Skip to main content
Date: 2026-07-17 · For: honest-scholar literature capability, position mode (sub-spec 2). · Status: verified-source digest; migrates to the plugin’s resources/references/. For a committed paper: systematically establish novelty, select baselines, situate the contribution. Every named method has a verified identifier below.

1. Named methodologies

  • Systematic Literature Review (SLR) — pre-registered, auditable: review questions → search → inclusion/exclusion → screen → extract → synthesize. Kitchenham & Charters (2007), Guidelines for Performing SLRs in Software Engineering, EBSE-2007-01. The backbone the others plug into (plan/conduct/report).
  • PRISMA (reporting standard, not a search method) — 27-item checklist + flow diagram (identified → deduped → screened → excluded-with-reasons → included). Page et al. (2021), BMJ 372:n71, DOI:10.1136/bmj.n71; Moher et al. (2009), PLoS Med 6(7):e1000097, DOI:10.1371/journal.pmed.1000097. Durable value = the flow diagram’s audit trail (the anti-cherry-picking defense).
  • Snowballing — Wohlin (2014), EASE, DOI:10.1145/2601248.2601268. Diverse start set (Google Scholar to reduce publisher bias); backward (references → foundations/precedent) + forward (citations → newer competitors/SOTA); iterate to saturation.
  • Concept-centric matrix — Webster & Watson (2002), MIS Quarterly 26(2):xiii–xxiii. Do not organize author-by-author; organize concept × source (rows=papers, cols=attributes your delta turns on). This matrix is the comparison table you ship.
  • Review typology — Cooper (1988), Knowledge in Society 1(1):104–126, DOI:10.1007/BF03177550 [page range unverified]. Choose the organizing spine; for a methods paper a conceptual/thematic taxonomy beats chronological.
  • Vote-counting / feature matrix — weak for evidence claims, legitimate for a capability comparison table (method × feature, ✓/✗/partial); the artifact reviewers scan first.
  • Related-Work section structure — thematic grouping, one paragraph per approach-family, each with an explicit delta: “Unlike ⟨family⟩ we ⟨difference⟩, which yields ⟨consequence⟩.” Contribution as a bulleted delta in intro, mirrored in related work.

2. Process: committed topic → positioning.md

  1. Frame the claim(s) as 1–3 falsifiable delta statements before searching (the review questions).
  2. Seed set — 3–6 diverse anchors spanning communities/terminologies.
  3. Snowball to saturation — backward + forward (S2/OpenAlex for forward); log every candidate with include/exclude reason (PRISMA counts). Stop when no new relevant methods appear.
  4. Screen against explicit inclusion criteria; record exclusions with reasons.
  5. Extract into the concept matrix — rows = included methods; cols = the concepts your delta turns on.
  6. Derive the taxonomy from matrix column clusters (the section spine).
  7. Write the delta — per taxonomy branch, one sentence grounded in a matrix cell.
  8. Derive baselines from the taxonomy, not from convenience.
  9. Ship positioning.md: taxonomy, comparison table, PRISMA search log, novelty/delta bullets, baseline list with justification.
Baseline selection: one strong tuned representative per taxonomy branch; current SOTA on target datasets; the most-likely-cited-against-you; a simplest reasonable floor. Match conditions (splits, tuning budget, compute). Under-tuned baselines are the #1 reviewer complaint (Dacrema et al. 2019; Musgrave et al. 2020; Lin 2019). Defend against “already done”: adversarial precedent search per contribution bullet (forward-snowball nearest prior work + keyword on the mechanism); maintain a “closest prior work” paragraph with the precise testable delta; encode the delta as an ablation; keep the PRISMA log.

3. Anti-patterns → safeguards

Venue expectation (NeurIPS/ICML): reviewers judge novelty/significance + relation to prior work; checklists require limitations + reproducibility. The positioning doc should pre-answer “delta over closest work?” and “baselines strong and fairly tuned?” — the two rejection levers.

4. Level split — same method, two configurations

→ one skill, mode/depth switch; hypothesis = a fast adversarial precedent rapid review. Ship a PRISMA-style log in both modes.

5. Tooling (free/scriptable emphasized)

  • Forward snowball + metadata: Semantic Scholar API, OpenAlex, Crossref (DOIs). Google Scholar = best unbiased start set, no API.
  • Screening/active learning: ASReview LAB (van de Schoot et al. 2021, Nat. Mach. Intell. 3:125–133, DOI:10.1038/s42256-020-00287-7); Rayyan.
  • Reference mgmt/dedup: Zotero (+ Better BibTeX); Publish or Perish.
  • PRISMA artifacts: PRISMA2020 R package / official flow-diagram generator.

Sources

Kitchenham & Charters 2007 (EBSE-2007-01) · Page et al. 2021 (bmj.n71) · Moher et al. 2009 (pmed.1000097) · Wohlin 2014 (2601248.2601268) · Webster & Watson 2002 (MISQ 26(2)) · Cooper 1988 (BF03177550) · van de Schoot et al. 2021 (s42256-020-00287-7) · Dacrema et al. 2019 (RecSys, 3298689.3347058) · Musgrave et al. 2020 (arXiv:2003.08505) · Lin 2019 (SIGIR Forum, 3308774.3308781).

Caveats

Cooper 1988 page range not verified in primary this session. NeurIPS/ICML checklist wording paraphrased (changes yearly) — verify current-year text before quoting.