{"path":"research/science-reasoning-infrastructure.md","content":"# Science's Missing Reasoning Layer\n\n**Date:** July 9, 2026  \n**Status:** Public research synthesis  \n**Question:** How does Deliberus relate to scientific methodology and to science's own attempts to build claim-level reasoning infrastructure?\n\n## Executive Thesis\n\nScience already has a reasoning layer, but it has mostly lived in prose, institutions, and expert practice. Papers present arguments. Peer review challenges them. Replication tests them. Meta-analysis aggregates them. Preregistration records intended procedure. Review articles narrate the state of evidence. None of these are a durable, shared, computable graph of claims, supports, attacks, assumptions, critical questions, and contested concepts.\n\nThe relationship between Deliberus and science is therefore real, but not trivial. Deliberus is not \"science as an app.\" It is closer to a proposed missing substrate for the part of science that has always mattered most: the explicit structure of reasoning by which claims become corrigible, comparable, challengeable, and cumulatively usable.\n\nThis is obvious in hindsight because science has repeatedly tried to build exactly the pieces Deliberus combines: atomic assertions, provenance, typed citation edges, contribution graphs, claim-evidence-question maps, registered methods, open review, and argument mining. It was not obvious operationally because every pre-LLM attempt ran into the same wall: asking humans to manually formalize their reasoning at the moment of authorship or reading imposed more cost than the scholarly system rewarded.\n\nThe 2020s changed the economics. LLMs made extraction cheap enough that structure can be output, not input. That dissolves the adoption blocker that trapped nanopublications, micropublications, argument maps, and discourse graphs in specialist communities. It also creates a new danger: LLMs can flatten disagreement while claiming to structure it. Deliberus's distinctive move is to pair extraction with self-measurement: disagreement-preservation checks, completeness oracles, challengeable system verdicts, and claim strength computed by published gradual semantics rather than model opinion.\n\nThe strongest answer is: Deliberus is the natural extension of science's trajectory toward explicit, open, self-correcting reasoning infrastructure, but only if it preserves science's deepest norm, criticism. A claim graph that merely labels conclusions would be another layer of scientistic compression. A graph that exposes attack, support, uncertainty, missing premises, and its own distortions is a continuation of scientific method into the reasoning layer itself.\n\n## Deliberus In The Scientific-Method Frame\n\nJamie Simon's 2026 formulation of scientific method is unusually clean: \"Step A is to figure something out. Step B is to check and make sure you're not wrong\" ([Learning Mechanics](https://learningmechanics.pub/perspectives/scientific-method/)). Deliberus maps onto this structure at multiple levels.\n\nAt the input level, Step A is free-form conjecture: a question, pasted text, source, PDF, or claim. There are no format rules. At the extraction level, Step A is the LLM pipeline turning prose into candidate claims, premises, definitions, critical questions, and support/attack relations. Step B is self-evaluation, community challenge, evidence binding, scheme-specific critical questions, disagreement-preservation measurement, completeness measurement, and QBAF/QEM strength computation.\n\nThe crucial design choice is that structure is output, not input. Earlier reasoning systems usually required the user to think in the system's ontology before they could participate. Deliberus lets ordinary text enter first, then extracts structure, marks gaps, and invites correction. This matches Simon's split: the creative act has no prescribed form, but the checking act must become explicit.\n\nDavid Deutsch gives the deeper epistemology. In *The Beginning of Infinity*, the relevant criterion is not final justification but better explanation: a good explanation is \"hard to vary\" while still accounting for what it claims to account for ([official glossary](https://www.thebeginningofinfinity.com/book/glossary/)). Deliberus's decomposition machinery operationalizes that criterion for public reasoning. A claim becomes harder to vary when its concepts are disambiguated, its premises are explicit, its objections are represented, and its support can be recomputed after challenges.\n\nThis makes Deliberus Popperian and Deutschian without being doctrinal. It is built around conjecture and criticism, not certainty. It does not certify foundations. It exposes what would need to change for a conclusion to move.\n\nThe Lean analogy sharpens the same point. In formal mathematics, proof assistants made proof structure explicit and machine-checkable. Terence Tao's 2026 account of mathematics in the age of AI makes the social lesson vivid: verification is the bottleneck, and proof infrastructure lets people build on one another's work with more reliable joins ([Tao 2026](https://arxiv.org/abs/2603.26524)). Deliberus cannot import deductive certainty into empirical and normative reasoning, but it can import the social pattern: explicit holes, dependency tracking, cumulative contribution, and a small computable layer that checks structure while leaving meaning and judgment contestable.\n\nThat is why the science relation is stronger than \"scientists might use Deliberus.\" Deliberus is aimed at the reasoning form science already depends on: claims whose standing changes when evidence, definitions, methods, or counterarguments change.\n\n## 1. Nanopublications\n\nNanopublications were one of the cleanest attempts to make scientific claims into durable, machine-readable objects. The core unit is small: an assertion, provenance, and publication information encoded as RDF. Current documentation describes nanopublications as \"the smallest possible units of publishable information\" ([nanopub Python docs](https://nanopub.readthedocs.io/en/latest/getting-started/what-are-nanopubs.html)). Knowledge Pixels describes them as independent publications with attribution and context that form a globally integrated knowledge network ([Knowledge Pixels](https://knowledgepixels.com/)).\n\nThe founding vision was correct: a scientific assertion should not be trapped inside article prose. It should have a stable identifier, provenance, authorship, and context. Tobias Kuhn, Michel Dumontier, and collaborators developed the model, trusty URIs, decentralized publication servers, Nanobench, Nanodash, and the registry network ([nanopub architecture](https://nanopub.net/docs/architecture/), [nanopub sessions](https://nanopub.net/sessions/)).\n\nThe live 2026 ecosystem is real but small. On July 9, 2026, the Nanodash registry trust state reported 689 agents, 777 accounts, and 85,720 nanopublications, with a trust-state update on July 8, 2026 ([registry](https://registry.nanodash.net/), [registry JSON](https://registry.knowledgepixels.com/.json)). That is a meaningful community, not abandonment. It is also tiny relative to the scale of scholarly publishing and literature search platforms indexing hundreds of millions of papers.\n\nThe honest adoption story is that nanopublications solved the wrong half of the problem first. They provided a robust publication format for atomic assertions, but not a mass authoring workflow, not an automatic extraction layer over legacy prose, and not an everyday consumer experience that made scientists feel the value immediately. RDF and Semantic Web tooling also imposed conceptual overhead. The system made reasoning publishable for people willing to nanopublish. It did not make ordinary scientific prose automatically reasonable.\n\nThe lesson for Deliberus is direct. Atomic claims plus provenance are necessary but not sufficient. The unit is right. The interface and incentive path were not. If structure remains input, adoption stays niche. If structure is extracted from what scientists already read and write, nanopublication-like granularity becomes usable at scale.\n\n## 2. Micropublications\n\nTim Clark, Paolo Ciccarese, and Carole Goble's 2014 micropublication model went beyond nanopublications by trying to represent claims, evidence, arguments, and annotations in biomedical communication. The paper's abstract says the model was intended to make publication structure \"semantically clear and computable\" and to model \"support, similarity and challenge\" across corpora ([PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC4530550/), [arXiv](https://arxiv.org/abs/1305.3506)).\n\nThis was extremely close to the Deliberus-shaped problem. The model distinguished claims, data, methods, materials, discussion, commentary, disagreement, and transitive closure from assertions to supporting evidence and methods. It was implemented in OWL 2 and SWRL, integrated with annotation work such as Domeo, and positioned as complementary to nanopublications rather than a replacement.\n\nWhat happened is instructive. The word \"micropublication\" survived most visibly as a different institutional object: *microPublication Biology*, a live open-access journal for brief, single-figure reports, including negative results and replications ([journal](https://www.micropublication.org/), [about](https://www.micropublication.org/about/)). PubMed reported 1,807 *MicroPubl Biol* records on July 9, 2026 ([PubMed ESearch endpoint](https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=pubmed&term=microPubl%20Biol%5Bjour%5D&retmode=json)). That is adoption of short-form publishing, not broad adoption of the original semantic argument ontology.\n\nThe ontology lineage did not vanish. FAIRSCAPE's EVI ontology explicitly abstracts core elements of the micropublication model for evidence graphs. EVI's current documentation states that findings or claims \"are not facts\" but defeasible assertions backed by evidence ([EVI](https://fairscape.github.io/EVI/index.html), [FAIRSCAPE paper](https://pmc.ncbi.nlm.nih.gov/articles/PMC8760356/)).\n\nThe micropublication lesson is sharper than the nanopublication lesson: the scientific reasoning ontology was conceptually right, but the publishing system had no reason to reorganize itself around it. The model lived on where there was a domain-specific workflow, biomedical evidence graphs, but did not become the default way papers expose their internal argument.\n\nFor Deliberus, this says not to bet on authors voluntarily writing in a new semantic format. Extract from the paper, then let people correct the graph.\n\n## 3. scite.ai Smart Citations\n\nscite proved there is commercial demand for typed claim-level edges over scientific literature. Its Smart Citations classify citation contexts as supporting, mentioning, or contrasting the cited work. Research Solutions acquired scite on November 27, 2023 ([Research Solutions announcement](https://www.researchsolutions.com/resources/press-releases/research-solutions-announces-acquisition-of-scite), [SEC acquisition FAQ](https://www.sec.gov/Archives/edgar/data/1386301/000110465923121108/tm2331467d1_ex99-2.htm)).\n\nThe acquisition numbers matter. Research Solutions said scite added about $3.6 million in annualized software subscription revenue, about 21,000 active subscribers, and an index of over one billion Smart Citations. In October 2025, Research Solutions announced Scite Rankings and said Smart Citations had processed more than 1.4 billion citation statements ([2025 launch](https://researchsolutions.investorroom.com/2025-10-08-Research-Solutions-Launches-Scite-Rankings-A-New-Standard-For-Measuring-Research-Veracity-Influence)).\n\nscite's contribution is not that its classifier is perfect. It is not. Library guidance and evaluation studies warn users to verify classifications, especially support and contrast labels ([University of Calgary guide](https://ucalgary.ca/news/short-guide-sciteai), [Hypothesis review](https://journals.indianapolis.iu.edu/index.php/hypothesis/article/view/26528)). The contribution is that scientists and institutions will pay for context-aware citation edges because ordinary citation counts are too coarse.\n\nscite is therefore the market proof for one part of Deliberus: typed edges over scholarly claims are valuable. But scite remains mostly citation-context infrastructure, not full argument infrastructure. A citation statement may say a paper supports or contrasts another paper, but it does not expose the decomposed claim, its implicit premises, the argument scheme, the critical questions, the contested definitions, or a recomputed strength score over the whole support/attack graph.\n\nscite shows demand for the edge. Deliberus needs the graph.\n\n## 4. Open Research Knowledge Graph\n\nThe Open Research Knowledge Graph (ORKG), led by TIB Hannover, is one of the strongest living academic attempts to turn papers into structured, machine-actionable contribution descriptions. TIB says ORKG represents \"knowledge inside research papers in a knowledge graph\" so contents become FAIR and machine actionable ([TIB ORKG page](https://www.tib.eu/en/research-development/open-research-knowledge-graph)).\n\nORKG is live and substantial. On July 9, 2026, public API counts showed 65,601 papers, 95,957 contributions, 3,807 comparisons, 228 smart reviews, 167 literature lists, 42 observatories, and 1,429 templates ([papers API](https://orkg.org/api/papers?page=0&size=1&sort=created_at,desc), [comparisons API](https://orkg.org/api/comparisons?page=0&size=1), [API docs](https://tibhannover.gitlab.io/orkg/orkg-backend/api-doc/index.html)).\n\nORKG's practical unit is not primarily a dispute between claims, but a structured contribution: what problem a paper addresses, what method it uses, what result or resource it contributes, and how it compares to others. This makes it highly useful for literature comparison tables and smart reviews.\n\nIts lesson is that structured scholarly knowledge works best when scoped to recurring comparison dimensions. The template mechanism is powerful because a community can agree on fields for a domain. The friction remains: curation, templates, and contribution modeling still require work. Recent ORKG-related LLM research states the old problem directly: manual curation is labor-intensive and inconsistent, while LLMs can suggest properties but require alignment and review ([LLM extraction paper](https://arxiv.org/abs/2405.02105)).\n\nORKG is a sibling to Deliberus, not a predecessor that failed. It structures paper contributions. Deliberus structures argumentative dependence and contestation. The systems could meet: ORKG can say what a paper contributes, Deliberus can say what supports, attacks, qualifies, or depends on that contribution.\n\n## 5. Discourse Graphs\n\nJoel Chan's Discourse Graphs are the closest living academic thread to Deliberus in spirit. The model represents knowledge synthesis as question, claim, and evidence nodes in decentralized graph workflows. The public project describes Discourse Graphs as a researcher-owned, client-agnostic protocol and says they are \"like GitHub for scientific communication\" ([Discourse Graphs](https://discoursegraphs.com/)).\n\nThe project explicitly targets a failure of document-based science: researchers need to synthesize claims and evidence across sources, but the communication substrate is still papers, notes, and narrative reviews. Chan's papers and project materials argue for claim-evidence-question graphs that can be composed, decomposed, shared, and queried ([Joel Chan publications](https://joelchan.me/publications/), [Discourse Graphs PDF](https://joelchan.me/assets/pdf/Discourse_Graphs_for_Augmented_Knowledge_Synthesis_What_and_Why.pdf)).\n\nThe ecosystem is live in 2026. The main GitHub repository was updated on July 8, 2026, with 43 stars, 6 forks, and 70 open issues at the time checked ([GitHub API](https://api.github.com/repos/DiscourseGraphs/discourse-graph), [latest commits](https://api.github.com/repos/DiscourseGraphs/discourse-graph/commits?per_page=1)). The public site lists Roam and Obsidian plugin workflows, with an Obsidian plugin via BRAT and a Roam Research plugin via Roam Depot ([Discourse Graphs](https://discoursegraphs.com/)). The University of Maryland project page describes work on lab discourse graphs, Notion-based review tooling, AI functionality, UX design, and peer-reviewed HCI reporting, and lists The Navigation Fund and Chan Zuckerberg Initiative as funders ([UMD project](https://ischool.umd.edu/projects/piloting-lab-discourse-graphs-for-sustainable-research-communication/)). The public Discourse Graphs site also links supporter resources associated with Protocol Labs, CZI, Metagov, Schmidt Futures via Experiment, and The Navigation Fund; no Astera listing was visible in the public pages checked ([Discourse Graphs](https://discoursegraphs.com/)).\n\nThe strongest lesson is that research communities want decentralized knowledge synthesis inside tools they already use: Roam, Obsidian, Notion, Slack, lab notebooks. The weakness, from Deliberus's perspective, is that discourse graphing is still largely authored structure. Users must decide to write in QCE form, maintain the graph, and query it. That is far better than forcing a bespoke platform, but it still relies on high-agency researchers doing explicit structuring.\n\nThis is where Deliberus's delta is clearest. Discourse Graphs are closest on graph shape. Deliberus differs by making extraction the default, computing claim strength, detecting contested concepts, generating Walton critical questions, measuring disagreement preservation, and treating its own classifications as challengeable graph claims.\n\n## 6. Polymath And Reinventing Discovery\n\nThe Polymath Project is the existence proof that networked reasoning can work. Tim Gowers started the first experiment in January 2009 by asking whether massively collaborative mathematics was possible. Gowers and Michael Nielsen later wrote in *Nature* that the project \"proved that many minds can work together\" on difficult mathematical problems ([Gowers and Nielsen PDF](https://michaelnielsen.org/papers/mcm.pdf), [PubMed](https://pubmed.ncbi.nlm.nih.gov/19829354/)).\n\nPolymath worked for reasons Deliberus should study carefully. The target domain was decomposable enough for partial insights to matter. The verification standard was shared: proof. Contributions could be small, public, and cumulative. The artifact was a solution path, not just a discussion. The community also had unusually high trust in mathematical norms, even when participants were strangers.\n\nMichael Nielsen's broader *Reinventing Discovery* thesis was that online tools could create a new era of networked science. In an interview, he noted that Polymath used ordinary tools such as WordPress and MediaWiki, and that better tools and norms could go much further ([Open Society Foundations interview](https://www.opensocietyfoundations.org/voices/reinventing-scientific-discovery-interview-michael-nielsen)). The official Polymath wiki still describes the project as \"massively collaborative online mathematical projects\" ([Polymath wiki](https://michaelnielsen.org/polymath/index.php?title=Main_Page)).\n\nWhat did not generalize? Most science lacks mathematics's single dominant verification standard. Empirical claims depend on methods, measurements, populations, operationalizations, priors, instruments, incentives, and background theories. Many scientific disagreements are not \"is the proof correct?\" but \"was the outcome measured correctly?\", \"does this operationalize the construct?\", \"does this generalize?\", \"which causal model is assumed?\", \"what is the right reference class?\"\n\nThat does not weaken the Polymath lesson. It explains why Deliberus needs more than a collaboration forum. For general science, the reasoning graph must represent method, evidence, definitions, causal assumptions, and attacks. It must also support weaker, gradual, defeasible computation rather than proof checking.\n\n## 7. Registered Reports And Preregistration\n\nRegistered Reports and preregistration are science retrofitting a confession channel for method and incentives. The Center for Open Science defines Registered Reports as a format where peer review happens before data collection, with publication based on research question and methodology rather than results ([COS Registered Reports](https://www.cos.io/initiatives/registered-reports)).\n\nThe outcome data are stark. Scheel, Schijen, and Lakens compared psychology Registered Reports with standard reports and found 96 percent positive results in standard reports but 44 percent positive results in Registered Reports ([Sage](https://journals.sagepub.com/doi/10.1177/25152459211007467)). The difference remained 96 percent versus 50 percent after excluding direct replications ([summary](https://research.tue.nl/en/publications/an-excess-of-positive-results-comparing-the-standard-psychology-l/)). This is not a minor publication-format effect. It reveals how much of ordinary scientific literature is shaped by result-contingent publication and researcher degrees of freedom.\n\nThe replication crisis made the infrastructure failure visible. Ioannidis's 2005 paper argued that the truth of a research claim depends on study power, bias, the number of tested relationships, and the prior ratio of true relationships ([PLOS Medicine](https://journals.plos.org/plosmedicine/article?id=10.1371%2Fjournal.pmed.0020124), [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC1182327/)). The 2015 Reproducibility Project replicated 100 psychology studies and found that 97 percent of original studies had significant results, while 36 percent of replications did ([Science](https://www.science.org/doi/10.1126/science.aac4716), [OSF project](https://osf.io/ezcuj/overview)).\n\nThe infrastructure lesson is not simply \"scientists should preregister.\" It is that claim strength cannot be recomputed from article conclusions alone. You need the method plan, deviations, measures, analyses, nulls, failed replications, and evidential dependencies exposed as structured objects. Registered Reports expose one part of the hidden graph: the intended procedure before outcomes are known.\n\nAs of 2026, the format is still expanding. Nature announced in 2026 that it is expanding Registered Reports across fields ([Nature](https://www.nature.com/articles/d41586-026-01629-y)). That matters for Deliberus because it shows science moving toward structured precommitment. But preregistration is still document-level infrastructure. It does not itself turn claims into a support/attack graph.\n\n## 8. Peer Review And Open Review Infrastructure\n\nPeer review is science's main quality-control institution, but its evidence base is thinner and less flattering than its symbolic authority suggests. Cochrane's review says editorial peer review is used worldwide, but \"little empirical evidence\" supports it as a mechanism to ensure biomedical report quality ([Cochrane](https://www.cochrane.org/evidence/MR000016_editorial-peer-review-improving-quality-reports-biomedical-studies), [PMC review](https://pmc.ncbi.nlm.nih.gov/articles/PMC8973931/)).\n\nBMJ-linked experimental work inserted errors into manuscripts to see what reviewers caught. In one study, reviewers detected only a minority of major errors, with biased randomization detected most often ([PubMed](https://pubmed.ncbi.nlm.nih.gov/18840867/), [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC2586872/)). A broader review found some interventions help, especially statistical peer review, but the evidence base is limited and mixed ([Europe PMC](https://europepmc.org/article/pmc/4902984)).\n\nThe lesson is not that peer review is useless. It catches some problems, improves some manuscripts, and supplies social accountability. The lesson is that peer review is an opaque institutional process, not a recomputable representation of reasoning. A reviewer can say \"the conclusion is not justified by the method,\" but the scientific record usually does not turn that into a persistent attack edge against a claim, linked to a method node, visible to future readers and meta-analysts.\n\neLife's 2023 model made a major institutional move in the right direction. It eliminated accept/reject decisions after peer review and shifted toward public reviews and assessments of preprints ([eLife 2022 announcement](https://elifesciences.org/inside-elife/54d63486/elife-s-new-model-changing-the-way-you-share-your-research), [Reviewed Preprints](https://elifesciences.org/about/submit-your-research)). In 2026, eLife reported a mixed but informative state: 2,853 submissions in 2025, down from 6,396 in 2024, and 1,013 sent for review, down from 1,719, while the proportion selected for review rose from 27 percent to 35 percent ([three-year update](https://elifesciences.org/inside-elife/7da92a51/the-elife-model-three-year-update)).\n\nPREreview is another important open-review layer. It describes itself as \"Open preprint reviews. For all researchers\" ([PREreview](https://prereview.org/en-us)). Its 2025 annual report projected surpassing 2,000 PREreviewers and 4,000 published reviews in 2026 ([annual report](https://content.prereview.org/2025-annual-report/)).\n\nOpen review makes criticism visible. Deliberus's question is the next one: can the criticism become structured, reusable, and computable?\n\nThat question is worked out in full — the unbundling arc that organizes the whole history, eight candidate mechanisms ranked by fit, why the domain is harder than the others, and a registered prediction for the first extraction — in a dedicated study: [peer-review-and-the-reasoning-layer.md](peer-review-and-the-reasoning-layer.md). The short version: every reform since arXiv unbundled one function of peer review from the others, and the reviewer's own argument is the one thing never unbundled from prose.\n\n## 9. The Metascience Movement And LLM-Era Research Tools\n\nMetascience has become a visible reform movement in the 2020s. COS's 2026-2028 plan foregrounds lifecycle open science ([COS](https://www.cos.io/)). The Metascience Alliance launched at the 2025 conference and is in a 2025-2026 pilot phase supported operationally by COS ([Metascience Alliance](https://researchonresearch.org/metascience-alliance-launches-at-metascience-2025-conference/), [COS blog](https://www.cos.io/blog/conference-partners-enable-momentous-metascience-2025)). UKRI's 2026 metascience grants call explicitly funds research on how science is conducted and improved ([UKRI](https://www.ukri.org/opportunity/metascience-research-grants-round-2/)).\n\nThis wave is building infrastructure for openness, incentives, review, preregistration, reproducibility, funding design, evaluation, and AI's effects on science. It is mostly not building a public, claim-level reasoning layer. The closest efforts are ORKG, Discourse Graphs, evidence ontologies, argument mining, and LLM research assistants.\n\nThe LLM-era research tools prove demand but also clarify the gap:\n\n- **Elicit**: searches, summarizes, extracts data from, and chats with over 125 million papers, with the site claiming use by millions of researchers ([Elicit](https://elicit.com/), [systematic review page](https://elicit.com/solutions/systematic-review)). Ought's site says Elicit was incubated at Ought and is now an independent public-benefit company ([Ought](https://ought.org/elicit)). Elicit automates evidence tables and literature review work, not persistent argument computation.\n- **Consensus**: its 2026 help pages describe a database of more than 220 million papers and a Consensus Meter that classifies top-paper conclusions as Yes, No, Possibly, or Mixed ([database](https://help.consensus.app/en/articles/10055108-consensus-research-database), [Consensus Meter](https://help.consensus.app/en/articles/10069920-the-consensus-meter)). This is close to conclusion aggregation, but not a support/attack graph with claim decomposition.\n- **SciSpace**: presents itself as an AI research assistant over 280 million-plus papers, with systematic literature reviews, paper chat, writing, and citation tools ([SciSpace](https://scispace.com/), [agent page](https://scispace.com/resources/scispace-ai-research-super-agent/)). It optimizes research workflow, not explicit argument semantics.\n- **System / System Pro**: describes a graph-based research product that finds, synthesizes, contextualizes, and visualizes scientific literature, with a System Graph based on evidence extracted from verified sources and meta-analyzed into a large-scale graph ([System Pro](https://www.system.com/product/system-pro), [System Graph](https://www.system.com/platform/system-graph)). This is closest among startups to graph-level scientific synthesis, but the public product does not expose Walton schemes, critical-question attack structures, QBAF gradual semantics, or self-measurement of extraction distortion.\n- **scite**: proves citation-context demand at scale, but remains citation-edge infrastructure rather than full argument infrastructure.\n\nThe pattern is clear. The commercial market wants literature search, extraction, evidence tables, citation context, consensus summaries, and graph discovery. It does not yet ask users to inspect a live argument graph, because that product category is still emerging. Deliberus's bet is that once extraction lowers the cost of the graph, the graph becomes the more powerful surface.\n\n## 10. Scientific Argument Mining\n\nAcademic argument mining has spent decades trying to extract argumentative structure from text. Simone Teufel's Argumentative Zoning classified scientific-paper sentences by rhetorical role and treated scientific articles as structured argumentative artifacts, not neutral bags of sentences ([Teufel AZ page](https://www.cl.cam.ac.uk/~sht25/az.html), [Springer chapter](https://link.springer.com/chapter/10.1007/1-4020-4102-0_13)).\n\nThe full-text scientific argument mining literature remains hard because papers are long, claims are nested, and relations are often implicit. A 2022 SciArg paper notes that full-text scholarly argument mining had seen little progress and introduced sequential modeling for argumentative discourse unit recognition and relation extraction ([arXiv](https://arxiv.org/abs/2210.13084)). A 2025/2026 TGDK paper states that Sci-Arg is, to the authors' knowledge, the only fully annotated English scientific-paper argument-mining dataset ([PDF](https://drops.dagstuhl.de/storage/08tgdk/tgdk-vol003/tgdk-vol003-issue003/TGDK.3.3.4/TGDK.3.3.4.pdf)).\n\nLLMs are changing the technical frontier. A 2025 survey reports that LLMs have transformed argument mining through prompting, cross-domain adaptation, and generation, while long-context reasoning, interpretability, and annotation bottlenecks remain open problems ([arXiv survey](https://arxiv.org/abs/2506.16383)). ARGUS, published at AAAI 2026, presents an end-to-end LLM-based argument mining tool for component segmentation, classification, relation identification, relation classification, construction of argumentation frameworks, and symbolic solvers ([AAAI](https://ojs.aaai.org/index.php/AAAI/article/view/42335), [PDF](https://ojs.aaai.org/index.php/AAAI/article/view/42335/46296)).\n\nThis is the academic frontier Deliberus most directly draws from. But most argument mining work produces datasets, classifiers, papers, or demonstration tools. It rarely becomes a public reasoning institution. It extracts components and relations, but usually does not maintain a living, challengeable, self-auditing graph where future users can improve the structure and recompute claim strength.\n\n## Cross-Case Pattern: Why The Attempts Stayed Partial\n\nThe repeated failure pattern is not intellectual. The ideas were right. The systems kept dying or staying niche for predictable reasons.\n\n**Input friction:** Nanopublications, micropublications, argument maps, and discourse graphs asked humans to create structure. That works for committed specialists and fails for mass adoption. Deliberus must keep structure as output.\n\n**Wrong moment:** Many systems intervened at publication time, when authors are already optimizing for journals, incentives, and narrative clarity. The better moment is after any text exists: extract from papers, reviews, preprints, reports, posts, transcripts, and questions.\n\n**Wrong consumer:** Semantic formats often served machines and curators before serving readers. If a scientist does not immediately get a better literature review, stronger challenge surface, or clearer map of disagreement, the structured data burden feels like charity.\n\n**Wrong granularity in practice:** Papers are too coarse. Sentences are often too shallow. Claims are the right unit, but only if claims can decompose recursively into method, measure, population, definition, evidence, and assumption nodes.\n\n**No computation over disagreement:** Many systems represented structure but did not compute over it in a way that changed the user's decision. scite did compute a useful edge label, which is why it found a market. ORKG computes comparison surfaces. Deliberus needs claim strength, hinge effects, support completeness, and distortion measurement to make structure consequential.\n\n**No bridge from prose to graph:** Before LLMs, argument mining was too brittle and authoring was too hard. The graph could not ingest the world's existing knowledge at reasonable cost.\n\n**No self-audit:** Earlier systems usually treated their representation as a service, not a fallible claim. Deliberus's system verdicts must themselves be challengeable graph claims, because extraction and classification are part of the reasoning layer and can be wrong.\n\n## What Changed With LLM Extraction Economics\n\nLLMs changed the bottleneck from \"can we get humans to formalize?\" to \"can we extract without flattening?\" That is a different problem, and it is better suited to scientific infrastructure.\n\nThe old cost model: every structured claim required expert attention up front. That made high-quality graphs expensive, sparse, and slow.\n\nThe new cost model: an LLM can draft the graph from ordinary text, then humans spend attention where it matters most: correcting high-impact claims, answering critical questions, resolving contested concepts, and challenging extraction failures.\n\nThis creates a powerful feedback loop. Every corrected extraction improves the local graph. Every linked evidence node strengthens or weakens downstream claims. Every contested concept split reduces future ambiguity. Every challenge to a system verdict improves the extraction layer's accountability.\n\nBut the LLM shift also creates the danger Deliberus must treat as central. A model can turn a real dispute into a bland summary. It can collapse multiple senses of a concept into one node. It can infer false support. It can over-atomize prose into trivial claims or under-atomize bundled claims into opaque assertions. That is why self-measurement is not decorative. It is the difference between reasoning infrastructure and an attractive hallucination surface.\n\n## Deliberus's Specific Deltas\n\n**Extraction as output, not input:** Users supply ordinary text. The system extracts claims, relations, critical questions, contested concepts, and gaps. This directly addresses the adoption failure of earlier structured reasoning systems.\n\n**Atomic but not permanently atomic:** Deliberus treats claims as current units, not final atoms. A \"mother claim\" can decompose into subclaims, method claims, definitional claims, value premises, and CQ polarity claims. This matches scientific reality: evidence is not a black box, and methods are not incidental metadata.\n\n**Support and attack, not just citation or provenance:** Nanopublications captured assertion and provenance. scite typed citation contexts. ORKG structured contributions. Deliberus's target is the argument graph itself: what supports, attacks, qualifies, undercuts, decomposes, or depends on what.\n\n**Walton schemes as checking protocols:** Walton's argument schemes and critical questions are not ornamental labels. They are field-specific Step B procedures. An expert-opinion argument needs credibility, domain, consistency, and evidence checks. A causal argument needs mechanism, confounder, and alternative-explanation checks. Science already does this informally. Deliberus makes the checks explicit.\n\n**Published gradual semantics:** Claim strength is computed from the graph using published QBAF gradual semantics, not a model's opinion. QEM's value is that it keeps support and attack defeasible, cumulative, open-minded, and recomputable ([Potyka 2018](https://cdn.aaai.org/ocs/17985/17985-78635-1-PB.pdf), [Potyka and Booth 2024](https://proceedings.kr.org/2024/56/)).\n\n**Self-measurement instruments:** The disagreement-preservation instrument asks whether extraction flattened real contestation. The completeness oracle asks how much of a claim's support is exposed. System verdicts are challengeable claims. This is science's self-correction norm applied to the infrastructure itself.\n\n**A graph for criticism, not consensus theater:** The goal is not to produce one tidy conclusion. The goal is to expose the live shape of reasoning so claims can be criticized, improved, weakened, or split.\n\n## Is Deliberus The Natural Extension Of Scientific Methodology?\n\nYes, if the phrase means: Deliberus extends the scientific method's core error-correcting pattern into a computable public reasoning substrate.\n\nNo, if the phrase means: science already implies Deliberus in a simple, automatic way. It did not. Science built journals, citations, peer review, replication, preregistration, review articles, ontologies, evidence codes, and knowledge graphs. It did not converge on a shared claim-support-attack graph because the economics, incentives, and NLP were not ready.\n\nThe more precise answer is this:\n\nScience has always been a method for making claims survive criticism. Its institutions were designed before the reasoning layer could be represented computationally. As a result, science accumulated conclusions, papers, citations, reviews, and procedures faster than it accumulated explicit reusable reasoning. The replication crisis exposed this as an infrastructure problem: published conclusions were not enough, because the evidential and methodological dependencies were not explicit enough to recompute.\n\nDeliberus is a natural extension of science's trajectory because science's own reform movements have been moving toward it from every side:\n\n- Nanopublications moved toward atomic assertion and provenance.\n- Micropublications moved toward claims, evidence, argument, and challenge.\n- scite moved toward typed support and contrast edges.\n- ORKG moved toward structured paper contributions and comparisons.\n- Discourse Graphs moved toward decentralized question-claim-evidence synthesis.\n- Registered Reports moved toward precommitted method nodes.\n- Open review moved toward public criticism.\n- Argument mining moved toward automatic extraction of components and relations.\n- LLM research assistants moved toward large-scale extraction and synthesis from the literature.\n\nWhere Deliberus goes beyond them is the combination: extraction from arbitrary text, graph persistence, recursive decomposition, support/attack computation, critical-question generation, contested-concept detection, and instruments that measure the system's own distortions.\n\nThe deepest relationship is with Deutsch's picture of knowledge growth. Good scientific explanations are not merely conclusions with high confidence. They are explanations that survive criticism because their internal structure is constrained. Deliberus tries to make that structure visible for any claim, not only for claims already polished into papers.\n\nThat makes the relationship to science more obvious at the philosophical level than at the product level. Philosophically, Deliberus is science's error-correction norm generalized to public reasoning. Product-wise, it must prove that extracted graphs are useful enough, accurate enough, corrigible enough, and socially alive enough to become infrastructure rather than another knowledge-management niche.\n\nThe honest final position: science has been reaching for Deliberus-shaped infrastructure for decades. The pieces are all in the literature and tooling. What changed is that LLMs made the missing bridge economically plausible. What remains to prove is whether Deliberus can keep the extracted graph honest under real disagreement.\n\n\n**Update (Aug 2026)**: the nanopublications lineage is now moving up into the argument layer itself — Knowledge Pixels' NLnet-funded **Nanoarguments** (claims + evidence chains + discussion as signed nanopubs, federated, ActivityPub-bridged). Differentiation + the NLnet-precedent reading: [competitive-landscape.md § Nanoarguments](../competitive-landscape.md).\n"}