sota-literature-review

SOTA Literature Review

Build literature review artifacts that are reproducible, inspectable, and easy to continue later. This skill is the orchestrator for autonomous SOTA work: from a rough idea, a seed paper, a few references, or a query, it should drive source discovery, citation-graph expansion, full-text acquisition, linear reading, per-source synthesis, bibliography hygiene, claim audit, and survey production.

Read First

• references/repository-contract.md
• references/source-ledger.md
• references/mcp-catalog.md
• references/workflow-stage-contracts.md
• references/citation-chasing.md
• references/per-source-synthesis-template.md
• references/survey-production.md

Review Scale

Before searching, declare the intended weight of the review and record it in sota/search-strategy.md.

Scale	Typical Use	Target Included Papers	Core Full Reads	Citation Chasing
quick-scan	orient an idea or proposal	8-15	3-5	seeds plus one high-yield hop
focused-sota	defensible related work and gap map	20-40	8-15	backward and forward until local saturation
full-survey	survey paper or major SOTA chapter	50+	20+	iterative frontier expansion with saturation evidence

The target count is a planning budget, not a quality target. Stop because the recorded search frontier reaches saturation or because the declared budget is exhausted and blind spots are documented; do not pad with weak papers.

Autonomous Workflow

1. Frame the review objective, research idea, contribution hypothesis, and inclusion/exclusion criteria. If the idea is underspecified, use research-design-positioning first.
2. Run npm run workflow:literature in create-academic-research projects when available; otherwise inspect MCP capability state with academic-mcp-tooling.
3. Build the initial seed set from user references, proposal citations, local PDFs, known seminal papers, and high-relevance API hits.
4. Create search concepts, synonyms, adjacent terms, exclusion terms, and negative-control queries. Record them in sota/search-strategy.md.
5. Run source-specific search across the active stack, saving raw metadata under sources/metadata/.
6. Expand through citation chasing. Record every round in sota/citation-chasing-log.csv.
7. Deduplicate and screen candidates into sota/screening-decisions.csv.
8. Promote included works into sources/source-ledger.csv and sota/literature-matrix.csv.
9. For every core or supporting source, acquire full text through source-ingestion and document-conversion, store a linear reading copy in sources/markdown-linear/, and update sota/reading-log.csv.
10. Read core/supporting sources linearly before writing source claims. Store per-paper syntheses in sota/paper-syntheses/ using references/per-source-synthesis-template.md.
11. Normalize and verify bibliography entries in sources/bib/references.bib with citation-bibliography-tooling.
12. Write theme-first synthesis in sota/synthesis.md, gap analysis in sota/gaps.md, and when requested a LaTeX survey draft at reports/paper/sota-survey.tex using references/survey-production.md.
13. Promote durable SOTA claims into sota/sota-claim-ledger.csv with allowed wording, limitations, contradictions, source IDs, BibTeX keys, review status, and downstream status.
14. Run citation-claim-audit before promoting claims into survey, agenda, framing, or manuscript prose.
15. When the research idea involves code/data/evaluation artifacts, connect gaps and evidence plans to ACM artifact badge readiness through artifact-open-science.

Keep the Project Quality Contract active across the loop: search metadata, reading copies, syntheses, bibliography, LaTeX drafts, and artifact evidence belong in separate durable zones, and only audited full-text evidence should be promoted into trusted SOTA or manuscript claims.

Review Loop And Handoff

For each search batch: ingest, screen, read, extract claims, audit citations, fix gaps, and re-review until no blocking issue remains. Handoff reviewed SOTA claims to survey-synthesis and research-agenda; do not write survey or agenda outputs as hidden side effects of SOTA work.

Search Strategy

Create or update sota/search-strategy.md with:

• research question
• review scale and paper budget
• seed set
• search concepts, synonyms, adjacent terms, and exclusion terms
• inclusion and exclusion criteria
• databases queried
• query strings
• date range
• filters
• screening decisions
• citation-chasing depth and saturation notes
• known blind spots

Prefer API-backed sources: arXiv, Semantic Scholar, OpenAlex, Crossref, PubMed, DBLP metadata, and Zotero. Use Google Scholar as a discovery fallback, not the primary evidence layer.

Retrieval Protocol

1. Decompose the topic into 2-4 search concepts and synonyms before querying.
2. Run source-specific searches instead of one vague web search.
3. Use short, high-signal queries for arXiv and technical indexes; put field, date, and venue constraints in filters when the tool supports them.
4. Search both recent work and known seminal work. Do not let citation count alone define importance.
5. Expand beyond keyword hits with citation chasing: follow references (backward) and citing papers (forward) from core seeds and proposal references. See references/citation-chasing.md. Skipping this step is the main cause of missed relevant work.
6. Capture raw metadata under sources/metadata/ before synthesis.
7. Deduplicate by DOI, arXiv ID, PMID, OpenAlex ID, Semantic Scholar ID, title, and author/year.
8. Rank candidates by relevance, methodological fit, recency, venue/source authority, and citation/context signals. Record why core papers were chosen.
9. If coverage is weak, state the gap and next database, query, or seed to try instead of padding the review with marginal sources.
10. Re-run search with new concepts learned from full readings when the matrix reveals missing terminology, benchmarks, venues, or method families.

Anti-Echo-Chamber Rules

• Maintain at least three seed types when possible: project-specific, seminal, and recent frontier.
• Search adjacent communities and alternative terminology, not only papers that cite the same seed.
• Include negative or contrastive queries that can disconfirm the initial idea.
• Sample excluded near-misses and high-citation outliers to check whether the screening rule is too narrow.
• Do not let one author group, venue, benchmark, or citation cluster dominate without recording why.
• Treat citation count, PageRank-style centrality, recommendations, and graph proximity as discovery signals only. Evidence still requires full text.

Full-Text Policy

For every core and supporting source, acquire the full text, not just the abstract:

• store the native PDF and a converted reading copy via source-ingestion and document-conversion
• read the full text linearly before writing its synthesis
• record progress in sota/reading-log.csv
• if no legal full text is found, mark it as a tracked limitation on the source and matrix row, not a silent drop; abstract-only records cannot carry durable claims

Literature Matrix

Maintain sota/literature-matrix.csv with at least:

• source_id
• title
• authors
• year
• venue
• method
• dataset/sample
• task/problem
• key claim
• metric/result
• limitations
• relevance
• full_text_status
• reading_status
• synthesis_path
• bib_key
• citation_count_or_signal
• identifiers

Synthesis

Write synthesis in sota/synthesis.md, not only in chat. Separate:

• established findings
• contested claims
• methodological families
• dataset or benchmark differences
• missing baselines
• open gaps
• implications for this project

Also maintain sota/gaps.md when the review affects project direction. A gap is not valid unless it names the evidence base, nearest prior work, why existing methods fail or leave space, and what evidence would close the gap.

Output Modes

• search-plan: query strategy, sources, filters, and expected blind spots.
• annotated-bibliography: source cards without broad claims.
• matrix: structured cross-paper comparison.
• synthesis: theme-first SOTA narrative with claim support.
• per-source-synthesis: uniform deep record per core/supporting source using references/per-source-synthesis-template.md.
• related-work: manuscript-ready draft after claims pass citation audit.
• survey: full survey narrative plus research agenda using references/survey-production.md.
• gap-map: defensible opportunities and failure conditions for this project.

Completion Gates

A serious SOTA task is not complete until:

• search scale, seeds, queries, and blind spots are recorded
• citation chasing has a logged stopping rule in sota/citation-chasing-log.csv
• core/supporting papers have full-text status and reading status
• each core/supporting paper has a sota/paper-syntheses/ record
• cited papers appear in sources/bib/references.bib
• survey or related-work claims pass citation-claim-audit
• ACM artifact implications are recorded when the SOTA motivates code, data, benchmark, model, reproduction, or evaluation artifacts

Quality Rules

• No unstructured paper-by-paper laundry list. For core/supporting sources use the structured per-source synthesis template; reserve plain cards for the annotated-bibliography mode.
• Claims need inline source IDs or wiki links.
• SOTA claims need date, benchmark, metric, and scope.
• Contradictions go to wiki/contradictions.md.
• Reusable findings should be linked from wiki/index.md.
• For each durable claim, record source IDs and allowed wording; do not promote abstract-only or unverified webpage summaries into paper text.
• Surface disagreements and negative results instead of smoothing them away.