| name | evidence-grounded-paper-deep-read |
| description | Generate detailed, evidence-grounded Markdown deep-reading reports from research paper PDFs. Use when Codex needs to analyze academic papers, extract figures/tables, explain formulas and technical terms for beginners, compare claims against paper evidence, or produce structured reading reports covering what the authors did, why they did it, how they did it, and how they validated it. |
Evidence-Grounded Paper Deep Read
Purpose
Create a detailed Markdown deep-reading report from a research paper PDF. The report must be specific enough that a beginner can understand the paper's motivation, method, formulas, figures, experiments, and technical terms without reading the original paper first.
Default to the user's language for the final report. If the user does not specify a language, use the language of the request.
Maintain strict language consistency. Section titles, table headers, checklist labels, image captions, and explanatory prose must use the report language. Keep English only for paper titles, author names, dataset names, model names, abbreviations, code identifiers, equations, and technical terms that are normally written in English.
Required Workflow
Never start by writing a polished summary. Work through these phases in order:
- Prepare a per-paper workspace.
- Extract text, figures, tables, captions, and formulas when possible. Prefer pdffigures2 for figure/table extraction. Use local fallback extraction and verified cropping only for visuals that pdffigures2 misses or renders incorrectly.
- Build a paper index that maps sections, claims, methods, formulas, figures, and experiments.
- Create evidence cards for all important claims and method details.
- Draft the report from the evidence cards.
- Insert figures with local relative Markdown paths.
- Run the quality checklist before final delivery.
Use this directory layout for each paper unless the user specifies another output structure:
paper-output/
source.pdf
extracted/
full_text.md
sections.json
figure_manifest.json
figures/
figure-001.png
figure-002.png
paper_index.json
evidence_cards.md
report.md
Tools And Resources
- Use
scripts/extract_pdf_text.py to extract page-level text into Markdown and JSON.
- Prefer the project-level pdffigures2 workflow for figure/table extraction. It should produce cropped figure/table images plus metadata containing captions, pages, and bounding boxes.
- Use
scripts/extract_pdf_figures.py to inspect embedded image objects and render page previews only when pdffigures2 misses a visual.
- Use
scripts/crop_pdf_region_verified.py as a fallback recovery tool for exact figure/table regions with safety margin and edge-tightness checks.
- Use
scripts/build_paper_index.py to create a paper index scaffold from extracted text and figures.
- Use
scripts/make_report_scaffold.py to create a report scaffold from assets/report_template.md.
- Use
scripts/validate_report.py to catch missing sections and weak report coverage.
- Read
references/reading_protocol.md before analyzing a paper.
- Read
references/report_schema.md before drafting the report.
- Read
references/figure_analysis_protocol.md before explaining figures and tables.
- Read
references/terminology_explanation.md when the paper uses specialized terms, formulas, or field-specific methods.
- Read
references/quality_checklist.md before claiming the report is complete.
The Python extraction scripts prefer PyMuPDF and automatically fall back to pdfplumber when available. If both extraction backends are unavailable, continue manually, but keep the same artifacts and report structure. Full-page previews are inspection aids only and should not be used as final report figures.
Evidence Rules
Treat the paper as the source of truth.
- Attach section names, page numbers, paragraph numbers, figure numbers, table numbers, or equation numbers to important claims whenever available.
- Do not invent motivations, baselines, datasets, metrics, ablations, or limitations.
- Mark uncertain items as
Not clearly stated in the paper instead of guessing.
- When a claim depends on interpretation, label it as an inference and explain the evidence behind it.
- Prefer concrete details over evaluative adjectives. Replace vague phrases such as "improves performance" with the exact mechanism, metric, dataset, and result when available.
Required Report Structure
The final report.md must use section titles in the report language. For Chinese reports, use:
# 论文精读报告:<paper title>
## 0. 一句话总览
## 1. 论文基本信息
## 2. 作者做了什么
## 3. 作者为什么要做这件事
## 4. 作者具体怎么做
## 5. 作者如何验证
## 6. 公式与关键技术细节
## 7. 创新点逐条拆解
## 8. 局限性与开放问题
## 9. 初学者背景补充
## 10. 复现与进一步阅读建议
## 11. 完整性自检
For English reports, translate the same structure fully into English. Do not mix labels such as WHAT, WHY, HOW, or EXPERIMENT into a Chinese report heading.
For each innovation point, include:
- The problem it addresses.
- The author's idea.
- The exact implementation or algorithmic procedure.
- The relevant formula, figure, or module if present.
- Why the design may work.
- How it differs from prior or baseline methods.
- Evidence location in the paper.
For each experiment, include:
- Research question tested by the experiment.
- Dataset or benchmark.
- Metric.
- Compared methods or baselines.
- Setup details.
- Main result.
- What the result proves and what it does not prove.
- Evidence location in the paper.
Figure Handling
Save cropped figures under extracted/figures/ or figures/ and reference them with relative paths:

Use pdffigures2 output as the first-choice visual source. Inspect its extracted images and metadata before writing. If pdffigures2 misses a figure/table or produces an incomplete crop, recover that visual with fallback page preview plus verified cropping. When cropping manually, add margin, inspect edge warnings, and recrop until the image contains all panels, labels, axes, legends, and caption elements needed for understanding.
Do not create a standalone figure/table explanation section. Integrate each figure or table into the section where it is used as evidence:
- Use task/problem definition figures in
作者做了什么.
- Use motivation or gap figures in
作者为什么要做这件事.
- Use architecture, algorithm, and formula figures in
作者具体怎么做.
- Use result tables, ablation tables, and qualitative examples in
作者如何验证.
For every major figure or table, explain naturally in the surrounding prose:
- What the figure/table shows.
- How to read each region, arrow, axis, row, or column.
- Which claim or method step it supports.
- What a beginner might misunderstand.
- Whether the visual evidence is strong, incomplete, or only illustrative.
Full-page previews are inspection aids, not final report figures. If pdffigures2 and fallback cropping both fail, mark the visual as an unresolved extraction limitation and do not pretend the page preview is the figure.
Formula Handling
For each important formula:
- Reproduce the formula if it can be extracted reliably.
- Define every symbol.
- Explain the computation in plain language.
- Explain where it appears in the method pipeline.
- Explain why the formula is needed.
- Give a toy intuitive example when useful.
- Mark any symbol that is not defined by the paper.
Batch Processing
When processing many papers, create one output folder per paper and keep a batch index:
reports/
batch_index.md
paper-001-short-title/
report.md
paper-002-short-title/
report.md
batch_index.md should list title, filename, status, report path, and any extraction problems.
Completion Standard
A report is not complete until it passes these checks:
- The task, motivation, method, and experiment sections are all substantive.
- All stated contributions are decomposed into problem, idea, implementation, and evidence.
- Major figures and tables are embedded or explicitly marked as unavailable.
- Figures/tables are precise crops and are integrated into the report's argument instead of isolated in a figure-reading section.
- Key formulas and technical terms are explained for beginners.
- Experiments include datasets, metrics, baselines, settings, and interpretation.
- The report distinguishes paper evidence from agent inference.
- The completeness checklist is filled honestly.