{"path":"research/ai-safety-and-tao-augmentation-research.md","content":"# AI Safety, Deliberative Alignment, and Tao's Verification Thesis\n\n**Date**: March 31, 2026\n**Purpose**: How Deliberus connects to AI safety/alignment research, the 80,000 Hours gap, sensemaking infrastructure, and Terence Tao's recent work on AI-augmented reasoning.\n\n---\n\n## 1. AI Alignment Through Structured Human Values\n\n### 1.1 The Convergence Thesis and Value Alignment\n\nDeliberus's \"convergence thesis\" — that decomposing any human value system far enough reveals shared bedrock (consciousness, suffering, shared humanity) — has direct analogues in current alignment research.\n\n**\"What Are Human Values, and How Do We Align AI to Them?\"** (Benkler et al., arXiv:2404.10636, 2024) decomposes the alignment problem into three parts: (1) eliciting values from people, (2) reconciling those values into an alignment target, and (3) training the model. The paper proposes **Moral Graph Elicitation (MGE)**, where an LLM interviews participants about their values in particular contexts, then maps directional edges between values — \"in this context, value A is wiser than value B.\" Crucially, participants overwhelmingly converge on the directionality of these transitions. This is structurally identical to Deliberus's vision: a graph where moral reasoning is decomposed, mapped, and navigable.\n\n> **Key quote**: \"Moral graph edges represent broad agreement amongst participants that one value is wiser than another for a particular context, and participants overwhelmingly converge on the directionality of these transitions.\"\n\n**ValueCompass** (Zhang et al., arXiv:2409.09586, 2024) grounds value alignment in psychological theory (Schwartz's value theory), identifying fundamental value dimensions and measuring human-LLM misalignment. Finding: humans endorse \"National Security\" values that LLMs largely reject, and values differ across scenarios — highlighting the need for context-aware alignment, not a single fixed constitution.\n\n**Superalignment with Dynamic Human Values** (arXiv:2503.13621, 2025) proposes algorithms inspired by Iterated Amplification that decompose complex value judgments into subtasks simple enough for an aligned human-level AI to evaluate. This mirrors Deliberus's pipeline: decompose claims into atomic assertions, classify them, then compose back into a navigable structure.\n\n### 1.2 Constitutional AI and Its Limits\n\nConstitutional AI (CAI) aligns models with explicitly stated normative principles via self-critique. But recent research reveals fundamental limitations:\n\n**\"Does Claude's Constitution Have a Culture?\"** (Pourdavood, arXiv:2603.28123, March 2026) finds that constitutions embed cultural assumptions — they are not universal. **\"How Well Do Models Follow Their Constitutions?\"** (MATS 9.0 / Neel Nanda, AI Alignment Forum, March 2026) finds compliance is inconsistent.\n\n**\"Moral Disagreement and the Limits of AI Value Alignment\"** (AI & Society, June 2025) argues that value alignment faces a dual challenge: epistemic justification (whose values?) and political legitimacy (by what authority?). Static constitutions cannot resolve genuine moral disagreement.\n\n> **Deliberus connection**: A constitution is a top-down list of principles. Deliberus is a bottom-up *process* for deriving, decomposing, and testing principles through structured argumentation. Constitutional AI is a snapshot; Deliberus is the camera. The platform could serve as the deliberative process that *generates* constitutions — a \"constitutional convention\" infrastructure for AI values.\n\n### 1.3 Deliberative Alignment\n\nThe term \"deliberative alignment\" has emerged as a research direction (AryaXAI, June 2025): \"aligning AI systems with human values — not through static assumptions or abstract moral theories, but through inclusive, structured, and democratic deliberation.\" It advocates embedding democratic deliberation directly into AI development and oversight, drawing from civic traditions like citizen assemblies.\n\n**\"Resolving Value Conflicts in Public AI Governance: A Procedural Justice Framework\"** (de Fine Licht, Chalmers University, 2025) proposes procedural justice as the mechanism for legitimating AI value choices when substantive agreement is impossible. This aligns with Deliberus's worldview-filter UX: the platform doesn't force consensus but makes the full landscape of argument visible and navigable.\n\n### 1.4 Debate-Based Alignment (Irving et al.)\n\n**Original proposal** (Irving, Christiano, Amodei, 2018): Two AI systems debate, with a human judge deciding the winner. The key insight: if honesty is incentivized at equilibrium, even a weak human judge can oversee a superhuman AI.\n\n**Recent evolution — Prover-Estimator Debate** (Brown-Cohen, Irving et al., AI Alignment Forum, June 2025): A new scalable oversight protocol that addresses the \"obfuscated arguments\" problem — where a dishonest debater adversarially chooses how to recurse, requiring an honest opponent to solve computationally intractable problems. The new protocol proves honesty is incentivized at equilibrium even when both AIs have similar compute.\n\n**\"An Alignment Safety Case Sketch Based on Debate\"** (UK AISI Alignment Team — Marie_DB, Pfau, Hilton, Irving, May 2025): Develops a concrete safety case for deploying debate-based alignment in practice, with formal conditions under which debate provides safety guarantees.\n\n> **Deliberus as judge infrastructure**: The debate alignment literature assumes a human judge who can evaluate arguments. But the quality of judgment scales with the quality of argument decomposition. A Deliberus-like system — where arguments are atomically decomposed, scheme-classified, and critically questioned — could serve as the structured \"courtroom\" that makes human judgment more reliable in AI debates. The platform provides exactly what the debate literature needs: transparent argument structure that helps weak judges make strong decisions.\n\n### 1.5 From Debate to Deliberation (DCI)\n\n**\"From Debate to Deliberation: Structured Collective Reasoning with Typed Epistemic Acts\"** (Prakash, arXiv:2603.11781, March 2026) — a paper published during the same week Deliberus was being built. It introduces Deliberative Collective Intelligence (DCI), which formalizes:\n\n- **4 reasoning archetypes** (analyst, advocate, critic, synthesizer)\n- **14 typed epistemic acts** (claim, evidence, objection, qualification, reframe, etc.)\n- **DCI-CF**: a convergent flow algorithm guaranteeing termination with a **decision packet** containing: selected option, residual objections, minority report, and reopen conditions\n\nKey finding: \"On non-routine tasks (n=40), DCI significantly improves over unstructured debate (+0.95, 95% CI [+0.41, +1.54]). DCI excels on hidden-profile tasks requiring perspective integration (9.56, highest of any system on any domain) while failing on routine decisions (5.39), confirming task-dependence.\"\n\n> **Deliberus connection**: DCI's typed epistemic acts map closely to Deliberus's Walton scheme classification + critical question generation. The \"decision packet\" (option + residual objections + minority report + reopen conditions) is exactly the kind of structured output Deliberus's sorry model produces. DCI validates that structured deliberation outperforms unstructured debate for complex decisions — the entire Deliberus thesis.\n\n### 1.6 ARGSBASE: Multi-Agent Structured Human-AI Deliberation\n\n**\"ARGSBASE: A Multi-Agent Interface for Structured Human-AI Deliberation\"** (Turkstra et al., EACL 2026 System Demonstrations, March 2026) presents a working system for structured human-AI deliberation using multi-agent architectures. Published at EACL 2026 (March 24-29, 2026).\n\n---\n\n## 2. The 80,000 Hours Gap\n\n### 2.1 The Problem Profile\n\n80,000 Hours published **\"Using AI to Enhance Societal Decision Making\"** (Zershaaneh Qureshi, September 2025, promoted via Substack and EA Forum in February-March 2026) as a formal problem profile. It identifies five categories of AI decision-making tools:\n\n1. **Fact-checking tools** — \"AI fact checkers may be more reliable and impartial evaluators of information than humans\"\n2. **Forecasting systems** — \"could help institutions make better predictions about world events\"\n3. **Negotiation tools** — \"could find mutually beneficial agreements... by rapidly simulating thousands of hours of negotiation\"\n4. **Coordination tools** — \"help groups work together and make better collective decisions\"\n5. **Moral progress tools** — \"more speculatively, could help us reason through complex ethical questions\"\n\nThe profile concludes: **\"Only a handful of projects are currently building the kinds of AI decision-making tools described — a drop in the ocean compared to the billions invested in developing more broadly capable AI agents.\"**\n\n### 2.2 The Gap Deliberus Fills\n\nThe 80,000 Hours profile does NOT mention:\n- **Structured argumentation platforms** (argument mapping, claim decomposition, scheme classification)\n- **Deliberation infrastructure** (the structured process of reasoning together, not just coordinating)\n- **Epistemic graph navigation** (worldview-as-filter, navigable argument landscapes)\n\nThis is a critical gap. The five categories above address *symptoms* of poor collective reasoning: wrong facts, bad predictions, failed negotiations, coordination failures, moral confusion. Deliberus addresses the *root cause*: the absence of infrastructure for structured reasoning itself.\n\n> **Framing for 80,000 Hours**: Deliberus is not a fact-checker, forecaster, negotiation tool, coordination tool, or moral progress tool. It is the **reasoning substrate** that makes all five more effective. A fact-check is only as good as the argument structure connecting evidence to claims. A forecast is only as good as the decomposition of assumptions. Moral progress requires making the structure of moral reasoning *visible* — which is exactly what argument decomposition does.\n\nThe profile's own logic supports this: \"most harmful decisions stem from failures in fact checking, prediction, negotiation, or other failures under the wide umbrella of 'epistemics and coordination.'\" Deliberus IS epistemics infrastructure.\n\n### 2.3 Neglectedness Assessment\n\nThe profile rates this area as \"highly neglected.\" Deliberus sits at the intersection of the most neglected sub-areas: structured argumentation (almost no funded projects) + AI-augmented reasoning (a few academic prototypes, no production systems at scale) + epistemic commons (the Consilience Project wound down, Polis is limited to opinion clustering without argument structure).\n\n---\n\n## 3. Schmachtenberger's Metacrisis and Sensemaking Infrastructure\n\n### 3.1 The Consilience Project — Current State\n\nThe Consilience Project (a publication of the **Civilization Research Institute**, co-founded by Schmachtenberger) has **slowed publication cadence significantly**. From their site: \"Although The Consilience Project is no longer releasing new articles at a regular cadence, it will continue to publish related content at varying intervals.\"\n\nThe project published several foundational articles on sensemaking, information warfare, and epistemic infrastructure, but has not achieved the scale or tooling originally envisioned. The CRI remains active as a 501(c)(3) research nonprofit focused on \"preventing civilizational risks and improving social coordination.\"\n\n### 3.2 The Sensemaking Infrastructure Vacuum\n\nSchmachtenberger's metacrisis framing identifies the **epistemic commons** as civilizational infrastructure: without shared mechanisms for establishing what is true and what matters, coordination fails at every level. The Consilience Project attempted to address this through high-quality analysis articles — but articles are a one-directional medium. They inform but don't enable structured collective reasoning.\n\nRecent metacrisis writing (Van Peborgh, \"The Thunder of the Metacrisis,\" March 2026) frames AI, ecological limits, energy decline, and institutional breakdown as converging into a single systemic turning point. But the proposed responses remain abstract — \"new coordination,\" \"better sensemaking\" — without concrete infrastructure.\n\n> **Deliberus as the missing tool**: Schmachtenberger diagnosed the disease (epistemic commons collapse) but the Consilience Project's prescription (better articles) didn't match the scale of the problem. Deliberus provides the *infrastructure* — a persistent, navigable, community-maintained argument graph where sensemaking is structural, not just narrative. It is the tooling that the metacrisis diagnosis demands: transparent decomposition of claims, visible assumption structures, argumentation scheme classification, and critical question generation. Not \"read better articles\" but \"participate in structured reasoning.\"\n\n### 3.3 The AI Safety Community's Sensemaking Gap\n\n**\"Field-Building and the Epistemic Culture of AI Safety\"** (First Monday, 2024) critiques the AI safety field's narrow focus and deference to corporate interests, arguing that safety \"becomes less a collective obligation and more a strategic discourse shaped by actors with the resources to steer public understanding.\"\n\nThis is precisely the problem structured deliberation addresses: when reasoning about AI safety happens in unstructured forums (Twitter threads, blog posts, EA Forum), the structure of arguments is invisible, and discourse is shaped by rhetorical skill rather than argument quality. Deliberus-style infrastructure would make the *structure* of AI safety reasoning transparent and navigable.\n\n---\n\n## 4. Collective Intelligence for AI Governance\n\n### 4.1 Stanford's AI for Deliberation and Consensus\n\nThe **Stanford Digital Economy Lab** (HAI) partnered with MIT GOV/LAB to create **DELiberationIO**, an online platform for AI-facilitated deliberation. Key features:\n- Conversational AI engages in Socratic dialogue with users\n- Deployed in Washington D.C. city proceedings (July 2025)\n- AI mediator generates consensus statements and next-round discussion questions\n- Results: \"the final petition created by both groups is overwhelmingly positive, suggesting that AI facilitation can help produce higher quality results in large group deliberations\"\n\n### 4.2 The Habermas Machine (Google DeepMind)\n\n**\"AI Can Help Humans Find Common Ground in Democratic Deliberation\"** (Tessler et al., Science, October 2024): An AI mediator iteratively generates and refines statements expressing common ground among groups on social/political issues.\n- N = 5,734 participants\n- Participants preferred AI-generated group statements to human mediator statements\n- Rated as more informative, clear, and unbiased\n- Replicated in a virtual citizens' assembly with a demographically representative UK sample\n\n### 4.3 Recent Academic Work (2025-2026)\n\n- **\"From Efficiency to Deliberation: Rethinking AI's Role in Institutionalizing Democratic Innovations\"** (Politics and Governance, January 2026) — argues for AI that supports deliberative processes rather than replacing them\n- **\"The Case for Using Generative AI to Run Deliberation Simulations\"** (Journal of Deliberative Democracy, February 2026) — AI-simulated deliberation as a tool for policy design\n- **\"Human/AI Collective Intelligence for Deliberative Democracy: A Human-Centred Design Approach\"** (arXiv:2603.16260, March 2026) — framework for human-centered AI-augmented deliberation\n- **\"Argumentative Human-AI Decision-Making: Toward AI Agents That Reason With Us, Not For Us\"** (arXiv:2603.15946, March 2026) — AI agents engaging in dialectical reasoning with humans through computational argumentation\n- **\"Democratic Governance through DAO-based Deliberation and Voting for Inclusive Decision Making in AI Models\"** (Nature Scientific Reports, March 2026) — DAO-based deliberation for AI governance\n- **\"AIMED: Towards a Philosophically Legitimated AI-assisted Iterative Method for Ethical Deliberation\"** (Rivelli, PhilSci-Archive, 2025) — formal philosophical grounding for AI-assisted ethical deliberation\n\n> **The landscape**: There is an explosion of academic interest in AI-augmented deliberation (March 2026 alone produced 5+ papers). But most focus on AI-as-mediator (generating consensus statements, summarizing arguments). Almost none build persistent, evolving argument graphs with formal scheme classification and critical question generation. Deliberus's differentiation is structural: it doesn't just mediate a conversation — it builds a *persistent knowledge structure* from it.\n\n---\n\n## 5. Terence Tao: Ideas Are Free, Verification Is the Bottleneck\n\n### 5.1 The Core Thesis\n\nTao articulated a paradigm shift across multiple venues in March 2026:\n\n**Mastodon post** (March 2026, widely quoted):\n> \"AI has driven the cost of idea generation down to almost zero... suddenly people can generate thousands of theories... Now we have to verify them, evaluate them.\"\n\n**Dwarkesh Podcast** (\"Terence Tao – Kepler, Newton, and the True Nature of Mathematical Discovery,\" March 20, 2026, ~84 min):\n- AI is \"like jumping machines that can jump two meters in the air, higher than any human. Sometimes they jump in the wrong direction, and sometimes they crash, but sometimes they can reach the tops of the lowest walls that we couldn't reach before.\"\n- But AI \"cannot build up cumulatively from partial progress\" — \"what they can't do is jump a little bit, reach some handhold, stay there, pull other people up, and then try to jump from there. There isn't this cumulative process which is built up interactively.\"\n- The distinction between \"artificial cleverness\" (brute-force search in solution space) and \"artificial intelligence\" (cumulative, contextual understanding)\n- In 2023, Tao predicted AI would be \"like a trustworthy co-author\" by 2026. By March 2026 at an IPAM conference, he says it \"saves more time than it wastes\" and is \"ready for primetime.\"\n\n**OpenAI Academy blog** (\"Terence Tao: AI Is Ready for Primetime in Math and Theoretical Physics,\" March 6, 2026):\n- How Tao came to trust AI as an assistant\n- His assessment shifted from \"mediocre grad student\" (September 2024) to \"ready for primetime\" (March 2026)\n\n### 5.2 The Lean Connection\n\nTao's answer to the verification problem is **formal verification using Lean**:\n- Lean verifies proofs line by line and \"can keep AI honest\"\n- Most working math is still \"informal mathematics\" (prose with equations) — room for subtle mistakes\n- AI can make this worse: \"producing arguments that look polished while hiding the weak step\"\n\n**Mathematics Distillation Challenge** (March 13, 2026): Tao launched the first SAIR Foundation competition — distilling 22 million algebra results from the Equational Theories Project (a Lean-formalized crowd-sourced effort) into a \"cheat sheet\" that enables weaker AI models to solve problems that currently stump them. The challenge structure:\n- Stage 1: Design a <=10KB cheat sheet maximizing accuracy of smaller models on universal algebra problems\n- Stage 2: Top 1,000 entries must generate formal proofs or valid counterexamples, not just answers\n\n**ChatGPT collaboration** (March 23, 2026): Tao posted a new paper to arXiv (\"Local Bernstein Theory, and Lower Bounds for Lebesgue Constants\") where one key inequality was proved by ChatGPT — a concrete example of AI-augmented mathematical research.\n\n### 5.3 \"Mathematical Methods and Human Thought in the Age of AI\" (March 29, 2026)\n\nThe most significant piece: a 40+ page philosophical essay by Tao and Tanya Klowden ([arXiv:2603.26524](https://arxiv.org/abs/2603.26524)), solicited for a forthcoming Blackwell Companion to the Philosophy of Mathematics. Published the same weekend Deliberus shipped its landing page text. Tao rarely writes philosophical essays; this one addresses AI's impact on the nature, purpose, and practice of knowledge itself.\n\n**Key arguments with Deliberus implications:**\n\n**The humanitarian lens**: \"AI tools and applications should not be viewed purely through the technical lens... but also through the macroscopic humanitarian lens of how our society, our shared body of knowledge and understanding, and our species benefits as a whole.\" This is Deliberus's civilizational vision in Tao's words.\n\n**Verification goes deeper than formal checking**: Lean certifies that \"a formalized argument establishes a formal mathematical statement, but does not rule out errors in translation between the formal statement and the original intended statement.\" Translation errors remain human-dependent. For Deliberus: the system can decompose arguments, but whether the decomposition *captures the meaning* is the discourse-layer problem Emanuel identified in 2013. Formal structure without semantic fidelity is hollow.\n\n**The \"penumbra\" of heuristic reasoning** — why arguments work, whether they generalize, what motivated them — resists formalization. \"Mathematical value transcends mechanical correctness; proofs must convey understanding, not merely deduce conclusions.\" This IS the attunement pole: analytical structure alone isn't enough; you need to understand WHY.\n\n**AI collapse — cumulative knowledge under threat**: AI trained on recursively generated AI outputs degrades over iterations. \"Without a sufficient amount of genuine content, AI becomes ungrounded from reality.\" A human-maintained argument graph is the antidote — genuine human reasoning, structured and persistent, immune to the recursive degradation that afflicts AI-generated content.\n\n**The \"red team\" framework**: Use AI for verification and testing (where errors are catchable), not primary knowledge generation (blue team). Deliberus does exactly this: AI extracts structure (red team), humans verify and create meaning (blue team).\n\n**The \"Copernican view\"**: \"Both human and artificial intelligences exist in the same ontological category, though with many distinctive differences and complementarities.\" This maps to the hybrid AGI thesis from Ivan Phillips (2011) and validates the convergence thesis's premise: human and AI reasoning can share a verification infrastructure built on common foundations.\n\n**Core tension articulated by Tao**: Mathematical truth is objective (formally verifiable), but mathematical *value* — what problems matter, which proofs illuminate — remains irreducibly human. For Deliberus: argument structure is formally decomposable, but argument *significance* — what matters, what's worth attending to — is the attunement layer that humans provide.\n\n### 5.5 The \"Creation to Filtration\" Shift\n\nMultiple sources (HowAIWorks, The Decoder, Awesome Agents) describe Tao's framing as a shift **from creation to filtration**:\n- Pre-AI: The bottleneck was generating ideas (expensive, slow, requires expertise)\n- Post-AI: Ideas are cheap. The bottleneck is verifying, evaluating, and filtering them\n- This demands new infrastructure for verification at scale\n\n> **Deliberus connection — the verification graph**: Tao's thesis maps perfectly onto Deliberus. Replace \"mathematical theorems\" with \"claims about the world\" and \"Lean proof checker\" with \"structured argument decomposition\":\n>\n> - **Ideas are free**: Anyone can make claims. LLMs can generate thousands. Social media already produces millions daily.\n> - **Verification is the bottleneck**: Which claims are supported by evidence? Which assumptions are they built on? Which are contested? Which survive critical questioning?\n> - **Lean is the answer for math**: A formal verification system where every step must be justified.\n> - **Deliberus is the answer for reasoning about the world**: A structured argumentation system where every claim is decomposed, scheme-classified, critically questioned, and connected to supporting/attacking evidence.\n>\n> The sorry model is the direct analogy: in Lean, a `sorry` marks an unproven assertion — a gap in the proof that must be filled. In Deliberus, a sorry marker (missing evidence, unanswered critical question, undecomposed value premise) marks a gap in the argument that is a functional entry point for contribution. Both systems make incompleteness *visible and actionable*.\n\n### 5.6 The Cumulative Reasoning Gap\n\nTao's observation that AI \"cannot build up cumulatively from partial progress\" — that it jumps but cannot establish handholds — is exactly the gap that a persistent argument graph fills. Individual AI interactions are ephemeral. A Deliberus-style graph provides the *persistent substrate* that enables cumulative reasoning:\n- Each extraction adds claims and connections to the existing graph\n- The `@[simp]` flywheel: each new source discovers connections to existing knowledge\n- Critical questions from one extraction inform the analysis of the next\n- The graph grows wiser over time — unlike any single AI conversation\n\n---\n\n## 6. Synthesis: What This Means for Deliberus\n\n### 6.1 Deliberus as Alignment Infrastructure\n\nThe research reveals that Deliberus occupies a unique position at the intersection of multiple high-priority research directions:\n\n| Research Direction | What They Need | What Deliberus Provides |\n|---|---|---|\n| Constitutional AI | Better constitutions | The deliberative *process* for generating them |\n| Debate-based alignment | Better judges | Structured argument decomposition for judge support |\n| Value alignment | Structured value elicitation | Moral Graph Elicitation at scale, with scheme classification |\n| Scalable oversight | Honest decomposition of reasoning | Transparent claim graphs with visible assumption structures |\n| AI governance | Democratic input on AI decisions | Persistent, navigable deliberation infrastructure |\n\n### 6.2 The Convergence Thesis Validated\n\nThe MGE paper (arXiv:2404.10636) provides empirical support for Deliberus's convergence thesis: when people's values are elicited in structured contexts, they *do* converge on directional relationships. This convergence is not imposed — it emerges from structured decomposition. This is exactly the mechanism Deliberus hypothesizes: decompose far enough, and shared bedrock becomes visible.\n\n### 6.3 The Analysis-Attunement Dialectic in Current Research\n\nThe research landscape reflects Deliberus's core dialectic:\n- **Analysis pole**: DCI's typed epistemic acts, Walton's 96 schemes, formal debate protocols, Lean verification\n- **Attunement pole**: The Habermas Machine's empathic mediation, perspective-taking in hidden-profile tasks, worldview navigation, the convergence thesis itself\n\nMost existing systems choose one pole. Kialo is pure analysis (no attunement). Polis is attunement-adjacent (opinion clustering) without argument structure. The Habermas Machine mediates but doesn't decompose. DCI formalizes but doesn't persist.\n\nDeliberus's vision — attunement-informed analysis on a persistent graph — has no competitor.\n\n### 6.4 The Sorry Model as Verification Infrastructure\n\nTao's verification thesis provides the strongest external validation for the sorry model:\n\n- **Lean**: `sorry` = unproven assertion, visible gap, actionable contribution point\n- **Deliberus**: sorry marker = missing evidence, unanswered CQ, undecomposed premise, visible gap, actionable contribution point\n- **Tao's thesis**: Verification is the bottleneck, and infrastructure for structured verification at scale is the solution\n\nThe sorry model is not a UI feature — it is **verification infrastructure for human reasoning**, analogous to what Lean provides for mathematical reasoning.\n\n### 6.5 The 80,000 Hours Opportunity\n\nDeliberus should position itself explicitly within the 80,000 Hours framework:\n- **High importance**: \"Most harmful decisions stem from failures under the wide umbrella of epistemics and coordination\" — Deliberus IS epistemics infrastructure\n- **High neglectedness**: \"Only a handful of projects are currently building these kinds of tools\" — and none with Deliberus's structural approach\n- **High tractability**: The extraction pipeline already works. 318 claims, 93 cross-extraction connections, 372 CQs. The `@[simp]` flywheel is validated\n\n### 6.6 The Sensemaking Infrastructure Vacuum\n\nThe Consilience Project diagnosed the problem (epistemic commons collapse) but prescribed the wrong medicine (articles). The sensemaking infrastructure vacuum remains. Deliberus provides the structural alternative: not better content, but better *infrastructure for reasoning about content*.\n\n### 6.7 Tao's Distillation Challenge as Template\n\nThe SAIR Mathematics Distillation Challenge offers a template for Deliberus: take a large body of structured knowledge (the argument graph), and challenge AI systems to distill it — to extract the essential reasoning patterns, identify the load-bearing assumptions, and present them in navigable form. This is precisely what Deliberus's extraction pipeline does for natural language arguments.\n\n---\n\n## 7. Specific Quotes and Citations for Deliberus Docs\n\n### For vision.md or landing page:\n> \"AI has driven the cost of idea generation down to almost zero... Now we have to verify them, evaluate them.\" — Terence Tao, March 2026\n\n> \"Most harmful decisions stem from failures in fact checking, prediction, negotiation, or other failures under the wide umbrella of 'epistemics and coordination.'\" — 80,000 Hours, \"Using AI to Enhance Societal Decision Making\"\n\n> \"Only a handful of projects are currently building the kinds of AI decision-making tools described — a drop in the ocean compared to the billions invested in developing more broadly capable AI agents.\" — 80,000 Hours\n\n### For the Lean analogy docs:\n> \"Lean verifies a proof line by line and can keep AI honest.\" — Terence Tao on formal verification\n\n> \"What AI tools can't do is jump a little bit, reach some handhold, stay there, pull other people up, and then try to jump from there. There isn't this cumulative process which is built up interactively.\" — Terence Tao, Dwarkesh Podcast, March 2026\n\n### For competitive positioning:\n> \"On non-routine tasks, DCI significantly improves over unstructured debate... DCI excels on hidden-profile tasks requiring perspective integration while failing on routine decisions, confirming task-dependence.\" — Prakash, \"From Debate to Deliberation,\" March 2026\n\n> \"Moral graph edges represent broad agreement amongst participants that one value is wiser than another for a particular context, and participants overwhelmingly converge on the directionality of these transitions.\" — Benkler et al., 2024\n\n### For AI safety framing:\n> \"Deliberative alignment: aligning AI systems with human values — not through static assumptions or abstract moral theories, but through inclusive, structured, and democratic deliberation.\" — AryaXAI, 2025\n\n> \"Recursive proposals have suffered from the obfuscated arguments problem: a dishonest system can adversarially choose how to recurse... meaning that an honest debater might need exponentially more compute than their dishonest opponent.\" — Brown-Cohen & Irving, 2025\n\n---\n\n## Sources\n\n### AI Alignment and Values\n- Benkler et al., \"What Are Human Values, and How Do We Align AI to Them?\" (arXiv:2404.10636, 2024) — https://arxiv.org/abs/2404.10636\n- Zhang et al., \"ValueCompass\" (arXiv:2409.09586, 2024) — https://arxiv.org/abs/2409.09586\n- \"Superalignment with Dynamic Human Values\" (arXiv:2503.13621, 2025) — https://arxiv.org/pdf/2503.13621\n- Pourdavood, \"Does Claude's Constitution Have a Culture?\" (arXiv:2603.28123, March 2026) — https://arxiv.org/html/2603.28123v1\n- MATS 9.0, \"How Well Do Models Follow Their Constitutions?\" (AI Alignment Forum, March 2026) — https://www.alignmentforum.org/posts/Tk4SF8qFdMrzGJGGw/\n- \"Moral Disagreement and the Limits of AI Value Alignment\" (AI & Society, June 2025) — https://link.springer.com/article/10.1007/s00146-025-02427-2\n- de Fine Licht, \"Resolving Value Conflicts in Public AI Governance\" (Chalmers, 2025) — https://research.chalmers.se/publication/546532\n\n### Debate-Based Alignment\n- Irving, Christiano, Amodei, \"AI Safety via Debate\" (2018, foundational)\n- Brown-Cohen, Irving et al., \"Prover-Estimator Debate\" (AI Alignment Forum, June 2025) — https://www.alignmentforum.org/posts/8XHBaugB5S3r27MG9/\n- UK AISI, \"An Alignment Safety Case Sketch Based on Debate\" (May 2025) — https://www.alignmentforum.org/posts/iELyAqizJkizBQbfr/\n- \"On Scalable Oversight with Weak LLMs Judging Strong LLMs\" (ICML 2025) — https://www.proceedings.com/content/079/079017-2395open.pdf\n\n### Deliberation and Collective Intelligence\n- Prakash, \"From Debate to Deliberation: DCI\" (arXiv:2603.11781, March 2026) — https://arxiv.org/abs/2603.11781\n- Turkstra et al., \"ARGSBASE\" (EACL 2026) — https://aclanthology.org/2026.eacl-demo.39.pdf\n- Tessler et al., \"AI Can Help Humans Find Common Ground\" (Science, October 2024) — https://www.science.org/doi/10.1126/science.adq2852\n- Stanford DEL, \"AI for Deliberation and Consensus\" — https://digitaleconomy.stanford.edu/project/ai-for-deliberation-and-consensus/\n- \"Human/AI Collective Intelligence for Deliberative Democracy\" (arXiv:2603.16260, March 2026) — https://arxiv.org/html/2603.16260\n- \"Argumentative Human-AI Decision-Making\" (arXiv:2603.15946, March 2026)\n- \"From Efficiency to Deliberation\" (Politics and Governance, January 2026) — https://www.cogitatiopress.com/politicsandgovernance/article/view/10632\n- \"The Case for Using Generative AI to Run Deliberation Simulations\" (Journal of Deliberative Democracy, February 2026) — https://delibdemjournal.org/article/id/1625/\n\n### 80,000 Hours\n- Qureshi, \"Using AI to Enhance Societal Decision Making\" (80,000 Hours, September 2025) — https://80000hours.org/problem-profiles/ai-enhanced-decision-making/\n- EA Forum discussion (February 2026) — https://forum.effectivealtruism.org/posts/659FtQYTESH2ec9Dt/\n\n### Metacrisis and Sensemaking\n- Consilience Project — https://consilienceproject.org/\n- Civilization Research Institute — https://civilizationresearchinstitute.org/\n- Van Peborgh, \"The Thunder of the Metacrisis\" (March 2026) — https://ernestopvanpeborgh.substack.com/p/the-thunder-of-the-metacrisis\n- \"Field-Building and the Epistemic Culture of AI Safety\" (First Monday, 2024) — https://firstmonday.org/ojs/index.php/fm/article/view/13626\n\n### Terence Tao\n- Tao, Mastodon/Mathstodon posts, March 2026 — https://mathstodon.xyz/@tao\n- Dwarkesh Podcast, \"Kepler, Newton, and the True Nature of Mathematical Discovery\" (March 20, 2026) — https://www.dwarkesh.com/p/terence-tao\n- OpenAI Academy, \"AI Is Ready for Primetime in Math and Theoretical Physics\" (March 6, 2026) — https://academy.openai.com/public/blogs/terence-tao-ai-is-ready-for-primetime-in-math-and-theoretical-physics-2026-03-06\n- Tao, \"Mathematics Distillation Challenge – Equational Theories\" (March 13, 2026) — https://terrytao.wordpress.com/2026/03/13/mathematics-distillation-challenge-equational-theories/\n- Tao, \"Local Bernstein Theory\" (arXiv, March 23, 2026) — https://terrytao.wordpress.com/2026/03/23/local-bernstein-theory-and-lower-bounds-for-lebesgue-constants/\n- The Decoder, \"Tao Says AI Drives Idea Generation Cost to Near Zero\" — https://the-decoder.com/terence-tao-says-ai-drives-idea-generation-cost-to-near-zero-but-shifts-the-bottleneck-to-verification/\n- HowAIWorks, \"Terence Tao: The Future of AI in Mathematics\" — https://howaiworks.ai/blog/terence-tao-dwarkesh-patel-interview\n- Awesome Agents, \"Ideas Are Now Free\" — https://awesomeagents.ai/news/terence-tao-ai-verification-bottleneck-math/\n\n### AI Governance\n- \"Democratic Governance through DAO-based Deliberation\" (Nature Scientific Reports, March 2026) — https://nature.com/articles/s41598-026-40180-8\n- Rivelli, \"AIMED: AI-assisted Iterative Method for Ethical Deliberation\" (2025) — https://philsci-archive.pitt.edu/26676/1/Rivelli_2025_AIMED_Preprint.pdf\n- AryaXAI, \"Deliberative Alignment\" (June 2025) — https://aryaxai.com/article/deliberative-alignment-building-ai-that-reflects-collective-human-values\n"}