{"path":"research/cross-thread-synthesis.md","content":"# Cross-Thread Synthesis and Gap Analysis\n\n*Critical friend document. Compiled March 28, 2026.*\n\n*This document exists to surface real problems, not to be encouraging. The 8 conceptual threads in `conceptual-threads.md` represent genuine intellectual work. They also have blind spots, emergent tensions when threads interact, and scaling assumptions that have not been stress-tested. What follows is the honest audit.*\n\n---\n\n## 1. Unexplored Thread Intersections\n\n### 1.1 Thread 1 (Adoption) × Thread 7 (Gaming): The Gamification Trap\n\nThe current design proposes gamification to solve the adoption problem — calibration scores, steelman rewards, belief-change tracking, argument quality ratings. Thread 1 notes that the entry point must feel effortless; Thread 7 proposes a reputation architecture with four independent signals. These threads have not been read together critically.\n\n**The emergent tension**: Gamification solves participation by creating extrinsic motivation, but extrinsic motivation systematically corrupts the behavior being rewarded. This is not a hypothesis — it is among the most replicated findings in behavioral psychology (Deci & Ryan's Self-Determination Theory, Overjustification Effect). Stack Overflow's experience is instructive: a reputation system designed to reward quality answers evolved into a system rewarding fast, confident answers to common questions. Hard questions by inexperienced users — exactly the questions most in need of good answers — get ignored because they earn fewer points. The metric captured a proxy for quality, not quality itself.\n\nFor Deliberus, the specific failure mode would look like this: users optimizing for calibration scores learn to make predictions about easy-to-verify claims, avoiding genuinely uncertain contested territory where calibration is hard and scores risky. Users optimizing for \"argument quality ratings\" from opponents learn to write arguments that *sound* well-reasoned to an opponent without actually challenging their premises — the rhetorical performance of good faith without the epistemic substance. Users optimizing for \"intellectual honesty record\" learn to publicly change their minds on low-stakes questions, preserving their ability to defend high-stakes positions without penalty.\n\n**The progressive disclosure version of this problem is worse**: Thread 1's resolution relies on a casual layer (vote, read summaries, trust the scaffolding) sitting above an expert layer. But gamification mechanics almost universally pull users *toward* higher engagement — more points, more badges, more status. If the gamification works, it succeeds by pulling casual users into the structured layer. If it succeeds too well with the wrong users (those motivated by status rather than epistemic virtue), the structured layer fills with reputation-optimizers who pollute exactly the high-rigor content that makes the platform valuable. You can only have a high-quality expert layer if you can somehow gate it against people who want expert-layer status without expert-layer epistemic standards.\n\n**Does gamifying the casual layer degrade the expert layer?** Probably yes, unless there is a credible gate between them. The current design does not have one. The four independent reputation signals (Thread 7) are an attempt at this gate — making it hard to game all four simultaneously. But \"hard to game all four simultaneously\" is not the same as \"impossible to perform well on all four without genuine epistemic virtue.\" A sufficiently motivated reputation-optimizer with enough time will solve this optimization problem.\n\n**What this means for the design**: The gamification and progressive disclosure layers need to be designed with specific failure modes anticipated, not just success modes. The safest gamification mechanics for epistemic quality are ones that reward *outcomes that can only be achieved through genuine engagement* — e.g., having an opponent explicitly say \"I can't refute this\" — rather than outcomes achievable through strategic behavior.\n\n---\n\n### 1.2 Thread 2 (Hybrid Intelligence) × Thread 8 (Consensus): Does the System Develop Ideology?\n\nThread 2 argues that Deliberus + its participants = a distributed reasoning entity. Thread 8 argues the system should seek truth, not consensus. These threads have a collision that neither addresses.\n\n**The emergent question**: If the system is a reasoning entity, does it develop its own cognitive biases? And does it have a mechanism to detect and correct them?\n\nThis is not metaphorical. Every system that aggregates human opinion and uses that aggregation to surface more content creates feedback loops between the content people see and the opinions they form. Algorithmic content curation — even truth-seeking curation — shapes the opinion distribution of its user base. The system's output influences the inputs that create future outputs. This is a closed loop. Closed loops in complex systems tend toward fixed points or oscillation, not toward objective truth.\n\nThe specific mechanisms through which Deliberus could develop ideological drift:\n\n1. **Early adopter bias**: The rationalist/EA/AI safety community is the proposed beachhead. This community has strong priors on epistemology, political economy, and what constitutes a \"well-reasoned\" argument. If the initial argument graph is seeded by this community's norms, those norms become the baseline against which all future contributions are evaluated. The \"quality\" judgments that populate the reputation system are judgments made by early adopters, who are not representative of humanity.\n\n2. **The argument quality feedback loop**: High-rated arguments attract more engagement, which generates more evidence links, which raises their quality scores further. Arguments that begin in the graph with low quality scores (because they are poorly expressed, from low-reputation contributors, or challenge cherished premises of the early community) may never accumulate the engagement needed to improve their quality scores, even if they are genuinely correct.\n\n3. **LLM-as-extraction-layer bias**: Thread 1 proposes that LLMs absorb the structuring burden. But LLMs have documented systematic political biases — multiple studies in 2024-2025 find that major models (GPT-4, Claude, Gemini) consistently score left-of-center on political values. One EMNLP 2024 study found that in debate simulations, attitudes converged toward a left-leaning stance across models and assigned identities regardless of the assigned political identity. If LLMs are doing claim extraction, argument structure identification, and quality scoring, they are not neutral — they are applying a consistent political lens to which arguments surface as \"well-structured\" and which claims get extracted as \"key points.\"\n\n**The memetic immune system (Thread 8) doesn't help here**: That research documents how ideological immune systems *protect* beliefs from challenge. Deliberus's problem is the inverse — the system may *inadvertently build* an ideological immune system against certain types of arguments without any actor intending it. The immune system emerges from the aggregate of design choices, early adopter composition, and LLM biases, not from intentional censorship.\n\n**What this means for the design**: The hybrid intelligence metaphor is compelling but it demands the next question: what is this entity's self-correction mechanism? The scientific method's answer is reproducibility and peer review — the community can check each other's work. Deliberus needs an equivalent: some mechanism to detect systematic bias in the argument graph, not just individual argument quality. This is a genuinely hard unsolved design problem.\n\n---\n\n### 1.3 Thread 3 (Fact/Value) × Thread 5 (Probabilistic): Can You Assign Probability to \"We Should Do X\"?\n\nThis intersection deserves more careful analysis than either thread provides. Thread 3 establishes that normative claims get a different scoring axis (agreement + importance, not true/false). Thread 5 proposes QBAFs with numeric strength values as the formal backbone. The question: can a normative claim have a probability score in a QBAF?\n\n**What the philosophical literature actually says**: The moral uncertainty literature (Ord, MacAskill, Bykvist; formalized in MacAskill's *Moral Uncertainty*, 2020) is explicit that you can apply probabilistic reasoning to normative claims, but with fundamental complications. The machinery works for *interpersonal comparisons of moral weight* (how confident am I that a utilitarian calculus is correct vs. a deontological one?) but breaks down for *cardinal comparisons across moral theories* (what does 70% probability that consequentialism is correct, combined with 30% probability that deontology is correct, actually imply you should do?). This is the \"intertheoretic value comparison\" problem — it is genuinely unsolved in moral philosophy.\n\n**What \"70% likely that we should ban X\" actually means**: It could mean several distinct things, none of which are equivalent:\n- *70% of participants (in this deliberation, at this time) support banning X* — a sociological fact, not a moral fact\n- *70% probability that banning X is the correct normative stance given contested empirical facts* — probability attached to empirical uncertainty, not normative uncertainty\n- *70% confidence that the value premises supporting a ban are correct* — probability attached to moral uncertainty in the MacAskill sense\n- *70% probability that, conditional on a specific ethical framework, banning X is the right call* — probability within a framework\n\nThese are different claims, and conflating them is exactly the kind of conceptual confusion that makes public discourse on normative issues intractable. Thread 3 identifies the fact/value boundary as a \"first-class classifier.\" The next step — which the research does not take — is to distinguish *types* of normative uncertainty and decide which of the above senses QBAF numeric strengths can legitimately represent.\n\n**The practical implication**: Putting a single probability number on a normative claim without specifying which of these senses it represents is not rigorous — it is a false precision that could make the system's output more misleading than free-form discussion. The display layer (Thread 6) needs to represent this ambiguity visually, not collapse it into a confidence interval on a bar.\n\n**The bridging opportunity**: The most defensible use of probability in normative argument is not \"probability that this ought-claim is correct\" but \"probability that this factual premise, on which this normative argument depends, is true.\" This separates the empirical uncertainty (quantifiable) from the normative uncertainty (philosophically contested). Users who agree on the facts can still disagree on the values — and making that distinction visible is more epistemically honest than a single probability on the normative claim.\n\n---\n\n### 1.4 Thread 4 (Dedup) × Thread 6 (Visualization): Showing Near-Duplicates Without Creating Confusion\n\nThe dedup research establishes a three-layer dedup pipeline. The visualization research proposes semantic zoom. These interact in a way neither thread addresses.\n\n**The visual representation problem**: When dedup identifies near-duplicate claims and merges them into a cluster, how does the graph show this? Options:\n\n- **Single node with \"N variants\" indicator**: Clean, but loses information. The variant wording often carries the substantive disagreement — \"The economy will collapse\" vs. \"The economy will face significant headwinds\" are near-duplicates in embedding space but make very different claims.\n- **Collapsed sub-graph**: The cluster is shown as a single node; expanding reveals the variants. This compounds the complexity that semantic zoom is already managing — now there are two layers of expansion (zoom level AND dedup cluster expansion).\n- **Fuzzy node**: A node whose visual representation encodes its certainty about dedup status — clear center for high-confidence match, blurry edges for uncertain near-duplicate. This is novel and interesting but requires design work that does not exist yet.\n\n**The false equivalence visualization problem**: Thread 4 notes the Emanuel insight — dedup that merges claims with identical surface forms but opposite operative meanings erases the most interesting disagreement. If the visualization shows a merged node labeled \"passenger safety must be guaranteed,\" and two groups endorsed it for opposite reasons, the graph gives users no signal that this node is a Trojan horse hiding a genuine unresolved conflict. The node looks resolved when it is not.\n\n**The practical implication**: Dedup and visualization must be co-designed, not designed independently and integrated afterward. The visualization needs a visual vocabulary for \"this node's apparent simplicity may be an artifact of deduplication — expand to see if the underlying variants are actually equivalent.\" Without this vocabulary, the visual clarity that Thread 6 promises as a cognitive prosthesis becomes cognitive deception.\n\n---\n\n### 1.5 Thread 7 (Gaming) × Thread 8 (Consensus): Can Reputation Manufacture False Consensus?\n\nThis is the adversarial design problem at its sharpest, and the current research does not take it seriously enough.\n\nThe proposal is that high reputation users' contributions carry more weight. High reputation requires demonstrated calibration, argument quality ratings from opponents, intellectual honesty tracking, and evidence contribution. This is more sophisticated than Stack Overflow's single karma number. But it still produces a single effective weight on each user's contributions.\n\n**Sybil attack dynamics**: The reputation system is vulnerable to coordinated Sybil attacks — creating multiple accounts that mutually reinforce each other's reputation through cross-adversarial ratings. \"I rate your argument well-reasoned; you rate mine.\" If 50 accounts do this for 3 months, they can each accumulate significant opponent-quality ratings without any of them ever actually engaging with genuine opponents. The cost of this attack is time and coordination, not any special epistemic virtue.\n\n**The consensus manufacturing failure mode**: A well-organized advocacy group with 50 high-reputation accounts — legitimately earned through years of genuine engagement — can systematically up-weight arguments supporting their position and down-weight opposing ones, not through any individual action visible as gaming, but through the aggregate effect of 50 people making reasonable-looking individual decisions. This is not a hypothetical: it is exactly what Wikipedia's ArbCom documented in 2025 with coordinated off-platform editing teams for Middle East articles.\n\n**The Polis anonymization insight (and its limits)**: Thread 7 notes that Polis uses anonymous voting to reduce social gaming pressure. But Polis's anonymity works because Polis is a one-shot deliberation tool — participants vote, clusters form, bridging statements surface, the deliberation ends. Deliberus is a persistent graph. Persistent graphs require persistent identities to build reputation systems. Reputation systems require persistent identities. Anonymity and reputation are structurally incompatible in a persistent system.\n\n**What this means for the design**: The two-axis voting system (agree vs. well-argued) is the right direction, but it needs adversarial modeling built into its architecture from day one, not added after the reputation system shows signs of gaming. Rate-limiting, anomaly detection on voting patterns, and clear governance procedures for reputation attacks need to be designed before the reputation system is built, not retrofitted.\n\n---\n\n## 2. The Scalability Paradox\n\nThe vision assumes that Deliberus scales to thousands of users on a single topic. Does it actually?\n\n**The graph growth problem is real**: Kialo's statistics are instructive. At 1M+ registered users and 720,000+ claims across 18,000+ debates, the average debate has 40 claims. But the high-engagement debates on contentious topics have far more — and they exhibit exactly the pathology Thread 4 describes: the same argument restated hundreds of times in different branches.\n\n**The Deliberatorium evidence**: Klein's Deliberatorium research (the ResearchGate paper on scaling e-deliberation) found a direct trade-off: \"the more structured the discussion becomes, the more argumentation occurs, but the less ideas are generated.\" This is not a UX problem — it is a cognitive property of structured deliberation. Structure focuses reasoning but constrains generativity. At 150+ participants on a single structured discussion, the system was already showing signs of strain.\n\n**Wikipedia's scaling lesson**: Wikipedia's peak in English-language active editors was approximately 2007, after which editor numbers declined even as content continued to grow. The cause is well-documented: as the encyclopedia approached completeness, coordination work grew faster than the editor base. Each additional editor required more governance overhead than they contributed in content. Wikipedia's ArbCom, AN/I boards, and dispute resolution processes are the governance apparatus that emerged to manage this — and they require significant editor attention that comes at the cost of article creation and improvement. Critically, this is with *asynchronous, non-real-time* collaboration on *factual* claims where right answers often exist.\n\nDeliberus adds two dimensions of complexity: real-time argumentation dynamics, and normative claims where \"right answers\" are contested by design. The governance overhead at scale will be larger than Wikipedia's, not smaller.\n\n**The semantic zoom doesn't solve this**: Semantic zoom (galaxy → constellation → star → planet) solves the *visualization* problem at scale. It does not solve the *maintenance* problem. Who decides which claims get promoted to the constellation level? How are star-level nodes that have been superseded by new evidence deprecated without creating orphaned sub-graphs? How are branching debates that have diverged irrecoverably merged or separated? These are governance questions, not visualization questions.\n\n**The honest assessment**: A single-topic argument graph with 10K active participants and 100K claims is probably unmanageable in any current architectural proposal. The path forward is not \"semantic zoom will handle it\" but rather: at what scale does a single argument graph become dysfunctional, and how should the system partition large debates into manageable sub-debates with cross-references? This requires intentional design, not deferred scalability.\n\n---\n\n## 3. The AI Dependency Risk\n\nThread 1's resolution — \"LLMs absorb the structuring burden\" — is the load-bearing premise of the entire adoption thesis. It deserves stress-testing.\n\n**What the LLM cost trajectory actually shows**: LLM inference costs have declined approximately 10x annually since 2022 — faster than PC compute or internet bandwidth. GPT-4 class performance now costs $0.40/million tokens versus $20 in late 2022. This trajectory is favorable. The concern about API costs \"remaining high\" is probably the weakest of the AI dependency risks.\n\n**What the LLM quality plateau scenario looks like**: The framing of Thread 1 — \"this time is different because LLMs change the friction equation\" — is implicitly a bet on continued LLM improvement. But the specific tasks Deliberus needs LLMs to perform (claim extraction, argument structure identification, deduplication verification, counter-argument generation, quality scoring) are tasks where current LLMs have known systematic failures:\n\n- Claim extraction produces false positives (extracting non-claims as claims) and false negatives (missing implicit claims that experienced reasoners would identify)\n- Argument structure identification struggles with multi-step inferences spread across a long text\n- Deduplication verification produces false equivalences (merging distinct claims with similar surface forms) and false distinctions (splitting equivalent claims expressed differently)\n- Counter-argument generation produces arguments that *look* like steelmen but address a different claim than the original\n- Quality scoring inherits the political biases documented above\n\nNone of these failure modes disappear at higher model capabilities — they shift. More capable models make more sophisticated errors. A GPT-5-class model might correctly identify that two claims are equivalent in surface form while missing that they carry different operative meanings in the Emanuel sense. The errors become harder to detect, not easier.\n\n**The authenticity rejection risk is underrated**: The adoption research documents that unguided LLM use reduces reasoning quality — users engage less critically when they feel AI has \"handled\" the thinking. For Deliberus, the specific risk is: users who see a \"Your argument has been structured by AI\" notice may disengage from the argument altogether (\"the AI got it right\") or reject the structuring as inauthentic (\"that's not what I meant\"). Both responses undermine the model. The Habermas Machine research found that AI-generated consensus statements were preferred over human mediators in an academic study setting — but academic study settings are not representative of contested political discourse where participants may be deeply skeptical of AI mediation.\n\n**The systematic bias risk is the most serious**: Research from 2024-2025 documents that major LLMs consistently exhibit left-leaning political biases across multiple metrics and methodologies. In debate simulations, LLM-assigned identities converge toward left-leaning stances regardless of their assigned political identity. For a truth-seeking platform that explicitly positions itself against political bias, using politically biased LLMs as the claim extraction and quality scoring layer is a structural integrity problem. It means the platform's claim to neutrality is undermined at the architectural level by its most fundamental component.\n\n---\n\n## 4. The Quality Bootstrap Problem\n\nHigh-quality argument maps attract users. Users produce argument maps. Early users produce low-quality maps. This is the classic content quality cold-start problem, and the research does not fully solve it.\n\n**The Wikipedia stub model works for facts, not arguments**: Wikipedia's stub model succeeds because stubs have a clear definition of \"improvement\": add verifiable facts, add sources, improve prose. The improvement path is determinate. For argument maps, \"improvement\" is contested: Does adding more premises improve the map, or does it obscure the core claim? Is a more nuanced argument better, or is it too complex for casual users to evaluate? The improvement path for argument maps is normatively contested in a way that factual stubs are not.\n\n**The LLM content seeding proposal has a specific failure mode**: Pre-populating 100-200 canonical topics with LLM-generated argument maps creates a library of LLM-generated content that early users are invited to improve. But the improvement process requires users to engage critically with the initial maps — identifying wrong structures, missing arguments, false equivalences. This requires users to have more argumentative sophistication than the LLM, or at least enough to recognize when the LLM is wrong. The early adopter community (rationalists, EA, AI safety) probably meets this bar. The general public does not.\n\n**The Quora quality collapse lesson**: Quora used invite-only exclusivity to seed high-quality content from known experts. It worked — briefly. When exclusivity was relaxed, quality declined faster than the additional participants contributed value. The relaxation of quality controls is structurally almost inevitable: if the platform succeeds, it will want to grow; if it grows, it must lower barriers; lower barriers bring lower-average-quality participants; lower-average-quality participants degrade the content that attracted the initial community.\n\n**The honest assessment**: There is no solved bootstrap solution that works for argument quality specifically. The closest analog might be LessWrong's moderation model — a karma-gated community that has maintained quality norms through aggressive curation by a small core community over many years. But LessWrong has never scaled to the general public, and it is not clear that its quality model transfers to a platform with broader ambitions. The design needs to make explicit choices about whether quality or scale is the primary objective in the first two years, because optimizing for both simultaneously may not be possible.\n\n---\n\n## 5. Cultural and Linguistic Challenges\n\nThe research is entirely English-language, Western-centric, and rooted in a Low German-Protestant epistemological tradition that values explicit disagreement, formal logic, and argumentative transparency. This is not the universal epistemological tradition.\n\n**The high-context culture problem is structural, not cosmetic**: High-context cultures — most of East Asia, much of the Middle East, significant parts of Latin America and sub-Saharan Africa — treat explicit disagreement as a face threat. Research on deliberation across individualist-collectivist cultural divides finds that deliberation among American individualists featured more arguments than among Korean collectivists and was rated as more \"deliberative\" by Western measures. The structured argumentation model that Deliberus proposes — explicit attack and support relations, named premises, public claim attribution — is a Low-Context communication format applied to a global problem.\n\n**This is not just a UX localization problem**: Adding Japanese-language support does not solve the structural incompatibility between explicit argument structure and high-context communication norms. A Japanese user who believes that publicly attributing a claim to a named person and attaching explicit counterarguments to it is a face violation will not use the system regardless of its language. The system's core UX primitive — \"your claim, my counterargument, publicly visible\" — is culturally specific.\n\n**What research on cross-cultural deliberation actually recommends**: Studies in cross-cultural deliberation find that effective leaders in interdependent, obligation-oriented (high-context) contexts do not eliminate dissent — they *structure it differently*. Specifically: dissent is assigned as a role-based expectation rather than a personal initiative. An \"assigned devil's advocate\" structure, where the platform explicitly invites certain users to challenge claims as a role rather than as a personal position, might mitigate the face-threat in high-context cultures. This is a substantively different UX pattern from the current design.\n\n**Oral tradition cultures**: Even beyond high-context/low-context distinctions, structured written argumentation is itself a culturally specific practice. In cultures with strong oral deliberation traditions — many indigenous cultures, much of rural sub-Saharan Africa — the assumption that knowledge claims should be written, attributed, structured, and publicly debatable is a foreign epistemological import. Voice-to-argument contribution (Thread 1's mobile unlock) partially addresses the input barrier, but not the cultural norms around knowledge authority and debate format.\n\n**Rhetorical structure divergence**: Aristotle's rhetorical tradition (logos, ethos, pathos; structured syllogisms) underlies the Western argumentative tradition that all existing argument mapping tools are built on. Chinese rhetorical tradition (qi-cheng-zhuan-he structure: introduce, develop, turn, conclude) produces arguments that look structurally incoherent to a DAG-based argument mapper but are internally coherent by their own conventions. Arabic rhetorical tradition emphasizes repetition and elaboration in ways that Western argument mapping treats as redundancy to be deduplicated. The dedup layer (Thread 4) would systematically flatten cross-cultural rhetorical variation.\n\n**The honest assessment**: If Deliberus is genuinely a global epistemic infrastructure project — \"a wiki where every word has been vetted\" — it cannot be designed solely for Western, low-context, Aristotelian-rhetorical epistemology. But genuinely universal design would require making the argumentation ontology itself culturally negotiable, which conflicts with the formalization and rigor that the platform's epistemic ambitions require. This is a real tension with no easy resolution. The minimum honest step is acknowledging it explicitly in the design philosophy, not pretending it is a localization problem to be solved later.\n\n---\n\n## Synthesis: The Five Most Critical Risks\n\nRanked by combination of likelihood and consequence:\n\n**1. LLM systematic bias becomes the platform's ideological fingerprint.** If claim extraction, quality scoring, and argument structuring are done by politically biased LLMs, the platform's epistemic neutrality is compromised at the infrastructure level. This is not a future risk — it is a present one. Mitigation: use multiple LLMs from different providers for structuring tasks, implement explicit bias monitoring on argument graph distributions, design for human override of LLM structuring decisions.\n\n**2. Reputation system gaming precedes the community it was designed to protect.** Sophisticated adversarial actors (political campaigns, corporations, coordinated advocacy groups) will invest in reputation optimization before good-faith users accumulate the reputation to outweigh them. Wikipedia's 2025 experience with coordinated off-platform editing teams is the preview. Mitigation: design adversarial models before the reputation system, not after.\n\n**3. Scalability assumptions fail at the governance layer, not the technical layer.** The visualization and database architecture can handle 100K nodes. The community governance of a 100K-node argument graph probably cannot — Wikipedia's experience shows governance overhead grows faster than content at scale. Mitigation: design argument graph partitioning and federated governance models explicitly, not as afterthoughts.\n\n**4. The quality bootstrap produces a high-quality ghost town.** If quality controls are maintained rigorously during the early phase, the platform may not achieve the network effects needed for self-sustaining growth before funding runs out. If quality controls are relaxed to achieve growth, the early community's quality norms are diluted irreversibly. Mitigation: make the quality-vs-scale tradeoff explicit and time-bounded — commit to a specific early-phase quality bar, a specific growth trigger, and a specific quality maintenance mechanism for the post-growth phase.\n\n**5. Cultural non-portability limits the global epistemic ambition.** The platform is designed for one epistemological tradition. The vision is global. Mitigation: either constrain the vision honestly (this is epistemic infrastructure for the English-speaking rationalist tradition) or invest in cross-cultural design research before building, not after.\n\n---\n\n*This document is a critical complement to [conceptual-threads.md](../conceptual-threads.md), not a replacement for it. The threads document the genuine intellectual progress. This document maps where the thinking stops and the difficult work begins.*\n\n*See also: [adoption-problem.md](adoption-problem.md) | [bridging-arguments.md](bridging-arguments.md) | [single-player-utility.md](single-player-utility.md) | [epistemic-gamification.md](epistemic-gamification.md) | [memetic-immune-system.md](memetic-immune-system.md)*\n"}