropriate mix?\n\n2. **How to handle the cold start problem?** New arguments have no vote history, no belief-update evidence, no exposure data. How does the feed surface genuinely new arguments that might be high-quality? (The HN approach: give new items a temporary boost; social-protocols proposes bias-corrected early-voting estimates.)\n\n3. **What is the right update frequency for the organic redraw?** The sketch implies continuous recomputation. What determines the batching interval? Topic velocity? User preference? Fixed schedule?\n\n4. **Can belief-update tracking be made low-friction enough to use?** The pre/post measurement model requires users to express beliefs before and after argument exposure. Is this feasible without feeling like a survey? Can voting behavior serve as an implicit proxy?\n\n5. **How does the fact/value boundary affect feed design?** Factual claims and normative claims have different epistemic quality criteria. Does the feed algorithm differ by claim type? (See [conceptual-threads.md §Thread 3](../conceptual-threads.md))\n\n6. **What is the relationship between the argument feed and the graph visualization?** The feed is a linearization of the argument graph. Should users move fluidly between the linear feed view and the graph view of the same content? How does the ranking algorithm translate across representations?\n\n---\n\n## References\n\n- [Bluesky algorithmic choice blog post (March 2023)](https://bsky.social/about/blog/3-30-2023-algorithmic-choice)\n- [Bluesky custom feeds / algorithmic marketplace (July 2023)](https://bsky.social/about/blog/7-27-2023-custom-feeds)\n- [Mastodon: Decentralized social media](https://joinmastodon.org/)\n- [Understanding Decentralized Social Feed Curation on Mastodon (arxiv, 2025)](https://arxiv.org/abs/2504.18817)\n- [How Hacker News ranking really works (righto.com, 2013)](http://www.righto.com/2013/11/how-hacker-news-ranking-really-works.html)\n- [social-protocols/quality-news: Towards a fairer ranking formula for Hacker News (GitHub)](https://github.com/social-protocols/quality-news)\n- [Wikipedia Help:Watchlist](https://en.wikipedia.org/wiki/Help:Watchlist)\n- [Wikimedia Foundation Annual Plan 2025-2026 OKRs](https://meta.wikimedia.org/wiki/Wikimedia_Foundation_Annual_Plan/2025-2026/Product_&_Technology_OKRs)\n- [Center for Humane Technology](https://www.humanetech.com/)\n- [Tristan Harris (Wikipedia)](https://en.wikipedia.org/wiki/Tristan_Harris)\n- [Knight-Georgetown Institute: Better Feeds: Algorithms That Put People First (March 2025)](https://kgi.georgetown.edu/research-and-commentary/better-feeds/)\n- [KGI Better Feeds full PDF](https://kgi.georgetown.edu/wp-content/uploads/2025/02/Better-Feeds_-Algorithms-That-Put-People-First.pdf)\n- [Epistemic welfare and algorithmic recommender systems, Communication Theory (2025)](https://academic.oup.com/ct/advance-article/doi/10.1093/ct/qtaf018/8240891)\n- [What is Deliberative Polling®? (Stanford Deliberative Democracy Lab)](https://deliberation.stanford.edu/what-deliberative-pollingr)\n- [America in One Room: Democratic Reform (FSI/Stanford, 2024)](https://cddrl.fsi.stanford.edu/news/america-in-one-room-democratic-reform-q-and-a)\n- [Topic-Level Bayesian Surprise and Serendipity for Recommender Systems (ACM RecSys 2023)](https://dl.acm.org/doi/10.1145/3604915.3608851)\n- [arxiv: Topic-Level Bayesian Surprise](https://arxiv.org/abs/2308.06368)\n- [Can Language Models Recognize Convincing Arguments? (arxiv, 2024)](https://arxiv.org/html/2404.00750v1)\n- [Blair et al., The Structure of Bridging (2025, Toronto/Harvard)](https://www.cs.toronto.edu/~nisarg/papers/bridging.pdf)\n- [Helberger et al.: Exposure diversity as a design principle for recommender systems (ResearchGate)](https://www.researchgate.net/publication/311957085_Exposure_diversity_as_a_design_principle_for_recommender_systems)\n- [Reducing echo chamber effects: allostatic regulator for recommendation algorithms (Tandfonline, 2025)](https://www.tandfonline.com/doi/full/10.1080/29974100.2025.2517191)\n- [Short-term exposure to filter-bubble recommendation systems has limited polarization effects (Harvard Gov, 2025)](https://www.gov.harvard.edu/2025/05/08/short-term-exposure-to-filter-bubble-recommendation-systems-has-limited-polarization-effects/)\n- [When automated fact-checking meets argumentation (HAL, 2025)](https://hal.science/hal-05017906/file/wang-et-al-2025-when-automated-fact-checking-meets-argumentation-unveiling-fake-news-through-argumentative-evidence.pdf)\n- [Lasser et al.: Designing social media content recommendation algorithms for societal good (Annals NYAS, 2025)](https://nyaspubs.onlinelibrary.wiley.com/doi/full/10.1111/nyas.15359)\n- [Value Alignment of Social Media Ranking Algorithms (arxiv, 2025)](https://arxiv.org/html/2509.14434v1)\n- [Thesephist: Epistemic calibration and searching the space of truth](https://thesephist.com/posts/epistemic-calibration/)\n\n---\n\n*See also: [polis-deep-dive.md](polis-deep-dive.md), [adoption-problem.md](adoption-problem.md), [habermas-machine.md](habermas-machine.md), [prediction-markets-argumentation.md](prediction-markets-argumentation.md), [conceptual-threads.md §Thread 3 (Fact/Value)](../conceptual-threads.md)*\n"}