Triage Paper

Add a new academic paper to the agentic-context research repo as a structured reference summary.

When to Use

• User provides an arxiv ID (e.g. 2310.08560), arxiv URL, or paper PDF path
• User says "triage this paper", "add this paper", or "analyse this paper"
• Evaluating whether a paper belongs in the repo

When Not to Use

• The paper has already been triaged (check REVIEWED.md first)
• The paper is clearly out of scope (not related to context window management)
• User wants a full deep-dive analysis — use triage-paper first, then promote to ANALYSIS-*.md

Mindset

Triage is a quality gate, not a data-entry task. The goal is a scannable, honest record.

• Evidence first: quote what the paper reports; never infer or embellish claims.
• Triage-then-promote: every paper enters via REVIEWED.md; promotion to ANALYSIS-*.md requires a deliberate user decision — never automatic.
• Scope over completeness: a well-reasoned rejection is as valuable as a full summary. If the paper is tangentially related, triage it and flag it; don't silently skip it.

Workflow

1. Resolve the source

• If given an arxiv ID or URL, fetch the abstract page to get title, authors, date, and abstract.
• If given a PDF path, read it to extract the same fields.
• Derive a stable slug: <firstauthor-surname>-<2-3-word-topic> (e.g. jiang-llmlingua, press-longchat).

2. Check for duplicates

• Search REVIEWED.md and references/REFERENCE_INDEX.md for the slug or arxiv ID.
• If already present, report it and stop.

3. Classify the paper

Assign one or more tags from this controlled list:

4. Fill in the reference summary

Create references/<slug>.md using templates/REFERENCE-paper.md. Populate every section:

• TL;DR: 3–8 bullets capturing what the paper does and why it matters for context management.
• What's novel: one paragraph — what does this do that adjacent work does not?
• Mechanism overview: describe context representation, compression/injection path, eviction policy, token budget model. Skip sections that are not applicable (mark N/A).
• Evaluation: quote benchmark name, baselines, key metric, and result — always mark (as reported).
• Implementation details worth stealing: concrete takeaways only.
• Open questions: gaps, risks, unverified claims.

Keep language precise. Do not pad. Quote all numbers with their source.

5. Update REVIEWED.md

Add a row to the summary table at the top (reverse-chronological):

| <today's date> | <slug> | paper | pending | <one-line description> |

Add a detailed section below the table:

## <slug> — <Full paper title>

- **arxiv**: <ID>
- **Authors**: <list>
- **Date**: <YYYY-MM-DD>
- **Tags**: <tags>
- **Summary**: <2–3 sentences>
- **Disposition**: pending — awaiting user decision on promotion

6. Update REFERENCE_INDEX.md

Add a row under the most relevant category table. If no category fits, add it to the closest one and note it for the user.

7. Report and offer next step

Summarise what was created. Then ask:

This paper is now in REVIEWED.md as pending. Would you like to:

• Promote it to a standalone ANALYSIS-arxiv-<id>-<slug>.md deep dive?
• Keep it in REVIEWED.md for now?
• Skip it (mark as not promoted with reasoning)?

Quick Commands

# Check for duplicate before starting
grep -i "<arxiv-id-or-slug>" REVIEWED.md references/REFERENCE_INDEX.md

# Fetch arxiv abstract (title, authors, date)
curl -s "https://arxiv.org/abs/<id>"

# Create reference file from template
cp templates/REFERENCE-paper.md references/<slug>.md

# Add triage entry to REVIEWED.md (reverse-chronological)
# | YYYY-MM-DD | <slug> | paper | pending | <one-line description> |

Anti-Patterns

NEVER invent benchmark numbers

WHY: Fabricated metrics corrupt the research record.

BAD: "Achieves 87% recall on LongMemEval." GOOD: "Reports 87% recall on LongMemEval (as reported, Table 3)."

NEVER skip the duplicate check

WHY: Re-triaging the same paper wastes effort and creates conflicting entries.

BAD: Creating a new references file without checking REVIEWED.md. GOOD: grep -i "<slug>" REVIEWED.md references/REFERENCE_INDEX.md first.

NEVER promote without user confirmation

WHY: Promotion to ANALYSIS-*.md is a quality gate, not automatic.

BAD: Creating ANALYSIS-*.md as part of triage. GOOD: Triage to REVIEWED.md, then ask the user.

References

• REFERENCE-paper template — frontmatter fields, section structure, and body conventions; use every time you create a new references/<slug>.md in the consuming project
• ANALYSIS-paper template — 3-stage deep-dive structure and frontmatter; use when promoting a paper to a standalone ANALYSIS-*.md