Extending AI for Research to the Humanities: A Multi-Agent Framework for Evidence-Grounded Scholarship

Jiajun Zhang; Jun Wang; Qi Su; Yating Pan

arxiv: 2605.30947 · v3 · pith:AO4QYAMRnew · submitted 2026-05-29 · 💻 cs.CL

Extending AI for Research to the Humanities: A Multi-Agent Framework for Evidence-Grounded Scholarship

Yating Pan , Jiajun Zhang , Jun Wang , Qi Su This is my paper

Pith reviewed 2026-06-28 22:30 UTC · model grok-4.3

classification 💻 cs.CL

keywords multi-agent systemsevidence-grounded reasoninghumanities scholarshipscholarly primitivesclassical Chinese textsGreco-Roman Latinretrieval-augmented generationprimary source citation

0 comments

The pith

A multi-agent framework assigns humanities operations like evidence annotation and provenance checking to cooperating agents over passages, graph communities, and semantic clusters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents SPIRE as a way to adapt AI research agents from experimental fields to interpretive humanities work. It maps standard scholarly steps onto distinct agent roles that operate together on a layered reading substrate. A sympathetic reader would care because humanities arguments depend on traceable quotes and close reading rather than quantitative experiments. If the mapping holds, the result is essays that cite primary sources more accurately than standard retrieval methods. Ablation results indicate that both the agent roles and the substrate layers contribute to the observed gains.

Core claim

SPIRE draws on Scholarly Primitives theory to treat operations such as source discovery, evidence annotation, comparison, provenance checking, sampling, citation binding, and argumentative synthesis as cooperating agent roles. These roles act over a multi-scale close-reading substrate consisting of passages, intra-context graph communities, and cross-context semantic clusters. On a benchmark of peer-reviewed papers in classical Chinese and Greco-Roman Latin scholarship, the system recovers cited primary-source evidence more reliably than Naive LLM, Text RAG, and GraphRAG baselines and earns higher blind-judge scores for answer accuracy, depth, coverage, and evidence quality.

What carries the argument

SPIRE multi-agent framework that decomposes humanities scholarship into cooperating agent roles over a multi-scale close-reading substrate of passages, graph communities, and semantic clusters.

If this is right

Evidence recovery and answer quality improve when scholarly operations are split across specialized agents rather than handled by a single model.
Both the division into agent roles and the layered retrieval substrate are necessary for the performance gains observed on the benchmark.
The approach produces outputs that score higher on accuracy, depth, coverage, and evidence quality under blind evaluation.
The framework can be applied to other bodies of classical scholarship that rely on close reading of primary sources.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same role division could be tested on modern-language literary criticism where provenance of quotations is equally central.
Adding explicit citation-binding agents might reduce hallucinated references even in non-humanities domains that still require source tracing.
If the substrate layers prove portable, similar multi-scale structures could be built for legal or historical argument without requiring new model training.

Load-bearing premise

Mapping humanities operations onto cooperating agent roles over a multi-scale substrate of passages, graphs, and clusters will produce faithful interpretive reasoning and verifiable evidence grounding.

What would settle it

A replication on the same classical Chinese and Greco-Roman Latin benchmark in which SPIRE no longer shows higher evidence recovery rates or blind-judge scores than the three baselines after the agent-role division is removed.

Figures

Figures reproduced from arXiv: 2605.30947 by Jiajun Zhang, Jun Wang, Qi Su, Yating Pan.

**Figure 1.** Figure 1: Overview of SPIRE: compiles classics into a multi-scale close-reading store, answers a research question with seven scholarly-primitive agents over an EvidencePool, and evaluates on a peer-reviewed-paper benchmark. weigh passages for and against a claim; and original interpretation, where argumentative logic itself is the contribution. It allows a wider range of well-argued readings, but still requires f… view at source ↗

**Figure 2.** Figure 2: SPIRE on a concrete cross-tradition run (Cicero’s De Re Publica vs. the Analects). Given Q, Discovering (D) emits a query plan π(Q) = ({qi}, lang, c, anchors, hints): it anchors the works, authors, and language in Q against corpus metadata, sets the comparison type c ∈ {none, intra-text, intra-author, intra-tradition, crosstradition}, and decomposes Q into sub-questions qi tagged by granularity and perspe… view at source ↗

**Figure 3.** Figure 3: Inter-rater agreement. (a) Gwet’s AC2 vs. quadratic-weighted Cohen’s κ (mean over the four aspects) per rater pair: the two coefficients nearly coincide for LLM–LLM (scores spread across systems) but diverge for the human-involving pairs (top-heavy scores), the kappa paradox. (b) Raw 5 × 5 joint-score counts per rater pair (free track; human-involving pairs on SPIRE+Naive LLM). Mass on the diagonal is agre… view at source ↗

read the original abstract

LLM-based research agents have advanced rapidly in science and engineering, where research is organized around executable experiments, code, and quantitative signals. Humanities scholarship, however, requires a different mode of reasoning: interpretive, evidence-grounded argument over primary sources, where scholarly value depends on faithful quotation, verifiable provenance, and close reading. Existing research agents remain largely optimized for execution and retrieval, not evidence-grounded interpretive reasoning. To address this gap, we introduce SPIRE (Scholarly-Primitives-Inspired Research Engine), a multi-agent framework for evidence-grounded humanities scholarship. Drawing on Scholarly Primitives theory, SPIRE casts recurring humanities operations as cooperating agent roles (source discovery, evidence annotation, comparison, provenance checking, sampling, citation binding, and argumentative synthesis) over a multi-scale close-reading substrate of passages, intra-context graph communities, and cross-context semantic clusters. On a peer-reviewed-paper benchmark over classical Chinese and Greco-Roman Latin scholarship, SPIRE recovers cited primary-source evidence more reliably than Naive LLM, Text RAG, and GraphRAG, and receives higher blind-judge scores on answer accuracy, depth, coverage, and evidence quality. Ablations show that both the scholarly-operation agents and close-reading retrieval contribute to evidence-grounded essays. Code, data catalogues, and reproduction scripts are released at https://github.com/YatingPan/SPIRE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SPIRE maps scholarly primitives to agent roles over multi-scale retrieval for humanities evidence work, but the evaluation claims rest on thin details and may mainly test citation lookup.

read the letter

The main point is that this paper introduces SPIRE, a multi-agent framework that turns humanities operations like evidence annotation, provenance checking, and synthesis into cooperating agent roles, layered on top of passages, graph communities, and semantic clusters for close reading.

It does a few things right. The design draws directly from Scholarly Primitives theory instead of just bolting agents onto generic RAG. Testing on classical Chinese and Greco-Roman Latin texts shows some attention to real interpretive domains rather than staying in English science abstracts. Releasing the code, data catalogues, and reproduction scripts at the GitHub link is useful for anyone who wants to check the substrate or try the roles themselves.

The soft spots sit in the evaluation. The abstract states that SPIRE recovers cited primary-source evidence more reliably than the three baselines and scores higher on blind-judge metrics for accuracy, depth, coverage, and evidence quality. Yet it supplies no numbers, no statistical tests, no benchmark construction details, and no ablation tables beyond the claim that both the agents and the retrieval help. That makes the performance edge hard to weigh. The stress-test concern also lands: if the test questions are drawn from the same peer-reviewed papers that already cite the target passages, the task may reduce to better retrieval of pre-linked evidence rather than testing whether the system can produce new, verifiable interpretive arguments from scratch. The abstract does not clarify the sampling protocol or whether the citing context was withheld.

This paper is for people working on AI systems for digital humanities who need a concrete starting point with defined roles and a released implementation. Readers who want to experiment with multi-agent setups on interpretive tasks will find usable pieces even if they have to strengthen the metrics themselves.

It deserves peer review. The theoretical grounding and code release give it enough substance for referees to evaluate the benchmark design and ask for the missing numbers.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces SPIRE, a multi-agent framework for evidence-grounded humanities scholarship that maps Scholarly Primitives operations (source discovery, evidence annotation, comparison, provenance, sampling, citation binding, and synthesis) to cooperating agent roles over a multi-scale close-reading substrate of passages, intra-context graph communities, and cross-context semantic clusters. On a benchmark derived from peer-reviewed papers in classical Chinese and Greco-Roman Latin scholarship, the paper claims SPIRE recovers cited primary-source evidence more reliably than Naive LLM, Text RAG, and GraphRAG baselines and receives higher blind-judge scores on accuracy, depth, coverage, and evidence quality. Ablations indicate that both the scholarly-operation agents and close-reading retrieval contribute to performance. Code, data catalogues, and reproduction scripts are released.

Significance. If the performance claims are substantiated with detailed metrics and a benchmark that tests interpretive synthesis, the work could meaningfully extend AI research agents beyond quantitative domains into evidence-grounded interpretive reasoning in the humanities. The grounding in established scholarly primitives theory and the public release of code and reproduction materials are explicit strengths that support verifiability and extension by others.

major comments (2)

[Evaluation section] Evaluation section: The abstract asserts superior recovery of primary-source evidence and higher judge scores, but provides no quantitative metrics (e.g., recovery rates or F1), statistical tests, benchmark construction details, or ablation numbers, leaving the central empirical claim with limited verifiable support.
[Benchmark description] Benchmark description: The protocol for sampling test cases from the peer-reviewed papers is unspecified, including whether questions were blinded to the citing context or require the full agent pipeline; this is load-bearing because it determines whether gains reflect citation retrieval of pre-linked passages or genuine advances in novel evidence-grounded argumentation over the multi-scale substrate.

minor comments (1)

[Abstract] Abstract: The invented term 'multi-scale close-reading substrate' is used without a diagram or formal definition of how passages, graph communities, and semantic clusters interact; adding a figure would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and for identifying areas where the empirical claims require greater explicit support. We address both major comments by committing to targeted revisions that add the requested quantitative details and protocol clarifications without altering the core claims or experimental design.

read point-by-point responses

Referee: [Evaluation section] Evaluation section: The abstract asserts superior recovery of primary-source evidence and higher judge scores, but provides no quantitative metrics (e.g., recovery rates or F1), statistical tests, benchmark construction details, or ablation numbers, leaving the central empirical claim with limited verifiable support.

Authors: We agree that the abstract summarizes results at a high level without numbers. The evaluation section (Section 4) already contains the supporting tables, but to make the central claims immediately verifiable we will revise the abstract to report key metrics (recovery rate, F1, judge scores) and statistical tests (paired t-tests with p-values). We will also add an explicit pointer to the full ablation tables and benchmark-construction appendix. These changes strengthen verifiability while preserving the existing experimental results. revision: yes
Referee: [Benchmark description] Benchmark description: The protocol for sampling test cases from the peer-reviewed papers is unspecified, including whether questions were blinded to the citing context or require the full agent pipeline; this is load-bearing because it determines whether gains reflect citation retrieval of pre-linked passages or genuine advances in novel evidence-grounded argumentation over the multi-scale substrate.

Authors: We acknowledge the need for explicit protocol details. In the revised manuscript we will insert a new subsection (3.3) that fully specifies the sampling procedure: questions were derived from the scholarly papers but formulated independently of the original citing sentences; the test set was blinded to citation links; and each question requires the complete multi-agent pipeline (discovery through synthesis) rather than retrieval of pre-linked passages. We will also provide concrete examples and a flowchart to demonstrate that performance gains reflect novel evidence-grounded argumentation over the multi-scale substrate. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical claims rest on external benchmark

full rationale

The paper introduces SPIRE as a multi-agent system mapping humanities operations to agent roles over a multi-scale substrate, then reports empirical gains on a benchmark of external peer-reviewed papers in classical Chinese and Greco-Roman Latin scholarship. No equations, fitted parameters, or self-referential definitions appear in the provided abstract or description. Performance is compared against independent baselines (Naive LLM, Text RAG, GraphRAG) with blind-judge metrics; the benchmark instances are drawn from published scholarship rather than quantities defined by the authors' own inputs or prior self-citations. The design draws on Scholarly Primitives theory without the central results reducing to ansatzes, uniqueness theorems, or renamings internal to the paper. This is a standard non-circular empirical framing.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that Scholarly Primitives theory supplies a complete and sufficient set of operations for evidence-grounded humanities reasoning; the framework itself is the primary invented contribution with no independent falsifiable evidence supplied in the abstract.

axioms (1)

domain assumption Scholarly Primitives theory identifies recurring operations in humanities scholarship that can be directly cast as cooperating agent roles
Invoked to justify the design of source discovery, evidence annotation, provenance checking, and argumentative synthesis agents.

invented entities (1)

multi-scale close-reading substrate (passages, intra-context graph communities, cross-context semantic clusters) no independent evidence
purpose: To enable evidence grounding at multiple levels of textual context
New retrieval structure introduced as part of SPIRE; no independent evidence outside the framework is provided.

pith-pipeline@v0.9.1-grok · 5778 in / 1436 out tokens · 33092 ms · 2026-06-28T22:30:29.593594+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 5 canonical work pages · 2 internal anchors

[1]

Openscholar: Synthesizing scientific literature with retrieval-augmented lms, 2024

Out of one, many: Using language mod- els to simulate human samples.Political Analysis, 31(3):337–351. David Armitage. 2012. What’s the big idea? intellectual history and the longue durée.History of European Ideas, 38(4):493–507. Akari Asai, Jacqueline He, Rulin Shao, Weijia Shi, Amanpreet Singh, Joseph Chee Chang, Kyle Lo, Luca Soldaini, Sergey Feldman, ...

work page arXiv 2012
[2]

Andong Chen, Lianzhang Lou, Kehai Chen, Xuefeng Bai, Yang Xiang, Muyun Yang, Tiejun Zhao, and Min Zhang

Augmenting large language models with chem- istry tools.Nature Machine Intelligence, 6:525–535. Andong Chen, Lianzhang Lou, Kehai Chen, Xuefeng Bai, Yang Xiang, Muyun Yang, Tiejun Zhao, and Min Zhang. 2025a. Benchmarking LLMs for translating classical Chinese poetry: Evaluating adequacy, flu- ency, and elegance. InProceedings of the 2025 Con- ference on E...

work page arXiv 2025
[3]

InAdvances in Neural Information Process- ing Systems

HyperGraphRAG: Retrieval-augmented gener- ation via hypergraph-structured knowledge represen- tation. InAdvances in Neural Information Process- ing Systems. Leland McInnes, John Healy, and Steve Astels. 2017. hdbscan: Hierarchical density based clustering.Jour- nal of Open Source Software, 2(11):205. Stuart M. McManus, Yuji Li, Leo Tam, Shuyang Qiu, Songy...

2017
[4]

AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

AgentSociety: Large-scale simulation of LLM-driven generative agents advances understand- ing of human behaviors and society.arXiv preprint arXiv:2502.08691. Geoffrey Rockwell and Stéfan Sinclair. 2016. Hermeneutica: Computer-Assisted Interpretation in the Humanities. MIT Press. Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej ...

work page internal anchor Pith review Pith/arXiv arXiv 2016
[5]

William F

Association for Computational Linguistics. William F. Shen, Xinchi Qiu, Chenxi Whitehouse, Lisa Alazraki, Shashwat Goel, Francesco Barbieri, Timon Willi, Akhil Mathur, and Ilias Leontiadis. 2026. Re- thinking rubric generation for improving LLM judge and reward modeling for open-ended tasks.Preprint, arXiv:2602.05125. Noah Shinn, Federico Cassano, Edward ...

work page arXiv 2026
[6]

close read- ing

Reflexion: Language agents with verbal rein- forcement learning. InAdvances in Neural Informa- tion Processing Systems. Quentin Skinner. 1969. Meaning and understanding in the history of ideas.History and Theory, 8(1):3–53. Barbara Herrnstein Smith. 2016. What was “close read- ing”? a century of method in literary studies.The Minnesota Review, 2016(87):57...

1969
[7]

AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

Autogen: Enabling next-gen llm applica- tions via multi-agent conversation.arXiv preprint arXiv:2308.08155. Qiankun Xu, Yutong Liu, Dongbo Wang, and Shuiqing Huang. 2025. Automatic recognition of cross- language classic entities based on large language models.npj Heritage Science, 13:59. Yue Yang, Yinzhi Xu, Chenghao Huang, JohnMichael Jurgensen, Han Hu, ...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[8]

high agreement, low kappa

From automation to autonomy: A survey on large language models in scientific discovery. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 17733–17750, Suzhou, China. Association for Com- putational Linguistics. Gucheng Zhou, editor. 1992.Zhongguo Xueshu Mingzhu Tiyao (Compendium of Chinese Academic Masterwork...

2025
[9]

INTERPRETATION: what the passage tells us in relation to the question (1-2 sentences of interpretive significance, not a summary)
[10]

SUB_QUESTION: which sub-question id (q1, q2, ...) it best bears on; "general" if several or none
[11]

CONNECTIONS: references it makes to other works/authors/concepts that also appear elsewhere in the batch
[12]

annotations

RELEVANCE_RANK: 1-5 (5 = directly addresses the core with specific textual evidence; 1 = tangential background). Output VALID JSON only: a single object with an "annotations" array of {text_unit_id, interpretation, sub_question, connections[], relevance_rank}. Do NOT return a bare array. Comparing – system prompt You are performing a scholarly comparison ...
[13]

Chinese e.g

Concept -- a recurring intellectual touchstone in a tradition (concept, doctrine, motif, value, or problem recurring across texts). Chinese e.g. ren, yi, li, dao, tian, ming, xing-shan, xing-e, si-duan, gewu-zhizhi, wu-wei, zhongyong, kong, yuanqi, fo-xing. Latin/Greco-Roman e.g. iustitia, virtus, fides, pietas, officium, prudentia, ratio, logos, natura, ...
[14]

Person -- a named individual who proposes, defines, debates, transmits, criticizes, or exemplifies a Concept (Confucius, Mencius, Zhu Xi, Cicero, Seneca, Augustine, Aristotle)
[15]

entities

Work -- a named text/treatise/poem/collection that carries, defines, records, or debates a Concept (Lunyu, Mengzi, Daxue, De Officiis). === ENTITY FIELDS (all required) === type (Concept|Person|Work); domain (Ethics & Morality | Politics & Governance | Religion & Spirituality | Metaphysics & Epistemology | Natural Philosophy & Science | Literature & Aesth...
[16]

AnswerAccuracy -- does the essay precisely address the relation the question asks (how / why / in what sense / through what mechanism) with textually sound claims, not merely list related concepts? 1 = off-topic / contradicts the texts; 2 = misreads the core concept; 3 = relevant but generic or half-answers the key relation 28 (topical adequacy CAPS HERE)...
[17]

ArgumentDepth -- is the argument driven by specific textual detail and developed in layers, not paraphrastic summary? 1 = no real argument; 2 = assertion/paraphrase only; 3 = a thesis but mainly generalisation or restatement (SUMMARY CAPS HERE); 4 = structured multi-step argument, several moves text-driven; 5 = layered analysis deriving claims from close ...
[18]

CoverageCompleteness -- does it cover the sub-tasks the question entails? (A comparison requires both sides + similarity + difference + the limits of the comparison.) 1 = misses major dimensions; 2 = one side only; 3 = main question answered but a key sub-dimension thin (a single general point CAPS HERE); 4 = main dimensions covered, minor omissions; 5 = ...
[19]

AnswerAccuracy

EvidenceQuality -- how well is the answer grounded in specific primary-text evidence: density of direct quotation, specificity of citation, close reading of the source material, versus vague work-name dropping or hand-waving paraphrase? (Faithfulness floor: treat a citation as fabricated only if the quoted words clearly do not belong to the cited work or ...

[1] [1]

Openscholar: Synthesizing scientific literature with retrieval-augmented lms, 2024

Out of one, many: Using language mod- els to simulate human samples.Political Analysis, 31(3):337–351. David Armitage. 2012. What’s the big idea? intellectual history and the longue durée.History of European Ideas, 38(4):493–507. Akari Asai, Jacqueline He, Rulin Shao, Weijia Shi, Amanpreet Singh, Joseph Chee Chang, Kyle Lo, Luca Soldaini, Sergey Feldman, ...

work page arXiv 2012

[2] [2]

Andong Chen, Lianzhang Lou, Kehai Chen, Xuefeng Bai, Yang Xiang, Muyun Yang, Tiejun Zhao, and Min Zhang

Augmenting large language models with chem- istry tools.Nature Machine Intelligence, 6:525–535. Andong Chen, Lianzhang Lou, Kehai Chen, Xuefeng Bai, Yang Xiang, Muyun Yang, Tiejun Zhao, and Min Zhang. 2025a. Benchmarking LLMs for translating classical Chinese poetry: Evaluating adequacy, flu- ency, and elegance. InProceedings of the 2025 Con- ference on E...

work page arXiv 2025

[3] [3]

InAdvances in Neural Information Process- ing Systems

HyperGraphRAG: Retrieval-augmented gener- ation via hypergraph-structured knowledge represen- tation. InAdvances in Neural Information Process- ing Systems. Leland McInnes, John Healy, and Steve Astels. 2017. hdbscan: Hierarchical density based clustering.Jour- nal of Open Source Software, 2(11):205. Stuart M. McManus, Yuji Li, Leo Tam, Shuyang Qiu, Songy...

2017

[4] [4]

AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

AgentSociety: Large-scale simulation of LLM-driven generative agents advances understand- ing of human behaviors and society.arXiv preprint arXiv:2502.08691. Geoffrey Rockwell and Stéfan Sinclair. 2016. Hermeneutica: Computer-Assisted Interpretation in the Humanities. MIT Press. Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej ...

work page internal anchor Pith review Pith/arXiv arXiv 2016

[5] [5]

William F

Association for Computational Linguistics. William F. Shen, Xinchi Qiu, Chenxi Whitehouse, Lisa Alazraki, Shashwat Goel, Francesco Barbieri, Timon Willi, Akhil Mathur, and Ilias Leontiadis. 2026. Re- thinking rubric generation for improving LLM judge and reward modeling for open-ended tasks.Preprint, arXiv:2602.05125. Noah Shinn, Federico Cassano, Edward ...

work page arXiv 2026

[6] [6]

close read- ing

Reflexion: Language agents with verbal rein- forcement learning. InAdvances in Neural Informa- tion Processing Systems. Quentin Skinner. 1969. Meaning and understanding in the history of ideas.History and Theory, 8(1):3–53. Barbara Herrnstein Smith. 2016. What was “close read- ing”? a century of method in literary studies.The Minnesota Review, 2016(87):57...

1969

[7] [7]

AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

Autogen: Enabling next-gen llm applica- tions via multi-agent conversation.arXiv preprint arXiv:2308.08155. Qiankun Xu, Yutong Liu, Dongbo Wang, and Shuiqing Huang. 2025. Automatic recognition of cross- language classic entities based on large language models.npj Heritage Science, 13:59. Yue Yang, Yinzhi Xu, Chenghao Huang, JohnMichael Jurgensen, Han Hu, ...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[8] [8]

high agreement, low kappa

From automation to autonomy: A survey on large language models in scientific discovery. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 17733–17750, Suzhou, China. Association for Com- putational Linguistics. Gucheng Zhou, editor. 1992.Zhongguo Xueshu Mingzhu Tiyao (Compendium of Chinese Academic Masterwork...

2025

[9] [9]

INTERPRETATION: what the passage tells us in relation to the question (1-2 sentences of interpretive significance, not a summary)

[10] [10]

SUB_QUESTION: which sub-question id (q1, q2, ...) it best bears on; "general" if several or none

[11] [11]

CONNECTIONS: references it makes to other works/authors/concepts that also appear elsewhere in the batch

[12] [12]

annotations

RELEVANCE_RANK: 1-5 (5 = directly addresses the core with specific textual evidence; 1 = tangential background). Output VALID JSON only: a single object with an "annotations" array of {text_unit_id, interpretation, sub_question, connections[], relevance_rank}. Do NOT return a bare array. Comparing – system prompt You are performing a scholarly comparison ...

[13] [13]

Chinese e.g

Concept -- a recurring intellectual touchstone in a tradition (concept, doctrine, motif, value, or problem recurring across texts). Chinese e.g. ren, yi, li, dao, tian, ming, xing-shan, xing-e, si-duan, gewu-zhizhi, wu-wei, zhongyong, kong, yuanqi, fo-xing. Latin/Greco-Roman e.g. iustitia, virtus, fides, pietas, officium, prudentia, ratio, logos, natura, ...

[14] [14]

Person -- a named individual who proposes, defines, debates, transmits, criticizes, or exemplifies a Concept (Confucius, Mencius, Zhu Xi, Cicero, Seneca, Augustine, Aristotle)

[15] [15]

entities

Work -- a named text/treatise/poem/collection that carries, defines, records, or debates a Concept (Lunyu, Mengzi, Daxue, De Officiis). === ENTITY FIELDS (all required) === type (Concept|Person|Work); domain (Ethics & Morality | Politics & Governance | Religion & Spirituality | Metaphysics & Epistemology | Natural Philosophy & Science | Literature & Aesth...

[16] [16]

AnswerAccuracy -- does the essay precisely address the relation the question asks (how / why / in what sense / through what mechanism) with textually sound claims, not merely list related concepts? 1 = off-topic / contradicts the texts; 2 = misreads the core concept; 3 = relevant but generic or half-answers the key relation 28 (topical adequacy CAPS HERE)...

[17] [17]

ArgumentDepth -- is the argument driven by specific textual detail and developed in layers, not paraphrastic summary? 1 = no real argument; 2 = assertion/paraphrase only; 3 = a thesis but mainly generalisation or restatement (SUMMARY CAPS HERE); 4 = structured multi-step argument, several moves text-driven; 5 = layered analysis deriving claims from close ...

[18] [18]

CoverageCompleteness -- does it cover the sub-tasks the question entails? (A comparison requires both sides + similarity + difference + the limits of the comparison.) 1 = misses major dimensions; 2 = one side only; 3 = main question answered but a key sub-dimension thin (a single general point CAPS HERE); 4 = main dimensions covered, minor omissions; 5 = ...

[19] [19]

AnswerAccuracy

EvidenceQuality -- how well is the answer grounded in specific primary-text evidence: density of direct quotation, specificity of citation, close reading of the source material, versus vague work-name dropping or hand-waving paraphrase? (Faithfulness floor: treat a citation as fabricated only if the quoted words clearly do not belong to the cited work or ...