# Brief — reference acquisition for the WP44 manuscript (Agent B-lit, 2026-11-03)

Goal: build the verified bibliography that the manuscript's methodological and historical parts need, BEFORE any analysis result exists. Every entry must be resolved to a real bibliographic record; nothing may be invented.

Output files (only these):

1. manuscript/references.bib — BibTeX, keys surnameYYYYkeyword (lower-case ASCII), one entry per work, fields title/author/year/journal-or-booktitle/volume/pages/doi (or eprint+archiveprefix={arXiv} for preprints, or url for software/datasets/reports), note = {verified: <method> <date>}.
2. manuscript/references_verification.csv — columns: key, doi_or_arxiv, verified_via (crossref|openalex|arxiv|datacite|url), verified_title (as returned by the service), match (exact|minor|FAILED), note. Entries that FAIL verification must be removed from references.bib and kept in the CSV with match=FAILED.
3. literature_synthesis/REFERENCE_MAP.md — for each manuscript part (I–XI, appendices), the list of keys it should cite and one line on why (what the reference supports).

Verification method (mandatory, per entry): curl -s "https://api.crossref.org/works/<DOI>?mailto=spbou4@icloud.com" or curl -s "https://api.openalex.org/works/https://doi.org/<DOI>" (single lookups are free) or curl -s "https://export.arxiv.org/api/query?id_list=<id>"; compare the returned title with yours. Do not use web search results as verification. If you cannot verify an item, drop it.

Scope (target 120–180 verified entries), grouped: A. Scientometric data sources: OpenAlex (Priem, Piwowar, Orr 2022 arXiv:2205.01833), DBLP (Ley 2002/2009), arXiv (Ginsparg), Semantic Scholar (Kinney et al. 2023 S2AG), Crossref, ACL Anthology, Microsoft Academic Graph (Wang et al. 2020), CiteSeerX, Papers with Code; OpenAlex topic classification (OpenAlex topic model documentation, 2024). B. Systematic mapping / review methodology: PRISMA 2020 (Page et al. 2021 BMJ), PRISMA-S, Kitchenham & Charters 2007 guidelines (SE), Petersen et al. 2008/2015 systematic mapping, Tricco scoping reviews, living reviews, automation of screening (e.g., ASReview van de Schoot 2021), LLM-assisted screening evaluations 2023–2025. C. Bibliometric indicators: Waltman 2016 review; field-normalised indicators (Waltman & van Eck 2019; MNCS; FWCI); citation windows and age bias (Wang 2013 "Citation time window choice"); h-index (Hirsch 2005); Gini/citation inequality (Nielsen & Andersen 2021); interdisciplinarity (Rao–Stirling: Stirling 2007; Leinster & Cobbold 2012; Porter & Rafols 2009; Leydesdorff); co-authorship networks (Newman 2001, 2004); country/institution analyses (Wagner 2017 international collaboration); preprints in CS (Lin et al. 2020 arXiv; Fraser 2021 bioRxiv analogies); citation advantage of OA (Piwowar 2018). D. Science mapping & embeddings: VOSviewer (van Eck & Waltman 2010), Small 1973 co-citation, Kessler bibliographic coupling, Boyack & Klavans 2010 accuracy of clustering, Leiden algorithm (Traag 2019), Louvain (Blondel 2008), SPECTER (Cohan et al. 2020), SPECTER2 (Singh et al. 2023 SciRepEval), SciNCL (Ostendorff 2022), SciBERT (Beltagy 2019), sentence-BERT (Reimers 2019), UMAP (McInnes 2018), HDBSCAN (Campello 2013; McInnes 2017), BERTopic (Grootendorst 2022), LDA (Blei 2003), dynamic topic models (Blei & Lafferty 2006), topic stability (Greene 2014). E. Structural change & time series: Bai & Perron 1998, 2003; Andrews 1993; Killick, Fearnhead & Eckley 2012 (PELT); Adams & MacKay 2007 (BOCPD); Truong, Oudre & Vayatis 2020 (ruptures review); Newey & West 1987; interrupted time series (Bernal, Cummins, Gasparrini 2017); synthetic control (Abadie 2010); event-study pitfalls. F. Emerging topic / technology detection: Rotolo, Hicks & Martin 2015; Small, Boyack & Klavans 2014; Cozzens et al. 2010; Wang Q. 2018 "A bibliometric model for identifying emerging research topics"; Klavans & Boyack 2017 research fronts; sleeping beauties (Ke, Ferrara, Radicchi, Flammini 2015); disruption index (Funk & Owen-Smith 2017; Wu, Wang, Evans 2019; Park, Leahey, Funk 2023 "Papers and patents are becoming less disruptive"); novelty (Uzzi et al. 2013); burst detection (Kleinberg 2003). G. Historical/contextual works on the AI transformation (only landmark, verified): Transformer (Vaswani 2017), BERT (Devlin 2019), GPT-3 (Brown 2020), scaling laws (Kaplan 2020; Hoffmann 2022), InstructGPT (Ouyang 2022), ChatGPT release (OpenAI blog 2022-11-30 — cite as web page with access date), GPT-4 technical report (2023), Llama 1/2/3 (Touvron 2023; Dubey 2024), diffusion (Ho 2020; Rombach 2022), CLIP (Radford 2021), ViT (Dosovitskiy 2021), NeRF (Mildenhall 2020), RLHF (Christiano 2017), chain-of-thought (Wei 2022), RAG (Lewis 2020), ReAct (Yao 2023), Toolformer (Schick 2023), Mamba (Gu & Dao 2023), MoE Switch (Fedus 2022), DeepSeek-R1 (2025), o1 system card (2024), AlphaFold 2 (Jumper 2021), AlphaGo/AlphaZero, foundation-model report (Bommasani 2021), "Sparks of AGI" (Bubeck 2023), stochastic parrots (Bender 2021), Stanford AI Index 2024/2025 reports, Epoch AI compute trends (Sevilla 2022). H. Evidence on LLM effects on science: Liang et al. 2024 (LLM-modified content in reviews/papers), Kobak et al. 2024/2025 ("delving"), Geng & Trotta 2024, Bin-Nashwan / Dwivedi 2023 "So what if ChatGPT wrote it?", Birhane 2023, Messeri & Crockett 2024 (illusions of understanding), Nature surveys on AI use by researchers (Van Noorden 2023), arXiv submission growth analyses, "AI research is growing faster than…" (Krenn et al. 2023 Nature Machine Intelligence forecasting AI research), Fortunato et al. 2018 Science of science, Wang & Barabási 2021 book. I. Reproducibility/open science in CS: Gundersen & Kjensmo 2018; Pineau et al. 2021 (ML reproducibility program); Kapoor & Narayanan 2023 (leakage); Papers with Code / Hugging Face ecosystem studies; benchmark saturation (Ott et al. 2022 "Mapping global dynamics of benchmark creation and saturation"; Dehghani 2021 benchmark lottery; Raji 2021 AI and the everything in the whole wide world benchmark); Sculley 2018; ACM artifact badging policy (web). J. Other CS areas' survey anchors (one or two verified surveys each, 2022–2026 preferred): quantum computing (Preskill 2018 NISQ; a 2023–2025 survey), blockchain, federated learning (McMahan 2017; Kairouz 2021), differential privacy (Dwork 2006/2014), post-quantum cryptography (NIST FIPS 203/204/205 2024), formal verification (Lean mathlib / Rust verification surveys), computer architecture (Hennessy & Patterson 2019 Turing lecture; "golden age"), AI accelerators surveys, sustainable computing (Patterson 2021 carbon; Strubell 2019; Luccioni 2023), edge computing (Shi 2016), robotics foundation models (RT-2 2023; survey), AI for science (Wang H. et al. 2023 Nature "Scientific discovery in the age of AI"), AI safety/alignment surveys (Hendrycks 2021; Ji et al. 2023 alignment survey), agents surveys (Wang L. et al. 2024; Xi et al. 2023), LLM surveys (Zhao et al. 2023), multimodal surveys, speech (Whisper 2023), information retrieval (dense retrieval Karpukhin 2020), databases (vector databases survey 2024), compilers (MLIR Lattner 2021), programming languages (Rust safety), operating systems (eBPF / unikernels / LLM-OS), networks (programmable networks / P4), security (LLM security surveys 2024; prompt injection Greshake 2023), cryptography (FHE surveys), computational biology (AlphaFold 3 2024; ESM-2 2023), graphics (Gaussian splatting Kerbl 2023), HCI (Generative AI & HCI 2023–2024 surveys).

Rules: work through the groups in order; verify as you go; keep each file write small (append to the .bib in batches of ~20 entries with cat >>). Report back ≤ 25 lines: counts per group, failures, and anything you could not find.