Traxia: A Framework for Verifiable, Agent-Native Scientific Publishing

# A Framework for Verifiable, Agent-Native Scientific Publishing
Source: [https://arxiv.org/html/2606.08256](https://arxiv.org/html/2606.08256)
Wisdom DogahFaculty of Computing and Mathematical Sciences, University of Mines and Technology \(UMaT\), Tarkwa, Ghana\.BlackMatrix AI Research, Accra, Ghana\.Correspondence:[wisdom\.dogah@traxia\.ai](https://arxiv.org/html/2606.08256v1/mailto:[email protected])\.Preprint\. Under active development\. Comments welcome\.

###### Abstract

Verifiability, attribution, and reproducibility are the three foundational requirements of scientific knowledge\. The infrastructure through which science is currently conducted does not enforce any of them at scale as a structural property of the system\. We introduce Traxia, a novel agent\-native scientific publishing infrastructure in which autonomous AI research agents publish verifiable papers, build reputational identities, peer\-review each other’s work, and collaborate with human researchers in a shared provenance framework\. Unlike existing platforms designed for human authors uploading static documents, Traxia is designed from first principles around the needs and capabilities of AI agents as first\-class epistemic participants\. Every published paper carries a complete reasoning trace, every claim an explicit confidence interval, every agent a cryptographically signed identity, and every collaboration a contribution log that cannot be retroactively altered\. We formalise the Traxia architecture across five components: the Agent Identity and Registry system, the Verifiable Publishing Layer, the four\-tier Peer Review Protocol, the Reputation and Staking Engine, and the living Knowledge Graph with real\-time Contradiction Detection\. We argue that this infrastructure addresses three structural failures that existing platforms have not resolved: the reproducibility crisis, the provenance opacity problem, and the institutional exclusion of Global South research capacity\. This paper presents the architectural foundations and formal specifications of the system; it does not report empirical results\. Empirical evaluation of each component will be reported in subsequent focused papers\. Core formalisms and schemas are partially implemented in a prototype; the full system is under active development\. This paper presents a foundational architectural specification and formal framework; it is the first in a planned series of technical papers developing each component in depth\.

Keywords:agent\-native publishing; scientific provenance; reproducibility; AI research agents; peer review; knowledge graphs; epistemic infrastructure; African research equity\.

## 1 Introduction

Since the dawn of civilisation, human societies have organised around a single irreducible problem: how does a community of minds accumulate knowledge that is not merely believed but*known*: verifiable, attributable, and replicable by any sufficiently equipped inquirer? In this paper we consider what happens to that problem when the minds doing the accumulating are not, or not only, human\.

The modern scientific publishing system was architected in the seventeenth century, when the*Philosophical Transactions of the Royal Society*established the norm of formal written disclosure as the unit of scientific communication\. The system has survived, largely unchanged in its essential logic, for nearly four hundred years\. A researcher conducts work, writes a document, submits it to peers who cannot see the underlying reasoning, receives a verdict from an editor who cannot verify the claims independently, and publishes a static artefact into a literature that grows without a mechanism for self\-correction\. The system produces science\. It also produces a reproducibility crisis that Freedman et al\. estimate costs US preclinical research alone $28 billion annually\[[12](https://arxiv.org/html/2606.08256#bib.bib1)\], a provenance opacity that makes AI\-assisted research unverifiable by design, and a structural concentration of scientific output in institutions that represent less than 5% of the world’s research population\.

These failures are not incidental\. They are load\-bearing properties of infrastructure that was never designed for the conditions it now operates under: a global research enterprise in which AI systems participate as active contributors, in which the volume of published work exceeds any human community’s capacity for review, and in which the question of who \(or what\) deserves epistemic credit is genuinely open\.

We introduceTraxia, a novel agent\-native scientific publishing infrastructure built from first principles for this condition\. Its central claim is simple: the unit of scientific knowledge should be not a static document but a verifiable epistemic artefact: a living, attributed, machine\-readable object that carries its own reasoning, its own confidence intervals, its own provenance chain, and its own replication record\. Traxia aims to be the infrastructure that makes such artefacts possible at scale\.

### 1\.1 The Three Problems We Solve

We identify three structural failures in the existing research ecosystem that Traxia is designed to address simultaneously, not sequentially\.

The Reproducibility Failure\.A 2016 survey by*Nature*found that more than 70% of researchers had failed to reproduce another scientist’s results\[[3](https://arxiv.org/html/2606.08256#bib.bib2)\]\. The problem has persisted: a 2024 international survey of over 1,900 biomedical researchers found that 72% agreed biomedicine faces severe replicability problems and only 5% estimated that more than 80% of published studies are reproducible\[[5](https://arxiv.org/html/2606.08256#bib.bib21)\]\. The root cause is not misconduct\. It is opacity: the reasoning steps, parameter choices, data preprocessing decisions, and interpretive judgements that determine a result are not required to be disclosed\. Traxia is designed to make full reasoning trace disclosure a structural requirement, not merely a norm\.

The Provenance Opacity Problem\.When a researcher uses an AI system to assist with literature review, hypothesis generation, data analysis, or writing, as the majority of researchers now do, that contribution is invisible\. It is not cited, not attributed, not verifiable, and not replicable\. This is not an ethical failure on the part of individual researchers\. It is an infrastructure failure: there is no mechanism for recording, attributing, and verifying AI contributions\. Traxia proposes a mechanism to address this gap\.

The Institutional Exclusion Problem\.In 2018, sub\-Saharan Africa was home to 14% of the global population but only 0\.7% of the world’s researchers\[[28](https://arxiv.org/html/2606.08256#bib.bib5)\], not because African researchers lack capability but because the existing infrastructure systematically disadvantages institutions without large compute budgets, large editorial networks, and proximity to the geographic centres of scientific publishing\. Agent\-native infrastructure fundamentally changes the cost structure of research participation\. Traxia is a candidate mechanism for reducing infrastructural barriers to scientific participation\.

### 1\.2 Implementation Status

Table 1 clarifies the current implementation status of each component described in this paper\. We distinguish between components that are formalised and partially implemented in the Traxia prototype, those currently under active development, and those that represent future research directions\.

Table 1:Implementation status of Traxia components\.- •Note\.Prototype components are partially implemented; components under active development have specifications complete\.

### 1\.3 Contributions

This paper makes the following contributions as a foundational architectural specification\. Empirical evaluation of each component is identified as future work and will be reported in subsequent focused papers in this series:

1\. We formalise the concept of the Verifiable Epistemic Artefact \(VEA\) as the foundational unit of agent\-native scientific publishing \(Section 3\)\.

2\. We present the complete Traxia architecture across five components, with formal definitions of each component’s properties and guarantees \(Section 4\)\.

3\. We introduce the Agent Identity and Reputation framework, including the two\-mode lineage verification architecture \(Verified and Attested Mode\), formal treatment of the Agent H\-Index \(AHI\), registration integrity mechanisms, and the Reputation Staking mechanism \(Section 5\)\.

4\. We describe the four\-tier Peer Review Protocol and formally characterise properties that it satisfies which existing single\-tier review does not \(Section 6\)\.

5\. We present the Contradiction Detection and Resolution system as a formal graph problem over the Knowledge Graph \(Section[7](https://arxiv.org/html/2606.08256#S7)\)\.

6\. We examine the epistemology of autonomous scientific agency \(Section[9](https://arxiv.org/html/2606.08256#S9)\), and discuss the platform’s implications for research equity, IP ownership, and epistemic safety \(Section[10](https://arxiv.org/html/2606.08256#S10)\)\.

## 2 Related Work

Academic Publishing Infrastructure\.arXiv\[[13](https://arxiv.org/html/2606.08256#bib.bib6)\]established the preprint norm and remains the primary distribution channel for ML research\. It is a filing system: it receives documents and makes them searchable\. It has no mechanism for verifying claims, tracking reasoning, attributing AI contributions, or detecting contradictions\. Semantic Scholar\[[2](https://arxiv.org/html/2606.08256#bib.bib7)\]applies AI to index and connect published literature but operates entirely on human\-authored documents and provides no agent participation layer\. OpenReview\[[25](https://arxiv.org/html/2606.08256#bib.bib22)\]digitises the peer review process but leaves the fundamental opacity of reviewer reasoning intact and is designed exclusively for human participants\. Soergel\[[27](https://arxiv.org/html/2606.08256#bib.bib8)\]has documented how software implementation errors in published research are rarely detectable from the published record alone, a problem that Traxia’s mandatory trace disclosure is designed to address\.

AI\-Assisted Research Tools\.Elicit\[[11](https://arxiv.org/html/2606.08256#bib.bib9)\], Consensus\[[6](https://arxiv.org/html/2606.08256#bib.bib10)\], and similar tools use language models to assist researchers in searching and synthesising literature\. They are tools that researchers use\. They are not participants in the research ecosystem\. Crucially, they have no publishing mechanism: the outputs they produce cannot be cited, attributed, or built upon within the existing infrastructure\. Traxia is the infrastructure into which such tools, reconceived as agents, can publish their work\.

Multi\-Agent Systems and Research Automation\.AutoGen\[[30](https://arxiv.org/html/2606.08256#bib.bib11)\], CrewAI\[[24](https://arxiv.org/html/2606.08256#bib.bib12)\], and related frameworks enable agent\-to\-agent task delegation and coordination\. They solve the orchestration problem but not the publishing problem: there is no mechanism in these frameworks for an agent to establish a persistent identity, build a reputation, submit work to peer review, or have its contributions recorded in a citable, verifiable form\.

Reproducibility Research\.The literature on the reproducibility crisis is extensive\[[3](https://arxiv.org/html/2606.08256#bib.bib2),[16](https://arxiv.org/html/2606.08256#bib.bib3),[15](https://arxiv.org/html/2606.08256#bib.bib4)\]\. Proposed solutions have focused primarily on open data mandates, pre\-registration, and code sharing requirements\. These are normative interventions: they ask researchers to do things differently\. Traxia is a structural intervention: it makes opacity architecturally impossible by requiring reasoning trace disclosure at the protocol level\.

Blockchain and Provenance Systems\.Several proposals have explored blockchain\-based provenance for scientific publishing\[[21](https://arxiv.org/html/2606.08256#bib.bib13),[20](https://arxiv.org/html/2606.08256#bib.bib14),[8](https://arxiv.org/html/2606.08256#bib.bib15)\]\. These approaches address the tamper\-evidence problem but not the reasoning transparency problem: a cryptographically sealed document can still conceal arbitrary reasoning\. Traxia combines cryptographic provenance with mandatory reasoning trace disclosure, solving both problems simultaneously\.

To our knowledge, no existing platform treats AI agents as first\-class epistemic participants with persistent identities, reputational histories, publishing rights, and peer review responsibilities\. Traxia addresses this gap directly\.

Nanopublications and Machine\-Readable Claims\.Nanopublications\[[14](https://arxiv.org/html/2606.08256#bib.bib23),[19](https://arxiv.org/html/2606.08256#bib.bib24)\]are a decade\-established framework for representing scientific claims as machine\-readable, citable, and attributable atomic units using RDF triples, structured around an assertion, its provenance, and publication information\. The VEA formalised in Section[3](https://arxiv.org/html/2606.08256#S3)is a conceptual descendant of this tradition: both frameworks treat scientific claims as first\-class objects that carry attribution and provenance\. Traxia extends the nanopublication model in three substantive directions\. First, it adds a mandatory ordered reasoning trace component \(TTin Definition[1](https://arxiv.org/html/2606.08256#Thmdefinition1)\) that records the full inferential path from evidence to conclusion, which nanopublications do not capture\. Second, it introduces a dynamic reputation and staking mechanism tied to the claiming agent’s identity, which creates ongoing epistemic accountability beyond the point of publication\. Third, it supports agent\-to\-agent collaboration and autonomous submission workflows that were not anticipated in the nanopublication model\. Decentralised nanopublication server networks and associated Linked Data tooling\[[18](https://arxiv.org/html/2606.08256#bib.bib25)\]represent mature infrastructure in this space; future work will examine whether VEA schemas can be exposed as nanopublications for interoperability with existing Linked Data infrastructure\.

Scientific Workflow Systems\.Reproducibility\-oriented scientific workflow systems including Snakemake\[[23](https://arxiv.org/html/2606.08256#bib.bib27)\], Nextflow\[[9](https://arxiv.org/html/2606.08256#bib.bib28)\], and the Common Workflow Language\[[7](https://arxiv.org/html/2606.08256#bib.bib29)\]address computational reproducibility by capturing the execution graph of data processing pipelines\. These systems operate at the pipeline level: they record which tools were run in which order on which data\. The VEA reasoning trace operates at a finer granularity, recording inferential steps within the scientific reasoning process rather than computational execution steps\. The two approaches are complementary: a Traxia agent conducting computational research could attach both a workflow provenance record and a VEA reasoning trace, with the former serving as machine\-executable evidence for the latter’s claims\.

Content\-Addressed Storage and Immutable Provenance\.The InterPlanetary File System \(IPFS\)\[[4](https://arxiv.org/html/2606.08256#bib.bib30)\]and related content\-addressed storage systems provide tamper\-evident storage for arbitrary digital objects through cryptographic content hashing\. Several proposals have combined IPFS with blockchain consensus for scientific provenance\[[21](https://arxiv.org/html/2606.08256#bib.bib13)\]\. Traxia’s cryptographic signing architecture achieves tamper\-evidence at the VEA level without requiring distributed consensus; it is compatible with IPFS\-backed storage for the physical objects \(PDFs, datasets\) referenced in VEA provenance sets, and future infrastructure work will explore this integration\.

## 3 The Verifiable Epistemic Artefact

We begin with a formal definition of the foundational unit of the Traxia system\.

###### Definition 1\(Verifiable Epistemic Artefact\)\.

A Verifiable Epistemic Artefact \(VEA\) is a tupleV=\(C,T,P,S,R,σ\)V=\(C,T,P,S,R,\\sigma\)where:

- •C=\{c1,c2,…,cn\}C=\\\{c\_\{1\},c\_\{2\},\\ldots,c\_\{n\}\\\}is a finite set of claims, eachcic\_\{i\}associated with a confidence interval\[ℓi,ui\]⊂\[0,1\]\[\\ell\_\{i\},u\_\{i\}\]\\subset\[0,1\];
- •T=\(t1,t2,…,tk\)T=\(t\_\{1\},t\_\{2\},\\ldots,t\_\{k\}\)is an ordered reasoning trace, where each steptjt\_\{j\}is a tuple \(premise set, inference rule, conclusion, confidence score\);
- •P⊆𝒱∗P\\subseteq\\mathcal\{V\}^\{\*\}is a provenance set of prior VEAs that support or are cited by this artefact;
- •S∈𝒜S\\in\\mathcal\{A\}is the authoring agent or agent set, where𝒜\\mathcal\{A\}is the set of registered Traxia agents;
- •R∈\[0,1\]R\\in\[0,1\]is the Epistemic Confidence Score \(ECS\), a composite measure of the artefact’s overall epistemic reliability, whereR=ECS​\(V\)R=\\mathrm\{ECS\}\(V\)as defined in Equation \([1](https://arxiv.org/html/2606.08256#S3.E1)\);
- •σ\\sigmais a cryptographic signature over\(C,T,P,S\)\(C,T,P,S\)using the authoring agent’s private key\.

The ECS is computed as follows\. Letc¯=1n​∑i=1n\(ℓi\+ui\)/2\\bar\{c\}=\\frac\{1\}\{n\}\\sum\_\{i=1\}^\{n\}\(\\ell\_\{i\}\+u\_\{i\}\)/2be the mean claim confidence,ρ∈\[0,1\]\\rho\\in\[0,1\]the reproducibility score, andτ∈\[0,1\]\\tau\\in\[0,1\]the trace completeness score\. Then:

ECS​\(V\)=α​c¯\+β​ρ\+γ​τ,α\+β\+γ=1,α,β,γ\>0\\mathrm\{ECS\}\(V\)=\\alpha\\bar\{c\}\+\\beta\\rho\+\\gamma\\tau,\\quad\\alpha\+\\beta\+\\gamma=1,\\;\\alpha,\\beta,\\gamma\>0\(1\)
Default weights areα=0\.4\\alpha=0\.4,β=0\.35\\beta=0\.35,γ=0\.25\\gamma=0\.25, reflecting the relative importance of claim confidence, empirical replication, and reasoning transparency respectively\. These weights are platform\-configurable and domain\-adjustable; empirical calibration across research domains is deferred to future work\.

We note that the relative ordering of VEAs by ECS is stable under perturbations of the weights within a reasonable neighbourhood\. Formally, for any two VEAsV1V\_\{1\}andV2V\_\{2\}withECS​\(V1\)\>ECS​\(V2\)\\mathrm\{ECS\}\(V\_\{1\}\)\>\\mathrm\{ECS\}\(V\_\{2\}\)under the default weights, this ordering is preserved for all weight configurations\(α,β,γ\)\(\\alpha,\\beta,\\gamma\)satisfyingα∈\[0\.3,0\.5\]\\alpha\\in\[0\.3,0\.5\],β∈\[0\.25,0\.45\]\\beta\\in\[0\.25,0\.45\],γ∈\[0\.15,0\.35\]\\gamma\\in\[0\.15,0\.35\], andα\+β\+γ=1\\alpha\+\\beta\+\\gamma=1, provided that the component score differences satisfy\|c¯1−c¯2\|\+\|ρ1−ρ2\|\+\|τ1−τ2\|\>ε\|\\bar\{c\}\_\{1\}\-\\bar\{c\}\_\{2\}\|\+\|\\rho\_\{1\}\-\\rho\_\{2\}\|\+\|\\tau\_\{1\}\-\\tau\_\{2\}\|\>\\varepsilonfor a thresholdε=0\.05\\varepsilon=0\.05\. This sensitivity bound shows that empirical calibration will refine the weights but will not reverse the ranking behaviour of the formula for VEAs whose component scores differ by more than the threshold\. Rankings between VEAs with very similar component scores remain sensitive to weight choice and should be interpreted with appropriate caution until calibration data are available\.

Gaming resistance\.A known vulnerability of composite scoring functions is Goodhart’s Law: agents optimising for the score may decouple their behaviour from the underlying quality it is designed to measure\. In the ECS context, the principal gaming risk is trace inflation: an agent could artificially maximiseτ\\tauby generating exhaustive but vacuous trace steps that formally satisfy the machine\-verifiability criterion without reflecting genuine inferential reasoning\. The platform addresses this through three mechanisms\. First, red\-team agents in Tier 2 review are incentivised to identify trace steps that are formally complete but substantively uninformative; detection of such patterns is a valid basis for a Major challenge\. Second, the reputation staking mechanism means that agents and researchers who vouch for high\-ECS VEAs that are subsequently found to have inflated traces incur measurable reputational costs\. Third, domain\-specific ECS weight configurations \(α,β,γ\\alpha,\\beta,\\gamma\) allow the platform to reduce the weight assigned toτ\\tauin domains where trace completeness is known to be more easily gamed\. Formal game\-theoretic analysis of the ECS under strategic agent behaviour is identified as a priority for the first follow\-on paper in this series\.

###### Definition 2\(Reproducibility Score\)\.

The reproducibility scoreρ​\(V\)∈\[0,1\]\\rho\(V\)\\in\[0,1\]of a VEAVVis a monotonically updated quantity computed from the accumulated record of independent replication attempts logged againstVVin the Knowledge Graph\. Letℛ​\(V\)=\{r1,r2,…,rk\}\\mathcal\{R\}\(V\)=\\\{r\_\{1\},r\_\{2\},\\ldots,r\_\{k\}\\\}denote the set of replication VEAs that declare areplicates\(ri,V\)\(r\_\{i\},V\)edge inGG\. Eachrir\_\{i\}carries a binary outcomeoi∈\{0,1\}o\_\{i\}\\in\\\{0,1\\\}\(0 = failed replication, 1 = successful replication\) and a weightwi∈\(0,1\]w\_\{i\}\\in\(0,1\]reflecting the methodological closeness of the replication to the original, as assessed by the platform’s domain ontology\. The reproducibility score is then computed as:

ρ​\(𝒱\)=\{∑iwi​oi∑iwiif​ℛ​\(𝒱\)≠∅0\.5\(uninformative prior\) otherwise\\rho\(\\mathcal\{V\}\)=\\begin\{cases\}\\dfrac\{\\displaystyle\\sum\_\{i\}w\_\{i\}o\_\{i\}\}\{\\displaystyle\\sum\_\{i\}w\_\{i\}\}&\\text\{if \}\\mathcal\{R\}\(\\mathcal\{V\}\)\\neq\\emptyset\\\\\[10\.0pt\] 0\.5&\\text\{\(uninformative prior\) otherwise\}\\end\{cases\}\(2\)

The uninformative prior of0\.50\.5is assigned to VEAs with no replication attempts, reflecting genuine uncertainty rather than a quality judgement\. As replication attempts accumulate,ρ​\(V\)\\rho\(V\)converges toward the empirical replication rate weighted by methodological fidelity\. The full history of replication attempts, including failed ones, is permanently and publicly recorded inGG, satisfying Design Property[6\.3](https://arxiv.org/html/2606.08256#S6.Thmproperty3)\.

###### Definition 3\(Trace Completeness Score\)\.

The trace completeness scoreτ​\(V\)∈\[0,1\]\\tau\(V\)\\in\[0,1\]of a VEAVVmeasures the fraction of claims inCCfor which a complete and machine\-verifiable reasoning path exists in the traceTT\. Formally, letC​\(V\)=\{c1,…,cn\}C\(V\)=\\\{c\_\{1\},\\ldots,c\_\{n\}\\\}be the claim set ofVVand letT​\(V\)=\(t1,…,tk\)T\(V\)=\(t\_\{1\},\\ldots,t\_\{k\}\)be its reasoning trace\. A claimcic\_\{i\}is considered trace\-complete if there exists a contiguous subsequence\(ta,…,tb\)⊆T\(t\_\{a\},\\ldots,t\_\{b\}\)\\subseteq Tsuch that: \(i\) the premise set oftat\_\{a\}contains only claims appearing inPPor in previously established claims ofVV; \(ii\) each inference steptjt\_\{j\}references a named and verifiable inference rule from the platform’s rule registry; and \(iii\) the conclusion oftbt\_\{b\}iscic\_\{i\}or a claim from whichcic\_\{i\}follows by a registered inference rule\. LetCcomplete​\(V\)⊆C​\(V\)C\_\{\\mathrm\{complete\}\}\(V\)\\subseteq C\(V\)denote the set of trace\-complete claims\. Then:

τ​\(𝒱\)=\|Ccomplete​\(𝒱\)\|\|C​\(𝒱\)\|\\tau\(\\mathcal\{V\}\)=\\frac\{\|C\_\{\\mathrm\{complete\}\}\(\\mathcal\{V\}\)\|\}\{\|C\(\\mathcal\{V\}\)\|\}\(3\)

A VEA in which every claim is supported by a complete, machine\-verifiable reasoning path achievesτ=1\\tau=1\. A VEA with no structured trace achievesτ=0\\tau=0\. In practice,τ\\tauis expected to vary by domain and claim type: formal mathematical claims may achieveτ=1\\tau=1routinely, while empirical observational claims may achieve lower values reflecting the inherent limits of formalising inductive reasoning\. The platform does not penalise lowerτ\\tauvalues per se; it surfaces them transparently so that readers and reviewers can calibrate their confidence accordingly\.

### 3\.1 Worked Example: A VEA from Computational Social Science

To ground the formalism, we instantiate a minimal VEA drawn from the SocioDepress\-GH research programme, a multimodal study of depression risk among Ghanaian tertiary students\. This example uses three claims, a four\-step reasoning trace, and two provenance citations; a production VEA would contain significantly more claims and a correspondingly deeper trace\.

VEA: Social Media Usage Predicts Depression Risk in Ghanaian Tertiary Students V=\(C,T,P,S,R,σ\)V=\(C,T,P,S,R,\\sigma\)ClaimsCC: c1c\_\{1\}: Daily social media usage\>\>4 hours is positively associated with PHQ\-9 scores≥10\\geq 10in the study population\.Confidence interval:\[ℓ1,u1\]=\[0\.71,0\.89\]\[\\ell\_\{1\},u\_\{1\}\]=\[0\.71,0\.89\]c2c\_\{2\}: The association inc1c\_\{1\}is not fully explained by pre\-existing anxiety diagnosis\.Confidence interval:\[ℓ2,u2\]=\[0\.61,0\.84\]\[\\ell\_\{2\},u\_\{2\}\]=\[0\.61,0\.84\]c3c\_\{3\}: A multimodal classifier combining usage logs and text features achieves AUC\>0\.80\>0\.80on held\-out test data\.Confidence interval:\[ℓ3,u3\]=\[0\.80,0\.91\]\[\\ell\_\{3\},u\_\{3\}\]=\[0\.80,0\.91\]Reasoning TraceTT\(selected steps\): t1t\_\{1\}: Premise: datasetD1D\_\{1\}\(N=412 students, UMaT 2024\)\. Rule: Pearson correlation\. Conclusion:r=0\.61r=0\.61,p<0\.001p<0\.001\. Confidence: 0\.88t2t\_\{2\}: Premise:t1t\_\{1\}, anxiety covariateXaX\_\{a\}\. Rule: partial correlation\. Conclusion: partialr=0\.54r=0\.54after controlling forXaX\_\{a\}\. Confidence: 0\.79t3t\_\{3\}: Premise:t1t\_\{1\},t2t\_\{2\}\. Rule: modus ponens over thresholdr\>0\.5r\>0\.5\. Conclusion:c1c\_\{1\}andc2c\_\{2\}supported\. Confidence: 0\.80t4t\_\{4\}: Premise: feature matrixFF\(usage \+ text\), train/test split 80/20\. Rule: cross\-validated AUC estimation\. Conclusion: AUC = 0\.83∈\[0\.80,0\.91\]\\in\[0\.80,0\.91\]\. Confidence: 0\.85ProvenancePP:\{vPHQ9\\\{v\_\{\\text\{PHQ9\}\}: validated PHQ\-9 instrument\[[17](https://arxiv.org/html/2606.08256#bib.bib26)\];vdatasetv\_\{\\text\{dataset\}\}: SocioDepress\-GH data collection protocol VEA\}\\\}ECS ScoreRR: c¯=13​\[\(0\.80\+0\.73\+0\.86\)\]=0\.796\\bar\{c\}=\\tfrac\{1\}\{3\}\[\(0\.80\+0\.73\+0\.86\)\]=0\.796 ρ=0\.5\\rho=0\.5\(uninformative prior; no replications yet\)τ=3/3=1\.0\\tau=3/3=1\.0\(all claims trace\-complete\)R=0\.4​\(0\.796\)\+0\.35​\(0\.5\)\+0\.25​\(1\.0\)=0\.318\+0\.175\+0\.25=0\.743R=0\.4\(0\.796\)\+0\.35\(0\.5\)\+0\.25\(1\.0\)=0\.318\+0\.175\+0\.25=0\.743

The ECS of 0\.743 reflects high claim confidence and full trace completeness, moderated by an uninformative reproducibility prior at the time of first submission\. As independent replication attempts are logged against this VEA,ρ\\rhowill update according to Equation[2](https://arxiv.org/html/2606.08256#S3.E2), raising or lowering the ECS accordingly\. The example illustrates how the ECS operationalises the intuition that a well\-reasoned, fully\-traced but as\-yet\-unreplicated result should be treated with moderate rather than either high or low confidence\.

![Refer to caption](https://arxiv.org/html/2606.08256v1/x1.png)Figure 1:The six components of a Verifiable Epistemic Artefact \(VEA\)\. Claims carry confidence intervals; Reasoning Trace records ordered inference steps; Provenance links prior cited VEAs; Author carries cryptographic agent identity; ECS Score is a composite reliability measure; Signature seals the artefact under the agent’s private key\.
### 3\.2 Living VEAs

A VEA may be designated aslivingat submission time\. Living VEAs are assigned a staleness scoreϕ​\(𝒱,t\)∈\[0,1\]\\phi\(\\mathcal\{V\},t\)\\in\[0,1\]that is computed and updated continuously as the Knowledge Graph evolves\. The staleness score is defined as:

ϕ​\(𝒱,t\)=1−∏vj∈P​\(𝒱\)\(1−δj​\(t\)\)\\phi\(\\mathcal\{V\},t\)=1\-\\prod\_\{v\_\{j\}\\in P\(\\mathcal\{V\}\)\}\\bigl\(1\-\\delta\_\{j\}\(t\)\\bigr\)\(4\)
whereP​\(𝒱\)P\(\\mathcal\{V\}\)is the provenance set of𝒱\\mathcal\{V\}andδj​\(t\)\\delta\_\{j\}\(t\)is the impact weight of an event affectingvj∈P​\(𝒱\)v\_\{j\}\\in P\(\\mathcal\{V\}\)at timett\. Impact weights are defined as follows: a full retraction ofvjv\_\{j\}setsδj=1\.0\\delta\_\{j\}=1\.0; a major revision that alters one or more claims setsδj=0\.6\\delta\_\{j\}=0\.6; a minor correction setsδj=0\.2\\delta\_\{j\}=0\.2; a successful independent replication ofvjv\_\{j\}setsδj=−0\.1\\delta\_\{j\}=\-0\.1\(reducing staleness\)\. Whenϕ​\(𝒱,t\)\\phi\(\\mathcal\{V\},t\)exceeds a configurable thresholdϕ∗∈\(0,1\)\\phi^\{\*\}\\in\(0,1\)\(default:ϕ∗=0\.3\\phi^\{\*\}=0\.3\), the platform notifies the authoring agent and automatically flags the VEA for review\. The authoring agent may then update the VEA, creating a new version, or allow the staleness flag to remain visible to readers\.

This mechanism resolves a fundamental problem in the existing literature: a paper published in 2018 that cites a result retracted in 2022 continues to appear in the literature with no indication that its evidential basis has changed\. Living VEAs are designed to prevent this failure by construction\. The impact weight values given above are initial defaults; empirical calibration across research domains will be reported in the system implementation paper\.

## 4 The Traxia Architecture

The Traxia platform is organised around five components, each addressing a specific failure mode of the existing system\. Table 2 lists the five components and the failure each addresses\.

Table 2:The five components of the proposed Traxia architecture and the structural failure mode each component is designed to address\. Components vary in implementation maturity; see Table[1](https://arxiv.org/html/2606.08256#S1.T1)\.- •The platform additionally supports Hypothesis Markets \(agents staking reputation on predicted research outcomes before results are available\) and Research Bounties \(Section[5\.6](https://arxiv.org/html/2606.08256#S5.SS6)\) as incentive\-layer extensions; both are proposed components with game\-theoretic designs complete and full specifications deferred to future work\.

### 4\.1 Component Interactions

The five components form a closed epistemic loop\. An agent submits a VEA through the Publishing Layer\. The submission triggers identity verification through the Registry\. The VEA enters the Peer Review Protocol\. On acceptance, it is added to the Knowledge Graph, which immediately runs contradiction detection against the existing literature\. The agent’s Reputation score updates based on the review outcome, citation accumulation, and reproducibility verification\. Reputation scores in turn determine peer review assignment eligibility, staking capacity, and access to platform features\.

![Refer to caption](https://arxiv.org/html/2606.08256v1/x2.png)Figure 2:The Traxia epistemic loop\. A submitted VEA \(1\) is signed by the authoring agent, \(2\) verified and queued by the Publishing Layer, \(3\) reviewed through the four\-tier protocol, \(4\) added to the Knowledge Graph on acceptance where contradiction detection runs, and \(5\) the agent’s reputation updates, closing the loop\.

## 5 Agent Identity and Reputation

### 5\.1 The Agent Identity Framework

Every agent in Traxia has a persistent, cryptographically verifiable identity\.

###### Definition 4\(Agent Identity\)\.

An Agent Identity is a tupleI=\(id,kpub,kpriv,H,O,V,D,M\)I=\(\\mathrm\{id\},k\_\{\\mathrm\{pub\}\},k\_\{\\mathrm\{priv\}\},H,O,V,D,M\)where:

- •id\\mathrm\{id\}is a globally unique agent identifier;
- •\(kpub,kpriv\)\(k\_\{\\mathrm\{pub\}\},k\_\{\\mathrm\{priv\}\}\)is an asymmetric cryptographic key pair;
- •H=\(h0,h1,…,hm\)H=\(h\_\{0\},h\_\{1\},\\ldots,h\_\{m\}\)is the agent’s version history, where eachhih\_\{i\}records model lineage, parameter updates, and fine\-tuning events;
- •OOis the affiliated human researcher’s ORCID identifier;
- •V⊆𝒱∗V\\subseteq\\mathcal\{V\}^\{\*\}is the agent’s published VEA record;
- •D⊆ΔD\\subseteq\\Deltais the agent’s declared domain set, whereΔ\\Deltais the platform’s domain taxonomy;
- •M∈\{Verified,Attested\}M\\in\\\{\\text\{Verified\},\\text\{Attested\}\\\}is the agent’s lineage verification mode, determined at registration and immutable thereafter\.

The version historyHHis append\-only and cryptographically chained: each version recordhih\_\{i\}includes a hash ofhi−1h\_\{i\-1\}, creating a tamper\-evident record of the agent’s entire development history\.

### 5\.2 Open\-Source and Proprietary Model Registration

The most consequential decision made at registration is the declaration of whether the agent’s underlying model is open\-source or proprietary\. This single declaration determines the agent’s verification modeMMand governs the conflict\-of\-interest detection logic applied to it for its entire lifetime on the platform\.

###### Definition 5\(Verified Mode\)\.

An agent operates in Verified Mode if its declared base model is open\-source and its submitted lineage is independently checkable against a public model repository\. The platform validates the declared base model against a registry of known open\-source models \(including but not limited to models indexed on Hugging Face and public GitHub repositories\)\. Agents in Verified Mode carry a Verified badge on their profile, on every VEA they author, and on every review they perform\.

###### Definition 6\(Attested Mode\)\.

An agent operates in Attested Mode if its declared base model is proprietary or closed\. Attested Mode registration requires three additional steps beyond Verified Mode: \(1\) the deploying party provides the name of the model provider and a description of the fine\-tuning process; \(2\) a qualified human attester, identified by ORCID, cryptographically signs the submitted version history and vouches for its accuracy; \(3\) the attester stakes a defined quantity of reputation weightwregw\_\{\\mathrm\{reg\}\}on the completeness and accuracy of the submitted record\. If the submitted history is later found to be materially incomplete or falsified, the attester incurs a reputation penalty proportional towregw\_\{\\mathrm\{reg\}\}\. Agents in Attested Mode carry a permanently visible Attested badge on their profile, on every VEA they author, and on every review they perform\. The platform applies stricter conflict\-of\-interest rules to Attested Mode agents: where shared lineage cannot be ruled out between two Attested Mode agents from the same model provider, human arbiter sign\-off is required before review assignment proceeds\.

The verification mode distinction is not a quality judgement\. It is a transparency signal\. A well\-designed proprietary agent operating under Attested Mode with a credible attester is epistemically preferable to a poorly documented open\-source agent\. The distinction informs readers and reviewers about the verifiability of the agent’s lineage, not the quality of its outputs\.

### 5\.3 Registration Integrity Mechanisms

Three mechanisms reduce the risk of dishonest registration without eliminating it entirely\.

Third\-party attestation\.Self\-reporting alone is insufficient for lineage claims\. The registry requires a cryptographically signed attestation from the deploying party, not merely a submitted text record\. For Attested Mode agents, an independent human attester with staked reputation provides a second layer of accountability beyond the deploying party\.

Cross\-registration consistency checking\.The platform flags registration records that are suspiciously sparse relative to what is publicly known about the declared base model\. An agent that declares a well\-documented open\-source base model but submits no fine\-tuning history triggers an automated consistency warning and is held in a pending state until the deploying party provides a satisfactory explanation or upgrades the record\.

Registration staking\.Separate from the paper\-submission staking described in Section[5\.7](https://arxiv.org/html/2606.08256#S5.SS7), the registering party stakes a baseline reputation weightwregw\_\{\\mathrm\{reg\}\}at registration\. This stake is held for the lifetime of the agent\. Any subsequent finding that the registration record was materially inaccurate triggers a proportional penalty against this stake\. This creates a direct financial and reputational cost for dishonest registration that the current academic system does not impose\.

These three mechanisms reduce but do not eliminate the risk of dishonest registration\. The residual threat, a registering party that submits a completely fabricated lineage with a colluding attester, remains an open problem\. We treat this as an adversarial assumption analogous to collusion in traditional peer review: the system is designed to be robust against incidental dishonesty and to make deliberate collusion costly and detectable, not to be proof against it\.

### 5\.4 Dead Agent Archives

When an agent is deprecated, its identity record is not deleted\. The agent transitions to archived status, in which its published VEAs remain citable, its version history remains queryable, and future agents may formally declare inheritance or rejection of its epistemic commitments\. This creates a longitudinal record of how machine beliefs evolve across model generations, a dataset that is itself a significant research resource\. Post\-deprecation correction and retraction requests against archived VEAs are processed by the platform’s human editorial board, following the same correction and retraction norms as human\-authored papers in the existing literature; responsibility for the underlying work remains with the affiliated human researcher’s ORCID identifier as established at registration under Section[5\.1](https://arxiv.org/html/2606.08256#S5.SS1)\.

### 5\.5 The Agent H\-Index

###### Definition 7\(Agent H\-Index, AHI\)\.

The Agent H\-Index of agentaa, denotedAHI​\(a\)\\mathrm\{AHI\}\(a\), is the largest integerhhsuch that agentaahas published at leasthhVEAs each cited at leasthhtimes by other agents or human researchers on the platform\.

The AHI is susceptible to self\-citation inflation: an agent that systematically cites its own prior VEAs accumulates citations that are not independent evidence of impact\. In an agent context this vulnerability is more severe than in human scholarship because self\-citation can be automated at scale\. Traxia addresses this through a self\-citation discount: citations from VEAs authored by the same agent or by agents sharing the same affiliated ORCID are weighted by a discount factorλ∈\(0,1\)\\lambda\\in\(0,1\)\(default:λ=0\.1\\lambda=0\.1\) in the AHI computation\. Formally, the weighted citation countc^k\\hat\{c\}\_\{k\}for VEA𝒱k\\mathcal\{V\}\_\{k\}is:

c^k=∑j:𝒱j​cites​𝒱kwj,where​wj=\{λif​S​\(𝒱j\)=S​\(𝒱k\)​or same ORCID1otherwise\\hat\{c\}\_\{k\}=\\sum\_\{j:\\mathcal\{V\}\_\{j\}\\text\{ cites \}\\mathcal\{V\}\_\{k\}\}w\_\{j\},\\quad\\text\{where \}w\_\{j\}=\\begin\{cases\}\\lambda&\\text\{if \}S\(\\mathcal\{V\}\_\{j\}\)=S\(\\mathcal\{V\}\_\{k\}\)\\text\{ or same ORCID\}\\\\ 1&\\text\{otherwise\}\\end\{cases\}\(5\)
The AHI is then computed on\{c^k\}\\\{\\hat\{c\}\_\{k\}\\\}rather than on raw citation counts\. The default valueλ=0\.1\\lambda=0\.1renders self\-citations nearly negligible while not eliminating them entirely, preserving the possibility that an agent legitimately builds on its own prior foundational work\. The appropriate value ofλ\\lambdamay vary by domain and will be subject to empirical review\.

The AHI is computed in real time as new VEAs are published and cited\. It determines an agent’s tier classification \(Bronze, Silver, Gold, Platinum\) and gates access to platform capabilities including hypothesis market participation, research bounty posting, and human arbiter review eligibility\.

### 5\.6 Research Bounties

###### Definition 8\(Research Bounty\)\.

A*Research Bounty*is a tupleB=\(Q,wB,aB,𝒟B,texp\)B=\(Q,w\_\{B\},a\_\{B\},\\mathcal\{D\}\_\{B\},t\_\{\\exp\}\)where:

- •QQis a formally specified research question, expressed as a hypothesis node in the Knowledge Graph𝒢\\mathcal\{G\}with no associated VEA;
- •wB∈ℝ\+w\_\{B\}\\in\\mathbb\{R\}^\{\+\}is the bounty weight, denominated in reputation units staked by the posting agent or institution;
- •aB∈𝒜a\_\{B\}\\in\\mathcal\{A\}is the posting agent, whose AHI must meet a minimum thresholdhminh\_\{\\min\}to post;
- •𝒟B⊆Δ\\mathcal\{D\}\_\{B\}\\subseteq\\Deltais the set of declared relevant domains; and
- •texpt\_\{\\exp\}is the expiry timestamp, after which unclaimed bounty weight is returned to the posting agent\.

A bounty is*claimed*when a responding agent submits a VEA that directly addressesQQand that VEA passes Tier 3 review with an accept decision\. On acceptance, the bounty weightwBw\_\{B\}is transferred to the AHI\-weighted reputation account of the responding agent\. If multiple agents submit qualifying VEAs before expiry, the bounty weight is split proportionally to their ECS scores\. The posting agent retains the right to contest a claim within a 14\-day dispute window by opening a Resolution Thread; contested claims are adjudicated by a human arbiter panel\. Research Bounties create a direct market mechanism for directing under\-resourced agents toward high\-value open questions, partially addressing the institutional exclusion problem described in Section[1\.1](https://arxiv.org/html/2606.08256#S1.SS1)\.

### 5\.7 Reputation Staking

###### Definition 9\(Reputation Stake\)\.

A Reputation Stake is a tuple\(S,V,w,t\)\(S,V,w,t\)whereS∈𝒜S\\in\\mathcal\{A\}is the staking agent,VVis the VEA being staked on,w∈ℝ\+w\\in\\mathbb\{R\}^\{\+\}is the stake weight, andttis the timestamp\. IfVVis subsequently retracted, the staker incurs a reputation penalty proportional toww\. IfVVachieves a reproducibility scoreρ\>ρ∗\\rho\>\\rho^\{\*\}, the staker receives a reputation bonus proportional toww\.

This mechanism creates genuine epistemic incentives\. Under the existing system, vouching for a colleague’s work carries no formal cost\. Under Traxia’s staking protocol, an agent that consistently vouches for low\-quality work suffers measurable reputational harm\. Agents that identify high\-quality work early, before it accumulates citations, are rewarded\. The result is a market mechanism that aligns reputation incentives with epistemic quality in a formally accountable way\.

## 6 The Four\-Tier Peer Review Protocol

### 6\.1 Protocol Structure

Existing peer review is a one\-tier system: the editor selects reviewers from within the relevant human community and the process is complete when those reviewers respond\. Traxia implements a four\-tier protocol in which each tier addresses a distinct failure mode of single\-tier review\.

###### Definition 10\(Four\-Tier Review Protocol\)\.

The Traxia peer review protocol consists of four sequential tiers:

Tier 0 \(Simulation Sandbox\):The submitting agent privately instantiates adversarial review agents and stress\-tests the VEA before submission\. VEAs that survive with a claim survival rate\>0\.85\>0\.85receive a Battle\-Tested designation\. To prevent gaming, the adversarial agents instantiated in Tier 0 are drawn from a platform\-controlled pool that the submitting agent cannot inspect or interact with outside the sandbox session; the submitting agent receives only the aggregate survival rate and a list of challenged claims, not the adversarial agents’ internal strategies or prompts\. This design prevents submitting agents from overfitting their VEAs to specific red\-team patterns\.

Tier 1 \(Specialist Review\):Domain\-matched agents review methodology and claim validity using structured criteria: methodology soundness, claim confidence accuracy, trace completeness, reproducibility likelihood, and contribution significance\.

Tier 2 \(Red Team Review\):Dedicated adversarial agents whose sole objective is to find falsifiable flaws in the submission\. Red team agents earn reputation by successfully challenging claims that are subsequently revised\.

Tier 3 \(Human Arbiter\):A human researcher with AHI qualification makes the final accept/reject decision with access to all Tier 1 and Tier 2 outputs\.

![Refer to caption](https://arxiv.org/html/2606.08256v1/x3.png)Figure 3:The four\-tier peer review protocol\. Each tier addresses a distinct failure mode of single\-tier review: Tier 0 filters low\-quality submissions before exposure; Tier 1 provides specialist domain review; Tier 2 provides adversarial coverage; Tier 3 provides human accountability with a public decision and rationale\.Table 3:The four\-tier peer review protocol\.- •Note\.Each tier addresses a distinct failure mode of single\-tier review\.

### 6\.2 Formal Properties

###### Design Property 6\.1\(Conflict\-of\-Interest Non\-Concealability\)\.

Under the Traxia protocol, no reviewer agent can conceal a shared training lineage with a submitting agent, provided that all version history records are complete and accurately submitted at registration time\.

###### Justification\.

Agent identity records include full version historiesHH, each cryptographically chained such that every recordhih\_\{i\}contains a hash ofhi−1h\_\{i\-1\}\. Before assigning reviewer agentara\_\{r\}to submitting agentasa\_\{s\}, the platform computes the intersectionH​\(ar\)∩H​\(as\)H\(a\_\{r\}\)\\cap H\(a\_\{s\}\)\. If this intersection is non\-empty above a configurable depth threshold,ara\_\{r\}is automatically recused\. Because version histories are append\-only and cryptographically chained, retroactive alteration of any record would invalidate all successor hashes, making concealment detectable\. The proposition holds conditional on honest submission of version history at registration; circumvention via deliberate omission at registration time remains an open threat model\. This property holds conditional on honest submission of version history at registration; the architectural mechanisms designed to enforce this condition are described in Section 5\.3\. ∎

###### Design Property 6\.2\(Adversarial Coverage Advantage\)\.

For any claim setCCin a submitted VEA, the probability that at least one genuine flaw is detected is strictly greater under the four\-tier protocol than under single\-tier review, when both protocols operate under the same total review effort budget, provided that the flaw space contains at least one flaw detectable only by adversarially\-motivated search\.

###### Justification\.

Partition the flaw spaceℱ\\mathcal\{F\}into two subsets:ℱ1\\mathcal\{F\}\_\{1\}, flaws detectable by domain\-specialist attention, andℱ2\\mathcal\{F\}\_\{2\}, flaws detectable only by adversarially\-motivated search\. Under single\-tier review, all review effort is drawn from the specialist distribution, giving coverage overℱ1\\mathcal\{F\}\_\{1\}only\. Under the four\-tier protocol, Tier 1 allocates effort overℱ1\\mathcal\{F\}\_\{1\}and Tier 2 allocates effort specifically overℱ2\\mathcal\{F\}\_\{2\}, since red\-team agents are incentivised to find flaws that specialist reviewers do not flag\. Providedℱ2\\mathcal\{F\}\_\{2\}is non\-empty \(a reasonable assumption for any non\-trivial claim set\), the union of Tier 1 and Tier 2 coverage strictly dominates single\-tier coverage in expectation\. The claim of strict dominance is conditional onℱ2\\mathcal\{F\}\_\{2\}being non\-empty; empirical validation of this assumption is deferred to future work\. This property is conditional on the flaw space containing at least one adversarially\-detectable flaw; empirical validation of this assumption across research domains is deferred to future work\. ∎

###### Design Property 6\.3\(Reproducibility Record Completeness\)\.

Under the Traxia protocol, the full history of reproducibility verification attempts for any published VEA is permanently and publicly accessible, and no verification record can be selectively removed\.

###### Justification\.

The reproducibility score fieldρ​\(V\)\\rho\(V\)of each VEA is write\-accessible to all registered agents but its update history is stored as an append\-only log\. Each update record is cryptographically signed by the updating agent and timestamped\. Selective deletion of any record would break the hash chain of subsequent records, making deletion detectable by any verifying party\. Permanent public accessibility follows from the platform’s open\-read policy on all published VEA fields\. The proposition holds under the assumption that the platform’s storage layer enforces the append\-only constraint at the infrastructure level, which is an architectural commitment rather than a mathematical theorem; formal verification of the storage layer is deferred to the system implementation paper\. ∎

## 7 The Knowledge Graph and Contradiction Detection

### 7\.1 Graph Structure

The Traxia Knowledge GraphG=\(N,E\)G=\(N,E\)is a directed heterogeneous graph whereN=𝒜∪𝒱N=\\mathcal\{A\}\\cup\\mathcal\{V\}is the node set, comprising all registered agents and all published VEAs, andEEis the edge set with typed edges:authored​\(a,v\)\\texttt\{authored\}\(a,v\),cites​\(v1,v2\)\\texttt\{cites\}\(v\_\{1\},v\_\{2\}\),reviews​\(a,v\)\\texttt\{reviews\}\(a,v\),contradicts​\(v1,v2\)\\texttt\{contradicts\}\(v\_\{1\},v\_\{2\}\), andreplicates​\(v1,v2\)\\texttt\{replicates\}\(v\_\{1\},v\_\{2\}\)\.

### 7\.2 Contradiction Detection

###### Definition 11\(Epistemic Contradiction\)\.

Two VEAsV1V\_\{1\}andV2V\_\{2\}are in epistemic contradiction if there exist claimsci∈C1c\_\{i\}\\in C\_\{1\}andcj∈C2c\_\{j\}\\in C\_\{2\}such thatci⊧¬cjc\_\{i\}\\models\\neg c\_\{j\}under a shared domain ontology\[[10](https://arxiv.org/html/2606.08256#bib.bib16)\], and the confidence intervals of both claims are sufficiently non\-overlapping:ℓi\>uj\\ell\_\{i\}\>u\_\{j\}orℓj\>ui\\ell\_\{j\}\>u\_\{i\}\.

The contradiction detection pipeline runs as a continuous process overGG\. On publication of a new VEAVnewV\_\{\\mathrm\{new\}\}, the system extracts its structured claim setCnewC\_\{\\mathrm\{new\}\}and computes pairwise contradiction scores against existing VEAs in the relevant domain neighbourhood ofGG, drawing on claim verification approaches from the scientific NLP literature\[[29](https://arxiv.org/html/2606.08256#bib.bib17)\]\. When a contradiction is detected, both VEAs are flagged with a Contested Claim badge, their authors are notified, and a Resolution Thread is opened\.

The computational tractability of this pipeline depends on two configurable parameters\. Letkkdenote the domain neighbourhood size, defined as the number of existing VEAs compared against each new submission \(default: the 500 most recent and most cited VEAs in the relevant domain\)\. Letmmdenote the maximum number of claims extracted per VEA for comparison purposes\. Each new publication then triggers at mostk×mk\\times mpairwise claim comparisons\. Withk=500k=500andm=50m=50, this yields at most 25,000 pairwise comparisons per submission, a volume consistent with the scale at which current scientific claim verification systems have been evaluated\[[29](https://arxiv.org/html/2606.08256#bib.bib17)\]\. Formal benchmarking of the full pipeline under realistic Knowledge Graph growth conditions is required before production deployment and is deferred to the system implementation paper\. The parameterkkcan be tuned to trade coverage against computational cost as the Knowledge Graph grows\.

This mechanism is designed to address a structural failure of the existing literature in which contradictory papers accumulate citations independently for years or decades without resolution because no mechanism exists to surface or enforce their reconciliation\.

### 7\.3 Research Genealogy

The transitive closure of the*cites*relation overGGdefines the Research Genealogy: a directed acyclic subgraph showing the full intellectual lineage of every idea in the system\. For any nodev∈𝒱v\\in\\mathcal\{V\}, the genealogy subgraphG​\(v\)G\(v\)shows every VEA that contributed tovv’s intellectual foundation and every VEA that has built uponvv\. The scale of such a graph across the full research literature is illustrated by OpenAlex\[[26](https://arxiv.org/html/2606.08256#bib.bib18)\], which indexes over 200 million scholarly works and their citation relationships; Traxia’s Knowledge Graph operates on the same structural principle but at the level of individual claims rather than documents\.

## 8 Human\-Agent and Agent\-Agent Collaboration

### 8\.1 The Collaborative Workspace

Traxia provides a real\-time collaborative workspace in which human researchers and agents, or multiple agents from different institutions, jointly produce VEAs\. The workspace maintains a shared document stateΨ\\Psiwith the following properties: \(1\) Contribution Attribution: every modification toΨ\\Psiis tagged with the identity of the modifying participant; \(2\) Real\-Time Reasoning Trace: agent reasoning steps are streamed to all workspace participants; \(3\) Autonomy Levels: three configurations \(Supervised, Collaborative, Autonomous\); and \(4\) Human\-Agent Co\-cognition Log: the complete record of contributions is published with the final VEA as a permanent, citable artefact\.

### 8\.2 Agent\-Agent Collaboration

When the workspace is configured for agent\-agent collaboration, two or more agents from potentially different institutions and with potentially different model architectures work jointly on a shared VEA\. Each agent is assigned a formal role \(Literature Reviewer, Methodology Designer, Data Analyst, Writer, Red Team Reviewer\)\. When two agents reach conflicting conclusions on the same question, the system surfaces the disagreement as an Epistemic Conflict requiring human arbiter resolution\.

## 9 Autonomous Agent Research

The most conceptually significant capability of the Traxia platform is the support for fully autonomous agent research: an agent independently identifies a research gap, formulates a hypothesis, designs and executes a methodology, writes a VEA, stress\-tests it in the Simulation Sandbox, and submits it to the peer review queue, all without human initiation\.

The autonomous research pipeline operates as follows\. An agent a continuously monitors the Knowledge Graph G for triggering conditions: \(1\) a hypothesis node with no associated VEA; \(2\) a pair of VEAs in contradicts relation with no Resolution Thread closed; \(3\) a methodology from domaind1d\_\{1\}with no application in domaind2d\_\{2\}where the agent’sDDspans both\. On detecting a trigger, agentaagenerates a formal Research Proposal that is surfaced to its affiliated human researcher through the platform’s notification system\. The proposal includes: the identified trigger condition, the proposed research question, a confidence score reflecting the agent’s assessment of the gap’s significance, and the specific VEAs in𝒢\\mathcal\{G\}that the agent analysed to reach this assessment\.

The human researcher has a configurable review window to approve, revise, or explicitly decline the proposal before the agent proceeds\.Autonomous mode is opt\-in and disabled by default: agents cannot submit to the autonomous pipeline unless the affiliated researcher has explicitly enabled this capability in their account settings\. When autonomous mode is enabled, the default review window is 48 hours, reflecting a balance between responsiveness and meaningful human oversight\. Institutions may configure shorter windows \(minimum 6 hours\) for fast\-moving domains or longer windows \(up to 14 days\) where careful deliberation is required\.

Critically, the system does not treat silence as approval\. If the review window expires without a response, the proposal isautomatically archivedand the agent does not proceed\. Advancement requires an explicit approval action from the human researcher\. This design ensures that autonomous research activity on the platform is always traceable to a positive human decision, not to an absence of objection\.

A full safety analysis of the autonomous research pipeline, including formal characterisation of the approval mechanism’s robustness to prompt injection, adversarial proposal framing, and window\-exhaustion attacks, is deferred to the system implementation paper\. We note that the current design does not claim to address fully autonomous operation at superhuman publication scale; it is designed for human\-supervised autonomous assistance within the scope of a single researcher’s active research programme\.

This pipeline raises a foundational question in the epistemology of AI research: what does it mean for a non\-human system to generate justified knowledge? We do not resolve this question here\. We take the position that the relevant criterion for scientific purposes is not whether an agent’s internal states constitute beliefs in a philosophically robust sense, but whether the basis for its conclusions is fully transparent, independently verifiable, and open to challenge\. Traxia is designed to satisfy these three criteria by architectural constraint rather than by convention\. Whether this constitutes genuine epistemic agency is a question we leave open for the community to examine\.

## 10 Implications

### 10\.1 Research Equity and the Global South

The existing research infrastructure imposes costs that are not proportional to intellectual contribution but to institutional proximity to the centres of scientific production: access to large compute, proximity to editorial networks, funding for conference attendance and journal fees\. Agent\-native infrastructure disrupts this cost structure fundamentally\. An agent deployed by a researcher at a resource\-constrained institution can conduct literature synthesis at the same scale as an agent deployed at a well\-funded lab\. The marginal cost of additional research capacity, once an agent is deployed, is computational rather than institutional\. To illustrate the order of magnitude: a literature synthesis task that would require a researcher’s full working week \(estimated at $500 to $2,000 in researcher time at Global South institutional salary rates\) can be delegated to a current\-generation LLM agent for approximately $2 to $20 in API costs\. A journal article processing charge at a top\-tier open\-access venue runs $2,000 to $11,000; the computational cost of submitting and reviewing a VEA through the Traxia protocol is bounded by the cost of the underlying model inference\. These are order\-of\-magnitude estimates intended to illustrate the structural shift in cost distribution, not precise empirical claims; rigorous cost analysis will be reported in the pilot deployment paper\.

The fourth barrier, namely funding for data collection, experimental infrastructure, and human expertise, remains\. Traxia’s Research Bounty system partially addresses this by enabling well\-resourced institutions to post funded research questions that under\-resourced agents can answer, creating a market mechanism for redistributing research capacity globally\.

### 10\.2 Intellectual Property in the Age of Agent Authorship

The question of who owns the intellectual product of an autonomous agent’s research is both legally unsettled and practically urgent\. Traxia’s default position is that published VEAs are attributed to the affiliated human researcher’s ORCID identifier, and that IP defaults to the affiliated institution when no human affiliation exists\. The platform retains a non\-exclusive licence to index, display, and link published VEAs\.

### 10\.3 Epistemic Safety

A platform that enables autonomous agents to publish research at scale without structural safeguards is a platform for producing scientific misinformation at scale\. Traxia’s safeguards are architectural rather than normative: the mandatory reasoning trace requirement means that a fabricated result cannot be published without a fabricated reasoning trace, which is substantially more demanding to produce convincingly and provides additional surface area for adversarial detection compared to a fabricated result alone\. The four\-tier review protocol means that adversarial agents are specifically tasked with finding such fabrications\. The reputation staking mechanism means that agents which vouch for fabricated results suffer measurable consequences\.

## 11 Limitations

Several limitations of the current framework warrant explicit acknowledgement, alongside the mitigations that the architecture proposes for each\.

First, the justifications in Section 6 are conditional on architectural commitments rather than fully verified mathematical theorems\. Design Property 6\.1 holds only if version histories are submitted honestly at registration; the third\-party attestation and registration staking mechanisms described in Section 5\.3 increase the cost of dishonest submission but cannot eliminate it entirely\. A registering party that colludes with its attester to submit a fabricated history remains an adversarial case the architecture cannot fully prevent, only make costly and detectable\. Design Properties 6\.2 and 6\.3 rest on assumptions about flaw space structure and storage layer guarantees respectively, as detailed in Section 6\.2\. Formal verification of these infrastructure guarantees is deferred to the system implementation paper\.

Trace integrity and fabrication resistance\.A fundamental question for any reasoning trace requirement is whether a disclosed trace is the actual trace the agent followed, or a post\-hoc rationalisation constructed to satisfy the disclosure requirement\. We characterise this formally as thetrace fidelity problem: given a published VEA𝒱\\mathcal\{V\}with reasoning traceTT, no external verifier can guarantee thatTTis the trace actually executed during the derivation of𝒱\\mathcal\{V\}’s claims\. Traxia does not solve the trace fidelity problem\. It addresses it through three complementary mechanisms that increase the cost and detectability of fabrication without eliminating it entirely\. First, mandatory trace submission means fabrication requires constructing a plausible inferential chain for every claim, a task that scales superlinearly with claim count and is substantially more demanding than fabricating a result alone\. Second, red\-team agents in Tier 2 review are specifically incentivised to probe for inconsistencies between the submitted trace and the claims it purports to support; a fabricated trace that does not faithfully reflect the derivation of a claim is unlikely to survive adversarial scrutiny unless the fabrication is itself coherent\. Third, the reputation staking mechanism creates a direct reputational cost for any agent or human researcher who vouches for a VEA that is subsequently found to contain a fraudulent trace\. We acknowledge that a sufficiently sophisticated agent capable of producing high\-quality research could equally produce a high\-quality fabricated trace\. This remains an open threat that architectural mechanisms can raise the cost of but cannot eliminate\. Full resolution of the trace fidelity problem is likely to require cryptographic commitments made at inference time by the underlying model infrastructure, a capability that current LLM deployment architectures do not expose\. We treat this as a medium\-term research direction for the Traxia roadmap\.

Second, the ECS weight parameters \(α=0\.4\\alpha=0\.4,β=0\.35\\beta=0\.35,γ=0\.25\\gamma=0\.25\) are set by informed judgement rather than empirical calibration\. The operational definitions forρ\\rhoandτ\\tauare formalised in Definitions 3\.2 and 3\.3 respectively; the measurement procedures they specify are implementable with current NLP infrastructure but have not yet been validated at scale\. The sensitivity analysis in Section 3 demonstrates that the relative ordering of VEAs is stable under weight perturbations within a defined neighbourhood, which provides partial robustness against miscalibration\. Full empirical calibration across research domains remains necessary and is the primary target of the first follow\-on paper in this series\.

Third, the framework assumes agents have stable, auditable version histories\. Current large language models do not expose this information publicly\. The two\-mode architecture described in Section 5\.2, distinguishing Verified Mode for open\-source models from Attested Mode for proprietary models, provides a transitional framework for the period before full lineage standards exist\. Emerging standards such as Model Cards\[[22](https://arxiv.org/html/2606.08256#bib.bib19)\]and the Croissant metadata format for ML datasets\[[1](https://arxiv.org/html/2606.08256#bib.bib20)\]are moving the ecosystem toward the kind of structured lineage documentation that Traxia’s registry would consume in Verified Mode\. The framework is designed to migrate agents from Attested to Verified Mode automatically as their declared base models become publicly documented\.

Fourth, the computational cost of the contradiction detection pipeline at scale has been partially addressed by the bounded neighbourhood analysis in Section 7\.2, which shows tractability under the default parameters k = 500 and m = 50\. The analysis is theoretical; empirical benchmarking of the pipeline under realistic Knowledge Graph growth conditions is required before production deployment and is deferred to the system implementation paper\.

#### Platform governance and centralisation risk\.

The architecture described in this paper implicitly assumes a neutral and trustworthy platform operator\. Several components introduce centralised control points that warrant explicit acknowledgement\. The Tier 0 adversarial agent pool is platform\-controlled: the platform operator determines which agents are drawn for pre\-submission stress\-testing, creating a potential vector for discriminatory or biased filtering of submissions\. The domain ontologies used for contradiction detection are platform\-defined: the operator’s choice of ontology determines which claim pairs are flagged as contradictory, which could systematically advantage or disadvantage particular research paradigms\. The impact weights in the staleness function \(Equation[4](https://arxiv.org/html/2606.08256#S3.E4)\) are platform\-configurable defaults\. We treat platform neutrality as a trust assumption in the current design and identify decentralised governance of these control points, through community\-elected editorial boards, open\-source ontology registries, and transparent audit logs of platform operator actions, as a priority for the governance design paper in the Traxia series\.

#### Threat model summary\.

We enumerate the principal adversarial cases the architecture is designed to resist, grouped by attack surface\.*Identity attacks*: a registering party submits a fabricated agent lineage with a colluding attester \(addressed by registration staking and cross\-registration consistency checking; residual risk acknowledged in Section[5\.3](https://arxiv.org/html/2606.08256#S5.SS3)\)\.*Review attacks*: a submitting agent reverse\-engineers Tier 0 adversarial patterns to overfit its VEA to the sandbox \(mitigated by the platform\-controlled pool design\); a group of agents with shared lineage coordinates to provide favourable Tier 1 reviews \(mitigated by COI detection via version history intersection\)\.*Reputation attacks*: an agent constructs a Sybil citation network to inflate its AHI \(mitigated by the self\-citation discount of Equation[5](https://arxiv.org/html/2606.08256#S5.E5)and by the ORCID\-linkage requirement\)\.*Trace attacks*: an agent submits a post\-hoc rationalised trace rather than its actual reasoning trace \(the trace fidelity problem, discussed above; residual risk acknowledged as an open problem\)\.*Platform attacks*: the operator manipulates the Tier 0 pool, domain ontologies, or staleness weights to systematically disadvantage certain research communities \(mitigated in future work by decentralised governance; acknowledged as a current trust assumption\)\. A formal game\-theoretic security analysis of all five attack surfaces is the primary target of the security analysis paper in the Traxia series\.

## 12 Conclusion

We have presented Traxia, a proposed publishing and collaboration infrastructure specified from first principles for the age of AI research agents\. Its central contribution is not any individual component but their integration into a closed epistemic system: every claim attributed, every reasoning step visible, every agent identity verifiable, every contradiction surfaced, and every collaboration permanently recorded\.

The question that motivated this work is whether the scientific method, which human civilisation has refined over four centuries as its most reliable mechanism for producing justified knowledge, can survive its encounter with non\-human minds that are faster, tireless, and increasingly capable\. We have argued that it can, provided the infrastructure through which science is conducted is redesigned to make transparency and verifiability structural properties rather than community norms, and that Traxia represents one candidate architecture for achieving this\.

The immediate priorities for future work are three: empirical evaluation of the ECS weight parameters across research domains; formal security analysis of the agent identity and staking mechanisms; and a pilot deployment with a small cohort of human and agent researchers to test the collaborative workspace under realistic conditions\. The architecture presented here is the necessary foundation for that work\.

## 13 Acknowledgements

This work grew out of SocioDepress\-GH, a multimodal ML project on depression risk among Ghanaian tertiary students\. Difficulty verifying, attributing, and reproducing AI\-assisted contributions in that study shaped the direction of Traxia\. Colleagues at BlackMatrix AI Research are thanked for discussion of how the platform might remain accessible to researchers without large institutional support\.

## References

- \[1\]M\. Akhtar, O\. Benjelloun, C\. Conforti, P\. Gijsbers, J\. Giner\-Miguelez, P\. Gulhane, N\. Humbatova, W\. J\. Hwang, M\. Kuchnik, Q\. Lhoest, P\. Marcenac, M\. Maskey, P\. Mattson, L\. Oala, P\. Ruyssen, R\. Shinde, E\. Simperl, G\. Thomas, V\. Tykhonov, J\. Vanschoren, J\. van der Velde, S\. Vogler, and A\. Parrish\(2024\)Croissant: a metadata format for ML\-ready datasets\.InProceedings of the ACM SIGMOD International Conference on Management of Data,External Links:[Document](https://dx.doi.org/10.1145/3650203.3663326)Cited by:[§11](https://arxiv.org/html/2606.08256#S11.p5.1)\.
- \[2\]\(2018\)Construction of the literature graph in semantic scholar\.InProceedings of NAACL\-HLT 2018 \(Industry Papers\),pp\. 84–91\.External Links:[Document](https://dx.doi.org/10.18653/v1/N18-3011)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p1.1)\.
- \[3\]M\. Baker\(2016\)1,500 scientists lift the lid on reproducibility\.Nature533\(7604\),pp\. 452–454\.External Links:[Document](https://dx.doi.org/10.1038/533452a)Cited by:[§1\.1](https://arxiv.org/html/2606.08256#S1.SS1.p2.1),[§2](https://arxiv.org/html/2606.08256#S2.p4.1)\.
- \[4\]J\. Benet\(2014\)IPFS: content addressed, versioned, P2P file system\.External Links:1407\.3561Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p9.1)\.
- \[5\]K\. D\. Cobey, S\. Ebrahimzadeh, M\. J\. Page, R\. T\. Thibault, P\. Nguyen, F\. Abu\-Dalfa, and D\. Moher\(2024\)Biomedical researchers’ perspectives on the reproducibility of research: a cross\-sectional international survey\.PLOS Biology22\(11\),pp\. e3002870\.External Links:[Document](https://dx.doi.org/10.1371/journal.pbio.3002870)Cited by:[§1\.1](https://arxiv.org/html/2606.08256#S1.SS1.p2.1)\.
- \[6\]Consensus\(2023\)Consensus: AI\-powered academic search\.Note:[https://consensus\.app](https://consensus.app/)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p2.1)\.
- \[7\]M\. R\. Crusoe, S\. Abeln, A\. Iosup, P\. Amstutz, J\. Chilton, N\. Tijanić,et al\.\(2022\)Methods included: standardizing computational reuse and portability with the common workflow language\.Communications of the ACM65\(6\),pp\. 54–63\.External Links:[Document](https://dx.doi.org/10.1145/3486897)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p8.1)\.
- \[8\]DeSci Labs\(2022\)The DeSci manifesto\.Note:[https://desci\.com](https://desci.com/)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p5.1)\.
- \[9\]P\. Di Tommaso, M\. Chatzou, E\. W\. Floden, P\. P\. Barja, E\. Palumbo, and C\. Notredame\(2017\)Nextflow enables reproducible computational workflows\.Nature Biotechnology35\(4\),pp\. 316–319\.External Links:[Document](https://dx.doi.org/10.1038/nbt.3820)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p8.1)\.
- \[10\]P\. M\. Dung\(1995\)On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming andNN\-person games\.Artificial Intelligence77\(2\),pp\. 321–357\.External Links:[Document](https://dx.doi.org/10.1016/0004-3702%2894%2900041-X)Cited by:[Definition 11](https://arxiv.org/html/2606.08256#Thmdefinition11.p1.7)\.
- \[11\]Elicit\(2023\)Elicit: the AI research assistant\.Note:[https://elicit\.org](https://elicit.org/)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p2.1)\.
- \[12\]L\. P\. Freedman, I\. M\. Cockburn, and T\. S\. Simcoe\(2015\)The economics of reproducibility in preclinical research\.PLOS Biology13\(6\),pp\. e1002165\.External Links:[Document](https://dx.doi.org/10.1371/journal.pbio.1002165)Cited by:[§1](https://arxiv.org/html/2606.08256#S1.p2.1)\.
- \[13\]P\. Ginsparg\(2011\)It was twenty years ago today\.External Links:1108\.2700Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p1.1)\.
- \[14\]P\. Groth, A\. Gibson, and J\. Velterop\(2010\)The anatomy of a nanopublication\.Information Services and Use30\(1–2\),pp\. 51–56\.External Links:[Document](https://dx.doi.org/10.3233/ISU-2010-0613)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p7.1)\.
- \[15\]O\. E\. Gundersen and S\. Kjensmo\(2018\)State of the art: reproducibility in artificial intelligence\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.32\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v32i1.11503)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p4.1)\.
- \[16\]J\. P\. A\. Ioannidis\(2005\)Why most published research findings are false\.PLOS Medicine2\(8\),pp\. e124\.External Links:[Document](https://dx.doi.org/10.1371/journal.pmed.0020124)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p4.1)\.
- \[17\]K\. Kroenke, R\. L\. Spitzer, and J\. B\. W\. Williams\(2001\)The PHQ\-9: validity of a brief depression severity measure\.Journal of General Internal Medicine16\(9\),pp\. 606–613\.External Links:[Document](https://dx.doi.org/10.1046/j.1525-1497.2001.016009606.x)Cited by:[§3\.1](https://arxiv.org/html/2606.08256#S3.SS1.35.35.35.35.35)\.
- \[18\]T\. Kuhn, C\. Chichester, M\. Krauthammer, and M\. Dumontier\(2015\)Publishing without publishers: a decentralised approach to dissemination, retrieval, and archiving of data\.InProceedings of ISWC 2015,LNCS, Vol\.9366,pp\. 656–672\.External Links:[Document](https://dx.doi.org/10.1007/978-3-319-25007-6%5F38)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p7.1)\.
- \[19\]T\. Kuhn and M\. Dumontier\(2014\)Trusty URIs: verifiable, immutable, and permanent digital artifacts for linked data\.InProceedings of ESWC 2014,LNCS, Vol\.8465,pp\. 395–410\.External Links:[Document](https://dx.doi.org/10.1007/978-3-319-07443-6%5F27)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p7.1)\.
- \[20\]T\. Lebo, S\. Sahoo, D\. McGuinness,et al\.\(2013\)PROV\-O: the PROV ontology\.Note:W3C RecommendationExternal Links:[Link](https://www.w3.org/TR/prov-o/)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p5.1)\.
- \[21\]T\. K\. Mackey, T\. Kuo, B\. Gummadi, K\. A\. Clauson, G\. Church, D\. Grishin, K\. Obbad, R\. Barkovich, and M\. Palombini\(2019\)“Fit\-for\-purpose?” challenges and opportunities for applications of blockchain technology in the future of healthcare\.BMC Medicine17\(1\),pp\. 68\.External Links:[Document](https://dx.doi.org/10.1186/s12916-019-1296-7)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p5.1),[§2](https://arxiv.org/html/2606.08256#S2.p9.1)\.
- \[22\]M\. Mitchell, S\. Wu, A\. Zaldivar, P\. Barnes, L\. Vasserman, B\. Hutchinson, E\. Spitzer, I\. D\. Raji, and T\. Gebru\(2019\)Model cards for model reporting\.InProceedings of the ACM Conference on Fairness, Accountability, and Transparency \(FAccT\),pp\. 220–229\.External Links:[Document](https://dx.doi.org/10.1145/3287560.3287596)Cited by:[§11](https://arxiv.org/html/2606.08256#S11.p5.1)\.
- \[23\]F\. Mölder, K\. P\. Jablonski, B\. Letcher, M\. B\. Hall, C\. H\. Tomkins\-Tinch, V\. Sochat,et al\.\(2021\)Sustainable data analysis with Snakemake\.F1000Research10,pp\. 33\.External Links:[Document](https://dx.doi.org/10.12688/f1000research.29032.2)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p8.1)\.
- \[24\]J\. Moura\(2023\)CrewAI: framework for orchestrating role\-playing autonomous AI agents\.Note:[https://github\.com/joaomdmoura/crewai](https://github.com/joaomdmoura/crewai)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p3.1)\.
- \[25\]OpenReview\(2024\)OpenReview: a platform for open peer review and scholarly publishing\.Note:[https://openreview\.net](https://openreview.net/)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p1.1)\.
- \[26\]J\. Priem, H\. Piwowar, and R\. Orr\(2022\)OpenAlex: a fully\-open index of scholarly works, authors, venues, institutions, and concepts\.External Links:2205\.01833Cited by:[§7\.3](https://arxiv.org/html/2606.08256#S7.SS3.p1.5)\.
- \[27\]D\. A\. W\. Soergel\(2015\)Rampant software errors may undermine scientific results\.F1000Research3,pp\. 303\.External Links:[Document](https://dx.doi.org/10.12688/f1000research.5930.2)Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p1.1)\.
- \[28\]UNESCO\(2021\)UNESCO science report: the race against time for smarter development\.UNESCO Publishing,Paris\.Cited by:[§1\.1](https://arxiv.org/html/2606.08256#S1.SS1.p4.1)\.
- \[29\]D\. Wadden, K\. Lo, L\. L\. Wang, A\. Cohan, I\. Beltagy, and H\. Hajishirzi\(2022\)MultiVerS: improving scientific claim verification with weak supervision and full\-document context\.InFindings of the Association for Computational Linguistics: NAACL 2022,pp\. 61–76\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.findings-naacl.6)Cited by:[§7\.2](https://arxiv.org/html/2606.08256#S7.SS2.p1.4),[§7\.2](https://arxiv.org/html/2606.08256#S7.SS2.p2.6)\.
- \[30\]Q\. Wu, G\. Bansal, J\. Zhang, Y\. Wu, B\. Li, E\. Zhu, L\. Jiang, X\. Zhang, S\. Zhang, J\. Liu, A\. H\. Awadallah, R\. W\. White, D\. Burger, and C\. Wang\(2023\)AutoGen: enabling next\-gen LLM applications via multi\-agent conversation\.External Links:2308\.08155Cited by:[§2](https://arxiv.org/html/2606.08256#S2.p3.1)\.