A structured literature workspace for coding-agent research, focused on 2025–2026 with selected earlier work. It keeps source PDFs, retrieval status, thematic indexes, and evidence notes together so a research question can be traced back to its sources.

Paper catalog · Topic index · Evidence notes · Unresolved sources

Start with a research question

1. Read the topic map.
2. Select the relevant topic and inspect its entries in papers/catalog.jsonl.
3. Read the original paper before adding a claim to papers/notes.md.
4. Use papers/papers-manifest.csv when you need retrieval status or source provenance.

This public repository is a literature and evidence workspace. A paper's presence in the catalog does not mean its results have been reproduced, and this checkout does not contain the newer local benchmark implementation suites.

Repository map

Location	Purpose
papers/pdfs/	Archived source PDFs
papers/paper_titles.txt	Targeted paper titles
papers/papers-manifest.csv	Machine-readable retrieval status
papers/catalog.jsonl	Compact paper catalog
papers/theme-overrides.tsv	Curated topic assignments
papers/agent-view/	Topic-based navigation
papers/notes.md	Evidence notes and hypotheses
papers/unresolved.txt	Sources needing manual follow-up
scripts/fetch_papers.sh	ArXiv-first source retrieval and manifest updates
scripts/build_agent_index.sh	Catalog and navigation generation

Add or refresh sources

# Add the target title to papers/paper_titles.txt, then:
bash scripts/fetch_papers.sh
bash scripts/build_agent_index.sh

Inspect the manifest and unresolved-source list after retrieval. Record the source, what was verified, and the limits of the interpretation in the evidence notes. See paper-folder guidance before changing the archive.

Contributor and agent workflow

How coding agents should work in this repo

1. Read the workflow files first:
   • papers/papers-manifest.csv to see current state.
   • papers/unresolved.txt for remaining manual targets.
   • papers/notes.md for current hypotheses and what has already been tried.
2. Choose one narrow task (for example: add 10 titles, clean up unresolved items, summarize new evidence).
3. Make changes only within your assigned scope.
4. Re-run:
   • bash scripts/fetch_papers.sh
5. Record what changed in papers/notes.md when you make evidence or interpretation updates.

Multi-agent parallel work contract

To reduce merge conflicts and duplicate effort:

• Use branch-per-agent or branch-per-task and keep changes scoped.
• One agent = one slice:
  • ingestion
  • metadata cleanup
  • evidence note synthesis
  • script/automation improvements
• Do not edit the same file in the same semantic section at the same time.
• If touching the same file, each agent must coordinate section ownership (for example: one agent updates only line-level entries in papers-manifest.csv, another only papers/notes.md).
• Before taking a task, update your active claim in papers/notes.md.
• After completion, remove your claim and add a short handoff with open items.

Frequent commit policy (required)

All work should be committed frequently:

• Minimum frequency: at least one commit every 30 minutes of active work, or sooner for any meaningful atomic change.
• Commit unit: one logical change only (e.g., “add titles + fetch PDFs” or “update notes for benchmark evidence”).
• Message format: include intent and scope in one line (e.g., Ingest: add papers and refresh manifest).
• Never mix evidence claims with file-format refactors in the same commit.
• Commit before context switch and before handing work to another agent.

Quality rules for all edits

• Keep changes evidence-oriented and minimal.
• Prefer deterministic artifacts (CSV entries, file paths, command outputs) over prose-only claims.
• Avoid editing papers/pdfs/* manually; only the fetch script or explicit manual archival actions should modify PDFs.
• When uncertain about a paper source or metadata, keep it in unresolved.txt and annotate reason in notes.

Repo-facing workflow for adding one paper

1. Add a title line to papers/paper_titles.txt.
2. Run bash scripts/fetch_papers.sh.
3. Confirm status in papers/papers-manifest.csv.
4. If not_found_on_arxiv, add the target venue/URL as a follow-up in unresolved.txt.
5. Document what was verified (or failed) in papers/notes.md if this is an evidence-bearing update.

Fast path for multiple agents

Use this read order before editing:

1. papers/agent-view/README.md (tree + navigation)
2. papers/catalog.jsonl (compact metadata)
3. Exactly one file in papers/agent-view/topics/

Avoid loading papers/papers-manifest.csv unless needed for raw ingestion debugging.

_{Original README preserved for reference.}