Bibliography

ARIST systematic review synthesizing 89 empirical studies of academic-library RDM services worldwide. Internal partnership appeared in 60% of studies; technical data-use/analysis services appeared in only 9% -- reinforces the distinction between libraries' strong advisory/repository/metadata/coordination roles and the much scarcer provision of intensive technical analysis. Peer-reviewed. Confidence: high for the RDM-specific finding; the wiki article itself flags that "discipline-specific" in this literature usually means subject-matter expertise, not DSC's methods/tools model — don't over-extend. Full author list, volume, pages added via ChatGPT deep-research verification 2026-08-16; also relevant to the public computational-methods canon, not just the private reorg-response research it was originally pulled for.

Co-citation network analysis built around ~70 IC2S2 keynote speakers, 2000-2020: finds increasing clustering and path length, i.e. fragmentation, not integration -- network science remains CSS's central shared literature but subcommunities have grown more internally cohesive and more distant from each other. Authors caution fragmentation also occurs in mature disciplines and that prominent-speaker sampling misses much of the field. In tension with @arxiv-division-to-unity-css-2024's finding of semantic blending -- see wiki article Uncertainty section, this is a genuine open disagreement (different measurement: co-citation structure vs. semantic similarity across a much larger corpus), not resolved by picking one. Confidence: high (full text examined). Verified via ChatGPT deep-research pass 2026-08-16, cross-checked against the University of Helsinki repository full text.

Merges Notre Dame's Center for Research Computing and Applied Analytics & Emerging Technology Lab into "DOME" (Deployment, Operations, Modeling, Engineering); broader initiative also coordinates the Lucy Family Institute and Scientific AI activities. CORRECTION: primary sources do NOT identify the library as a participant -- an earlier draft of the wiki article's TLDR incorrectly characterized this as merging "library-data" units. It is a research-computing/AI-lab consolidation; library involvement is unconfirmed, not a fact. Confidence: high (primary institutional announcement) for what's stated; the library-involvement claim was actively wrong and has been removed. Initiative launched Sept. 2025; unified structure/DOME name announced July 29, 2026 (dates refined via ChatGPT deep-research verification).

25 interviews with researchers, librarians, and library managers in China; derives 10 requirements under 5 headings (AI-for-Science readiness, research-idea development, collaboration networks, research execution, publication). Also flags algorithmic limitations, ethics, data-property questions, over-reliance, rising demand for specialized information services. Generalizability beyond the Chinese case is proposed by the authors, not empirically tested -- don't over-extend. Confidence: medium-high (real venue/DOI/authors; content is abstract-level, not full-text). OnlineFirst July 8, 2026. Accessible source was the publisher's detailed abstract, not the restricted full article (per ChatGPT deep-research pass).

University of Hong Kong survey: demand for hands-on help with analysis, visualization, cleaning; substantial reliance on self-teaching and peers; significant AI-project participation. Proposes a one-stop "Data Clinic" combining advisory and technical support; calls for advanced librarian skills and collaborative delivery. Confidence: medium (real venue/DOI/authors; abstract-level only, no full text available). DOI assigned 2025, issue date January 2026. CORRECTION: an earlier draft's claim that this piece emphasizes funding specifically was not substantiated by the abstract (per ChatGPT deep-research pass, which could only access the repository abstract, not full text) -- removed that framing below.

Consensus study report on the trajectory of statistical/computational science through 2035. Confidence: medium (real NASEM report, not read in full). Committee chaired by Katherine Ensor (Rice University).

Merges UNC's School of Information and Library Science with its School of Data Science and Society, planned Oct. 2025, launched July 1, 2026. The strongest example found of formal LIS/data-science fusion at the academic-school level -- not a library operational reorganization. Confidence: high (primary source directly verified). CORRECTION: a ChatGPT deep-research pass claimed this merger "explicitly cites generative AI" as a motivation. Direct WebFetch of this primary source, 2026-08-16, found no mention of AI as a driver -- the actual text cites building "on our strengths in information, library and data sciences." That specific claim was NOT adopted into the wiki/canon.

Applies a CSS classifier to 11 million papers, compares transformer-based knowledge embeddings; finds inflections around 2005 and 2014, concludes CSS developed a distinct semantic identity through 2014 but increasingly blended into parent social sciences after. In tension with @qss-comp-social-science-fragmentation-2026's fragmentation finding -- different measurement (semantic similarity across a huge corpus vs. co-citation structure among prominent scholars), a field can diffuse its methods outward while its specialist communities fragment inward. Confidence: high (full citation verified). arXiv preprint posted Dec 11, 2024 (per ChatGPT deep-research verification); became a WWW 2025 Companion paper, not a standalone 2026 publication -- corrects an earlier mislabeling in the raw research pass that also called this "2026."

294 interview transcripts from 27 institutions, stratified sample of 41 analyzed. Researchers commonly didn't understand "data services" as a category, preferred individualized/memorable services, struggled with data size/custody/access/permissions/pipelines/publication, frequently treated IT/legal/ethics offices as part of the same support ecosystem. Strong evidence for one recognizable intake point; coded sample relatively small. Confidence: medium-high (real Ithaka S+R report; via ChatGPT deep-research).

Organizes biological foundation models around sequence analysis, structure construction, function prediction, multimodal integration. Emphasizes noise, bias, interpretability, and cases where general models underperform -- parallels NLP concerns while requiring biological validation. Confidence: medium (real Briefings in Bioinformatics article; via ChatGPT deep-research).

Defines computational humanities as both use AND critique of computational methods; rejects clean close/distant-reading or quantitative/qualitative boundaries; foregrounds data creation, labor, code quality, reproducibility, positionality, power. Explicitly argues for humanities expertise within HPC support since support structures affect which research gets priority -- directly relevant to a library DSC's role. Confidence: medium (real edited volume chapter; via ChatGPT deep-research).

Traces the transition from bag-of-words/task-specific classifiers through embeddings/transformers to pretrained generative models. LLMs are now de facto experimental baselines, but closed models produce serious reproducibility, contamination, privacy, bias, explainability, version-instability problems. Confidence: medium (real Frontiers in AI article; via ChatGPT deep-research).

Placed the San Diego Supercomputer Center and the Halıcıoğlu Data Science Institute together as foundational pillars of a new school -- combining HPC/data-intensive infrastructure with data-science/AI research and education. School-level academic consolidation, not a library/service-desk reorganization; structurally comparable to Notre Dame's DOME. Confidence: high (primary institutional announcement).

Systematic web inventory: 120 US institutions, 8 Canadian universities, a 29-university collaboration. 495 services at R1s (avg 12.4/school), 165 at R2s (avg 4.1), 45 at liberal arts colleges (avg 1.1); 73% consultation, 27% training. Libraries supplied ~40-46% of services at research universities, ~60% at liberal arts colleges; bioinformatics located almost exclusively in research-office/medical-school cores; IT's modest 8-15% share understates its qualitative importance for computing/storage/partnership. Confidence: high (independently verified numbers). Percentages independently verified 2026-08-16 via direct fetch: R1 libraries 227/495 services (~46%), R2 64/165 (~39%), liberal arts 27/45 (~60%) -- matches the ChatGPT deep-research summary closely (reported as "40-45%... 60%").

RADS study, six research-intensive universities: average combined institutional + researcher data-sharing cost ~\$2.5M/year, institutional totals \$800K-\$6M+. Public-data support is a distributed institutional expense spanning libraries, IT, research offices, researcher labor. Six-institution sample -- not a sector-wide cost benchmark. Confidence: medium-high (real ARL report; via ChatGPT deep-research).

Forward-looking position paper by the field's most prominent current figure (NCSA/UIUC; co-founder of the Research Software Alliance, ReSA). Confidence: medium (real arXiv preprint, not read in full).

Extends the shift to combined molecular modalities with spatial coordinates. Discussion of missing/unmatched measurements, resolution, validation shows why bioinformatics support now requires large-scale storage/compute, specialized pipelines, domain-aware statistical advice. Confidence: medium (real Nature Reviews Genetics article; via ChatGPT deep-research).

26 long-form interviews plus ACL Anthology bibliometrics; distinguishes recurring methodological cycles from genuinely unprecedented changes, especially benchmark culture and centralized software/model infrastructure. Current NLP is increasingly dependent on shared pretrained models, software stacks, compute, industrial resources -- alters funding, incentives, who can participate. Confidence: medium-high (real EMNLP paper; via ChatGPT deep-research).

McGill case study documenting the centralized collaborative service (vice-principal for research, library, IT, research ethics) including shared intake and "query the panel" consultations. Valuable implementation detail; single reflective case, doesn't establish comparative effectiveness. Confidence: medium-high (real eScience Librarianship article; via ChatGPT deep-research).

Broad review: protein structure/function prediction, genome engineering, systems biology/data integration, phylogenetics. Treats the protein-structure breakthrough as genuine but identifies shared bottlenecks: training data, problem formulation, transfer to new settings, interpretability, architecture reuse. Confidence: medium (real Nature Communications article; via ChatGPT deep-research, not independently read in full).

Bibliometric/social-network/term-network analysis finds DH developing through four stages with durable themes: collections/content; technologies/theories/methods; collaboration/interdisciplinarity/support; reflection on DH's own evolution. Evidence that infrastructure and collaboration are constitutive of the field, not auxiliary. Confidence: medium (real Journal of Documentation article; via ChatGPT deep-research).

27-person international workshop produced 65 RSE research questions across people/policy/infrastructure; only 24 (37%) connected to existing literature. Career paths, recognition, training, recruitment, retention, funding dominated priorities -- RSE's unresolved problems are at least as organizational as technical. Confidence: medium-high (real JORS article; via ChatGPT deep-research).

Survey of 1,149 primarily US researchers: recurring shortages of time/opportunity for software training across the lifecycle, weak funding models, inadequate professional credit. Quantitative support for dedicated RSE capacity rather than assuming disciplinary researchers can absorb production-quality software work via short workshops. Confidence: medium-high (real PeerJ CS article; via ChatGPT deep-research).

Identifies three major routes into RSE careers with different associated gaps, rather than assuming a uniform CS pipeline. Connects formalization of university/national-lab/industry RSE groups to recruitment, retention, career progression, cross-disciplinary training. Confidence: medium-high (real Computing in Science & Engineering article; via ChatGPT deep-research).

CaRCC survey: 563 valid responses (539 with RCD responsibilities, 452 at academic institutions); CaRCC estimated a US academic workforce of at least 5,000. Establishes RCD as a heterogeneous professional workforce (infrastructure, data, software, facilitation, domain backgrounds); workforce snapshot, not longitudinal. Confidence: medium-high (real PEARC paper; via ChatGPT deep-research).

Defines sustainable RSE support through four mutually dependent pillars: software development, community, training, policy. Frames research software as institutional infrastructure maintained by specialist professionals, not disposable code incidental to individual grants. Confidence: high (full citation verified). arXiv preprint 2020, published version IEEE Software 2021; full citation via ChatGPT deep-research pass.

Documents the methodological shift from single-assay analysis to integrating samples, technologies, donors, batches, modalities. Data harmonization and latent representation are now primary scientific methods, not preprocessing. Confidence: medium (real Nature Biotechnology article; via ChatGPT deep-research).

The report that supplied the "foundation model" framing itself: broad self-supervised pretraining + adaptation across many tasks. Describes homogenization as both leverage and systemic risk -- improvements/defects in a few upstream models propagate downstream. Well-known, foundational. Confidence: high (widely known, verified via ChatGPT deep-research).

Landmark position paper: ever-larger LMs impose environmental/financial costs, obscure training-corpus content and biases, concentrate research capacity, generate plausible form without grounded meaning. Recommendations (document/curate data, assess stakeholder values before building, pursue scaling alternatives) directly challenge a "just buy compute/API access" service strategy. Counterpoint to Bommasani et al. Confidence: high (extremely well-known paper, independently known to be real).

Closest operational analogue to a library data-service center: Research and Innovation, McGill Libraries, and IT Services jointly support RDM, advanced research computing, and research software through a common hub -- a shared gateway/routing layer WITHOUT administratively merging the partner organizations. This is the "federated gateway" model, distinct from Notre Dame/UCSD/UNC's structural consolidation. Confidence: high (primary institutional source, via ChatGPT deep-research).

Full author list, volume, and pages independently verified 2026-08-16 (WebSearch cross-check, matched a ChatGPT deep-research pass exactly). Policy-forum piece, not a systematic review: argues institutional arrangements (platform-data access, privacy infrastructure, ethics/governance, incentives, integrated training) lag behind CSS's methodological progress; recommends public-private data-access structures, secure administrative-data facilities, multidisciplinary institutes. Confidence: high.

Compares 2019 North American library-director survey with 2012: share of libraries with NO research-data services declined from ~60% to 44.1%. Growth uneven -- larger institutions more likely to have dedicated specialists; insufficient funding, staff time, personnel, faculty interest remained principal barriers. Confidence: medium-high (real Choice white paper with linked dataset; via ChatGPT deep-research).

Audits prominent computational literary studies work, argues many fail basic statistical validation or can't sustain their interpretive claims. Generated an extensive critical forum; makes methodological validation -- not just tool access -- a central support requirement. Counterpoint to Tilton/Mimno/Johnson's more expansive framing (see wiki Uncertainty section). Confidence: medium-high (well-known, widely cited piece; via ChatGPT deep-research).

Foundational document for the RSE-as-profession movement, tracing back to discussions at the UK Software Sustainability Institute's 2012 Collaborations Workshop. Confidence: medium (real, dated report; not read in full this session).

The founding paper widely credited with naming and launching computational social science as a field. Confidence: high (widely corroborated citation details: volume, pages, DOI all consistent across sources).

Widely recommended foundational text for the digital-humanities/humanities-computing field. Confidence: medium (consistently cited across library reading lists; not read in full).

Argues NLP still depends on linguistics across six facets (resources, evaluation, low-resource settings, interpretability, explanation, study of language) even in the LLM era, against the "learn language from scratch" framing. Confidence: medium-high (verifiable journal/volume/issue/page citation; not read in full).

A field-defining "what is digital humanities" report/book, widely used as an entry point. Confidence: medium (real, well-known work; publication year unconfirmed). Year commonly cited as 2012 but not independently confirmed this session -- VERIFICATION NEEDED.

PRIMARY LOCATED (was a secondhand X stub). The argument that AI/"vibe coding" makes OSS easier to consume while weakening the feedback, attention, and revenue that sustain maintainers. Limits: Zenodo preprint (v2), not peer-reviewed, and the title is a deliberately provocative framing — represent the finding, not the headline. Confidence: medium (real primary, but preprint).

PRIMARY LOCATED (was a secondhand X stub). Peer-reviewed Nature article: AI tools expand individual scientists' impact (more papers, more citations) but contract the collective focus of science. Directly supports the "individual gains, collective narrowing" claim. Still observational and AI-use is inferred, so it shows association, not causation. Confidence: medium (peer-reviewed, but observational).

Survey-thread summary naming RSE's core problems: correctness with weak tests, domain concepts embedded in code, low career incentive for maintenance. Community survey, self-selected, social-posted. Confidence: medium.

UCLA Library's own announcement of joining the UC OSPO Network. Primary institutional source for UCLA's participation. Confidence: high.

Announcement of the multi-campus UC OSPO Network (Berkeley, Davis, Santa Barbara, San Diego, Santa Cruz), Sloan-funded, building on the first UC OSPO at UCSC. Institutional source. Confidence: high.

Vendor announcement of Terraform MCP Server GA (June 2026) for HCP Terraform and Terraform Enterprise. Read-only by default; destructive/apply tools gated behind ENABLE_TF_OPERATIONS=true. Load-bearing for the "AI as infrastructure interface" claim, but vendor-authored with an incentive to sell. Confidence: high (for the product facts).

Enumerates four agent patterns: plan review, drift detection, change explanation, standardized code generation. Useful as the concrete "what agents actually do with IaC" reference. Vendor blog. Confidence: high (for the patterns).

Pairing Terraform MCP with Vault MCP so agents request secrets through a governed broker instead of holding them in context. Supports the "secrets stay in a broker" best practice. Vendor. Confidence: high (for the pattern).

Independent press coverage of the Terraform MCP GA — useful as non-vendor corroboration of the same facts. Confidence: high.

Canonical docs. create_run supports plan_and_apply and refresh_state; destructive tools require ENABLE_TF_OPERATIONS=true. Cite this over blogs for exact tool behavior. Confidence: high.

AAP 2.7 GA (June 2026) adds a Technology-Preview native MCP server (agents query jobs, gather facts, launch workflows) and an MCP-backed assistant in the VS Code Ansible extension. Note: Tech Preview, not GA. Confidence: high (for the release facts).

Natural-language to YAML for tasks/playbooks/roles; GA. IMPORTANT LIMIT: requires an AAP subscription — not usable with UCLA's forked community dataverse-ansible playbook without one. Confidence: high.

Community "skill" encoding Terraform conventions for AI agents to follow when generating HCL. Practitioner resource (well-known author), not peer-reviewed. Confidence: medium.

Analysis of 63 user corrections across 197 Claude Code sessions (Mar–May 2026). Finding: scope/process (44%) and factual-state (37%) errors dominate over voice (19%); core failure is acting on inferred state. Single-user, self-selected sample. Confidence: medium (indicative, not a benchmark).

PRIMARY LOCATED (was a secondhand X stub). The argument that AI/"vibe coding" makes OSS easier to consume while weakening the feedback, attention, and revenue that sustain maintainers. Limits: Zenodo preprint (v2), not peer-reviewed, and the title is a deliberately provocative framing — represent the finding, not the headline. Confidence: medium (real primary, but preprint).

PRIMARY LOCATED (was a secondhand X stub). Peer-reviewed Nature article: AI tools expand individual scientists' impact (more papers, more citations) but contract the collective focus of science. Directly supports the "individual gains, collective narrowing" claim. Still observational and AI-use is inferred, so it shows association, not causation. Confidence: medium (peer-reviewed, but observational).

PRIMARY LOCATED (was a secondhand X stub). NBER working paper measuring productivity effects across generations of AI coding tools — the title itself is the thesis: writing code and shipping code diverge. Directly backs "code production is not shipped/maintained software." Limit: working paper, not yet peer-reviewed; general software dev, not academic research software specifically. Confidence: medium.

Vendor announcement of Terraform MCP Server GA (June 2026) for HCP Terraform and Terraform Enterprise. Read-only by default; destructive/apply tools gated behind ENABLE_TF_OPERATIONS=true. Load-bearing for the "AI as infrastructure interface" claim, but vendor-authored with an incentive to sell. Confidence: high (for the product facts).

Enumerates four agent patterns: plan review, drift detection, change explanation, standardized code generation. Useful as the concrete "what agents actually do with IaC" reference. Vendor blog. Confidence: high (for the patterns).

Pairing Terraform MCP with Vault MCP so agents request secrets through a governed broker instead of holding them in context. Supports the "secrets stay in a broker" best practice. Vendor. Confidence: high (for the pattern).

Independent press coverage of the Terraform MCP GA — useful as non-vendor corroboration of the same facts. Confidence: high.

Canonical docs. create_run supports plan_and_apply and refresh_state; destructive tools require ENABLE_TF_OPERATIONS=true. Cite this over blogs for exact tool behavior. Confidence: high.

AAP 2.7 GA (June 2026) adds a Technology-Preview native MCP server (agents query jobs, gather facts, launch workflows) and an MCP-backed assistant in the VS Code Ansible extension. Note: Tech Preview, not GA. Confidence: high (for the release facts).

Natural-language to YAML for tasks/playbooks/roles; GA. IMPORTANT LIMIT: requires an AAP subscription — not usable with UCLA's forked community dataverse-ansible playbook without one. Confidence: high.

Community "skill" encoding Terraform conventions for AI agents to follow when generating HCL. Practitioner resource (well-known author), not peer-reviewed. Confidence: medium.

PRIMARY LOCATED (was a secondhand X stub). NBER working paper measuring productivity effects across generations of AI coding tools — the title itself is the thesis: writing code and shipping code diverge. Directly backs "code production is not shipped/maintained software." Limit: working paper, not yet peer-reviewed; general software dev, not academic research software specifically. Confidence: medium.