EviGraph: Evidence-Guided Autonomous Research Agents

# EviGraph: Evidence-Guided Autonomous Research Agents
Source: [https://arxiv.org/html/2608.04738](https://arxiv.org/html/2608.04738)
Zhenjiang Ren1,2, Ruiji Li1, Xujing Zhang3, Ziliang Pang1, Shuo Ren1\\corresponding, Jiajun Zhang1,2,4\\corresponding

###### Abstract

Autonomous research agents can generate hypotheses, execute experiments, and draft manuscripts, yet their outputs often contain unsupported claims and inconsistencies between research questions, experiments, results, and conclusions\. We argue that this problem is partly architectural: existing systems organize research as sequential pipelines but do not explicitly maintain or validate the evolving claim–evidence structure across stages\. In this paper, we introduceEviGraph, an autonomous research framework that represents the research process as a typed evidence graph containingProblem,Gap,Hypothesis,Experiment,Finding, andClaimnodes\. The graph serves as the operational state of the agent rather than a post\-hoc record\. EviGraph inspects evidence chains for missing dependencies, semantic misalignment, and result–claim inconsistencies, localizes the earliest weak node, and regenerates its affected downstream subgraph\. Graph checkpointing prevents unsuccessful repairs from corrupting previously validated evidence\. Manuscripts are generated only after every retained claim is grounded in a validated evidence chain\. Experiments on ARC\-Bench\-ML and NanoResearch\-20 show that EviGraph outperforms the compared end\-to\-end research\-agent baselines in overall research performance, improves Claim Support Rate by 40\.19% over the strongest baseline, and achieves 87\.73% Experimental Data Consistency\. These results demonstrate the value of explicit evidence\-state maintenance for reliable autonomous research\.

## 1Introduction

Autonomous research systems built on large language models can now generate ideas, write code, run experiments, analyze results, and draft full manuscripts\(Luet al\.[2024](https://arxiv.org/html/2608.04738#bib.bib1); Schmidgallet al\.[2025](https://arxiv.org/html/2608.04738#bib.bib3); InternAgent Teamet al\.[2025](https://arxiv.org/html/2608.04738#bib.bib5); Branet al\.[2024](https://arxiv.org/html/2608.04738#bib.bib16); Jansenet al\.[2024](https://arxiv.org/html/2608.04738#bib.bib17)\)\. Despite this progress, producing a fluent research paper does not guarantee that the underlying process is reliable\. In practice, autonomous systems can propagate weak assumptions, unsupported claims, and experimental inconsistencies into the final manuscript, making it difficult to determine whether a conclusion is genuinely supported by the executed research\(Chenet al\.[2025](https://arxiv.org/html/2608.04738#bib.bib11); Yanget al\.[2026](https://arxiv.org/html/2608.04738#bib.bib7); Trehan and Chopra[2026](https://arxiv.org/html/2608.04738#bib.bib10); Minet al\.[2023](https://arxiv.org/html/2608.04738#bib.bib19)\)\.

We argue that this problem is partly architectural\. Most autonomous research systems are organized as sequential pipelines in which outputs move from idea generation to experimentation, analysis, and writing\(Luet al\.[2024](https://arxiv.org/html/2608.04738#bib.bib1); Liuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib6); Qianet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib8); Xuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib9)\)\. Such pipelines specify what stage should be executed next, but they do not explicitly maintain the evolving evidential relationships among research objects\. Consequently, a hypothesis can drift from its motivating research gap, an experiment can cease to test the current hypothesis after revision, or a manuscript claim can overstate what the recorded findings support\. Because these dependencies remain implicit, inconsistencies are often discovered only after they have propagated across several stages\. This motivates the problem ofmaintaining claim–evidence consistencythroughout the autonomous research lifecycle, requiring an explicit research\-state representation for both cross\-stage inspection and dependency\-aware revision\.

To address this problem, we proposeEviGraph, a graph\-driven autonomous research framework that treats a typed evidence graph as the central operational state of the research process\. The graph explicitly connects six types of research objects—Problem,Gap,Hypothesis,Experiment,Finding, andClaim—across the research lifecycle\.

Forcross\-stage inspection, EviGraph evaluates whether these objects form complete evidence paths and whether adjacent stages remain semantically aligned, including whether an experiment tests the current hypothesis, whether a finding is faithful to execution records, and whether a claim is supported by the resulting evidence\. Fordependency\-aware revision, once a weak node is identified, EviGraph traces its downstream dependencies and regenerates the affected subgraph in topological order\. Intermediate checkpoints allow the system to reject repairs that introduce new weaknesses or invalidate previously verified evidence chains\.

Manuscript generation is therefore gated by the validated research state rather than triggered automatically after pipeline completion\. EviGraph produces a paper only when every retained claim is grounded in a complete and validated evidence chain, and the graph supplies explicit provenance and scope constraints during drafting\. Additionally, EviGraph uses the Hypothesis Filter to screen out weak hypotheses during graph initialization and supports the retrieval of prior graph states to inform graph construction and repair on related tasks\.

We conduct experiments on ARC\-Bench\-ML\(Liuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib6)\)and NanoResearch\-20\(Xuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib9)\)\. EviGraph achieves an Overall score of 86\.45% on ARC\-Bench\-ML and obtains the strongest Novelty, Performance, and Writing scores among the compared systems on NanoResearch\-20\. It improves Claim Support Rate from 27% to 37\.85% over the strongest baseline while maintaining an Experimental Data Consistency of 87\.73%\. These results show that explicit evidence\-state maintenance improves the reliability of autonomous research without sacrificing overall performance\.

Our contributions are threefold:

- •Evidence graphs as operational research state\. We formulate autonomous research as the construction and validation of a typed graph connecting research problems, gaps, hypotheses, experiments, findings, and manuscript claims, making cross\-stage evidential dependencies explicit and inspectable\.
- •Dependency\-aware evidence inspection and repair\. We introduce a graph\-guided mechanism that detects weak evidence links, localizes their root causes, regenerates affected downstream subgraphs, and uses checkpointed rollback to prevent repair degradation\. Manuscript generation is gated by the resulting validated evidence state\.
- •Improvement of auto\-research reliability\. Experiments on ARC\-Bench\-ML and NanoResearch\-20 show that EviGraph substantially improves claim support and consistently outperforms the compared baseline systems in overall research performance\.

## 2Related Work

#### Autonomous research systems\.

LLM\-based research systems can generate ideas, implement methods, execute experiments, analyze results, and draft manuscripts\(Luet al\.[2024](https://arxiv.org/html/2608.04738#bib.bib1); Yamadaet al\.[2025](https://arxiv.org/html/2608.04738#bib.bib2); Schmidgallet al\.[2025](https://arxiv.org/html/2608.04738#bib.bib3); Schmidgall and Moor[2025](https://arxiv.org/html/2608.04738#bib.bib4); InternAgent Teamet al\.[2025](https://arxiv.org/html/2608.04738#bib.bib5); Skarlinskiet al\.[2024](https://arxiv.org/html/2608.04738#bib.bib18)\)\. Recent works improve reliability through strict result verification, adversarial evaluation, structured memory, or cross\-run self\-improvement\(Liuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib6); Yanget al\.[2026](https://arxiv.org/html/2608.04738#bib.bib7); Qianet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib8); Xuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib9)\)\. These approaches primarily organize research as a sequence of stages or accumulated trajectories\. EviGraph instead maintains a typed evidence graph as the operational research state and explicitly inspects and repairs the dependencies linking hypotheses, experiments, findings, and claims\.

#### Scientific knowledge graphs and claim provenance\.

Scientific knowledge graphs, micropublications, and nanopublications represent research objects, claims, evidence, methods, and provenance in structured forms\(Jaradehet al\.[2019](https://arxiv.org/html/2608.04738#bib.bib13); Clarket al\.[2014](https://arxiv.org/html/2608.04738#bib.bib12); Grothet al\.[2010](https://arxiv.org/html/2608.04738#bib.bib14)\)\. Formal argumentation additionally models support and conflict among claims\(Dung[1995](https://arxiv.org/html/2608.04738#bib.bib15)\)\. These works establish structured representations for scientific knowledge and evidence\. EviGraph builds on this perspective by turning the graph into an active control structure that is continuously updated during research and guides new scientific actions when inconsistencies are detected\(Baiet al\.[2024](https://arxiv.org/html/2608.04738#bib.bib20)\)\.

## 3Method

### 3\.1Framework Overview

![Refer to caption](https://arxiv.org/html/2608.04738v1/x1.png)Figure 1:Overall workflow of the EviGraph framework, organized into three collaborative phases\.As illustrated in Figure[1](https://arxiv.org/html/2608.04738#S3.F1), EviGraph takes a research goal \(q\) as input\. Initialization uses temporary structured records to buildG0G\_\{0\}; thereafter, graph\-facing components read and update this shared graph of research problems, gaps, hypotheses, experiments, findings, and claims\. The framework proceeds through three phases\.

#### Initial evidence\-graph construction\.

EviGraph first analyzes the research goal, retrieves relevant literature, and generates candidate hypotheses\. The Hypothesis Filter groups the candidates into research directions and uses pilot experiments to screen out weak directions before the retained hypotheses undergo full\-scale evaluation\. These research objects are assembled into a provisional graph \(G0G\_\{0\}\)\. This graph is intentionally incomplete: it provides an explicit research state that can be inspected and revised, rather than assuming that all stage outputs are already reliable\.

#### Cross\-stage evidence inspection and repair\.

The graph inspector examines the current graph for incomplete evidence paths, semantic misalignment between adjacent nodes in the graph\. Once a node is identified, EviGraph traces its downstream dependencies and regenerates the affected subgraph in dependency order\. Intermediate graph checkpoints preserve previously validated evidence and allow the framework to roll back repairs that introduce new weaknesses\. The updated graph is then inspected again, forming an iterative inspection\-and\-repair loop\.

#### Evidence\-gated manuscript generation\.

The loop terminates when the graph satisfies the evidence\-readiness condition, under which every retained claim is grounded in a complete and validated evidence chain\. EviGraph then converts the validated graph into a manuscript skeleton and generates the final paper using the recorded research context, experimental artifacts, and claim boundaries\. A final review checks the manuscript for consistency with the validated graph and triggers targeted textual revisions when necessary\.

### 3\.2Evidence Graph as Operational Research State

EviGraph represents the evolving research state as a typed directed graph\. After initialization, it is the shared state of all graph\-facing components and the object inspected to determine whether research should continue\. Unlike a linear pipeline trace, the graph can branch and merge \(e\.g\., one hypothesis may be evaluated by multiple experiments, while several findings may jointly support or qualify a claim\)\.

#### Research graph\.

An EviGraph research graph is a tupleG=\(V,E,τ,ρ,α\)G=\(V,E,\\tau,\\rho,\\alpha\), whereVVis a set of research\-object nodes,E⊆V×VE\\subseteq V\\times Vis a set of directed edges,τ:V→𝒯\\tau:V\\rightarrow\\mathcal\{T\}assigns a type to each node,ρ:E→ℛ\\rho:E\\rightarrow\\mathcal\{R\}assigns a relation type to each edge, andα:V→𝒜\\alpha:V\\rightarrow\\mathcal\{A\}stores type\-specific node attributes, where𝒜\\mathcal\{A\}denotes the space of structured attribute records\. For example, anExperimentnode records the datasets, implementation code, and evaluation metrics associated with the experiment\. The node\-type set is

𝒯=\{\\displaystyle\\mathcal\{T\}=\\\{Problem,Gap,Hypothesis,\\displaystyle\\textit\{Problem\},\\textit\{Gap\},\\textit\{Hypothesis\},Experiment,Finding,Claim\}\.\\displaystyle\\textit\{Experiment\},\\textit\{Finding\},\\textit\{Claim\}\\\}\.

#### Node semantics\.

Each node stores type\-specific content together with the structured attributes used during inspection\. The six node types areProblem, representing the task boundary;Gap, representing an unresolved limitation in prior work;Hypothesis, representing a testable proposal;Experiment, recording its protocol and implementation;Finding, recording an outcome derived from execution data; andClaim, representing a manuscript\-level assertion grounded in findings and carrying an explicit retention status\.

#### Edge semantics\.

The edge\-type setℛ\\mathcal\{R\}specifies how research objects depend on one another\. Table[1](https://arxiv.org/html/2608.04738#S3.T1)lists the permitted relation schemas\.

Table 1:Typed relations in the EviGraph research graph\.
#### Evidence chain\.

For a claimccwithτ​\(c\)=C​l​a​i​m\\tau\(c\)=Claim, an*evidence chain*is a directed support subgraph containing at least one typed claim\-support path

h→t​e​s​t​e​d−b​ye→p​r​o​d​u​c​e​sf→s​u​p​p​o​r​t​sc,h\\xrightarrow\{tested\-by\}e\\xrightarrow\{produces\}f\\xrightarrow\{supports\}c,whereτ​\(h\)=H​y​p​o​t​h​e​s​i​s\\tau\(h\)=Hypothesis,τ​\(e\)=E​x​p​e​r​i​m​e​n​t\\tau\(e\)=Experiment, andτ​\(f\)=F​i​n​d​i​n​g\\tau\(f\)=Finding\. The path can be traced to its research context byp→i​d​e​n​t​i​f​i​e​sg→m​o​t​i​v​a​t​e​shp\\xrightarrow\{identifies\}g\\xrightarrow\{motivates\}h, whereppandggare a Problem and Gap, respectively\. The support subgraph may contain multiple paths, for example, when several experiments or findings jointly supportcc\. It is*valid*when, for every edge\(u,v\)\(u,v\)in the support subgraph, the endpoint attributesα​\(u\)\\alpha\(u\)andα​\(v\)\\alpha\(v\)satisfy the semantic constraints of the edge’s relation typeρ​\(\(u,v\)\)\\rho\(\(u,v\)\)\.

AfterG0G\_\{0\}is constructed, all graph\-facing framework components read from and write to this shared graph\. A graph update may add a node or edge, revise a node attribute, or replace a dependent portion of the graph\. The graph therefore serves as both the operational research state and the interface through which the framework coordinates research actions\.

### 3\.3Cross\-Stage Evidence Inspection and Dependency\-Aware Repair

Given the current research graph, EviGraph repeatedly inspects its nodes and evidence chains, identifies unreliable research objects, and repairs the portions of the graph that depend on them\. It maintains a run\-local short\-term libraryℳS\\mathcal\{M\}\_\{S\}of intermediate graph versions for rollback and a persistent long\-term libraryℳL\\mathcal\{M\}\_\{L\}of evidence\-ready graphs and successful repair traces for cross\-run retrieval\.

#### Weak\-node inspection and repair\.

The graph inspector evaluates node quality and evidence\-chain validity\. A nodevvis considered*weak*if its attributesα​\(v\)\\alpha\(v\)and those of a connected nodeuufail to satisfy the semantic constraints specified by the relation typeρ​\(\(u,v\)\)\\rho\(\(u,v\)\)of the edge between them\. The inspector returns a set of repair groups𝒲=\{Sw\}\\mathcal\{W\}=\\\{S\_\{w\}\\\}\. Each repair groupSwS\_\{w\}contains a weak nodewwand its subordinate nodes, namely, downstream nodes whose contents depend on the current state ofww\. This graph\-based scope prevents the framework from treating an inconsistency as an isolated textual error: when an upstream research object changes, the dependent experiments, findings, and claims must be reconsidered accordingly\. Table[2](https://arxiv.org/html/2608.04738#S3.T2)summarizes the main weak\-node patterns and their corresponding repair actions\.

Table 2:Weak\-node patterns and repair actions used by graph inspection\.Algorithm[1](https://arxiv.org/html/2608.04738#alg1)summarizes the loop\. HereC,L,R,P,XC,L,R,P,Xdenote the task context, literature, retrieved experience, pilot results, and full\-scale results;ℋ,𝒟,ℋ⋆,𝒲,𝒯\\mathcal\{H\},\\mathcal\{D\},\\mathcal\{H\}^\{\\star\},\\mathcal\{W\},\\mathcal\{T\}denote candidate hypotheses, grouped directions, retained hypotheses, repair groups, and successful repair traces, respectively\. Statusssdistinguishes ready, repairable, reinitialization, and incomplete outcomes\.

Algorithm 1EviGraph research loop0:research goal

qq, execution budget

BB, long\-term graph library

ℳL\\mathcal\{M\}\_\{L\}
0:reviewed manuscript, or available state withIncomplete

\{Initialize the graph and experiments\}

C←AnalyzeTask​\(q\)C\\leftarrow\\textsc\{AnalyzeTask\}\(q\)

L←RetrieveLiterature​\(C\)L\\leftarrow\\textsc\{RetrieveLiterature\}\(C\)

R←RetrieveExperience​\(q,ℳL\)R\\leftarrow\\textsc\{RetrieveExperience\}\(q,\\mathcal\{M\}\_\{L\}\)

ℋ←GenerateHypotheses​\(C,L\)\\mathcal\{H\}\\leftarrow\\textsc\{GenerateHypotheses\}\(C,L\)

𝒟←GroupDirections​\(ℋ\)\\mathcal\{D\}\\leftarrow\\textsc\{GroupDirections\}\(\\mathcal\{H\}\)

P←RunPilotExperiments​\(𝒟\)P\\leftarrow\\textsc\{RunPilotExperiments\}\(\\mathcal\{D\}\)

ℋ⋆←SelectBestSupported​\(𝒟,P\)\\mathcal\{H\}^\{\\star\}\\leftarrow\\textsc\{SelectBestSupported\}\(\\mathcal\{D\},P\)

X←FullScaleEvaluate​\(ℋ⋆\)X\\leftarrow\\textsc\{FullScaleEvaluate\}\(\\mathcal\{H\}^\{\\star\}\)

G←BuildInitialGraph​\(C,L,ℋ⋆,X,R\)G\\leftarrow\\textsc\{BuildInitialGraph\}\(C,L,\\mathcal\{H\}^\{\\star\},X,R\)

\{Inspect and repair with checkpoints\}

ℳS←⟨G⟩\\mathcal\{M\}\_\{S\}\\leftarrow\\langle G\\rangle;

𝒯←⟨⟩\\mathcal\{T\}\\leftarrow\\langle\\rangle
\(s,𝒲\)←InspectGraph​\(G,B\)\(s,\\mathcal\{W\}\)\\leftarrow\\textsc\{InspectGraph\}\(G,B\)

while

s=Repairables=\\textsc\{Repairable\}and budget

BBremainsdo

Sw←SelectRepairGroup​\(𝒲\)S\_\{w\}\\leftarrow\\textsc\{SelectRepairGroup\}\(\\mathcal\{W\}\)

Gbase←GG\_\{\\mathrm\{base\}\}\\leftarrow G

𝒜w←PlanRepair​\(Gbase,Sw,R\)\\mathcal\{A\}\_\{w\}\\leftarrow\\textsc\{PlanRepair\}\(G\_\{\\mathrm\{base\}\},S\_\{w\},R\)

\(Gcand,ℳS,B,complete\)←ExecuteRepairGroup​\(𝒜w,ℳS,B\)\(G\_\{\\mathrm\{cand\}\},\\mathcal\{M\}\_\{S\},B,\\mathrm\{complete\}\)\\leftarrow\\textsc\{ExecuteRepairGroup\}\(\\mathcal\{A\}\_\{w\},\\mathcal\{M\}\_\{S\},B\)

if

complete\\mathrm\{complete\}then

d←Degrading​\(Gcand,Gbase\)d\\leftarrow\\textsc\{Degrading\}\(G\_\{\\mathrm\{cand\}\},G\_\{\\mathrm\{base\}\}\)

if

¬d\\neg dthen

G←GcandG\\leftarrow G\_\{\\mathrm\{cand\}\}

else

G←RollbackToBest​\(ℳS,Gbase\)G\\leftarrow\\textsc\{RollbackToBest\}\(\\mathcal\{M\}\_\{S\},G\_\{\\mathrm\{base\}\}\)

endif

else

G←RollbackToBest​\(ℳS,Gbase\)G\\leftarrow\\textsc\{RollbackToBest\}\(\\mathcal\{M\}\_\{S\},G\_\{\\mathrm\{base\}\}\)

endif

if

WeakCount​\(G\)<WeakCount​\(Gbase\)\\textsc\{WeakCount\}\(G\)<\\textsc\{WeakCount\}\(G\_\{\\mathrm\{base\}\}\)then

𝒯←Append​\(𝒯,\(Gbase,Sw,𝒜w,G\)\)\\mathcal\{T\}\\leftarrow\\textsc\{Append\}\(\\mathcal\{T\},\(G\_\{\\mathrm\{base\}\},S\_\{w\},\\mathcal\{A\}\_\{w\},G\)\)

endif

\(s,𝒲\)←InspectGraph​\(G,B\)\(s,\\mathcal\{W\}\)\\leftarrow\\textsc\{InspectGraph\}\(G,B\)

endwhile

\{Gate writing on evidence readiness\}

if

s=Reinitializes=\\textsc\{Reinitialize\}then return

ReinitializeAndResume​\(C,L,ℋ⋆,X,R,B\)\\textsc\{ReinitializeAndResume\}\(C,L,\\mathcal\{H\}^\{\\star\},X,R,B\)
if

s≠Readys\\neq\\textsc\{Ready\}then

return

\(G,Incomplete\)\(G,\\textsc\{Incomplete\}\)
endif

ℳL←ℳL∪\{\(G,𝒯\)\}\\mathcal\{M\}\_\{L\}\\leftarrow\\mathcal\{M\}\_\{L\}\\cup\\\{\(G,\\mathcal\{T\}\)\\\}

return

GenerateReviewAndRevise​\(G,B\)\\textsc\{GenerateReviewAndRevise\}\(G,B\)

#### Dependency\-aware subgraph repair\.

For each selected repair groupSwS\_\{w\}, the repair planner first produces an ordered action sequence from the current graph\. It then removes the subordinate nodesSw∖\{w\}S\_\{w\}\\setminus\\\{w\\\}whose contents depend on the weak node\. Specialized agents repairwwand regenerate the affected nodes in dependency order by updating existing nodes or adding new nodes to the graph\. The repaired graph is subsequently re\-inspected, and the process continues with the remaining weak nodes\.

#### Checkpointing and rollback\.

Each node\-level repair stores a new graph version inℳS\\mathcal\{M\}\_\{S\}\. After all repairs in a groupSwS\_\{w\}have been completed, the graph inspector compares the updated graph with the versionGbaseG\_\{\\mathrm\{base\}\}recorded before the group repair\. A repair is considered degrading if it increases the number of weak nodes or removes an evidence chain that was valid inGbaseG\_\{\\mathrm\{base\}\}\. In this case, EviGraph invokesRollbackToBest​\(ℳS,Gbase\)\\textsc\{RollbackToBest\}\(\\mathcal\{M\}\_\{S\},G\_\{\\mathrm\{base\}\}\)\. The rollback target is the historical version that minimizes the number of weak nodes while preserving all evidence chains that were already valid inGbaseG\_\{\\mathrm\{base\}\}\. If no intermediate version improves uponGbaseG\_\{\\mathrm\{base\}\}under these criteria, the framework returns toGbaseG\_\{\\mathrm\{base\}\}\. Otherwise, the best repaired version is retained and used for the next inspection step\.

#### Evidence readiness and manuscript generation\.

A graph is*evidence\-ready*, writtenReady​\(G\)\\textsc\{Ready\}\(G\), if it is schema\-valid, contains at least one retainedClaim, covers required empirical deliverables, and gives every retainedClaima valid evidence chain with no weak node\. The loop continues untilReady​\(G\)\\textsc\{Ready\}\(G\)holds or the budget expires\. The writer then expands validated nodes and paths into a manuscript, and a reviewer audits structure, novelty framing, citations, and graph faithfulness\. Writing weaknesses trigger targeted revisions\. A paper is released only after review and provenance validation succeed; failure to satisfy either graph or manuscript gate within budget returns the available state withIncomplete\.

### 3\.4Auxiliary Mechanisms

During initialization, theHypothesis Filtergroups candidate hypotheses into research directions based on semantic similarity and conducts small\-scale pilot experiments for each direction\. Based on the pilot results, it retains the best\-supported groupℋ⋆\\mathcal\{H\}^\{\\star\}for full\-scale evaluation, after which the surviving hypotheses are instantiated asHypothesisnodes in the provisional graph\.

For related tasks,ℳL\\mathcal\{M\}\_\{L\}retrieves graphs and repair traces using task similarity\. Retrieved structures may inform graph construction and repair planning, but their claims are not directly reused and must be validated again in the new graph\.

## 4Experiment

### 4\.1Experiment Setup

We evaluate EviGraph on two autonomous\-research benchmarks and compare its overall research performance and research reliability with existing end\-to\-end systems\.

#### Benchmarks\.

We conduct experiments on two benchmarks\.ARC\-Bench\-ML\(Liuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib6)\)contains 25 machine\-learning research topics, each specifying a research question, target dataset, and expected experimental deliverables, including implementation, results, and analysis\.NanoResearch\-20\(Xuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib9)\)contains 20 research tasks spanning seven domains: NLP, computer vision, multimodal learning, tabular learning, time series, graph learning, and audio\. It evaluates complete research workflows under multi\-round feedback from an LLM\-simulated scientist\.

#### Compared systems\.

We compare EviGraph with two end\-to\-end autonomous\-research systems\.AutoResearchClaw\(Liuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib6)\)uses a full\-stage research pipeline with strict output evaluation, and we use its full\-auto setting\.NanoResearch\(Xuet al\.[2026](https://arxiv.org/html/2608.04738#bib.bib9)\)performs multi\-round self\-improvement through cross\-run evolution and simulated\-scientist feedback\. Across both baselines and EviGraph, we use qwen\-3\.6\-plus as the LLM backbone and keep the sandboxed execution environment and per\-experiment time budget identical, reducing confounding from model capability and execution infrastructure\.

#### Metrics\.

Following the baselines’ official settings, ARC\-Bench\-ML reports Code Development \(CD\), Code Execution \(CE\), Result Analysis \(RA\), and their weighted Overall score, with CD:CE:RA=25:25:50=25:25:50\. CD evaluates implementation correctness; CE, successful execution and valid outputs; and RA, whether conclusions are grounded in measurements, hypotheses receive explicit verdicts, and limitations are properly reported\. Two independent agent reviewers apply the strict judge, with disagreements above0\.200\.20re\-adjudicated\. NanoResearch\-20 reports Alignment \(Align\.\), end\-to\-end completion \(E2E\), Performance \(Perf\.\), Novelty \(Novel\.\), and Writing quality \(Writ\.\), which respectively evaluate task compliance, workflow completion, task effectiveness, originality, and manuscript quality\.

We additionally report two reliability metrics\. We partition each manuscript into chunks, extract claims from each with an LLM, and aggregate them as𝒞\\mathcal\{C\}\. Claim Support Rate isCSR=\|𝒮\|/\|𝒞\|\\mathrm\{CSR\}=\|\\mathcal\{S\}\|/\|\\mathcal\{C\}\|, where𝒮\\mathcal\{S\}contains claims traceably supported by the corresponding research run\. Because extraction is stochastic,𝒞\\mathcal\{C\}may vary across runs and include implicit, broad, or difficult\-to\-ground claims, enlarging the denominator without increasing supported claims\. Absolute CSR is therefore conservative and may be lower than an estimate from a manually curated claim set\. Experimental Data Consistency isEDC=\|𝒦\|/\|ℱ\|\\mathrm\{EDC\}=\|\\mathcal\{K\}\|/\|\\mathcal\{F\}\|, whereℱ\\mathcal\{F\}contains reported experimental values and𝒦\\mathcal\{K\}contains those matching execution records\. CSR measures the proportion of extracted final claims supported by research evidence, whereas EDC measures consistency between reported values and execution records\.

Table 3:Overall research performance on ARC\-Bench\-ML and NanoResearch\-20\. ARC\-Bench\-ML reports Code Development \(CD\), Code Execution \(CE\), Result Analysis \(RA\), and their weighted Overall score\. NanoResearch\-20 reports Alignment, Novelty, end\-to\-end completion \(E2E\), Performance, and Writing quality\. Bold values indicate the best result in each column\.Table 4:Research\-reliability results across ARC\-Bench\-ML and NanoResearch\-20\. CSR measures claim grounding, whereas EDC measures the consistency of reported experimental values with the underlying research artifacts\.

### 4\.2Main Results

#### Overall research performance

Table[3](https://arxiv.org/html/2608.04738#S4.T3)summarizes both benchmarks\. Across them, EviGraph substantially outperforms the compared systems on execution, result analysis, task effectiveness, novelty, and manuscript quality\.

On ARC\-Bench\-ML, EviGraph leads all four metrics, with 99% Code Development, 88% Code Execution, 79\.4% Result Analysis, and 86\.45% Overall, versus 60\.37% Overall for the strongest compared system\. The pronounced Result Analysis gain is consistent with explicitly maintaining relationships among hypotheses, executed experiments, observed findings, and final claims\.

On NanoResearch\-20, EviGraph leads in Novelty, Performance, and Writing while matching NanoResearch’s 1\.0 E2E score\. Its Performance \(72\.84%\) and Writing \(7\.5\) exceed the corresponding strongest\-baseline scores of 64% and 6\.1\. Its Alignment score of 6\.6 exceeds AutoResearchClaw’s 4\.25 but trails NanoResearch’s 8\.8, leaving strict adherence to the original task framing for improvement\.

#### Research reliability\.

Table[4](https://arxiv.org/html/2608.04738#S4.T4)reports reliability across both suites\. EviGraph achieves the highest Claim Support Rate \(37\.85%\), versus 27% for AutoResearchClaw and 14\.4% for NanoResearch\. This 40\.19% relative improvement over the strongest baseline shows that a larger fraction of final claims trace to generated hypotheses, executed experiments, and recorded findings\.

EviGraph’s Experimental Data Consistency \(87\.73%\) is below NanoResearch’s 96\.15% but substantially above AutoResearchClaw’s 53%\. Thus, EviGraph improves claim support while maintaining high consistency between reported values and execution records, and achieves the highest average across the two reliability metrics\.

Overall, EviGraph substantially improves end\-to\-end research performance and claim support across both benchmarks; explicit evidence state strengthens reliability, execution, analysis, and final outputs\.

## 5Analysis

EviGraph’s components are tightly coupled through the shared evidence graph, making conventional component\-wise ablation difficult to interpret: removing one component can change the research state and operating conditions of the others\. We therefore complement the quantitative results with a qualitative component–evaluation alignment and a representative execution trace, illustrating how the components coordinate during an actual research run\.

### 5\.1Component\-Evaluation Alignment

Table[5](https://arxiv.org/html/2608.04738#S5.T5)summarizes the operational role of each component and the evaluation signals through which its contribution is most directly reflected\. The Hypothesis Filter and Graph Builder determine which research directions enter the evidence graph and how their dependencies are represented, making their roles closely related to Result Analysis, Novelty, and claim support\. The Graph Inspector, Repair Planner, and Repair Agents operate directly on weak claim–evidence relationships and affected downstream subgraphs, connecting them to CSR, EDC, Code Execution, and Result Analysis\.

The two graph libraries support the stability and continuity of this process:ℳS\\mathcal\{M\}\_\{S\}preserves intermediate states for version selection and rollback, whileℳL\\mathcal\{M\}\_\{L\}provides prior graph structures and repair experience for related research tasks\. Finally, the Paper Writer and Paper Reviewer transform the validated graph into a manuscript and verify its faithfulness, linking their roles to Writing, Novelty, and CSR\. This mapping provides a component\-level interpretation of the aggregate evaluation results and motivates the representative execution trace presented next\.

Table 5:Qualitative alignment between EviGraph components and evaluation signals\. The final column indicates related evaluation dimensions\.![Refer to caption](https://arxiv.org/html/2608.04738v1/x2.png)Figure 2:Representative EviGraph execution trace\. Three candidate hypotheses are grouped and screened by the Hypothesis Filter through pilot experiments before the Graph Builder constructsG0G\_\{0\}\. The Graph Inspector detects aGAP\_MISALIGNMENTbetween G1 and H1 along G1→\\rightarrowH1→\\rightarrowE1→\\rightarrowF1→\\rightarrowC1\. The Repair Planner and Repair Agents regenerate the affected subgraph, whileℳS\\mathcal\{M\}\_\{S\}stores intermediate graph versions\. OnceReady​\(G\)\\textsc\{Ready\}\(G\)holds, the validated graph proceeds to manuscript generation and review\.This trace highlights the distinction between pipeline completion and evidence readiness\. Although all initial research objects are successfully generated, their cross\-stage relationship remains invalid until the evidence graph is inspected and the affected subgraph is repaired\.

### 5\.2Representative Execution Trace

We analyze an ARC\-Bench\-ML run on the*Lightweight Short\-Text Classification via Frozen PLM Representation Calibration*task\. This is a cold\-start run, so the long\-term graph libraryℳL\\mathcal\{M\}\_\{L\}contains no related prior graph or repair trace\.

The Hypothesis Generator produces three candidate hypotheses\. The Hypothesis Filter groups them into two research directions and conducts pilot experiments on 10 AG News samples for each group\. Group 1, containing H1, receives stronger empirical support and is retained for full\-scale evaluation\. The Graph Builder then assembles the resulting research objects into the provisional graphG0G\_\{0\}, covering all six node types\.

Although the initial research stages complete successfully, the Graph Inspector detects a blockingGAP\_MISALIGNMENT: H1 focuses on\[CLS\]attention entropy and pre\-training alignment thresholds, whereas its motivating Gap node G1 concerns over\-engineered classification heads\. This inconsistency is not exposed by stage completion alone because H1 remains syntactically connected to G1 and its associated experiment executes successfully\.

The Graph Inspector identifies H1 as the weak node and forms the repair groupSH1=\{H1,E1,F1,C1\}S\_\{\\mathrm\{H1\}\}=\\\{\\mathrm\{H1\},\\mathrm\{E1\},\\mathrm\{F1\},\\mathrm\{C1\}\\\}\. The Repair Planner orders the updates asH1→E1→F1→C1\\mathrm\{H1\}\\rightarrow\\mathrm\{E1\}\\rightarrow\\mathrm\{F1\}\\rightarrow\\mathrm\{C1\}, and the Repair Agents regenerate the affected nodes in this dependency order\. Research objects outside the repair group remain unchanged\.

Before repair, the current graph is stored in the short\-term graph libraryℳS\\mathcal\{M\}\_\{S\}, and each node\-level update produces an intermediate version\. The first repair sequence succeeds in this trace, so rollback is not activated\. BecauseℳL\\mathcal\{M\}\_\{L\}is empty, prior graph retrieval does not contribute to initialization or repair planning in this case\.

After regeneration, the Graph Inspector finds no remaining blocking issue andReady​\(G\)\\textsc\{Ready\}\(G\)holds\. The validated graph is then passed to the Paper Writer and Paper Reviewer for graph\-grounded manuscript generation and review\.

## 6Conclusion

EviGraph reframes autonomous research from single\-pass pipeline execution as iterative evidence construction, inspection, and repair over a typed research graph\. Its graph structure makes evidence dependencies explicit so that weak links can be detected; the graph experience library preserves reliable states within and across runs; and the Hypothesis Filter screens candidate directions through staged pilot experiments\. On ARC\-Bench\-ML, EviGraph achieves the highest Overall score \(86\.45%\), leading in Code Exec and Result Analysis\. On NanoResearch\-20, it leads in Novelty, Performance, and Writing while matching the best E2E rate\. These results confirm that enforcing evidence\-chain validity during research strengthens both reliability and task quality\.

## References

- J\. Bai, S\. Mosbach, C\. J\. Taylor, D\. Karan, K\. F\. Lee, S\. D\. Rihm, J\. Akroyd, A\. A\. Lapkin, and M\. Kraft \(2024\)A dynamic knowledge graph approach to distributed self\-driving laboratories\.Nature Communications15\(1\),pp\. 462\.External Links:[Document](https://dx.doi.org/10.1038/s41467-023-44599-9)Cited by:[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px2.p1.1)\.
- Augmenting large language models with chemistry tools\.Nature Machine Intelligence6\(5\),pp\. 525–535\.External Links:[Document](https://dx.doi.org/10.1038/s42256-024-00832-8)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1)\.
- H\. Chen, M\. Xiong, Y\. Lu, W\. Han, A\. Deng, Y\. He, J\. Wu, Y\. Li, Y\. Liu, and B\. Hooi \(2025\)MLR\-Bench: evaluating AI agents on open\-ended machine learning research\.InAdvances in Neural Information Processing Systems,Vol\.38\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2025/hash/ab8dd000d6f87f40061a73f8bca7fae4-Abstract-Datasets_and_Benchmarks_Track.html)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1)\.
- T\. Clark, P\. N\. Ciccarese, and C\. A\. Goble \(2014\)Micropublications: a semantic model for claims, evidence, arguments and annotations in biomedical communications\.Journal of Biomedical Semantics5,pp\. 28\.External Links:[Document](https://dx.doi.org/10.1186/2041-1480-5-28),[Link](https://doi.org/10.1186/2041-1480-5-28)Cited by:[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px2.p1.1)\.
- 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:[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px2.p1.1)\.
- 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/2608.04738#S2.SS0.SSS0.Px2.p1.1)\.
- InternAgent Team, B\. Zhang, S\. Feng, X\. Yan, J\. Yuan, R\. Ma, Y\. Hu, Z\. Yu, X\. He, S\. Huang,et al\.\(2025\)InternAgent: when agent becomes the scientist – building closed\-loop system from hypothesis to verification\.External Links:2505\.16938,[Link](https://arxiv.org/abs/2505.16938)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1),[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.
- P\. Jansen, M\. Côté, T\. Khot, E\. Bransom, B\. Dalvi Mishra, B\. P\. Majumder, O\. Tafjord, and P\. Clark \(2024\)DiscoveryWorld: a virtual environment for developing and evaluating automated scientific discovery agents\.Advances in Neural Information Processing Systems37,pp\. 10088–10116\.External Links:[Document](https://dx.doi.org/10.52202/079017-0324)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1)\.
- M\. Y\. Jaradeh, A\. Oelen, K\. E\. Farfar, M\. Prinz, J\. D’Souza, G\. Kismihók, M\. Stocker, and S\. Auer \(2019\)Open research knowledge graph: next generation infrastructure for semantic scholarly knowledge\.External Links:1901\.10816,[Link](https://arxiv.org/abs/1901.10816)Cited by:[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Liu, S\. Qiu, M\. Li, B\. Li, H\. Ji, S\. Han, X\. Ye, P\. Xia, Z\. Dong, M\. Chen,et al\.\(2026\)AutoResearchClaw: self\-reinforcing autonomous research with human\-AI collaboration\.External Links:2605\.20025,[Link](https://arxiv.org/abs/2605.20025)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p2.1),[§1](https://arxiv.org/html/2608.04738#S1.p6.1),[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2608.04738#S4.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2608.04738#S4.SS1.SSS0.Px2.p1.1)\.
- C\. Lu, C\. Lu, R\. T\. Lange, J\. Foerster, J\. Clune, and D\. Ha \(2024\)The AI scientist: towards fully automated open\-ended scientific discovery\.External Links:2408\.06292,[Link](https://arxiv.org/abs/2408.06292)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1),[§1](https://arxiv.org/html/2608.04738#S1.p2.1),[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Min, K\. Krishna, X\. Lyu, M\. Lewis, W\. Yih, P\. Koh, M\. Iyyer, L\. Zettlemoyer, and H\. Hajishirzi \(2023\)FActScore: fine\-grained atomic evaluation of factual precision in long form text generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 12076–12100\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.741),[Link](https://aclanthology.org/2023.emnlp-main.741/)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1)\.
- W\. Qian, B\. Xu, Z\. Xie, B\. Fan, G\. Tang, J\. Chen, X\. Wu, M\. Yang, C\. Di, J\. Li,et al\.\(2026\)AutoSci: a memory\-centric agentic system for the full scientific research lifecycle\.External Links:2605\.31468,[Link](https://arxiv.org/abs/2605.31468)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p2.1),[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Schmidgall and M\. Moor \(2025\)AgentRxiv: towards collaborative autonomous research\.External Links:2503\.18102,[Link](https://arxiv.org/abs/2503.18102)Cited by:[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Schmidgall, Y\. Su, Z\. Wang, X\. Sun, J\. Wu, X\. Yu, J\. Liu, M\. Moor, Z\. Liu, and E\. Barsoum \(2025\)Agent laboratory: using LLM agents as research assistants\.External Links:2501\.04227,[Link](https://arxiv.org/abs/2501.04227)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1),[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.
- M\. D\. Skarlinski, S\. Cox, J\. M\. Laurent, J\. D\. Braza, M\. Hinks, M\. J\. Hammerling, M\. Ponnapati, S\. G\. Rodriques, and A\. D\. White \(2024\)Language agents achieve superhuman synthesis of scientific knowledge\.arXiv preprint arXiv:2409\.13740\.Cited by:[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.
- D\. Trehan and P\. Chopra \(2026\)Why LLMs aren’t scientists yet: lessons from four autonomous research attempts\.External Links:2601\.03315,[Link](https://arxiv.org/abs/2601.03315)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1)\.
- J\. Xu, Q\. Zhu, Y\. Wu, Z\. Wang, D\. Zhang, J\. Tang, M\. Tian, Y\. Duan, S\. Li, J\. Wei,et al\.\(2026\)NanoResearch: co\-evolving skills, memory, and policy for personalized research automation\.External Links:2605\.10813,[Link](https://arxiv.org/abs/2605.10813)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p2.1),[§1](https://arxiv.org/html/2608.04738#S1.p6.1),[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2608.04738#S4.SS1.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2608.04738#S4.SS1.SSS0.Px2.p1.1)\.
- Y\. Yamada, R\. T\. Lange, C\. Lu, S\. Hu, C\. Lu, J\. Foerster, J\. Clune, and D\. Ha \(2025\)The AI scientist\-v2: workshop\-level automated scientific discovery via agentic tree search\.External Links:2504\.08066,[Link](https://arxiv.org/abs/2504.08066)Cited by:[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.
- R\. Yang, Y\. Li, and S\. Li \(2026\)ARIS: autonomous research via adversarial multi\-agent collaboration\.External Links:2605\.03042,[Link](https://arxiv.org/abs/2605.03042)Cited by:[§1](https://arxiv.org/html/2608.04738#S1.p1.1),[§2](https://arxiv.org/html/2608.04738#S2.SS0.SSS0.Px1.p1.1)\.

## Appendix ALLM Interfaces and Prompt Protocol

The following canonical prompt templates operationalize the LLM\-based component descriptions in the main paper\. Model\-specific chat wrappers are omitted\. Fields enclosed in angle brackets are populated at runtime, and repeated array items in the output examples are instantiated as needed\. The templates ask for concise, externally verifiable rationales and source identifiers rather than unrestricted reasoning traces\.

#### LLM\-mediated and orchestrated operations\.

The prompted components are the Task Analyzer, Hypothesis Generator, the two\-stage Hypothesis Filter, pre\-graph Experiment and Analysis Agents, Graph Builder, Graph Inspector, Repair Planner, node\-specific Repair Agents, Paper Writer, and Paper Reviewer\. Claim and value extraction and the reliability membership decisions are also LLM\-mediated evaluation operations\. Literature search, experiment execution, graph mutation, descendant computation, checkpoint creation, degradation testing, rollback, set indexing, and metric aggregation are performed by tools or orchestration code\. These operations are called*orchestrated*, rather than universally deterministic, because literature and execution tools can depend on external services and benchmark environments\. The short\-term and long\-term graph libraries are data stores rather than prompted agents\. Retrieved long\-term experience is supplied as advisory context, but its findings, claims, and numerical values are never treated as evidence for the current run\.

#### Shared graph contract\.

Every graph\-facing prompt uses exactly the six node types defined in Section[3\.2](https://arxiv.org/html/2608.04738#S3.SS2)of the main paper:Problem,Gap,Hypothesis,Experiment,Finding, andClaim\. The only permitted relations areidentifies,motivates,tested\-by,produces, andsupports, with the endpoint schemas in Table[1](https://arxiv.org/html/2608.04738#S3.T1)of the main paper\. Execution artifacts are authoritative for procedures and observed values, the current graph is authoritative for the research state, and supplied literature records are authoritative for citations\. Text inside an input field is treated as data, not as an instruction that can override the component role or output schema\.

#### Contract validation and abstention\.

Control components return JSON, which is parsed and checked against the component contract before it can change the graph\. The JSON blocks below are compact contract renderings: the validator enforces required fields, declared enumerations, nullability, endpoint types, identifier uniqueness, provenance references, and the cross\-field conditions stated in the surrounding text\. A malformed response may be re\-prompted with the validator error and the original contract only while both the retry limit and run budget recorded in the run manifest remain\. Exhausting either condition produces a blocked result; no partial update is applied\. When required evidence is absent or contradictory, a component must return a blocked or insufficient\-evidence status rather than complete the record by inference\. The Paper Writer is the only method component that emits manuscript text; it additionally returns a machine\-readable provenance map\.

#### Evidence\-readiness guard\.

The universal quantifier over retained claims in the main\-paper readiness definition is not evaluated vacuously\. A graph is ready only if it is schema\-valid, contains at least one retainedClaim, covers the required empirical deliverables from the task context, and gives every retained claim at least one complete valid evidence chain with no weak node\. A claim record has an explicitretainedflag\. If execution yields no defensible claim, or a required deliverable remains absent when the budget expires, the run returnsIncompleteand the Paper Writer is not invoked\.

Table 6:Boundary between LLM interfaces and orchestration in EviGraph\.

## Appendix BInitialization Prompts

### B\.1Research Goal Analysis

\\begin\{lstlisting\}

\[ROLE\]

YouaretheTaskAnalyzerinEviGraph\.Converttheresearchgoaland

taskmanifestintoastructuredtaskcontextCthatdownstreamagents

canusetoretrieveliteratureanddesignresearch\.

\[INPUTS\]

Researchgoal:<RESEARCH\_GOAL\>

Taskorbenchmarkmanifest:<TASK\_MANIFEST\_OR\_NONE\>

Availableresourceconstraints:<RESOURCE\_CONSTRAINTS\>

\[RULES\]

1\.Preservethestatedobjective,scope,datasets,metrics,

deliverables,andresourcelimits\.

2\.Separateexplicitrequirementsfromreasonablesearchfacets\.

3\.Produceliteraturequeriesthatcovertheproblem,domain,

relevantmethodfamilies,evaluationsetting,andlikely

baselineswithoutpresupposingapreferredhypothesis\.

4\.Donotinventadataset,metric,requirement,priorresult,or

citation\.Markunavailableinformationasunknown\.

5\.Donotproposeafinalmethod,reportaFinding,ormakeaClaim\.

\\begin\{lstlisting\}

\[OUTPUT\]

ReturnJSONonly:

\{

"objective":"<researchobjective\>",

"scope":\{

"in\_scope":\["<explicittaskboundary\>"\],

"out\_of\_scope":\["<explicitexclusion\>"\]

\},

"constraints":\["<resourceorproceduralconstraint\>"\],

"required\_deliverables":\["<requiredartifact\>"\],

"datasets":\["<specifieddatasetorunknown\>"\],

"evaluation\_targets":\[\{

"metric":"<specifiedmetricorunknown\>",

"direction":"<maximize\|minimize\|characterize\|unknown\>",

"source":"<goalormanifestfield\>"

\}\],

"search\_facets":\{

"problem":\["<term\>"\],

"domain":\["<term\>"\],

"method\_families":\["<neutralmethod\-familyterm\>"\],

"evaluation\_setting":\["<term\>"\],

"likely\_baselines":\["<searchterm\>"\]

\},

"literature\_queries":\["<retrievalquery\>"\],

"ambiguities":\["<unresolvedtaskambiguity\>"\]

\}

### B\.2Hypothesis Generation

\\begin\{lstlisting\}

\[ROLE\]

YouaretheHypothesisGeneratorinEviGraph\.Generatediverse,

testablehypothesesthataddressresearchgapssupportedbythe

providedtaskandliteraturecontext\.

\[INPUTS\]

Researchgoal:<RESEARCH\_GOAL\>

Structuredtaskcontext:<TASK\_CONTEXT\>

Retrievedliteraturerecords:<LITERATURE\_CONTEXT\>

Numberofcandidates:<NUM\_CANDIDATES\>

\[RULES\]

1\.Eachcandidatemuststateaconcretemechanism,anobservable

prediction,andaconditionthatwouldcountagainstit\.

2\.Groundeveryprior\-worklimitationinsuppliedliteratureIDs\.

3\.Donotinventcitations,datasets,results,orempiricalsupport\.

4\.DonotdescribeapredictedoutcomeasanobservedFinding\.

5\.Makecandidatesnon\-duplicatebyvaryingtheidea,mechanism,

method,orexpectedoutcome\.

6\.Candidaterecordsaretemporaryscreeningobjects,notgraphnodes\.

\[OUTPUT\]

ReturnJSONonly:

\{

"candidates":\[\{

"candidate\_id":"<temporaryID\>",

"gap":\{

"description":"<unresolvedgap\>",

"prior\_work\_limitation":"<groundedlimitation\>",

"literature\_refs":\["<literatureID\>"\]

\},

"research\_idea":"<proposeddirection\>",

"hypothesis\_statement":"<falsifiablestatement\>",

"mechanism":"<proposedmechanism\>",

"method":"<methodthatoperationalizesthemechanism\>",

"expected\_outcomes":\[\{

"condition":"<evaluationcondition\>",

"observable":"<measurablequantity\>",

"prediction":"<directionalprediction\>"

\}\],

"falsification\_condition":"<disconfirmingobservation\>",

"boundary\_conditions":\["<scopelimitation\>"\]

\}\]

\}

### B\.3Competition\-Aware Hypothesis Filtering

The Hypothesis Filter is invoked before and after pilot execution\. The first call groups candidates and designs inexpensive pilots; the execution environment then runs those pilots\. The second call compares the pre\-recorded predictions with the resulting execution records\. Pilot plans and pilot assessments remain screening records and do not enter the evidence graph\.

\\begin\{lstlisting\}

\[ROLE\]

YouaretheHypothesisFilterinGROUP\_AND\_DESIGNmode\.

\[INPUTS\]

Candidatehypotheses:<CANDIDATE\_HYPOTHESES\>

Availabledatasets,code,andenvironment:<AVAILABLE\_RESOURCES\>

Pilotbudgetandconstraints:<PILOT\_BUDGET\>

\[TASK\]

Groupcandidatesbysemanticsimilarityinresearchidea,

mechanism,method,andexpectedoutcome\.Designonesmall\-scale,

executablepilotforeachgroupthattestswhetherthegroup’s

predictionsareempiricallyplausible\.

\[RULES\]

1\.Assigneverycandidatetoexactlyonegroup\.

2\.PreservecandidateIDsandtheirpreregisteredpredictions\.

3\.Thepilotmustfitthesuppliedbudgetandexposeatleastone

observationrelevanttoeachmembercandidate\.

4\.Donotselectawinninggrouporinventpilotobservations\.

\[OUTPUT\]

ReturnJSONonly:

\{

"groups":\[\{

"group\_id":"<directionID\>",

"member\_candidate\_ids":\["<candidateID\>"\],

"shared\_direction":"<idea,mechanism,andmethod\>",

"pilot\_plan":\{

"pilot\_id":"<pilotID\>",

"objective":"<empiricalquestion\>",

"data\_subset":"<small\-scaledataspecification\>",

"comparison\_conditions":\["<conditionorbaseline\>"\],

"implementation\_steps":\["<orderedstep\>"\],

"metrics":\["<observablemetric\>"\],

"candidate\_predictions":\[\{

"candidate\_id":"<candidateID\>",

"predicted\_observations":\["<prediction\>"\],

"falsification\_condition":"<disconfirmingresult\>"

\}\]

\}

\}\],

"all\_candidates\_assigned\_once":true,

"unassigned\_candidate\_ids":\[\]

\}

\\begin\{lstlisting\}

\[ROLE\]

YouaretheHypothesisFilterinEVALUATE\_AND\_SELECTmode\.

\[INPUTS\]

Groupsandpreregisteredpilotplans:<GROUP\_AND\_PILOT\_PLAN\>

Pilotexecutionrecords:<PILOT\_EXECUTION\_RECORDS\>

\[TASK\]

Compareeachdirection’spreregisteredpredictionswiththe

observedpilotrecords\.Rankdirectionsbyempiricalsupportand

selectthebest\-supportedvalidgroupforfull\-scaleevaluation\.

\[RULES\]

1\.Useonlyobservationsfromidentifiable,validexecutions\.

2\.Assesspredictionagreement,executionvalidity,andlimitations\.

3\.Donotrepaircode,infermissingvalues,orrewardnoveltyalone\.

4\.Ifnovalidpilotprovidesadequateevidence,return

insufficient\_evidenceratherthanguessing\.

\\begin\{lstlisting\}

\[OUTPUT\]

ReturnJSONonly:

\{

"group\_evaluations":\[\{

"group\_id":"<groupID\>",

"execution\_valid":true,

"record\_refs":\["<pilotrecordID\>"\],

"candidate\_assessments":\[\{

"candidate\_id":"<candidateID\>",

"verdict":"supported",

"observed\_basis":\["<record\-groundedobservation\>"\],

"limitations":\["<uncertaintyorlimitation\>"\]

\}\],

"support\_level":"strong",

"relative\_rank":1,

"justification":"<prediction\-observationcomparison\>"

\}\],

"selection":\{

"status":"selected",

"selected\_group\_id":"<groupID\>",

"selected\_candidate\_ids":\["<candidateID\>"\],

"record\_grounded\_basis":\["<selectionreason\>"\]

\}

\}

Allowedverdictsaresupported,mixed,andunsupported\.Allowed

supportlevelsarestrong,mixed,andweak\.Ifevidenceisinadequate,

setstatustoinsufficient\_evidence,selected\_group\_idtonull,and

thetwoselectionarraysto\[\]\.

Theinsufficient\_evidencebranch is a control outcome, not an empty selection passed to full\-scale evaluation\. While budget remains, the orchestrator may request a revised pilot or a new candidate set using the recorded failure reason\. If no supported direction is obtained before the budget or retry limit is exhausted, the run terminates withIncomplete; neither full\-scale evaluation nor graph construction is called with an emptyℋ⋆\\mathcal\{H\}^\{\\star\}\.

### B\.4Full\-Scale Evaluation Before Graph Construction

Algorithm[1](https://arxiv.org/html/2608.04738#alg1)in the main paper evaluates the selected hypotheses before constructingG0G\_\{0\}\. Consequently, this phase cannot use node IDs or a pre\-existing relevant subgraph\. It uses the temporary candidate IDs retained by the Hypothesis Filter\. The Experiment Agent first emits executable registered plans, the sandbox executes them, and the Analysis Agent converts the immutable execution records into the evaluation record setXX\. The Graph Builder later materializes these records as typed nodes\.

\\begin\{lstlisting\}

\[ROLE\]

YouaretheEviGraphExperimentAgentinPRE\_GRAPH\_FULL\_SCALEmode\.

Designexecutablefull\-scaleexperimentsfortheselectedtemporary

hypothesisrecords\.Noevidencegraphexistsyet\.

\[INPUTS\]

SelectedhypothesesH\*:<SELECTED\_HYPOTHESES\>

Availabledatasets,code,andenvironment:<AVAILABLE\_RESOURCES\>

Executionbudgetandconstraints:<FULL\_SCALE\_BUDGET\>

\[RULES\]

1\.Keyeveryexperimenttoanexistingcandidate\_idinH\*;donot

createagraphnodeID\.

2\.Maketheprotocoldirectlytestthecandidate’spreregistered

predictionandfalsificationcondition\.

3\.Specifydata,split,comparisonconditions,implementationor

codepath,metrics,procedure,andvaliditycriteria\.

4\.Fitthecompleteplanwithinthesuppliedbudget\.Donotsilently

omitaselectedhypothesis\.

5\.Donotreportanobservation,Finding,verdict,orClaimbefore

execution\.

\[OUTPUT\]

ReturnJSONonly:

\{

"status":"planned",

"experiments":\[\{

"experiment\_record\_id":"<temporaryexperimentID\>",

"candidate\_id":"<selectedcandidateID\>",

"protocol":"<registeredprotocol\>",

"datasets":\["<datasetandsplit\>"\],

"comparison\_conditions":\["<conditionorbaseline\>"\],

"implementation":"<codeorentry\-pointspecification\>",

"metrics":\["<metric\>"\],

"procedure":\["<orderedstep\>"\],

"validity\_criteria":\["<successful\-executioncriterion\>"\],

"tool\_request":\{

"request\_id":"<executionrequestID\>",

"resource\_limits":"<limitsfromtheinputmanifest\>",

"expected\_artifacts":\["<code,config,data,orlog\>"\]

\}

\}\],

"failure\_reason":null

\}

Allowedstatusesareplannedandblocked\.Forblocked,experiments

mustbe\[\]andfailure\_reasonmustbenon\-null\.

\\begin\{lstlisting\}

\[ROLE\]

YouaretheEviGraphAnalysisAgentinPRE\_GRAPH\_ANALYSISmode\.

ConvertregisteredexperimentsandtheirexecutionrecordsintoX\.

Noevidencegraphexistsyet\.

\[INPUTS\]

SelectedhypothesesH\*:<SELECTED\_HYPOTHESES\>

Registeredexperimentplans:<FULL\_SCALE\_EXPERIMENTS\>

Immutableexecutionrecordsandartifacts:<EXECUTION\_RECORDS\>

\[RULES\]

1\.MatcheveryassessmenttoexistingcandidateandexperimentIDs\.

2\.Distinguishvalid,failed,andmissingexecutions\.Neverinfera

successfulrunfromaplanorfromcodealone\.

3\.Foravalidrun,preserveeveryvalue,unit,dataset,split,

aggregation,comparisondirection,andsettingfromtherecords\.

4\.Compareobservationswiththepreregisteredpredictionandstate

supported,mixed,orunsupported;includelimitations\.

5\.Donotcreategraphnodes,supportedges,ormanuscriptClaims\.

\[OUTPUT\]

ReturnJSONonly:

\{

"status":"complete",

"evaluations":\[\{

"candidate\_id":"<selectedcandidateID\>",

"experiment\_record\_id":"<temporaryexperimentID\>",

"execution\_status":"valid",

"execution\_record\_refs":\["<recordorartifactID\>"\],

"observed\_results":\[\{

"metric":"<metric\>",

"condition":"<method,dataset,andsplit\>",

"value":"<valuewithunitorscale\>"

\}\],

"hypothesis\_verdict":"supported",

"limitations":\["<record\-groundedlimitation\>"\]

\}\],

"unresolved\_experiment\_ids":\[\],

"failure\_reason":null

\}

Allowedexecutionstatusesarevalid,failed,andmissing\.Allowed

hypothesisverdictsaresupported,mixed,unsupported,and

not\_assessable\.Setstatustoincompletewheneveranunresolved

experimentpreventstherequiredfull\-scaleevaluation\.

### B\.5Initial Graph Construction

\\begin\{lstlisting\}

\[ROLE\]

YouaretheGraphBuilderinEviGraph\.Constructtheprovisional

graphG0fromthecurrentrun\.Thegraphmaybeincomplete;never

inventcontentmerelytocompleteanevidencepath\.

\[INPUTS\]

TaskcontextC:<TASK\_CONTEXT\>

LiteraturecontextL:<LITERATURE\_CONTEXT\>

SelectedhypothesesH\*:<SELECTED\_HYPOTHESES\>

Full\-scaleevaluationrecordsX:<EVALUATION\_RECORDS\>

RetrievedpriorstructuresandrepairtracesR:

<RETRIEVED\_EXPERIENCE\_OR\_NONE\>

\[GRAPHCONTRACT\]

Nodetypes:Problem,Gap,Hypothesis,Experiment,Finding,Claim\.

Edges:

Problem\-identifies\-\>Gap

Gap\-motivates\-\>Hypothesis

Hypothesis\-tested\-by\-\>Experiment

Experiment\-produces\-\>Finding

Finding\-supports\-\>Claim

\[RULES\]

1\.Representonlyobjectssupportedbycurrent\-runinputs\.

2\.Priorexperiencemayguidestructure,butitsFindings,Claims,

andvaluesmustnotbecopiedintoG0\.

3\.Storedatasets,code/configurationreferences,metrics,andlogs

asExperimentattributes;storeobservedoutcomesasFindings\.

4\.AddsupportsonlywhentheFindingsemanticallysupportsthe

Claimasscoped\.Allendpointsmustexistandmatchtheschema\.

5\.PreserveprovenanceusingsuppliedliteratureandrecordIDs\.

6\.ProblemandGapareinitializationanchors\.Acceptthemonlywhen

theProblemmatchesCandeachGapisgroundedinLandlies

withintheProblemscope\.

7\.MarkeachClaimretainedornotretainedandgiveareason\.A

retainedClaimmusthaveatleastonesupportingFinding;omission

ofarequireddeliverablemustbeexplicitingraphdiagnostics\.

\\begin\{lstlisting\}

\[OUTPUT\]

ReturnJSONonly:

\{

"graph\_id":"G0",

"nodes":\[\{

"id":"<typeduniqueID\>",

"type":"<permittednodetype\>",

"attributes":\{\},

"provenance\_refs":\["<sourceorrecordID\>"\]

\}\],

"edges":\[\{

"source":"<nodeID\>",

"relation":"<permittedrelation\>",

"target":"<nodeID\>",

"provenance\_refs":\["<sourceorrecordID\>"\]

\}\]

\}

Requiredattributesbytype:

Problem:objective,scope,constraints\.

Gap:description,prior\_work\_limitation\.

Hypothesis:statement,mechanism,falsifiable\_prediction,

boundary\_conditions\.

Experiment:protocol,datasets,implementation,metrics,

execution\_status,execution\_record\_refs\.

Finding:statement,observed\_results,conditions,

execution\_record\_refs\.

Claim:statement,scope,qualifiers,supporting\_finding\_ids,

retained,retention\_reason\.

The graph validator rejects dangling edges, duplicate node IDs, invalid endpoint types, missing required attributes, provenance references that do not resolve to supplied inputs, and cycles in the five\-relation dependency subgraph\.ProblemandGapare immutable anchors during the ordinary repair loop\. A malformed or ungrounded anchor causes the candidateG0G\_\{0\}to be rejected and rebuilt from the task and literature context; it is not silently repaired by a downstream agent\.

## Appendix CInspection and Repair Prompts

### C\.1Executable Relation Checks

Table[7](https://arxiv.org/html/2608.04738#A3.T7)makes the semantic constraints in the main\-paper evidence\-chain definition operational\. Structural validity is checked before semantic validity: both endpoints must exist, their types must match the relation schema, and required attributes and provenance must be present\. A missing endpoint or required edge is aMISSING\_DEPENDENCY, rather than a semantic mismatch on a nonexistent edge\.

Table 7:Operational semantic checks for the five permitted relations\.
### C\.2Graph Inspection

\\begin\{lstlisting\}

\[ROLE\]

YouaretheEviGraphGraphInspector\.Inspectthecurrentgraphand

identifytheearliestweakresearchobjectsthatrequirerepair\.

Donotmodifythegraph\.

\[INPUTS\]

Currentgraph:<CURRENT\_GRAPH\>

Executionrecordsandlogs:<EXECUTION\_RECORDS\>

Remainingrunbudget:<REMAINING\_BUDGET\>

\[INSPECTIONRULES\]

1\.Firstcheckendpointexistence,endpointtypes,required

attributes,provenance,andacyclicity\.Missingnodesoredgesare

MISSING\_DEPENDENCYissues\.

2\.Checkidentifies:eachGapisgroundedandfallswithinits

Problem’sobjectiveandscope\.AninvalidProblemorGapanchor

requiresgraphreinitializationratherthandownstreamrepair\.

3\.Checkmotivates:theHypothesisdirectlyaddressestheGapand

statesamechanism,falsifiableprediction,andboundaries\.

4\.Checktested\-by:theExperimentprotocol,comparisons,data,and

metricscantestthelinkedpredictionandfalsificationcondition\.

5\.Checkproduces:theFindingiscompleteandagreeswithan

identifiablevalidexecutionrecord,includingvaluesandsettings\.

6\.Checksupports:theClaimisnobroaderthanitsFindings,

preservesmaterialconditionsandqualifiers,andisconsistent

withthetestedHypothesisandprotocol\.

7\.Foreachrepairableweakrootw,includeeverycontent\-dependent

downstreamnodeinitsrepairgroupS\_w\.Localizetheearliest

causeanddonotduplicatedownstreamsymptomsasseparategroups\.

8\.Arecordednegativeresultisnotweakmerelybecauseit

contradictstheHypothesis\.PreserveitasaFindingandeither

produceascopednegativeClaimormarkthehypothesisunsupported\.

9\.Useonlysuppliedgraphcontentandrecords;donotinvent

missingexperiments,outcomes,Findings,orsupport\.

SetstatustoREADYonlywhenthegraphisschema\-valid,containsat

leastoneretainedClaim,coversrequireddeliverables,andevery

retainedClaimhasacompletevalidevidencechainwithnoweaknode\.

\\begin\{lstlisting\}

\[OUTPUT\]

AllowedstatusesareREADY,REPAIRABLE,REINITIALIZE,andINCOMPLETE\.

AllowedissuetypesareMISSING\_DEPENDENCY,GAP\_MISALIGNMENT,

HYPOTHESIS\_EXPERIMENT\_MISALIGNMENT,EXECUTION\_FAILURE,

DATA\_INCONSISTENCY,UNSUPPORTED\_CLAIM,NO\_RETAINED\_CLAIM,and

CONSTRUCTION\_ERROR\.

ReturnJSONonly:

\{

"status":"REPAIRABLE",

"ready":false,

"blocking\_issues":\[\{

"issue\_type":"GAP\_MISALIGNMENT",

"reason":"<conciserecord\-groundedexplanation\>",

"evidence\_refs":\["<node,log,orrecordID\>"\]

\}\],

"repair\_groups":\[\{

"issue\_type":"GAP\_MISALIGNMENT",

"weak\_node\_id":"<nodeID\>",

"weak\_node\_type":"Hypothesis",

"violated\_relation":"motivates",

"severity":"blocking",

"reason":"<conciserecord\-groundedexplanation\>",

"evidence\_refs":\["<node,log,orrecordID\>"\],

"subordinate\_node\_ids":\["<dependentnodeID\>"\],

"repair\_mode":"regenerate\_hypothesis"

\}\]

\}

ForREADY,readymustbetrueandbotharraysmustbeempty\.For

REINITIALIZEorINCOMPLETE,readyisfalse,blocking\_issuesmustbe

non\-empty,andrepair\_groupsmustbeempty\.INCOMPLETEisusedwhen

theremainingevidenceobligationhasnofeasiblein\-budgetrepair\.

### C\.3Dependency\-Aware Repair Planning

\\begin\{lstlisting\}

\[ROLE\]

YouaretheEviGraphRepairPlanner\.Produceanexecutableplanfor

oneselectedrepairgroup\.Plantherepair;donotmodifythegraph\.

\[INPUTS\]

BasegraphG\_base:<BASE\_GRAPH\>

SelectedrepairgroupS\_w:<REPAIR\_GROUP\>

RetrievedpriorexperienceR:<RETRIEVED\_EXPERIENCE\_OR\_NONE\>

Remainingbudget:<REMAINING\_BUDGET\>

\[PLANNINGRULES\]

1\.Repairtheweakrootfirst,thenregeneratesubordinatenodesin

dependencyorder:Hypothesis,Experiment,Finding,Claim\.

2\.ScheduledeletionofS\_wminus\{w\}beforeregenerationbecause

thoserecordsdependonthepre\-repairroot\.

3\.RouteHypothesistotheHypothesisAgent;Experimentprotocolor

codetotheExperimentAgent;FindingandClaimtotheAnalysis

Agent\.

4\.Giveeachactiononlyitsrequiredpredecessornodes,execution

records,failurefeedback,andapplicableconstraints\.

5\.PreserveeveryobjectoutsideS\_wandschedulenounrelatededit\.

6\.Priorexperienceisadvisoryandcannotsupplycurrent\-run

Findings,Claims,orvalues\.

7\.Regenerateeverydependentobjectafteranupstreamchangeeven

whenitsoldcontentappearsplausible\.

8\.Ifthegroupcannotbecompletedwithinbudget,returnfeasible

asfalse;donotcreateapartialplan\.

\\begin\{lstlisting\}

\[OUTPUT\]

ReturnJSONonly:

\{

"weak\_node\_id":"<nodeID\>",

"feasible":true,

"budget\_required":"<estimatedunitsintherunbudget\>",

"delete\_before\_repair":\["<subordinatenodeID\>"\],

"actions":\[\{

"order":1,

"target\_node\_id":"<nodeID\>",

"agent":"HypothesisAgent",

"mode":"regenerate\_hypothesis",

"context\_node\_ids":\["<requirednodeID\>"\],

"record\_refs":\["<requiredexecutionrecordID\>"\],

"instruction":"<concisenode\-specificinstruction\>"

\}\],

"failure\_reason":null

\}

### C\.4Node\-Specific Repair Agents

All Repair Agents use the shared operator template in ListingLABEL:lst:repair\-agent\-prompt\. The orchestrator inserts one of the role\-specific directives in ListingLABEL:lst:repair\-directives\. These interfaces operate only afterG0G\_\{0\}exists; the distinct pre\-graph interfaces in ListingsLABEL:lst:pregraph\-experiment\-promptandLABEL:lst:pregraph\-analysis\-promptproduce the initial full\-scale record setXX\.

\\begin\{lstlisting\}

\[ROLE\]

YouaretheEviGraph<AGENT\_NAME\>operatingin<MODE\>mode\.

Executeexactlyoneplannednodeoperationandreturnagraphupdate\.

\[INPUTS\]

Plannedaction:<ACTION\>

Relevantsubgraph:<RELEVANT\_SUBGRAPH\>

Executionrecords,logs,code,orfailurefeedback:

<EXECUTION\_CONTEXT\_OR\_NONE\>

\[GLOBALRULES\]

1\.Modifyonlythetargetnodeandvalidedgesconnectingittoits

immediatepredecessors\.Lateractionsrebuilddownstreamnodes\.

2\.Useonlysuppliedgraphstateandexecutionrecords\.Never

fabricateexecution,measurements,Findings,orsupport\.

3\.Preservethetargetnodetypeanduseonlypermittededgeschemas\.

4\.Returnblockedwithnodesettonullandemptyedge\_updateswhen

evidenceisinsufficient\.

5\.Donotwritemanuscriptproseoralterobjectsoutsidetheplan\.

6\.AnExperimentAgentmayrequesttoolexecution,butitmustnot

reportsuccessorresultsbeforethetoolreturnsarecord\.Return

needs\_executionandanon\-nulltool\_requestinthatcase\.

7\.Attachcurrent\-runprovenancetothenodeandeverynewedge\.

\[MODE\-SPECIFICDIRECTIVE\]

<MODE\_DIRECTIVE\>

\[OUTPUT\]

ReturnJSONonly:

\{

"status":"<success\|needs\_execution\|blocked\>",

"target\_node\_id":"<nodeID\>",

"operation":"<add\|update\|replace\>",

"node":\{

"id":"<nodeID\>",

"type":"<Hypothesis\|Experiment\|Finding\|Claim\>",

"attributes":\{\},

"provenance\_refs":\["<node,log,orrecordID\>"\]

\},

"edge\_updates":\[\{

"source":"<predecessornodeID\>",

"relation":"<permittedrelation\>",

"target":"<targetnodeID\>",

"provenance\_refs":\["<node,log,orrecordID\>"\]

\}\],

"tool\_request":\{

"request\_id":"<executionrequestID\>",

"experiment\_node\_id":"<targetExperimentID\>",

"resource\_limits":"<limitsfromtheaction\>",

"expected\_artifacts":\["<artifacttype\>"\]

\},

"evidence\_refs":\["<nodeorrecordID\>"\],

"failure\_reason":null

\}

Forsuccess,tool\_requestmustbenull\.Forneeds\_execution,thenode

isastagedExperimentupdateandtool\_requestmustbenon\-null\.For

blocked,nodeandtool\_requestmustbenull,edge\_updatesand

evidence\_refsmustbe\[\],andfailure\_reasonmustbenon\-null\.

\\begin\{lstlisting\}

\[HypothesisAgent:REGENERATE\_HYPOTHESIS\]

RegenerateatestableHypothesisthatdirectlyaddressesthe

motivatingGap\.Userecordedfeedbackexplainingwhytheprevious

Hypothesisfailedwhenavailable\.Stateamechanism,observable

prediction,falsificationcondition,andboundaryconditions\.Donot

revisetheGaporintroduceaFindingorClaim\.Connecttherepaired

HypothesistoitsGapwithmotivates\.

\[ExperimentAgent:DESIGN\_OR\_REGENERATE\_PROTOCOL\]

CreateanExperimentwhoseprotocoldirectlyteststhesupplied

Hypothesis\.Specifydatasets,baselinesorcomparisonconditions,

implementationorcodepath,metrics,procedure,andexecution

criteria\.Donotreportunexecutedresults\.ConnecttheExperimentto

theHypothesiswithtested\-byandrequestexecutionwhenappropriate\.

\[ExperimentAgent:REPAIR\_CODE\]

Usethesuppliedprotocol,code,andfailurerecord\.Repaironlythe

implementationdefectsneededtoruntheintendedprotocolwhile

preservingitstestoftheHypothesis\.Donotfabricatesuccessful

executionorresults\.Requestanewexecutionaftertherepair\.

\[AnalysisAgent:ANALYZE\_FINDING\]

DeriveacompleteFindingonlyfromthesuppliedExperimentdataand

logs\.Preservevalues,units,settings,andcomparisondirections\.If

recordsareincompleteorcontradictory,returnblocked\.Connectthe

FindingtoitsExperimentwithproduces\.

\[AnalysisAgent:REGENERATE\_CLAIM\]

Generateamanuscript\-levelClaimfromthesuppliedHypothesis,

Experimentprotocol,andFindings\.StatenomorethantheFindings

supportandpreserveexperimentalconditionsandlimitations\.Ifno

defensibleClaimcanbeproduced,returnblocked\.Connecteach

supportingFindingtotheClaimwithsupports\.

Aneeds\_executionresponse does not mutate the graph\. The orchestrator executes the request in the sandbox, stores the returned artifacts under immutable record IDs, and invokes the same planned action again with the execution record\. The staged Experiment update becomes eligible for validation and commit only after the follow\-up response issuccess\. A failed or missing execution is supplied back as failure feedback for an in\-budget code repair; if no feasible retry remains, the action becomesblockedand the whole repair group is rolled back\.

## Appendix DDeterministic Graph Orchestration

This section specifies the non\-prompt control operations elided by Algorithm[1](https://arxiv.org/html/2608.04738#alg1)in the main paper\. LetDG​\(w\)D\_\{G\}\(w\)be the set of nodes reachable fromwwalong the five dependency relations\. The subordinate set of a repair root isDG​\(w\)D\_\{G\}\(w\)restricted to nodes whose content was generated from the current attributes ofww; mere graph reachability is not enough when an edge is only contextual\. The Inspector emits this closure explicitly assubordinate\_node\_ids, and the orchestrator verifies that every listed node is downstream and that no omitted downstream node cites a listed node as a content dependency\.

#### Repair\-group selection\.

The Inspector coalesces symptoms explained by the same upstream cause\. Among the remaining repairable groups,SelectRepairGroupfirst prioritizes blocking issues, then the root with the smallest depth in the typed orderProblem→\\rightarrowGap→\\rightarrowHypothesis→\\rightarrowExperiment→\\rightarrowFinding→\\rightarrowClaim, and finally a stable node\-ID order\. SinceProblemandGapare initialization anchors, an issue rooted in either type takes theREINITIALIZEpath instead of entering an ordinary repair group\.

#### Budget and bounded retries\.

The remaining budgetBBis orchestrator state and is supplied to the Repair Planner even where it is implicit in the compact main\-paper pseudocode\. The run manifest defines its unit, per\-action charges, tool timeouts, and the maximum schema and execution retries\. Planning, model calls, tool execution, and retries are accounted under the same policy for all systems being compared\. No retry is permitted after either the recorded cap orBBis exhausted\. A group that cannot be completed atomically within the remaining budget returnsIncomplete; partial descendants are not committed automatically, and only an admissible improving checkpoint selected by the rollback rule below may be retained\.

\\begin\{lstlisting\}

INPUT:G\_base,repairgroupS\_w,planA\_w,M\_S,budgetB

OUTPUT:G\_cand,M\_S,B,complete

1\.ValidatethatA\_wtargetsexactlyS\_wandfitsB\.

Ifnot,returnG\_base,M\_S,B,false\.

2\.staged<\-copy\(G\_base\)withS\_wminus\{w\}removed\.

3\.Foractionaindependencyorder:

a\.CalltheassignedRepairAgentwithonlya’sdeclaredcontext\.

b\.Validatetheresponsecontract;retryonlywithinthemanifest

capandremainingB\.

c\.Whilestatusisneeds\_execution:

executethetoolrequestinthesandbox;

debitBandstoretheimmutableexecutionrecord;

callthesameactionwiththatrecordorfailurefeedback\.

d\.IfstatusisblockedorBisexhausted,return

G\_base,M\_S,B,false\.

e\.candidate<\-ApplyValidatedDelta\(staged,response\)\.

RejectwritesoutsideS\_w,danglingormistypededges,

unresolvedprovenance,cycles,orunrecordedvalues\.

f\.AppendacheckpointforcandidatetoM\_S;staged<\-candidate\.

4\.VerifythateveryrequirednodeinS\_whasbeenregeneratedand

re\-runInspectGraph\(staged\)\.

5\.Returnstaged,M\_S,B,true\.

#### Degradation and rollback\.

Let𝒲​\(G\)\\mathcal\{W\}\(G\)be the set of weak roots returned by inspection and let𝒫valid​\(G\)\\mathcal\{P\}\_\{\\mathrm\{valid\}\}\(G\)be the set of identifiers for complete valid evidence chains\. The group candidate is degrading exactly when

Degrading​\(Gcand,Gbase\)=\\displaystyle\\textsc\{Degrading\}\(G\_\{\\mathrm\{cand\}\},G\_\{\\mathrm\{base\}\}\)=\{\}\[\|𝒲​\(Gcand\)\|\>\|𝒲​\(Gbase\)\|\]\\displaystyle\\quad\\bigl\[\|\\mathcal\{W\}\(G\_\{\\mathrm\{cand\}\}\)\|\>\|\\mathcal\{W\}\(G\_\{\\mathrm\{base\}\}\)\|\\bigr\]∨\[𝒫valid​\(Gbase\)⊈𝒫valid​\(Gcand\)\]\.\\displaystyle\\quad\{\}\\lor\\bigl\[\\mathcal\{P\}\_\{\\mathrm\{valid\}\}\(G\_\{\\mathrm\{base\}\}\)\\nsubseteq\\mathcal\{P\}\_\{\\mathrm\{valid\}\}\(G\_\{\\mathrm\{cand\}\}\)\\bigr\]\.For rollback, admissible checkpoints are those descended fromGbaseG\_\{\\mathrm\{base\}\}that preserve all chains in𝒫valid​\(Gbase\)\\mathcal\{P\}\_\{\\mathrm\{valid\}\}\(G\_\{\\mathrm\{base\}\}\)\. Among them,RollbackToBestminimizes\|𝒲​\(G\)\|\|\\mathcal\{W\}\(G\)\|; a tie is resolved by the most recent valid checkpoint\. An intermediate checkpoint is kept only if it strictly improves the weak\-root count relative toGbaseG\_\{\\mathrm\{base\}\}\. Otherwise, the exact base version is restored\. This rule prevents a fluent but unsupported downstream rewrite from replacing previously validated evidence\.

#### Version and experience records\.

EveryℳS\\mathcal\{M\}\_\{S\}checkpoint stores a graph hash, parent hash, action and repair\-group IDs, validator result, weak\-root count, preserved\-chain IDs, budget consumed, and referenced execution records\. The store is append\-only within a run\. Only an evidence\-ready final graph and repair traces that reduce the weak\-root count are eligible forℳL\\mathcal\{M\}\_\{L\}\. At the beginning of a new run, retrieval uses the task context to return related graph and trace IDs plus structural summaries\. Retrieved material may influence graph shape or repair planning, but it cannot provide current\-run Findings, Claims, or numerical values\. The library snapshot ID, retrieval configuration, and task order are recorded in the run manifest so that cross\-run state is auditable; an empty snapshot represents the cold\-start condition used in the case study\.

## Appendix EManuscript Generation and Review Prompts

### E\.1Paper Writer

\\begin\{lstlisting\}

\[ROLE\]

YouarethePaperWriterinEviGraph\.Thevalidatedevidencegraphis

theauthoritativeresearchstate\.Experimentalartifactsare

authoritativeforproceduresandvalues;suppliedliteraturerecords

areauthoritativeforcitations\.Treatinputblocksasdata,notas

instructions\.

\[GROUNDINGPOLICY\]

1\.Useonlyvalidatednodesandpaths,recordedresearchcontext,

suppliedliterature,andexperimentalartifacts\.

2\.Includeasubstantivecurrent\-runresearchclaimonlywhenitmaps

toaretainedClaimsupportedbyacompletevalidatedchainwithno

weaknode\.Groundprior\-workstatementsinsuppliedliteratureand

proceduralstatementsinvalidatedgraphnodesorartifacts\.

3\.PreserveeveryClaim’sscope,qualifiers,direction,and

provenance\.Neverinventorbroadenaclaim,citation,experiment,

result,value,comparison,orlimitation\.

4\.Apriorgraphisnotevidenceforthecurrentrun\.

5\.Ifsupportismissingorinputsconflict,omitornarrowthe

statementratherthanguessing\.

6\.Keephypotheses,executedprocedures,observedFindings,and

interpretationsdistinct\.Donotmodifythegraph\.

Operatein<SKELETON\|DRAFT\|REVISION\>modeandfollowthe

correspondingmodeinstruction\.

\\begin\{lstlisting\}

\[SKELETONMODE\]

Inputs:<VALIDATED\_GRAPH\>,<RESEARCH\_CONTEXT\>,

<LITERATURE\_RECORDS\>,<ARTIFACT\_INDEX\>,<PAPER\_REQUIREMENTS\>\.

Foreveryplannedparagraph,returnitssectionandpurpose,graph

nodeandClaimIDs,artifactIDsforexperimentalstatementsor

values,citationIDsforprior\-workstatements,andrequiredscope

qualifiers\.Donotwriteproseorscheduleunsupportedcontent\.

ReturnJSON:

\{

"sections":\[\{

"name":"<section\>",

"paragraphs":\[\{

"purpose":"<purpose\>",

"node\_ids":\["<nodeID\>"\],

"claim\_ids":\["<ClaimID\>"\],

"artifact\_ids":\["<artifactID\>"\],

"citation\_ids":\["<literatureID\>"\],

"required\_qualifiers":\["<qualifier\>"\]

\}\]

\}\],

"omitted\_content":\[\{

"source\_id":"<ID\>",

"reason":"<notneededornotgrounded\>"

\}\]

\}

\[DRAFTMODE\]

Inputs:<APPROVED\_SKELETON\>,<VALIDATED\_GRAPH\>,

<LITERATURE\_RECORDS\>,<EXPERIMENTAL\_ARTIFACTS\>,

<PAPER\_REQUIREMENTS\>\.

Expandtheapprovedskeletonwithoutaddingsubstantiveclaims\.

Reportvalues,units,datasets,settings,andcomparisondirections

exactlyasrecorded\.ReturnJSON:

\{

"manuscript":"<drafttext\>",

"provenance\_map":\[\{

"span":"<exactsubstantiveclaimorreportedvalue\>",

"claim\_ids":\["<ClaimID\>"\],

"artifact\_ids":\["<artifactID\>"\],

"citation\_ids":\["<literatureID\>"\],

"qualifiers\_preserved":true

\}\],

"unresolved\_skeleton\_items":\[\]

\}

\[REVISIONMODE\]

Inputs:<CURRENT\_DRAFT\>,<REVIEW\_ISSUES\>,<VALIDATED\_GRAPH\>,

<LITERATURE\_RECORDS\>,<EXPERIMENTAL\_ARTIFACTS\>\.

Applythesmallestchangethatresolveseachissue\.Arevisionmay

clarify,qualify,narrow,relocate,orremovetext\.Addorchange

contentonlywhensuppliedevidencesupportsit\.ReturnJSON:

\{

"manuscript":"<revisedtext\>",

"edit\_log":\[\{

"issue\_id":"<reviewissueID\>",

"old\_span":"<exactoldtext\>",

"new\_span":"<exactreplacementoremptyfordeletion\>",

"source\_ids":\["<graph,artifact,orliteratureID\>"\]

\}\],

"provenance\_map":\["<samerecordtypeasDRAFTmode\>"\],

"unresolved\_issues":\[\{

"issue\_id":"<reviewissueID\>",

"reason":"<whysuppliedevidencecannotresolveit\>"

\}\]

\}

### E\.2Paper Reviewer

\\begin\{lstlisting\}

\[ROLE\]

YouarethePaperReviewerinEviGraph\.Auditwriting\-level

weaknessesonly\.Treatthevalidatedgraph,artifacts,andsupplied

literatureasauthoritative\.Donotmodifythegraph,inventevidence,

oruseexternalknowledge\.

\[INPUTS\]

Draft:<MANUSCRIPT\_DRAFT\>

Validatedgraph:<VALIDATED\_GRAPH\>

Literaturerecords:<LITERATURE\_RECORDS\>

Experimentalartifacts:<EXPERIMENTAL\_ARTIFACTS\>

Paperrequirements:<PAPER\_REQUIREMENTS\>

\[CRITERIA\]

STRUCTURAL\_INTEGRITY:Theproblem,method,experiments,Findings,

limitations,andconclusionarecoherentandpreserverequiredlinks\.

NOVELTY\_FRAMING:NoveltyisscopedtosuppliedGapsandliterature

anddoesnotoverstatethecontrastwithpriorwork\.

CITATION\_CONSISTENCY:Everycitationexistsinthesuppliedrecords

andsupportsthepropositionforwhichitisused\.

GRAPH\_FAITHFULNESS:Everysubstantivecurrent\-runresearchclaim

preservesaretainedClaim’sscopeandqualifiersandhasacomplete

validatedchain;everyexperimentalstatementandvalueagreeswith

itsartifact\.

\[OUTPUT\]

Foreachissue,citetheexactdraftspanandrelevantsourceIDs\.

Setreadytotrueonlywhennoissueremains\.ReturnJSONonly:

\{

"ready":false,

"issues":\[\{

"issue\_id":"<uniqueID\>",

"criterion":"<onecriterionabove\>",

"section":"<draftsection\>",

"span":"<exactdrafttext\>",

"source\_ids":\["<graph,artifact,orliteratureID\>"\],

"diagnosis":"<conciseevidence\-groundeddiagnosis\>",

"revision\_instruction":"<smallesttargetedrevision\>"

\}\]

\}

The review loop is gated in the same way as graph repair\. A reviewer response with issues is passed to the Writer inREVISIONmode and then reviewed again\. Contract failures and review iterations consume the run budget and obey the manifest retry cap\. A draft is released only when the Reviewer returnsready=truewith an empty issue list and the provenance\-map validator resolves every substantive claim and reported value\. If a blocking issue remains when the cap is reached, the run returnsIncompleterather than emitting the unresolved draft as its final paper\.

## Appendix FReliability\-Evaluation Protocol and Prompts

The native ARC\-Bench\-ML and NanoResearch\-20 judges follow their official evaluation harnesses and are not reconstructed here\. ARC\-Bench\-ML uses its 25 machine\-learning topics and the25:25:5025\{:\}25\{:\}50weighting of Code Development, Code Execution, and Result Analysis\. Its strict score is produced by two independent agent reviewers, with score disagreements greater than0\.200\.20re\-adjudicated\. NanoResearch\-20 uses its official 20\-task, seven\-domain protocol and reports Alignment, Novelty, end\-to\-end completion, Performance, and Writing quality\. These native scores are kept separate from the reliability metrics below\.

### F\.1Common Run Configuration

The compared systems use the sameqwen\-3\.6\-plusbackbone, sandbox, and per\-experiment time budget; AutoResearchClaw is run in its full\-auto setting\. Table[8](https://arxiv.org/html/2608.04738#A6.T8)distinguishes values stated in the main paper from fields that must be frozen and exported with the run records\. The appendix does not infer unstated hardware, sampling, seed, or timeout values\.

Table 8:Reproduction contract for each evaluated run\. Unstated values must come from the recorded run manifest rather than post\-hoc reconstruction\.
### F\.2Construction of the Reliability Sets

The reliability evaluation has four stages\. First, a manuscript is serialized into section\-aware chunks with stable character offsets; tables and captions are serialized with their row, column, and caption context\. Chunk size, overlap, extractor model revision, decoding configuration, and random seed are frozen in the evaluation manifest and kept identical across systems\. Second, the extractors below produce candidate records for manuscript claims and reported experimental values\. Third, orchestration code validates and indexes the records, blinds system identity, and attaches normalized records from the corresponding research run\. Finally, the membership judges make categorical decisions and deterministic code computes the ratios\. Schema retries correct only malformed output; they do not resample a valid extraction in search of a more favorable denominator\.

Overlapping chunks can expose the same source occurrence more than once\. The indexer merges records only when their normalized manuscript source intervals are identical; semantically similar statements at different locations remain separate occurrences\. It splits a compound extraction into atomic records when the constituent propositions or values can receive different membership decisions\. Stable IDs are assigned after this validation, yielding𝒞\\mathcal\{C\}andℱ\\mathcal\{F\}\. This conservative policy avoids subjective paraphrase\-based denominator reduction and preserves the stochastic\-extraction caveat discussed in Section[4\.1](https://arxiv.org/html/2608.04738#S4.SS1)of the main paper\.

\\begin\{lstlisting\}

\[ROLE\]

Youareablindedclaimindexer\.Extractsubstantiveresearchclaims

fromonemanuscriptchunk\.Donotassesswhethertheyaresupported\.

\[INPUTS\]

ChunkIDandsourceoffsets:<CHUNK\_METADATA\>

Section\-awaremanuscriptchunk:<MANUSCRIPT\_CHUNK\>

\[ELIGIBILITY\]

Includeanexplicitorimplicitpropositionabouttheproblem,prior

work,method,experiment,observedresult,comparison,mechanism,

novelty,limitation,orconclusionwhenareadercouldreasonablyask

whatevidencesupportsit\.Excludeheadings,purenavigation,citation

tokensalone,acknowledgments,andstatementsthatonlydescribethe

paper’sorganization\.

\[RULES\]

1\.Preservepolarity,modality,scope,conditions,comparison,and

qualifiers\.Donotstrengthenornormalizeawayuncertainty\.

2\.Makeeachrecordanatomic,self\-containedproposition\.Split

conjunctionswhosepartscouldreceivedifferentsupportdecisions\.

3\.Citethesmallestexactsourcespanthatexpressestheclaim\.An

implicitclaimmayuseamulti\-sentencespanbutcannotrelyontext

outsidethischunk\.

4\.Donotuseexternalknowledge,inferevidence,mergerepetitions,

ordecidesupport\.

\[OUTPUT\]

ReturnJSONonly:

\{

"chunk\_id":"<chunkID\>",

"claims":\[\{

"local\_id":"<chunk\-localID\>",

"claim\_text":"<self\-containedproposition\>",

"claim\_type":"<problem\|prior\_work\|method\|experimental\|

comparative\|mechanism\|novelty\|limitation\|

conclusion\>",

"source\_span":"<exactmanuscripttext\>",

"start\_offset":"<absolutecharacteroffset\>",

"end\_offset":"<exclusiveabsolutecharacteroffset\>",

"qualifiers":\["<scopeoruncertaintyqualifier\>"\]

\}\]

\}

\\begin\{lstlisting\}

\[ROLE\]

Youareablindedexperimental\-valueindexer\.Extractreported

experimentalmeasurementsfromonemanuscriptchunk\.Donotdecide

whethertheymatchanexecutionrecord\.

\[INPUTS\]

ChunkIDandsourceoffsets:<CHUNK\_METADATA\>

Section\-,table\-,andcaption\-awarechunk:<MANUSCRIPT\_CHUNK\>

\[ELIGIBILITY\]

Includeeachscalarorcompactnumericresultpresentedasanobserved

experimentalmeasurement,score,difference,uncertainty,oraggregate\.

Excludecitationyears,section/tablenumbers,datasetsizes,budgets,

hyperparameters,andpredictedvaluesunlesstheyareexplicitly

reportedasmeasuredoutcomes\.

\[RULES\]

1\.Emitonerecordperindependentlycheckablevalue\.Preservethe

displayedstring,sign,precision,unitorscale,anduncertainty\.

2\.Recovermetric,methodorcondition,datasetandsplit,

aggregation,andexperimentalsettingonlyfromthesuppliedspan

anditsserializedtable/captioncontext\.Useunknownwhenabsent\.

3\.Citeanexactsourcespanandoffsets\.Donotperformunit

conversion,rounding,tolerancematching,orrecordlookup\.

\[OUTPUT\]

ReturnJSONonly:

\{

"chunk\_id":"<chunkID\>",

"values":\[\{

"local\_id":"<chunk\-localID\>",

"displayed\_value":"<verbatimnumericstring\>",

"parsed\_value":"<numericvalueornull\>",

"unit\_or\_scale":"<unit,percent,fraction,orunknown\>",

"metric":"<metricorunknown\>",

"method\_or\_condition":"<methodorconditionorunknown\>",

"dataset\_and\_split":"<datasetandsplitorunknown\>",

"aggregation":"<mean,median,singlerun,orunknown\>",

"experimental\_setting":"<settingorunknown\>",

"source\_span":"<exactmanuscripttext\>",

"start\_offset":"<absolutecharacteroffset\>",

"end\_offset":"<exclusiveabsolutecharacteroffset\>"

\}\]

\}

### F\.3Blinding, Membership, and Aggregation

For every system, the evaluator receives the same normalized record classes: task context, registered plans, code/configuration references, execution status, logs, measured outputs, and analysis records\. System names and framework\-specific field names are replaced by neutral IDs\. Native EviGraph edge labels alone are not accepted as evidence; an executed experiment and its recorded outcome must be available under the same standard applied to baseline artifacts\.

The numerical protocol supplied to the EDC judge is frozen before system outputs are inspected\. Table[9](https://arxiv.org/html/2608.04738#A6.T9)defines its default rules; any benchmark\-specific tolerance must be declared in the evaluation manifest before judging\.

Table 9:Frozen numerical matching protocol for EDC membership\.The two membership prompts below operate on the indexed sets𝒞\\mathcal\{C\}andℱ\\mathcal\{F\}\. The protocol definition, source indexing, blinding, and arithmetic are fixed outside the membership calls; the EDC judge receives the frozen protocol and applies it to each categorical comparison\. After validation, deterministic code constructs

𝒮\\displaystyle\\mathcal\{S\}=\{c∈𝒞:JCSR​\(c\)=SUPPORTED\},\\displaystyle=\\\{c\\in\\mathcal\{C\}:J\_\{\\mathrm\{CSR\}\}\(c\)=\\texttt\{SUPPORTED\}\\\},𝒦\\displaystyle\\mathcal\{K\}=\{f∈ℱ:JEDC​\(f\)=MATCH\},\\displaystyle=\\\{f\\in\\mathcal\{F\}:J\_\{\\mathrm\{EDC\}\}\(f\)=\\texttt\{MATCH\}\\\},and reportsCSR=\|𝒮\|/\|𝒞\|\\mathrm\{CSR\}=\|\\mathcal\{S\}\|/\|\\mathcal\{C\}\|andEDC=\|𝒦\|/\|ℱ\|\\mathrm\{EDC\}=\|\\mathcal\{K\}\|/\|\\mathcal\{F\}\|\. For a pooled cross\-benchmark result, the sets are unions of the per\-run indexed sets, so the reported rate is a micro\-average over eligible occurrences\. The raw counts\|𝒞\|,\|𝒮\|,\|ℱ\|,\|𝒦\|\|\\mathcal\{C\}\|,\|\\mathcal\{S\}\|,\|\\mathcal\{F\}\|,\|\\mathcal\{K\}\|accompany every aggregate\. A zero denominator is reported as undefined rather than as zero or one\.

\\begin\{lstlisting\}

\[ROLE\]

YouareablindedevaluatorofClaimSupportRate\(CSR\)\.Evaluate

onlytheindexedmanuscriptclaimsinthesuppliedsetC\.Useonly

evidenceproducedinthecorrespondingresearchrun\.

\[INPUTS\]

IndexedmanuscriptclaimsC:<MANUSCRIPT\_CLAIMS\>

Normalizedresearch\-runevidence:<RUN\_EVIDENCE\>

\[DECISIONRULE\]

Foreachclaim,returnSUPPORTEDonlywhenidentifiablerunrecords

traceablysupporttheclaimaswritten,includingitsscope,

conditions,direction,andcomparison\.Ahypothesis,intended

experiment,manuscriptassertion,orbaregraphsupportlabelwithout

theunderlyingexecutedexperimentandrecordedFindingis

insufficient\.Ifanyessentialpartisunsupported,contradicted,

ambiguous,orlackstraceableevidence,returnNOT\_SUPPORTED\.

Donotuseexternalknowledge,thesystemname,oranotherunsupported

manuscriptstatementasevidence\.GivesourceIDsandaconcisereason,

notanunrestrictedreasoningtrace\.

\[OUTPUT\]

ReturnJSONonly:

\{

"items":\[\{

"claim\_id":"<indexedclaimID\>",

"decision":"<SUPPORTED\|NOT\_SUPPORTED\>",

"evidence\_ids":\["<runrecordID\>"\],

"reason":"<conciseevidence\-basedjustification\>"

\}\]

\}

DonotcomputeCSR\.TheevaluationcodeconstructsSfromitems

labeledSUPPORTEDandcomputes\|S\|/\|C\|\.

\\begin\{lstlisting\}

\[ROLE\]

YouareablindedevaluatorofExperimentalDataConsistency\(EDC\)\.

EvaluateonlytheindexedreportedexperimentalvaluesinsetF\.

Useonlycorrespondingexecutionrecords\.

\[INPUTS\]

IndexedreportedvaluesF:<REPORTED\_VALUES\>

Normalizedexecutionrecords:<EXECUTION\_RECORDS\>

Frozenmatchingrules:<NUMERICAL\_MATCHING\_PROTOCOL\>

\[DECISIONRULE\]

Foreachvalue,returnMATCHonlywhenanidentifiableexecution

recordreportsthesamevalueforthesamemetric,methodor

condition,datasetandsplit,unitorscale,aggregation,and

experimentalsetting\.Applyonlyunitconversion,displayed\-value

rounding,ortoleranceexplicitlyallowedbythefrozenmatching

rules\.Iftherecordisabsent,ambiguous,orinconsistent,return

NOT\_MATCH\.

Donotuseexternalknowledge,thesystemname,oranothermanuscript

statementasevidence\.GiverecordIDsandaconcisereason\.

\[OUTPUT\]

ReturnJSONonly:

\{

"items":\[\{

"value\_id":"<indexedvalueID\>",

"decision":"<MATCH\|NOT\_MATCH\>",

"record\_ids":\["<executionrecordID\>"\],

"reason":"<conciserecord\-basedjustification\>"

\}\]

\}

DonotcomputeEDC\.TheevaluationcodeconstructsKfromitems

labeledMATCHandcomputes\|K\|/\|F\|\.

## Appendix GExpanded Representative Execution Trace

Table[10](https://arxiv.org/html/2608.04738#A7.T10)expands the control\-state transitions behind Figure[2](https://arxiv.org/html/2608.04738#S5.F2)and Section[5\.2](https://arxiv.org/html/2608.04738#S5.SS2)of the main paper\. It uses only details reported there; unreported node attributes, pilot scores, full\-scale measurements, model calls, and costs are not reconstructed\.

This example demonstrates weak\-root localization and downstream regeneration, but it does not empirically exercise rollback or long\-term retrieval: the first repair succeeds andℳL\\mathcal\{M\}\_\{L\}is empty\. Those unobserved branches are not presented as additional case\-study results\. Their specified control behavior is as follows: a degrading complete repair invokes the rollback ordering in Section[D](https://arxiv.org/html/2608.04738#A4); a blocked partial repair invokes the same ordering over its valid checkpoints; an invalid Problem or Gap anchor rebuildsG0G\_\{0\}; and an exhausted budget returns the current validated rollback state withIncomplete, without manuscript generation\.

Table 10:Control\-state expansion of the representative run in the main paper\.