paper-reading

name	paper-reading
description	Use when Codex needs to read open-access paper full text for Paper Radar, extract methods/evidence/limitations, identify affiliations, and find core figure or table candidates.

Paper Reading

Read selected papers closely using legal open sources.

Allowed sources:

• arXiv PDF or HTML
• OpenReview PDF or forum page
• author/lab/university open PDFs
• official conference open pages
• publisher pages that are explicitly open access

Do not:

• bypass paywalls
• use pirate paper sites
• save PDFs or unrelated large binary files to the repo
• copy long passages
• pretend a paper was fully read when only metadata was available

For each selected paper produce an internal reading note with:

• source_fulltext_url
• read_status: fulltext_read, partial_read, or metadata_only
• one_sentence_idea
• core_question
• method: concrete mechanism, pipeline, training signal, data construction, or evaluation design
• evidence: datasets, benchmarks, baselines, ablations, key tables/curves, and any source-backed numbers
• contribution
• limitation
• institutions
• figures: 1-3 candidates with label, image_url if reliable, otherwise source_hint, and caption_summary
• visual_claims: what each figure/table supports and what should be interpreted carefully
• relevance
• confidence

Figure guidance:

• Prefer the main figure or method overview first, then key result tables/curves, then important ablations or case studies.
• If a figure/table is central to understanding the paper, render the open PDF/HTML page or make a local screenshot/page crop to inspect it closely.
• Important screenshots may be committed only under images/paper-radar/YYYY-MM-DD-HHMM/.
• Use short, lowercase, hyphenated filenames such as futureworld-main-figure.png.
• Do not commit whole PDFs, unrelated screenshots, or oversized raw images.
• If arXiv HTML exposes stable figure image URLs, use remote URLs in the final Markdown.
• If no reliable image URL is available, write a short figure pointer such as Figure 2: ... or Table 1: ..., and summarize what the figure shows based on the temporary inspection.
• Never fabricate figures or captions.

Writing handoff:

• Do not expose raw read_status labels in the blog.
• If the reading was partial, explain the limitation naturally.
• Hand off enough detail for a mini explainer: the reader should grasp the core idea quickly, then see how the method works and what evidence actually supports it.
• Keep methods and evidence specific enough to show close reading; do not hand off abstract paraphrases as if they were analysis.