WP44: Appendices A (taxonomy), E (data dictionary), H (reproducibility) by Agent L

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Simon-Pierre Boucher committed 3 h ago (Nov 3, 2026) parent 05bf3a4 Browse files

11 changed files +2,944 −6

added agents/briefs/WRITER_APPENDICES_BC.md +22 −0

++# Brief — Appendix B (search strategies and sources) and Appendix C (mathematical derivations) — files manuscript/appendices/appB_search_and_sources.tex, appC_derivations.tex
++
++Follow agents/briefs/WRITER_COMMON.md. Files start with \section{...} (inside \begin{appendices}).
++- Appendix B (≈ 8–12 pages): the exact acquisition procedures as executed — read ingestion/*.py, data/raw/README.md,
++  reports/PHASE0_FEASIBILITY.md, RESEARCH_PLAN.md §2–3, manuscript/sections/part03_data.tex (do not repeat its prose; this
++  appendix is the operational record). For each source: endpoint/bucket, release/version, filters (OpenAlex field 17 rule,
++  year range, id-set second pass, API increment slicing by day × primary-CS/non-CS, arXiv OAI set and windows, ACL export, DBLP
++  dump + DTD), rate limits and budgets, licence, integrity checks (manifest totals, SHA-256 per file, MD5), what was NOT
++  obtained (Semantic Scholar, OpenAlex key, full texts), and the equivalent of a PRISMA-S search record. Include a table of
++  exclusion codes E1–E7 with their SQL predicates paraphrased, the keep/hard-exclude type lists and the generic-deposit list
++  (from systematic_review/define_corpus.py), and the linkage cascade with its thresholds.
++- Appendix C (≈ 10–14 pages): derivations and proofs referenced by Part IV (manuscript/sections/part04_methodology.tex — read it
++  first and match its notation): (i) the ITS model as a regression with the HAC covariance estimator and the placebo p-value
++  as a rank statistic; (ii) Bai–Perron dynamic programme: SSR additivity, the recursion, complexity O(n² m), BIC and LWZ
++  criteria as implemented in econometrics/structural_change.py; (iii) PELT cost and pruning condition, penalty 2σ² log n;
++  (iv) BOCPD: Normal–Inverse-Gamma conjugate updates, Student-t predictive, run-length recursion, the fact that with a
++  constant hazard the mass on r=0 equals H (hence the MAP-run-length detection statistic used in the code); (v) Gini as a
++  rank formula and its equivalence with the Lorenz definition; (vi) Rao–Stirling and Leinster–Cobbold (q=2) identity;
++  (vii) Newman–Girvan modularity for directed weighted graphs; (viii) Krippendorff's α with MASI distance; (ix) the
++  multi-label score rule and the tie-break as a total order; (x) the emerging-topic label and why the feature construction
++  is look-ahead-free (formal statement in terms of filtrations). Derivations must be correct and self-contained; cite the
++  original sources (verify any new citation; most keys exist in manuscript/references.bib and references_part04.bib — reuse them).

modified manuscript/appendices/appA_taxonomy.tex +684 −1

++% ============================================================================
++% Appendix A — The ontology of computer-science research (taxonomy)
++% Written by Agent L (appendices A/E/H instance), 2026-11-03.
++% Sources of truth: taxonomy/ONTOLOGY.md (hand-written §1–§3 and generated tables), taxonomy/ontology.yaml (v1.0,
++% 2026-11-03), taxonomy/topic_map.csv, taxonomy/classify.py, taxonomy/VALIDATION.md. Counts in this appendix are
++% properties of the ontology (numbers of signals, mapped topics and venues), computed from ontology.yaml; they are
++% not corpus results. Validation results are reported in Part IX.
++% ============================================================================
++
++\section{The Ontology of Computer-Science Research}
++\label{app:taxonomy}
++
++This appendix documents the classification scheme that assigns every eligible work of the corpus to one or
++more of the forty research categories listed in the mandate (Section~\ref{sec:m-classification} describes the
++scoring rule and its validation design). The scheme is an \emph{extensible hierarchical ontology} rather than a
++flat keyword list: forty leaf categories are grouped into eight areas, a small set of parent--child edges records
++where a leaf specialises another leaf, and a denser set of ``related'' edges records thematic adjacency. The
++mapping from bibliographic signals to leaves is data, not prose. It lives in one Python module per area
++(\texttt{taxonomy/src/*.py}), is validated and serialised by \texttt{taxonomy/build\_ontology.py} into the
++machine-readable file \texttt{taxonomy/ontology.yaml} (version~1.0, built on 2026-11-03 from the OpenAlex topic
++export of 2026-10-14 and the DBLP stream table of 2018--2026) and into an audit table of OpenAlex topics
++(\texttt{taxonomy/topic\_map.csv}). Everything stated below is derived from those files; the tables were
++regenerated from the YAML file rather than transcribed, and the counts they contain are properties of the
++ontology itself (how many topics, venues and patterns point to a category), not statistics of the corpus. The
++agreement of the rules with independent raters on the stratified validation sample is reported in Part~IX and is
++deliberately not anticipated here.
++
++\subsection{Structure: areas, leaves and edges}
++\label{app:taxonomy-structure}
++
++The ontology has three levels. The eight areas of Table~\ref{tab:appA-areas} are an organisational device for the
++manuscript (they correspond to the disciplinary reviews of Parts~V--VII and to the area-level flow statistics of
++Part~VIII); classification operates on leaves only, and every leaf belongs to exactly one area. The forty leaves
++are numbered as in the mandate (Table~\ref{tab:appA-leaves}). Seven \texttt{parent\_\allowbreak of}/\texttt{child\_\allowbreak of}
++edges record genuine specialisation:
++\texttt{foundation\_\allowbreak models} $\rightarrow$ \texttt{llm} and \texttt{multimodal\_\allowbreak ai};
++\texttt{programming\_\allowbreak languages} $\rightarrow$ \texttt{compilers};
++\texttt{distributed\_\allowbreak computing} $\rightarrow$ \texttt{cloud\_\allowbreak computing} and \texttt{edge\_\allowbreak computing};
++\texttt{computer\_\allowbreak architecture} $\rightarrow$ \texttt{ai\_\allowbreak accelerators};
++\texttt{privacy\_\allowbreak preserving\_\allowbreak computation} $\rightarrow$ \texttt{federated\_\allowbreak learning}.
++These edges have one operational consequence, the tie-breaking rule for the primary category (a child wins
++over its parent at equal score), and one interpretive consequence: a work carrying both labels is not
++double-counted as evidence of convergence between distinct fields, since the two labels are nested by
++construction. The \texttt{related} edges (236 directed entries in the YAML file, 144 distinct unordered pairs)
++carry no weight in classification; they document which categories the ontology's author expected to share
++works, and they are used in Part~IX only to separate expected from unexpected co-classification.
++
++\begin{table}[htbp]
++\begin{singlespace}
++\centering
++\caption{The eight areas and their leaves (\texttt{ontology.yaml}, \texttt{areas}).}
++\label{tab:appA-areas}
++\small
++\begin{tabularx}{\textwidth}{p{3.3cm} p{3.9cm} c X}
++\toprule
++Area id & Name & Leaves & Leaf identifiers \\
++\midrule
++\texttt{ai\_\allowbreak ml} & Artificial Intelligence and Machine Learning & 9 &
++  \texttt{llm}, \texttt{foundation\_\allowbreak models}, \texttt{multimodal\_\allowbreak ai}, \texttt{generative\_\allowbreak ai}, \texttt{ai\_\allowbreak agents},
++  \texttt{reinforcement\_\allowbreak learning}, \texttt{ml\_\allowbreak theory}, \texttt{deep\_\allowbreak learning\_\allowbreak architectures}, \texttt{ai\_\allowbreak safety} \\
++\texttt{perception\_\allowbreak interaction} & Perception, Language, Robotics and Interaction & 6 &
++  \texttt{computer\_\allowbreak vision}, \texttt{nlp}, \texttt{speech\_\allowbreak audio}, \texttt{robotics}, \texttt{hci}, \texttt{computer\_\allowbreak graphics} \\
++\texttt{software\_\allowbreak pl} & Software, Programming Languages and Formal Methods & 4 &
++  \texttt{software\_\allowbreak engineering}, \texttt{programming\_\allowbreak languages}, \texttt{compilers}, \texttt{formal\_\allowbreak verification} \\
++\texttt{theory} & Theory of Computation & 2 & \texttt{algorithms}, \texttt{computational\_\allowbreak complexity} \\
++\texttt{data\_\allowbreak distributed} & Data, Information and Distributed Systems & 6 &
++  \texttt{database\_\allowbreak systems}, \texttt{information\_\allowbreak retrieval}, \texttt{distributed\_\allowbreak computing}, \texttt{cloud\_\allowbreak computing},
++  \texttt{edge\_\allowbreak computing}, \texttt{computer\_\allowbreak networks} \\
++\texttt{security\_\allowbreak privacy} & Security, Privacy and Trust & 5 &
++  \texttt{cybersecurity}, \texttt{applied\_\allowbreak cryptography}, \texttt{blockchain}, \texttt{privacy\_\allowbreak preserving\_\allowbreak computation},
++  \texttt{federated\_\allowbreak learning} \\
++\texttt{hardware\_\allowbreak emerging} & Hardware, Architecture and Emerging Computing & 5 &
++  \texttt{computer\_\allowbreak architecture}, \texttt{ai\_\allowbreak accelerators}, \texttt{operating\_\allowbreak systems}, \texttt{quantum\_\allowbreak computing},
++  \texttt{sustainable\_\allowbreak computing} \\
++\texttt{computational\_\allowbreak science} & Computational Science and Discovery & 3 &
++  \texttt{scientific\_\allowbreak computing}, \texttt{computational\_\allowbreak biology}, \texttt{autonomous\_\allowbreak scientific\_\allowbreak discovery} \\
++\bottomrule
++\end{tabularx}
++\end{singlespace}
++\end{table}
++
++Two features of the grouping deserve comment because they are choices, not facts. First, \texttt{ai\_\allowbreak safety}
++is placed under the AI area rather than under security and privacy, although adversarial robustness and
++backdoors are scored in both \texttt{ai\_\allowbreak safety} and \texttt{cybersecurity}; the ontology treats safety as a
++property of learning systems, with security as the adjacent field. Second, \texttt{computer\_\allowbreak graphics} sits in
++the perception area next to vision and HCI rather than in a visual-computing area of its own, and
++\texttt{federated\_\allowbreak learning} sits under privacy rather than under distributed computing, which reflects the
++definitional emphasis of the leaf (training without centralising raw data) rather than its algorithmic
++kinship. Readers who would group the leaves differently can do so from the leaf-level tables of Part~IX
++without re-running anything, because areas are never used in scoring.
++
++\subsection{The forty categories}
++\label{app:taxonomy-leaves}
++
++Table~\ref{tab:appA-leaves} reproduces the definition of each leaf exactly as it is given to the classifier's
++author and to the validation raters (the raters receive these definitions and nothing else about the rules).
++Each definition states the object of the category, lists its typical sub-topics and names the neighbouring
++category to which excluded work belongs; the latter clause is what makes the definitions a partition rather than
++a list of keywords, and it is the part most often invoked in the ambiguity decisions of
++Section~\ref{app:taxonomy-ambiguities}. Identifiers are those of the YAML file; numbers are those of the
++mandate.
++
++\begin{singlespace}
++\begin{small}
++\begin{longtable}{@{}r p{3.4cm} p{10.6cm}@{}}
++\caption{The forty leaf categories: identifier, area and definition (\texttt{ontology.yaml}, \texttt{categories[].definition}). Parent/child edges are given under the identifier.}
++\label{tab:appA-leaves}\\
++\toprule
++No. & Category & Definition \\
++\midrule
++\endfirsthead
++\caption[]{The forty leaf categories (continued).}\\
++\toprule
++No. & Category & Definition \\
++\midrule
++\endhead
++\midrule
++\multicolumn{3}{r}{\footnotesize continued on next page}\\
++\endfoot
++\bottomrule
++\endlastfoot
++1 & \textbf{Large Language Models}\newline {\footnotesize \texttt{llm}\newline area \texttt{ai\_\allowbreak ml}; child of \texttt{foundation\_\allowbreak models}} & Research whose object is a transformer-based language model trained at scale: pre-training, instruction tuning, prompting and in-context learning, reasoning, retrieval augmentation, efficient inference, evaluation and benchmarks of LLMs. Excludes classical NLP tasks solved with small task-specific models (nlp) and non-language foundation models (foundation\_models). \\[2pt]
++2 & \textbf{Foundation Models}\newline {\footnotesize \texttt{foundation\_\allowbreak models}\newline area \texttt{ai\_\allowbreak ml}; parent of \texttt{llm}, \texttt{multimodal\_\allowbreak ai}} & Large pre-trained models intended for adaptation to many downstream tasks in any modality: scaling laws, pre-training objectives, self-supervised representation learning at scale, transfer, fine-tuning and adapter methods, model releases and their evaluation. Language-specific work is llm (a child); vision-language work is multimodal\_ai (a child). \\[2pt]
++3 & \textbf{Multimodal AI}\newline {\footnotesize \texttt{multimodal\_\allowbreak ai}\newline area \texttt{ai\_\allowbreak ml}; child of \texttt{foundation\_\allowbreak models}} & Models and systems that jointly process two or more modalities (vision-language, audio-visual, video-text, 3D-language, sensors with language): vision-language models and multimodal LLMs, captioning, visual question answering, cross-modal retrieval and grounding. Unimodal vision is computer\_vision; pure speech is speech\_audio. \\[2pt]
++4 & \textbf{Generative AI}\newline {\footnotesize \texttt{generative\_\allowbreak ai}\newline area \texttt{ai\_\allowbreak ml}} & Models that synthesise new content (images, video, audio, 3D, text, code, molecules) and their training, sampling, evaluation, detection and misuse: GANs, VAEs, diffusion and flow models, autoregressive generators, text-to-X synthesis, deepfakes and AI-generated-content detection. LLM text generation as such is llm unless the contribution concerns the generative model or its outputs' authenticity. \\[2pt]
++5 & \textbf{AI Agents and Autonomous Systems}\newline {\footnotesize \texttt{ai\_\allowbreak agents}\newline area \texttt{ai\_\allowbreak ml}} & Software agents that perceive, decide and act with some autonomy: LLM-based agents and tool use, multi-agent systems, planning and automated reasoning for acting, agent architectures, negotiation, agentic workflows and benchmarks, embodied agents in simulated environments. Physical robots are robotics; RL algorithms as such are reinforcement\_learning. \\[2pt]
++6 & \textbf{Reinforcement Learning}\newline {\footnotesize \texttt{reinforcement\_\allowbreak learning}\newline area \texttt{ai\_\allowbreak ml}} & Learning to act from reward signals: value- and policy-based methods, model-based RL, offline and inverse RL, multi-agent RL, bandits, exploration, reward design, RL from human feedback, and RL theory. Applications to robot control are shared with robotics. \\[2pt]
++7 & \textbf{Machine Learning Theory}\newline {\footnotesize \texttt{ml\_\allowbreak theory}\newline area \texttt{ai\_\allowbreak ml}} & Statistical and computational learning theory and principled ML methodology: generalisation and sample complexity, optimisation theory for learning, kernel and Bayesian methods, probabilistic models, causal inference, online learning, PAC learning, interpretability and fairness as formal problems. Architecture engineering is deep\_learning\_architectures. \\[2pt]
++8 & \textbf{Deep Learning Architectures}\newline {\footnotesize \texttt{deep\_\allowbreak learning\_\allowbreak architectures}\newline area \texttt{ai\_\allowbreak ml}} & Design, training and compression of neural network architectures as such: CNNs, RNNs, transformers and attention, graph neural networks, state-space models, normalisation and optimisers in practice, neural architecture search, pruning, quantisation and distillation, neuromorphic and spiking networks as models. Task-specific results go to the application category; hardware goes to ai\_accelerators. \\[2pt]
++9 & \textbf{Computer Vision}\newline {\footnotesize \texttt{computer\_\allowbreak vision}\newline area \texttt{perception\_\allowbreak interaction}} & Extracting information from images, video and 3D data: recognition, detection, segmentation, tracking, pose, depth and 3D reconstruction, image restoration and enhancement, medical and remote-sensing image analysis, visual representation learning. Image synthesis is generative\_ai; vision-language is multimodal\_ai; rendering is computer\_graphics. \\[2pt]
++10 & \textbf{Natural Language Processing}\newline {\footnotesize \texttt{nlp}\newline area \texttt{perception\_\allowbreak interaction}} & Computational processing of human language: machine translation, parsing and tagging, information extraction, sentiment and text classification, summarisation, question answering, dialogue, text mining, computational linguistics, low-resource and multilingual NLP, NLP evaluation and resources. LLM-centred work is llm (also classified here when the task framing dominates). \\[2pt]
++11 & \textbf{Speech and Audio Computing}\newline {\footnotesize \texttt{speech\_\allowbreak audio}\newline area \texttt{perception\_\allowbreak interaction}} & Processing and generation of speech, music and general audio: speech recognition and synthesis, speaker and language recognition, enhancement and separation, spoken language understanding, music information retrieval and generation, audio event detection, neural audio codecs. Text-only dialogue is nlp. \\[2pt]
++12 & \textbf{Robotics}\newline {\footnotesize \texttt{robotics}\newline area \texttt{perception\_\allowbreak interaction}} & Machines that sense and act in the physical world: manipulation, locomotion, motion and path planning, SLAM and localisation, robot learning, human-robot interaction, autonomous vehicles and drones as robotic systems, soft and swarm robotics, teleoperation. RL algorithms are reinforcement\_learning; robotic process automation and surgical-outcome studies are excluded. \\[2pt]
++13 & \textbf{Human-Computer Interaction}\newline {\footnotesize \texttt{hci}\newline area \texttt{perception\_\allowbreak interaction}} & Design, evaluation and understanding of interactive systems and their users: user studies and usability, interaction techniques, VR/AR/XR, accessibility and assistive technology, CSCW and social computing, human-AI interaction, ubiquitous and wearable computing, affective computing. Rendering is computer\_graphics; physical robots are robotics. \\[2pt]
++14 & \textbf{Software Engineering}\newline {\footnotesize \texttt{software\_\allowbreak engineering}\newline area \texttt{software\_\allowbreak pl}} & How software is specified, built, tested, maintained and studied: requirements, architecture, testing and debugging, program repair, mining software repositories, empirical studies of developers, DevOps, model-driven engineering, open-source ecosystems, and AI-assisted software development. Language design is programming\_languages; proofs of correctness are formal\_verification. \\[2pt]
++15 & \textbf{Programming Languages}\newline {\footnotesize \texttt{programming\_\allowbreak languages}\newline area \texttt{software\_\allowbreak pl}; parent of \texttt{compilers}} & Design, semantics, type systems and implementation principles of programming languages: type theory, operational and denotational semantics, functional and logic programming, effects, concurrency models, memory safety, domain-specific languages, program analysis as a language problem. Code generation and optimisation pipelines are compilers (a child); machine-checked correctness is formal\_verification. \\[2pt]
++16 & \textbf{Compilers}\newline {\footnotesize \texttt{compilers}\newline area \texttt{software\_\allowbreak pl}; child of \texttt{programming\_\allowbreak languages}} & Translation and optimisation of programs for execution: compiler construction, intermediate representations, optimisation passes, code generation, JIT compilation, auto-vectorisation and parallelisation, polyhedral methods, high-level synthesis, and compilers for machine-learning workloads and accelerators. A child of programming\_languages. \\[2pt]
++17 & \textbf{Formal Verification}\newline {\footnotesize \texttt{formal\_\allowbreak verification}\newline area \texttt{software\_\allowbreak pl}} & Mathematically rigorous establishment of system properties: model checking, theorem proving and proof assistants, SAT/SMT solving and automated reasoning, deductive and runtime verification, verified software and hardware, formal semantics as a verification tool, verification of neural networks, smart contracts and protocols, automata and temporal logics. Language semantics as design is programming\_languages. \\[2pt]
++18 & \textbf{Algorithms and Data Structures}\newline {\footnotesize \texttt{algorithms}\newline area \texttt{theory}} & Design and analysis of algorithms and data structures: graph, string, geometric, approximation, randomized, online, streaming, parallel and parameterized algorithms, combinatorial optimisation and metaheuristics, algorithmic game theory, algorithm engineering, and the combinatorics that underpins them. Lower bounds and hardness as such are computational\_complexity; numerical algorithms are scientific\_computing. \\[2pt]
++19 & \textbf{Computational Complexity}\newline {\footnotesize \texttt{computational\_\allowbreak complexity}\newline area \texttt{theory}} & The intrinsic difficulty of computational problems: complexity classes and their separations, hardness of approximation, circuit, proof, communication and query complexity, fine-grained and parameterized complexity, pseudorandomness and derandomization, computability and automata theory, algebraic complexity, and quantum complexity. Algorithm design is algorithms; cryptographic hardness assumptions are applied\_cryptography. \\[2pt]
++20 & \textbf{Database Systems}\newline {\footnotesize \texttt{database\_\allowbreak systems}\newline area \texttt{data\_\allowbreak distributed}} & Storage, querying and management of structured and semi-structured data: query processing and optimisation, transactions and concurrency control, indexing and storage engines, data integration, cleaning and provenance, big-data processing frameworks, data mining as a data-management problem, learned database components, vector and graph databases. Ranking of unstructured documents is information\_retrieval. \\[2pt]
++21 & \textbf{Information Retrieval}\newline {\footnotesize \texttt{information\_\allowbreak retrieval}\newline area \texttt{data\_\allowbreak distributed}} & Finding and ranking relevant information for a user need: search engines and ranking models, dense and sparse retrieval, recommender systems, query understanding, evaluation and test collections, web search and mining, scholarly and digital-library retrieval, retrieval components of RAG systems. Structured query processing is database\_systems; text understanding tasks are nlp. \\[2pt]
++22 & \textbf{Distributed Computing}\newline {\footnotesize \texttt{distributed\_\allowbreak computing}\newline area \texttt{data\_\allowbreak distributed}; parent of \texttt{cloud\_\allowbreak computing}, \texttt{edge\_\allowbreak computing}} & Computation across multiple processors or machines: distributed algorithms and consensus, fault tolerance and replication, parallel programming and HPC systems software, cluster and grid resource management, distributed storage and middleware, distributed and parallel ML training systems. Cloud service models are cloud\_computing and edge placement is edge\_computing (both children). \\[2pt]
++23 & \textbf{Cloud Computing}\newline {\footnotesize \texttt{cloud\_\allowbreak computing}\newline area \texttt{data\_\allowbreak distributed}; child of \texttt{distributed\_\allowbreak computing}} & Computing delivered as elastic, multi-tenant, on-demand services: IaaS/PaaS/SaaS and serverless, virtualisation and containers, resource provisioning and auto-scaling, SLAs and pricing, cloud-native architectures, data-centre resource management, cloud security and migration. A child of distributed\_computing; placement at the network edge is edge\_computing. \\[2pt]
++24 & \textbf{Edge Computing}\newline {\footnotesize \texttt{edge\_\allowbreak computing}\newline area \texttt{data\_\allowbreak distributed}; child of \texttt{distributed\_\allowbreak computing}} & Computation placed near data sources and users: edge/fog/MEC architectures, computation offloading, on-device and TinyML inference, edge-cloud collaboration and split computing, IoT platforms as computing substrates, latency-critical edge services, edge caching and orchestration. Radio access and protocols are computer\_networks; model compression as such is deep\_learning\_architectures or ai\_accelerators. \\[2pt]
++25 & \textbf{Computer Networks}\newline {\footnotesize \texttt{computer\_\allowbreak networks}\newline area \texttt{data\_\allowbreak distributed}} & Design, protocols, measurement and performance of communication networks: Internet architecture and transport, routing, wireless and mobile networks, 5G/6G and software-defined networking, network management, sensor and vehicular networks, content delivery and caching, network coding and information theory applied to networking. Physical-layer signal processing is included only when framed as networking. \\[2pt]
++26 & \textbf{Cybersecurity}\newline {\footnotesize \texttt{cybersecurity}\newline area \texttt{security\_\allowbreak privacy}} & Protecting computing systems, networks, software and users from attack: vulnerability discovery and exploitation, malware and intrusion detection, authentication and access control, system and hardware security, web and mobile security, digital forensics, security of AI systems, security operations and human factors of security. Cryptographic primitives and protocols are applied\_cryptography; data-protection techniques are privacy\_preserving\_computation. \\[2pt]
++27 & \textbf{Applied Cryptography}\newline {\footnotesize \texttt{applied\_\allowbreak cryptography}\newline area \texttt{security\_\allowbreak privacy}} & Design, analysis, implementation and deployment of cryptographic primitives and protocols: symmetric and public-key schemes, hash functions, signatures, key exchange, post-quantum cryptography, zero-knowledge proofs, secure multi-party computation and homomorphic encryption as primitives, cryptanalysis, side-channel-resistant implementations, cryptographic hardware, provable security, quantum key distribution. Deployment of these primitives for data protection is privacy\_preserving\_computation; ledgers are blockchain. \\[2pt]
++28 & \textbf{Blockchain Systems}\newline {\footnotesize \texttt{blockchain}\newline area \texttt{security\_\allowbreak privacy}} & Distributed ledgers and the systems built on them: consensus for permissioned and permissionless chains, smart contracts and their security, scalability (sharding, layer-2, rollups), cryptocurrencies and DeFi mechanisms, blockchain applications and governance, measurement of blockchain ecosystems. Generic BFT consensus theory is shared with distributed\_computing; primitives are applied\_cryptography. \\[2pt]
++29 & \textbf{Computer Architecture}\newline {\footnotesize \texttt{computer\_\allowbreak architecture}\newline area \texttt{hardware\_\allowbreak emerging}; parent of \texttt{ai\_\allowbreak accelerators}} & Organisation and design of processors, memory systems and digital hardware: microarchitecture, caches and memory hierarchies, interconnects and networks-on-chip, FPGAs and reconfigurable computing, embedded and real-time systems, hardware/software co-design, electronic design automation, VLSI and low-power design, processing-in-memory. Accelerators specialised for ML are ai\_accelerators (a child); kernels are operating\_systems. \\[2pt]
++30 & \textbf{Accelerators and AI Hardware}\newline {\footnotesize \texttt{ai\_\allowbreak accelerators}\newline area \texttt{hardware\_\allowbreak emerging}; child of \texttt{computer\_\allowbreak architecture}} & Hardware and hardware-aware software for efficient machine-learning computation: GPU/TPU/NPU design and programming, ASIC and FPGA DNN accelerators, neuromorphic and spiking hardware, in-memory and analog/photonic computing for AI, quantisation and sparsity exploited by hardware, efficient LLM inference and serving systems, ML compilers' hardware mapping, benchmarks such as MLPerf. A child of computer\_architecture. \\[2pt]
++31 & \textbf{Operating Systems}\newline {\footnotesize \texttt{operating\_\allowbreak systems}\newline area \texttt{hardware\_\allowbreak emerging}} & System software that manages hardware and provides execution abstractions: kernels and drivers, scheduling, memory management, file and storage systems, virtualisation and containers, OS security and isolation, I/O and networking stacks, real-time and embedded OS, systems support for new hardware (persistent memory, CXL, SmartNICs). Distributed coordination is distributed\_computing; hardware design is computer\_architecture. \\[2pt]
++32 & \textbf{Quantum Computing}\newline {\footnotesize \texttt{quantum\_\allowbreak computing}\newline area \texttt{hardware\_\allowbreak emerging}} & Computation and information processing exploiting quantum mechanics: quantum algorithms and complexity, quantum error correction and fault tolerance, qubit hardware platforms and control, quantum software, compilers and simulators, variational and NISQ algorithms, quantum machine learning, quantum communication and key distribution, quantum networking. Quantum physics without a computational or information-processing framing is out of scope; post-quantum cryptography is applied\_cryptography. \\[2pt]
++33 & \textbf{Scientific Computing}\newline {\footnotesize \texttt{scientific\_\allowbreak computing}\newline area \texttt{computational\_\allowbreak science}} & Numerical and computational methods and software for mathematical models of science and engineering: numerical linear algebra, discretisation of differential equations, solvers and preconditioners, simulation frameworks, uncertainty quantification, reduced-order and surrogate models, scientific machine learning (PINNs, neural operators), high-performance scientific software and performance portability, mathematical software, computational physics/chemistry/engineering tooling. Domain biology is computational\_biology; parallel systems software is distributed\_computing. \\[2pt]
++34 & \textbf{Computational Biology}\newline {\footnotesize \texttt{computational\_\allowbreak biology}\newline area \texttt{computational\_\allowbreak science}} & Computational methods for biological and biomedical data: bioinformatics and genomics, protein structure and design, molecular modelling and computational drug discovery, single-cell and multi-omics analysis, systems and synthetic biology, biomedical text mining, computational neuroscience, biomedical and health informatics methods, DNA/molecular computing. Medical image analysis is shared with computer\_vision; numerical methods as such are scientific\_computing. \\[2pt]
++35 & \textbf{Computer Graphics}\newline {\footnotesize \texttt{computer\_\allowbreak graphics}\newline area \texttt{perception\_\allowbreak interaction}} & Synthesis and manipulation of visual content from models: rendering, geometry processing and modelling, animation and simulation for graphics, neural rendering and 3D representations, visualisation, procedural content, game and VR engines. Image understanding is computer\_vision; pixel-space generative models are generative\_ai; user studies of VR are hci. \\[2pt]
++36 & \textbf{Privacy-Preserving Computation}\newline {\footnotesize \texttt{privacy\_\allowbreak preserving\_\allowbreak computation}\newline area \texttt{security\_\allowbreak privacy}; parent of \texttt{federated\_\allowbreak learning}} & Computing on, releasing or sharing data while limiting what is learned about individuals: differential privacy, homomorphic encryption, secure multi-party computation and TEEs applied to data analysis and ML, anonymisation and synthetic data, private information retrieval, privacy attacks on models and their defences, machine unlearning, privacy engineering and regulation (GDPR). Federated learning is a child; cryptographic constructions as such are applied\_cryptography. \\[2pt]
++37 & \textbf{Federated Learning}\newline {\footnotesize \texttt{federated\_\allowbreak learning}\newline area \texttt{security\_\allowbreak privacy}; child of \texttt{privacy\_\allowbreak preserving\_\allowbreak computation}} & Training models across multiple clients or institutions without centralising their raw data: federated averaging and optimisation under heterogeneity, communication efficiency, client selection, personalisation, secure aggregation and privacy of federated systems, robustness to malicious clients, vertical and cross-silo federated learning, federated fine-tuning of foundation models, decentralised/gossip learning. A child of privacy\_preserving\_computation; also linked to distributed\_computing and edge\_computing. \\[2pt]
++38 & \textbf{AI Safety and Alignment}\newline {\footnotesize \texttt{ai\_\allowbreak safety}\newline area \texttt{ai\_\allowbreak ml}} & Making AI systems safe, aligned, robust, fair, interpretable and governable: alignment and preference learning framed as a safety problem, red-teaming and jailbreaks, adversarial robustness and backdoors, hallucination and truthfulness, bias and fairness, mechanistic interpretability, evaluation of dangerous capabilities, AI governance and regulation. Generic system security is cybersecurity; XAI as a statistical method only is ml\_theory. \\[2pt]
++39 & \textbf{Sustainable Computing}\newline {\footnotesize \texttt{sustainable\_\allowbreak computing}\newline area \texttt{hardware\_\allowbreak emerging}} & The energy, carbon, water and material footprint of computing and how to reduce it: energy-efficient and carbon-aware systems, data-centre energy and cooling, Green AI and the cost of training and inference, life-cycle and embodied carbon of hardware, green software engineering and networking, energy harvesting and intermittent computing, computing for sustainability when the object studied is the computing system. Generic low-power circuit design is computer\_architecture; energy-efficient communication is computer\_networks. \\[2pt]
++40 & \textbf{Autonomous Scientific Discovery}\newline {\footnotesize \texttt{autonomous\_\allowbreak scientific\_\allowbreak discovery}\newline area \texttt{computational\_\allowbreak science}} & AI and automation systems that perform or substantially drive the scientific process: AI scientist and research-agent systems, automated hypothesis generation and literature synthesis, self-driving laboratories and closed-loop experimentation, AI-driven discovery of materials, molecules, equations, algorithms or theorems, autonomous experimental design, benchmarks for research agents, and the epistemology and evaluation of machine-generated science. Using ML inside a scientific pipeline without autonomy is scientific\_computing or computational\_biology; LLM agents in general are ai\_agents. \\[2pt]
++\end{longtable}
++\end{small}
++\end{singlespace}
++
++\subsection{Signal types and the assignment rule}
++\label{app:taxonomy-rule}
++
++Every leaf lists weighted signals of five types, and the five types are the only evidence the classifier ever
++sees. (i)~\emph{OpenAlex topics}: the up-to-three topic identifiers that OpenAlex attaches to a work, drawn
++from any of the 26 OpenAlex fields and not only from field~17; a topic may appear under several leaves with
++different weights. (ii)~\emph{OpenAlex subfields}: a coarse fallback used by 21 leaf--subfield entries over
++eleven subfields (the ten field-17 subfields and subfield~1312, Molecular Biology, for computational biology),
++always with a small weight. (iii)~\emph{arXiv categories}: the primary category and cross-lists of the linked
++arXiv record, 162 entries over 49 distinct categories, namely
++\texttt{cs.AI}, \texttt{cs.AR}, \texttt{cs.CC}, \texttt{cs.CE}, \texttt{cs.CG}, \texttt{cs.CL}, \texttt{cs.CR}, \texttt{cs.CV}, \texttt{cs.CY}, \texttt{cs.DB}, \texttt{cs.DC}, \texttt{cs.DL}, \texttt{cs.DM}, \texttt{cs.DS}, \texttt{cs.ET}, \texttt{cs.FL}, \texttt{cs.GR}, \texttt{cs.GT}, \texttt{cs.HC}, \texttt{cs.IR}, \texttt{cs.IT}, \texttt{cs.LG}, \texttt{cs.LO}, \texttt{cs.MA}, \texttt{cs.MM}, \texttt{cs.MS}, \texttt{cs.NA}, \texttt{cs.NE}, \texttt{cs.NI}, \texttt{cs.OS}, \texttt{cs.PF}, \texttt{cs.PL}, \texttt{cs.RO}, \texttt{cs.SD}, \texttt{cs.SE}, \texttt{cs.SI}, \texttt{cs.SY}, \texttt{eess.AS}, \texttt{eess.IV}, \texttt{eess.SP}, \texttt{eess.SY}, \texttt{math.IT}, \texttt{math.OC}, \texttt{math.ST}, \texttt{physics.comp-ph}, \texttt{q-bio.GN}, \texttt{q-bio.QM}, \texttt{quant-ph}, \texttt{stat.ML}. (iv)~\emph{DBLP streams}: the venue stream (journal or conference series,
++the prefix of the DBLP key) of the linked DBLP record, 1{,}320 entries over 771 distinct streams.
++(v)~\emph{Keyword patterns}: 273 regular expressions matched, in the present version of the ontology without
++exception, on the concatenation of title and abstract (the YAML schema allows title-only and abstract-only
++patterns, but none is used); 24 exclusion patterns veto a leaf when they match the same text.
++
++The rule itself is the one written out in equations~\eqref{eq:score} and~\eqref{eq:assign}: within each
++signal type the weights of the matched signals are summed and capped at~1.0; the capped values of the five
++types are summed; the leaf is assigned when the total reaches the threshold $\theta = 1.0$ and no exclusion
++pattern matches; the primary leaf is the argmax, ties going to the child before the parent and then to the
++lower category number. The weights are therefore to be read on a scale where 1.0 means ``this signal alone
++suffices''. Table~\ref{tab:appA-signals} reports, for each leaf, how many signals of each type it lists and
++how many of them carry weight~1.0; Table~\ref{tab:appA-area-signals} aggregates the same counts by area.
++Three regularities are visible. First, every one of the forty leaves has at least one keyword pattern of
++weight 1.0 (106 of the 273 patterns are at 1.0, the rest spread between 0.2 and 0.9), so that a title or
++abstract naming the object of the category assigns it without metadata support; the build script refuses to
++serialise an ontology in which this is false for the thirteen categories that the design singles out (the
++six AI categories that OpenAlex topics do not separate and seven small categories without a topic or venue of
++their own), and in version~1.0 the property happens to hold for all forty. Second, the number of topic
++signals varies by an order of magnitude across leaves, from 4 for \texttt{compilers},
++\texttt{reinforcement\_\allowbreak learning} and \texttt{federated\_\allowbreak learning} to 45 for \texttt{computer\_\allowbreak networks} and 44 for \texttt{computer\_\allowbreak vision}; this
++mirrors the granularity of the OpenAlex topic model, which has many clusters for networking and vision and
++none for federated learning, and not the importance of the categories. Third, venue signals are dense (291 of
++the 1{,}320 stream entries are at weight 1.0, almost all for single-purpose venues) but concentrated on the
++systems, networking and theory leaves, where the DBLP venue structure is informative, and sparse for the
++post-2022 AI leaves, where the same few venues host everything.
++
++\begin{singlespace}
++\begin{footnotesize}
++\begin{longtable}{@{}r p{4.3cm} r r r r r r r r r r@{}}
++\caption{Signals per category in \texttt{ontology.yaml}: number of weighted OpenAlex topics (T), of which with weight 1.0 (T$_1$); OpenAlex subfields (F); arXiv categories (X); DBLP streams (D), of which with weight 1.0 (D$_1$); keyword patterns (K), of which with weight 1.0 (K$_1$); exclusion patterns (Z); related edges (R). Computed from the YAML file.}
++\label{tab:appA-signals}\\
++\toprule
++No. & Category & T & T$_1$ & F & X & D & D$_1$ & K & K$_1$ & Z & R \\
++\midrule
++\endfirsthead
++\caption[]{Signals per category (continued).}\\
++\toprule
++No. & Category & T & T$_1$ & F & X & D & D$_1$ & K & K$_1$ & Z & R \\
++\midrule
++\endhead
++\midrule
++\multicolumn{12}{r}{continued on next page}\\
++\endfoot
++\bottomrule
++\endlastfoot
++1 & \texttt{llm} & 6 & 1 & 1 & 3 & 12 & 0 & 12 & 3 & 0 & 5 \\
++2 & \texttt{foundation\_\allowbreak models} & 5 & 0 & 1 & 4 & 8 & 0 & 10 & 2 & 0 & 5 \\
++3 & \texttt{multimodal\_\allowbreak ai} & 5 & 1 & 1 & 4 & 15 & 0 & 9 & 4 & 1 & 5 \\
++4 & \texttt{generative\_\allowbreak ai} & 5 & 1 & 0 & 6 & 9 & 0 & 11 & 4 & 2 & 6 \\
++5 & \texttt{ai\_\allowbreak agents} & 5 & 1 & 1 & 2 & 12 & 3 & 9 & 5 & 1 & 5 \\
++6 & \texttt{reinforcement\_\allowbreak learning} & 4 & 1 & 0 & 5 & 8 & 0 & 9 & 3 & 1 & 5 \\
++7 & \texttt{ml\_\allowbreak theory} & 20 & 1 & 1 & 4 & 51 & 2 & 11 & 1 & 0 & 6 \\
++8 & \texttt{deep\_\allowbreak learning\_\allowbreak architectures} & 11 & 2 & 1 & 4 & 21 & 0 & 12 & 1 & 1 & 6 \\
++9 & \texttt{computer\_\allowbreak vision} & 44 & 9 & 1 & 2 & 58 & 13 & 10 & 3 & 0 & 6 \\
++10 & \texttt{nlp} & 22 & 4 & 1 & 2 & 29 & 13 & 8 & 2 & 0 & 5 \\
++11 & \texttt{speech\_\allowbreak audio} & 10 & 3 & 1 & 4 & 16 & 5 & 5 & 3 & 0 & 4 \\
++12 & \texttt{robotics} & 23 & 6 & 0 & 3 & 41 & 27 & 4 & 1 & 2 & 5 \\
++13 & \texttt{hci} & 24 & 4 & 1 & 3 & 44 & 22 & 10 & 1 & 0 & 6 \\
++14 & \texttt{software\_\allowbreak engineering} & 20 & 7 & 1 & 1 & 39 & 18 & 5 & 3 & 0 & 6 \\
++15 & \texttt{programming\_\allowbreak languages} & 6 & 1 & 0 & 3 & 22 & 6 & 6 & 3 & 0 & 4 \\
++16 & \texttt{compilers} & 4 & 0 & 0 & 5 & 25 & 2 & 5 & 3 & 0 & 5 \\
++17 & \texttt{formal\_\allowbreak verification} & 10 & 1 & 0 & 4 & 43 & 10 & 7 & 5 & 0 & 6 \\
++18 & \texttt{algorithms} & 35 & 2 & 1 & 7 & 68 & 9 & 10 & 2 & 0 & 6 \\
++19 & \texttt{computational\_\allowbreak complexity} & 9 & 0 & 1 & 7 & 28 & 2 & 7 & 5 & 1 & 5 \\
++20 & \texttt{database\_\allowbreak systems} & 21 & 1 & 1 & 1 & 36 & 9 & 5 & 2 & 1 & 5 \\
++21 & \texttt{information\_\allowbreak retrieval} & 20 & 2 & 1 & 3 & 29 & 5 & 6 & 3 & 0 & 5 \\
++22 & \texttt{distributed\_\allowbreak computing} & 16 & 2 & 1 & 4 & 51 & 15 & 7 & 3 & 1 & 8 \\
++23 & \texttt{cloud\_\allowbreak computing} & 11 & 2 & 0 & 4 & 29 & 5 & 5 & 2 & 1 & 6 \\
++24 & \texttt{edge\_\allowbreak computing} & 13 & 1 & 0 & 4 & 38 & 1 & 5 & 2 & 1 & 7 \\
++25 & \texttt{computer\_\allowbreak networks} & 45 & 9 & 1 & 4 & 96 & 42 & 5 & 1 & 0 & 5 \\
++26 & \texttt{cybersecurity} & 34 & 7 & 0 & 1 & 29 & 16 & 6 & 2 & 1 & 8 \\
++27 & \texttt{applied\_\allowbreak cryptography} & 14 & 3 & 0 & 4 & 29 & 9 & 6 & 3 & 0 & 6 \\
++28 & \texttt{blockchain} & 6 & 1 & 0 & 3 & 19 & 3 & 4 & 2 & 1 & 5 \\
++29 & \texttt{computer\_\allowbreak architecture} & 21 & 4 & 1 & 5 & 59 & 6 & 4 & 2 & 1 & 7 \\
++30 & \texttt{ai\_\allowbreak accelerators} & 11 & 1 & 0 & 5 & 44 & 1 & 5 & 3 & 1 & 7 \\
++31 & \texttt{operating\_\allowbreak systems} & 12 & 0 & 0 & 5 & 39 & 6 & 4 & 2 & 1 & 6 \\
++32 & \texttt{quantum\_\allowbreak computing} & 16 & 2 & 0 & 6 & 27 & 4 & 3 & 2 & 1 & 5 \\
++33 & \texttt{scientific\_\allowbreak computing} & 41 & 4 & 1 & 6 & 57 & 10 & 4 & 2 & 0 & 8 \\
++34 & \texttt{computational\_\allowbreak biology} & 34 & 3 & 1 & 3 & 60 & 12 & 3 & 2 & 0 & 7 \\
++35 & \texttt{computer\_\allowbreak graphics} & 15 & 2 & 1 & 4 & 21 & 8 & 9 & 2 & 0 & 5 \\
++36 & \texttt{privacy\_\allowbreak preserving\_\allowbreak computation} & 13 & 1 & 0 & 4 & 26 & 1 & 5 & 3 & 1 & 7 \\
++37 & \texttt{federated\_\allowbreak learning} & 4 & 0 & 0 & 4 & 15 & 0 & 5 & 2 & 1 & 6 \\
++38 & \texttt{ai\_\allowbreak safety} & 8 & 1 & 0 & 5 & 13 & 4 & 14 & 7 & 1 & 7 \\
++39 & \texttt{sustainable\_\allowbreak computing} & 11 & 2 & 0 & 6 & 28 & 2 & 4 & 2 & 1 & 7 \\
++40 & \texttt{autonomous\_\allowbreak scientific\_\allowbreak discovery} & 12 & 0 & 0 & 8 & 26 & 0 & 4 & 3 & 1 & 8 \\
++\midrule
++& Total (entries) & 646 & 93 & 21 & 162 & 1320 & 291 & 273 & 106 & 24 & 236 \\
++\end{longtable}
++\end{footnotesize}
++\end{singlespace}
++
++\begin{table}[htbp]
++\begin{singlespace}
++\centering
++\caption{Signal entries aggregated by area (sums of the rows of Table~\ref{tab:appA-signals}; a topic or venue mapped to several leaves of the same area is counted once per leaf).}
++\label{tab:appA-area-signals}
++\small
++\begin{tabular}{l r r r r r}
++\toprule
++Area & Leaves & Topic entries & arXiv entries & DBLP entries & Keyword patterns \\
++\midrule
++\texttt{ai\_\allowbreak ml} & 9 & 69 & 37 & 149 & 97 \\
++\texttt{perception\_\allowbreak interaction} & 6 & 138 & 18 & 209 & 46 \\
++\texttt{software\_\allowbreak pl} & 4 & 40 & 13 & 129 & 23 \\
++\texttt{theory} & 2 & 44 & 14 & 96 & 17 \\
++\texttt{data\_\allowbreak distributed} & 6 & 126 & 20 & 279 & 33 \\
++\texttt{security\_\allowbreak privacy} & 5 & 71 & 16 & 118 & 26 \\
++\texttt{hardware\_\allowbreak emerging} & 5 & 71 & 27 & 197 & 20 \\
++\texttt{computational\_\allowbreak science} & 3 & 87 & 17 & 143 & 11 \\
++\bottomrule
++\end{tabular}
++\end{singlespace}
++\end{table}
++
++The weight scale has a second consequence that the reader of Part~IX must keep in mind. A signal with weight
++below 1.0 can only contribute through corroboration by a signal of \emph{another} type, because repetitions
++within a type are capped. The design relies on this in both directions: a general machine-learning venue at
++0.3 and the \texttt{cs.LG} category at 0.2 never assign \texttt{ml\_\allowbreak theory} by themselves, whereas a 0.6 topic
++and a 0.4 arXiv category do reach the threshold. One discrepancy between the documentation and the
++implementation of the within-type aggregation (maximum versus sum over matched topics of the same leaf) is
++recorded in Section~\ref{sec:m-score}; the implemented rule, the sum, is the one described here and the one
++fixed by the unit tests.
++
++\subsection{Coverage of the OpenAlex topic vocabulary}
++\label{app:taxonomy-openalex}
++
++The OpenAlex export of 2026-10-14 contains 4{,}516 topics, 302 of them in field~17 (Computer Science). The
++ontology maps 240 of the 302 (79.5\,\%) to at least one leaf and leaves 62 unmapped with an explicit reason
++each (Section~\ref{app:taxonomy-unmapped}). Table~\ref{tab:appA-subfields} breaks the field-17 coverage down
++by subfield. Coverage is complete for Computer Vision and Pattern Recognition, Hardware and Architecture,
++Human--Computer Interaction, Signal Processing and Software, and nearly complete for Computational Theory and
++Mathematics; it is lowest for Information Systems (45 of 78), the subfield into which OpenAlex places most of
++its educational-technology and media clusters, and for Artificial Intelligence (61 of 77), where the unmapped
++topics are physical-science, generic residual or empty clusters rather than AI research.
++
++\begin{table}[htbp]
++\begin{singlespace}
++\centering
++\caption{Coverage of the OpenAlex field~17 (Computer Science) topics by subfield: number of topics in the 2026-10-14 export, mapped to at least one category, and left unmapped with a stated reason (Table~\ref{tab:appA-unmapped}).}
++\label{tab:appA-subfields}
++\small
++\begin{tabular}{l l r r r}
++\toprule
++Subfield & Name & Topics & Mapped & Unmapped \\
++\midrule
++1702 & Artificial Intelligence & 77 & 61 & 16 \\
++1703 & Computational Theory and Mathematics & 24 & 23 & 1 \\
++1704 & Computer Graphics and Computer-Aided Design & 4 & 3 & 1 \\
++1705 & Computer Networks and Communications & 43 & 35 & 8 \\
++1706 & Computer Science Applications & 10 & 7 & 3 \\
++1707 & Computer Vision and Pattern Recognition & 35 & 35 & 0 \\
++1708 & Hardware and Architecture & 8 & 8 & 0 \\
++1709 & Human-Computer Interaction & 8 & 8 & 0 \\
++1710 & Information Systems & 78 & 45 & 33 \\
++1711 & Signal Processing & 10 & 10 & 0 \\
++1712 & Software & 5 & 5 & 0 \\
++\midrule
++& Field 17 total & 302 & 240 & 62 \\
++\bottomrule
++\end{tabular}
++\end{singlespace}
++\end{table}
++
++Topics outside field~17 matter because computer-science work is frequently assigned a primary topic from
++Engineering, Mathematics, Physics or the life sciences, and because several leaves (computational biology,
++scientific computing, quantum computing, speech and audio, robotics) have their natural OpenAlex home outside
++field~17. The ontology therefore maps 170 further topics from eighteen other fields
++(Table~\ref{tab:appA-fields}), Engineering (71) and Mathematics (20) being the largest contributors, for a
++total of 410 distinct mapped topics and 646 topic--leaf weights. These adjacent topics were selected by
++reading the OpenAlex keyword lists of about 120 candidate topics named in the area modules; the remaining
++roughly 4{,}000 topics of the export were not inspected, which is the main known gap of the mapping
++(Section~\ref{app:taxonomy-limits}). Of the 410 mapped topics, 146 point to more than one leaf, up to six for
++the broadest clusters (for instance T10054, \emph{Parallel Computing and Optimization Techniques}, scores in
++\texttt{computer\_\allowbreak architecture} at 0.8, \texttt{compilers} and \texttt{distributed\_\allowbreak computing} at 0.6,
++\texttt{scientific\_\allowbreak computing} at 0.4, \texttt{ai\_\allowbreak accelerators} at 0.3 and \texttt{operating\_\allowbreak systems} at
++0.2). The full audit list of multiply mapped topics is the last generated section of \texttt{ONTOLOGY.md};
++it is the first place to look when a co-classification pattern in Part~IX seems suspicious, because a topic
++mapped to two leaves at weight 1.0 produces that pair mechanically.
++
++\begin{table}[htbp]
++\begin{singlespace}
++\centering
++\caption{Distinct OpenAlex topics mapped to at least one category, by OpenAlex field of the topic (\texttt{topic\_map.csv}).}
++\label{tab:appA-fields}
++\small
++\begin{tabular}{r l r}
++\toprule
++Field & Name & Mapped topics \\
++\midrule
++17 & Computer Science & 240 \\
++22 & Engineering & 71 \\
++26 & Mathematics & 20 \\
++13 & Biochemistry, Genetics and Molecular Biology & 18 \\
++31 & Physics and Astronomy & 15 \\
++33 & Social Sciences & 10 \\
++18 & Decision Sciences & 8 \\
++27 & Medicine & 6 \\
++12 & Arts and Humanities & 5 \\
++28 & Neuroscience & 4 \\
++23 & Environmental Science & 3 \\
++19 & Earth and Planetary Sciences & 2 \\
++32 & Psychology & 2 \\
++11 & Agricultural and Biological Sciences & 1 \\
++14 & Business, Management and Accounting & 1 \\
++20 & Economics, Econometrics and Finance & 1 \\
++21 & Energy & 1 \\
++25 & Materials Science & 1 \\
++36 & Health Professions & 1 \\
++\midrule
++& All fields & 410 \\
++\bottomrule
++\end{tabular}
++\end{singlespace}
++\end{table}
++
++\subsection{Coverage of the DBLP venue vocabulary}
++\label{app:taxonomy-dblp}
++
++DBLP streams are the venue signal. The stream table from which the mapping was built lists every DBLP
++journal or conference series with at least one hundred records in 2018--2026 (3{,}569 streams); the ontology
++maps 771 of them. Measured on the DBLP records of 2018--2026 and excluding CoRR (the arXiv mirror stream,
++which carries no venue information), the mapped streams cover 2{,}017{,}910 of 3{,}319{,}884 records
++(60.8\,\%). Among the 599 largest streams, 496 are mapped (82.8\,\%); the 103 unmapped large streams
++(Table~\ref{tab:appA-dblp-unmapped} lists the 25 largest) are of two kinds. Some are out of scope for the
++forty categories although DBLP indexes them: control theory (\emph{IEEE Transactions on Automatic Control},
++\emph{Automatica}, \emph{IEEE Control Systems Letters}), instrumentation and industrial electronics,
++geo-information, agriculture, production research, information-systems and engineering-education
++conferences (HICSS, AMCIS, ICIS, FIE). Others are too generic to be a category signal at any weight, the
++largest being \emph{IEEE Access} with 108{,}838 records, followed by \emph{Information}, \emph{Complexity},
++\emph{Systems} and \emph{PeerJ Computer Science}; works published there are classified from their topics,
++arXiv categories and text alone. A record in an unmapped stream is therefore not excluded from the corpus
++(scope is defined in Part~III, not here); it simply receives no venue evidence. Because weights are assigned
++per stream, workshop volumes that share a stream key with a flagship conference inherit the flagship's
++weight, a known coarseness of the venue signal.
++
++\begin{table}[htbp]
++\begin{singlespace}
++\centering
++\caption{The 25 largest DBLP streams (2018--2026 records, CoRR excluded) deliberately left unmapped, out of 103 unmapped streams among the 599 largest (\texttt{ONTOLOGY.md}, generated section ``DBLP coverage'').}
++\label{tab:appA-dblp-unmapped}
++\small
++\begin{tabular}{l r l}
++\toprule
++Stream & Records & Venue \\
++\midrule
++\texttt{journals/access} & 108,838 & IEEE Access \\
++\texttt{journals/tim} & 13,470 & IEEE Trans. Instrum. Meas. \\
++\texttt{journals/tie} & 10,844 & IEEE Trans. Ind. Electron. \\
++\texttt{conf/hicss} & 7,816 & HICSS \\
++\texttt{conf/iecon} & 7,727 & IECON \\
++\texttt{journals/cea} & 7,291 & Comput. Electron. Agric. \\
++\texttt{conf/smc} & 6,048 & SMC \\
++\texttt{journals/complexity} & 5,786 & Complex. \\
++\texttt{journals/ress} & 5,732 & Reliab. Eng. Syst. Saf. \\
++\texttt{journals/tac} & 5,687 & IEEE Trans. Autom. Control. \\
++\texttt{journals/information} & 5,621 & Inf. \\
++\texttt{journals/taes} & 5,543 & IEEE Trans. Aerosp. Electron. Syst. \\
++\texttt{journals/jfi} & 5,432 & J. Frankl. Inst. \\
++\texttt{journals/ijgi} & 5,118 & ISPRS Int. J. Geo Inf. \\
++\texttt{journals/automatica} & 4,727 & Autom. \\
++\texttt{conf/eucc} & 4,160 & ECC \\
++\texttt{conf/iccsa} & 3,817 & ICCSA (4) \\
++\texttt{conf/fie} & 3,807 & FIE \\
++\texttt{conf/amcis} & 3,798 & AMCIS \\
++\texttt{journals/peerj-cs} & 3,768 & PeerJ Comput. Sci. \\
++\texttt{journals/systems} & 3,667 & Syst. \\
++\texttt{journals/csysl} & 3,647 & IEEE Control. Syst. Lett. \\
++\texttt{journals/aei} & 3,628 & Adv. Eng. Informatics \\
++\texttt{journals/ijpr} & 3,591 & Int. J. Prod. Res. \\
++\texttt{journals/sj} & 3,558 & IEEE Syst. J. \\
++\bottomrule
++\end{tabular}
++\end{singlespace}
++\end{table}
++
++\subsection{Unmapped field-17 topics and the reasons}
++\label{app:taxonomy-unmapped}
++
++OpenAlex places in field~17 a number of clusters that are not computer-science research. The ontology leaves
++62 field-17 topics unmapped and records a reason for each; Table~\ref{tab:appA-reasons} groups the reasons
++into eight codes and Table~\ref{tab:appA-unmapped} lists the topics. The dominant case, 29 topics, is
++education and e-learning research (pedagogy, learning outcomes, teacher development, educational
++technology studied as an intervention rather than as software); such work is not one of the forty categories,
++and educational technology \emph{as software} is still captured through software-engineering and HCI venues
++and keywords when the contribution is technical. Nine topics are physical-science or engineering clusters
++without computing content (sensor instrumentation, photovoltaics, geochemistry, seismology, nonlinear
++dynamics, control theory, transportation safety, chemical and environmental engineering). Five are generic
++residual labels (``Computer Science and Engineering'', ``Diverse Interdisciplinary Research Studies'',
++``Mathematics, Computing, and Information Processing'', ``Statistical and Computational Modeling'',
++``Advanced Technologies and Applied Computing'') whose keyword lists carry no specific content; works bearing
++them reach categories through their other topics, venue, arXiv category and text. Four are media,
++communication, literary or philosophical clusters, four are out-of-scope disciplines (psychiatry, economic
++growth, cognitive science, pandemic health care), one is the history of computing (kept out of the
++research-category map but relevant to Part~II), one is the small cognitive-informatics cluster that has no
++counterpart among the forty leaves and was not forced into \texttt{ai\_\allowbreak agents}, and nine are residual topics
++with zero works in the export. The works column of Table~\ref{tab:appA-unmapped} is the OpenAlex
++\texttt{works\_\allowbreak count} of the topic over all years and fields; it is reported to show the scale of what is set
++aside (the education cluster alone exceeds one million works, the media cluster T13745 alone 204{,}789) and is
++not a count of corpus records. The hand-written design note of \texttt{ONTOLOGY.md} speaks of 63 unmapped
++clusters; the generated tables, which are authoritative, list 62.
++
++\begin{table}[htbp]
++\begin{singlespace}
++\centering
++\caption{Reason codes for the unmapped field~17 topics (Table~\ref{tab:appA-unmapped}), with the number of topics and the sum of their OpenAlex \texttt{works\_\allowbreak count}.}
++\label{tab:appA-reasons}
++\small
++\begin{tabularx}{\textwidth}{c X r r}
++\toprule
++Code & Reason recorded in the ontology & Topics & Works (sum) \\
++\midrule
++E & Education / e-learning research; not a CS research category; technical ed-tech captured elsewhere & 29 & 1{,}027{,}591 \\
++P & Physical-science or engineering cluster without computing contribution, placed in field~17 by OpenAlex & 9 & 352{,}761 \\
++Z & Residual OpenAlex topic with zero works in the export; nothing to map & 9 & 0 \\
++G & Generic, uninformative residual label; works reach categories through other signals & 5 & 30{,}206 \\
++M & Media, communication, cultural, literary or philosophy-of-science cluster; not CS research & 4 & 275{,}159 \\
++O & Out-of-scope discipline (psychiatry, economics, cognitive science, pandemic health care) & 4 & 33{,}228 \\
++H & History of computing technologies (history of science); background for Part~II only & 1 & 72{,}582 \\
++C & Cognitive informatics / brain informatics; idiosyncratic cluster with no counterpart among the leaves & 1 & 13{,}053 \\
++\midrule
++& Total & 62 & 1{,}804{,}580 \\
++\bottomrule
++\end{tabularx}
++\end{singlespace}
++\end{table}
++
++\begin{singlespace}
++\begin{footnotesize}
++\begin{longtable}{@{}l p{7.4cm} l r l@{}}
++\caption{The 62 OpenAlex field~17 topics left unmapped, with the subfield, the number of works of the topic in the 2026-10-14 topic export (all years, all fields) and the reason code of Table~\ref{tab:appA-reasons}.}
++\label{tab:appA-unmapped}\\
++\toprule
++Topic & Name & Subfield & Works & Reason \\
++\midrule
++\endfirsthead
++\caption[]{Unmapped field~17 topics (continued).}\\
++\toprule
++Topic & Name & Subfield & Works & Reason \\
++\midrule
++\endhead
++\midrule
++\multicolumn{5}{r}{continued on next page}\\
++\endfoot
++\bottomrule
++\endlastfoot
++T13745 & Media and Digital Communication & 1705 & 204,789 & M \\
++T14496 & Educational Technology and E-Learning & 1710 & 163,032 & E \\
++T14516 & English Language Learning and Teaching & 1710 & 107,202 & E \\
++T14430 & Higher Education and Teaching Methods & 1710 & 103,711 & E \\
++T12797 & History of Computing Technologies & 1706 & 72,582 & H \\
++T12564 & Sensor Technology and Measurement Systems & 1705 & 71,005 & P \\
++T11276 & Solar Radiation and Photovoltaics & 1702 & 70,601 & P \\
++T14455 & Technology and Education Systems & 1705 & 70,588 & E \\
++T12157 & Geochemistry and Geologic Mapping & 1702 & 69,914 & P \\
++T12171 & Open Education and E-Learning & 1706 & 66,332 & E \\
++T14498 & Educational Technology in Learning & 1710 & 63,096 & E \\
++T14216 & Multimedia Learning Systems & 1710 & 62,005 & E \\
++T11187 & Nonlinear Dynamics and Pattern Formation & 1705 & 61,626 & P \\
++T14103 & Educational Management and Quality & 1710 & 57,729 & E \\
++T11370 & Mobile Learning in Education & 1710 & 55,961 & E \\
++T12408 & Educational Innovations and Challenges & 1710 & 53,673 & E \\
++T12395 & Educational Innovations and Technology & 1710 & 45,138 & E \\
++T14025 & Educational Technology and Assessment & 1710 & 37,519 & E \\
++T13602 & Educational Methods and Media Use & 1710 & 35,313 & E \\
++T13194 & ICT in Developing Communities & 1710 & 32,298 & M \\
++T13615 & Transportation Systems and Safety & 1704 & 26,785 & P \\
++T14215 & Chemical and Environmental Engineering Research & 1710 & 25,315 & P \\
++T13018 & Seismology and Earthquake Studies & 1702 & 23,104 & P \\
++T13081 & Digital Education and Society & 1710 & 22,995 & E \\
++T14029 & Thoreau and American Literature & 1710 & 19,567 & M \\
++T13386 & Educational Methods and Teacher Development & 1710 & 19,565 & E \\
++T14345 & Scientific Research and Philosophical Inquiry & 1710 & 18,505 & M \\
++T13016 & Innovative Educational Technologies & 1710 & 16,953 & E \\
++T13674 & Computer Science and Engineering & 1702 & 16,620 & G \\
++T13559 & Edcuational Technology Systems & 1702 & 16,507 & E \\
++T14381 & Psychiatry, Mental Health, Neuroscience & 1702 & 15,800 & O \\
++T14519 & Innovations in Education and Learning Technologies & 1710 & 14,509 & E \\
++T13062 & Cognitive Computing and Networks & 1702 & 13,053 & C \\
++T13867 & Economic Growth and Development & 1710 & 9,780 & O \\
++T13898 & Diverse Interdisciplinary Research Studies & 1702 & 6,495 & G \\
++T14246 & E-Learning and Knowledge Management & 1706 & 6,373 & E \\
++T13523 & Mathematics, Computing, and Information Processing & 1703 & 6,343 & G \\
++T12805 & Cognitive Science and Mapping & 1702 & 5,430 & O \\
++T13935 & Mathematical Control Systems and Analysis & 1702 & 4,409 & P \\
++T14409 & Educational Systems and Policies & 1710 & 3,650 & E \\
++T13565 & Education and Learning Interventions & 1710 & 2,867 & E \\
++T14071 & Healthcare during COVID-19 Pandemic & 1710 & 2,218 & O \\
++T14403 & Educational Challenges and Innovations & 1710 & 1,150 & E \\
++T13482 & Engineering Education and Technology & 1710 & 765 & E \\
++T14351 & Statistical and Computational Modeling & 1702 & 731 & G \\
++T13932 & Education, Technology, and Economics & 1710 & 577 & E \\
++T13873 & Educational Methods and Technology & 1710 & 329 & E \\
++T13676 & Educational and Technological Research & 1710 & 52 & E \\
++T13345 & Advanced Technologies and Applied Computing & 1705 & 17 & G \\
++T14387 & Environmental Engineering and Cultural Studies & 1710 & 2 & P \\
++T13514 & Experience-Based Knowledge Management & 1702 & 0 & Z \\
++T13567 & AI and Multimedia in Education & 1702 & 0 & E \\
++T13693 & Smart Systems and Machine Learning & 1705 & 0 & Z \\
++T13748 & Advanced Statistical Modeling Techniques & 1705 & 0 & Z \\
++T13836 & Educational Research and Pedagogy & 1705 & 0 & E \\
++T13964 & Educational Technology and Optimization & 1710 & 0 & E \\
++T14064 & Organizational and Employee Performance & 1702 & 0 & Z \\
++T14155 & Advanced Technology in Applications & 1710 & 0 & Z \\
++T14174 & Artificial Intelligence and Decision Support Systems & 1710 & 0 & Z \\
++T14175 & Intuitionistic Fuzzy Systems Applications & 1702 & 0 & Z \\
++T14413 & Advanced Technologies in Various Fields & 1702 & 0 & Z \\
++T14431 & Diverse Research and Applications & 1710 & 0 & Z \\
++\end{longtable}
++\end{footnotesize}
++\end{singlespace}
++
++\subsection{Known ambiguities and the decisions taken}
++\label{app:taxonomy-ambiguities}
++
++Forty categories drawn up for a mandate are not a partition of the literature, and several pairs or triples
++overlap by construction. The ontology's author resolved each overlap by a stated rule rather than by leaving
++it to the weights; the rules are recorded here verbatim in substance, because they determine how the
++multi-label statistics of Part~IX must be read.
++
++\paragraph{Large language models, foundation models, generative AI and NLP.}
++\texttt{foundation\_\allowbreak models} is the parent concept (any modality, adaptation-oriented); \texttt{llm} is its
++language child and the home of the ChatGPT/GPT/LLaMA-era literature; \texttt{generative\_\allowbreak ai} concerns the
++generative model and its outputs (diffusion, GANs, text-to-X synthesis, deepfakes and detection), not text
++generation by LLMs as such; \texttt{nlp} keeps task-framed language work. Multi-label assignment is expected
++and intended: an instruction-tuning paper scores in \texttt{llm} and \texttt{nlp}, a diffusion language model
++in \texttt{llm} and \texttt{generative\_\allowbreak ai}. The OpenAlex topic T10028, labelled ``Topic Modeling'', is in
++fact the cluster hosting BERT, pre-trained language models and retrieval-augmented generation and is weighted
++1.0 for \texttt{llm}; a reader who consults the OpenAlex label alone would misjudge this choice.
++
++\paragraph{Agents, reinforcement learning and robotics.}
++Software autonomy (LLM agents, multi-agent systems, planning) is \texttt{ai\_\allowbreak agents}; learning from reward
++is \texttt{reinforcement\_\allowbreak learning}; physical embodiment is \texttt{robotics}. The word ``agent'' never
++counts on its own: the patterns require a computing qualifier (``LLM agent'', ``autonomous agent'',
++``multi-agent system'') and an exclusion vetoes chemical, medical, commercial and other non-computing
++senses (reagent, contrast agent, real-estate agent).
++
++\paragraph{AI safety, cybersecurity and the statistical side of interpretability.}
++Adversarial examples, backdoors and prompt injection score in both \texttt{ai\_\allowbreak safety} and
++\texttt{cybersecurity} at 0.8, so that either a second signal or the text decides; fairness and governance
++belong to \texttt{ai\_\allowbreak safety}; explainability as a statistical method is \texttt{ml\_\allowbreak theory} (0.5), whereas
++interpretability of large models as a safety practice is \texttt{ai\_\allowbreak safety} (1.0).
++
++\paragraph{Cloud, distributed and edge computing.}
++\texttt{distributed\_\allowbreak computing} is the parent (algorithms, HPC systems software, consensus);
++\texttt{cloud\_\allowbreak computing} is the service model (virtualisation, serverless, data-centre resource management);
++\texttt{edge\_\allowbreak computing} is placement near users and devices (offloading, multi-access edge computing,
++on-device inference). Internet-of-things topics are shared between \texttt{edge\_\allowbreak computing} and
++\texttt{computer\_\allowbreak networks}.
++
++\paragraph{Cryptography, security and privacy.}
++Primitives and protocols are \texttt{applied\_\allowbreak cryptography}; attacks, defences and systems hardening are
++\texttt{cybersecurity}; using cryptography or statistics to protect data is
++\texttt{privacy\_\allowbreak preserving\_\allowbreak computation}, whose child \texttt{federated\_\allowbreak learning} has no OpenAlex topic
++and no venue of its own and is carried by keywords. The topic T10764 (``Privacy-Preserving Technologies in
++Data'') hosts both the parent (1.0) and the child (0.7).
++
++\paragraph{Architecture, accelerators and operating systems.}
++Accelerators are a child of architecture and include efficient-inference \emph{systems} (serving engines,
++attention kernels, quantisation for hardware); operating systems have no OpenAlex topic and are carried by
++the OSDI, SOSP, EuroSys, USENIX ATC and FAST streams, by \texttt{cs.OS} and by kernel and file-system keywords.
++
++\paragraph{Quantum computing and quantum physics.}
++Only computing and information-processing framings count. Physics topics on qubit platforms are weighted
++0.2--0.3, and an anchored exclusion removes texts about quantum dots, quantum chemistry, quantum field theory
++and similar physical senses whenever no computing term (qubit, quantum algorithm, error correction, QKD,
++NISQ, \ldots) appears anywhere in the title or abstract.
++
++\paragraph{Scientific computing, computational biology and autonomous discovery.}
++A numerical method is \texttt{scientific\_\allowbreak computing}; a biological object is \texttt{computational\_\allowbreak biology};
++AI that drives the scientific process itself (AI-scientist systems, self-driving laboratories, automated
++hypothesis generation, AI-discovered algorithms or theorems) is \texttt{autonomous\_\allowbreak scientific\_\allowbreak discovery},
++which is keyword-driven and vetoed when no AI or automation term is present.
++
++\paragraph{Algorithms and complexity.}
++STOC, FOCS and ICALP are shared; ECCC, \emph{ToCT} and \texttt{cs.CC} are complexity-only. Metaheuristics and
++evolutionary computation are counted as \texttt{algorithms} at 0.5--0.6 because the corpus is expected to
++contain many such works and no other leaf claims them.
++
++\paragraph{Education, media and other field-17 residues.}
++As documented in Section~\ref{app:taxonomy-unmapped}, the educational, media, physical-science and generic
++clusters that OpenAlex places in field~17 are left unmapped, with four exceptions mapped at low weight; this
++is a deliberate narrowing of ``computer science'' relative to the OpenAlex field, and its effect on corpus
++scope is examined in Part~III, where scope is defined by DBLP membership and OpenAlex primary field rather
++than by the ontology.
++
++\subsection{Limitations of the mapping and the validation design}
++\label{app:taxonomy-limits}
++
++The weights are expert priors set by reading keyword lists and venue scopes, not parameters fitted to
++labelled data, and the ontology's own documentation says so. Three limitations follow and are repeated here
++because they bound what the classification can show. Venue weights are per DBLP stream, so a workshop sharing
++a stream key with its main conference inherits the main conference's weight. OpenAlex keyword lists were read
++in full for the 302 field-17 topics and for about 120 CS-adjacent topics named in the area modules, but not for
++the remaining roughly 4{,}000 topics; relevant clusters from Medicine or Materials Science may be missing and
++are to be added when the corpus reveals them, through an amendment. And keyword patterns on abstracts fire on
++mentions as well as on contributions, a failure mode discussed with the others in
++Section~\ref{sec:m-classification}.
++
++The ontology was validated, after the classification run, on a stratified sample whose design is fixed in
++\texttt{taxonomy/VALIDATION.md}: fifty eligible works with an abstract per rule-assigned primary category
++(forty strata) plus two hundred works to which the rules assigned no category, drawn with seed~44; two
++independent LLM raters that never see the rule-based assignment and receive only the definitions of
++Table~\ref{tab:appA-leaves} with the title, abstract, venue and year of each work; the principal investigator
++as third rater on a 200-work subsample (checkpoint H-2). The pre-specified decision rule is that a category
++with consensus $F_1 < 0.60$ triggers a mapping revision, logged as a protocol amendment, and a re-run of the
++classification. The agreement statistics, the per-category precision and recall, the list of categories that
++met or failed the decision rule, and any amendment that resulted are reported in Part~IX; nothing about them
++is asserted in this appendix.

modified manuscript/appendices/appB_search_and_sources.tex +613 −1

++% ============================================================================
++% Appendix B — Search strategies, sources and acquisition record
++% Written by Agent L (Appendices B/C instance), 2026-11-03. Operational record
++% of what ingestion/*.py, systematic_review/define_corpus.py and the runbook
++% actually did; Part III gives the rationale and must not be repeated here.
++% Corpus numbers only through macros of tables/numbers.tex; sizes, hashes,
++% dates, file counts and parameter values are provenance identifiers and are
++% typed from data/raw/README.md, reports/PHASE0_FEASIBILITY.md and the ledgers.
++% ============================================================================
++
++\section{Search strategies, sources and acquisition record}
++\label{sec:appB}
++
++This appendix is the operational counterpart of Part~III\@. Part~III explains
++why four bibliographic sources were combined and what each contributes; the
++present appendix records, source by source, the endpoint or bucket that was
++read, the release or version that was read, the filters that were applied,
++the rate limits and budgets under which the reading took place, the licence
++of the material, the integrity checks performed on arrival, and what was
++sought but not obtained. It then gives the reporting items of the PRISMA-S
++extension for literature searches \citep{rethlefsen2021prismas} in the form
++they take for a scope-filtered mapping study, the exclusion codes of the
++eligibility stage with their predicates, the type and platform lists on which
++those predicates depend, and the record-linkage cascade with its thresholds.
++Every statement below can be checked against the scripts in
++\texttt{ingestion/} and \texttt{systematic\_review/}, the runbook
++(\texttt{reproducibility/RUNBOOK.md}) and the provenance note
++\texttt{data/raw/README.md}; where a count is a property of the corpus it is
++rendered by a macro of \texttt{tables/numbers.tex}, and where no macro exists
++the value is marked pending rather than typed.
++
++\subsection{Acquisition at a glance}
++\label{sec:appB-glance}
++
++Table~\ref{tab:appB-acq} lists the five acquisitions in the order of the
++runbook. All were performed from the compute node \texttt{M3U96b} between the
++evening of 2~November and 3~November~2026 (Eastern Time), with no credential
++of any kind: the OpenAlex key, the Semantic Scholar key and the Kaggle
++credentials were tested as absent before the first download
++(\texttt{reports/PHASE0\_FEASIBILITY.md}, §2). The absence of credentials is
++not incidental. It determined the choice of the snapshot over the API for
++OpenAlex (decision D1 of the plan), it excluded Semantic Scholar (D3), and it
++capped the post-snapshot increment at the public daily budget
++(Section~\ref{sec:appB-increment}).
++
++\begin{table}[htbp]
++\centering\footnotesize
++\caption{Acquisitions performed for the WP44 corpus, in runbook order.}
++\label{tab:appB-acq}
++\begin{tabularx}{\textwidth}{>{\raggedright\arraybackslash}p{1.8cm}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{2.2cm}>{\raggedright\arraybackslash}p{4.0cm}}
++\toprule
++Source & Object read & Obtained & Integrity record \\
++\midrule
++OpenAlex works & public snapshot, release 2026-10-14, \texttt{data/jsonl/works/}, 1{,}898 part files, 675{,}298{,}563{,}638 bytes gzip & 2026-11-02 21:19 to 23:05 ET, HTTPS, 8 parallel streams & manifest copy; content length and SHA-256 per file in \texttt{works/ledger.sqlite}; 1{,}898/1{,}898 done, 0 malformed lines; record total equal to the manifest total \\
++OpenAlex entities & snapshots of topics, subfields, fields, domains, sources, institutions, publishers, funders, concepts & 2026-11-02 & \texttt{entities\_ledger.json}: content length and SHA-256 per file \\
++OpenAlex API & works with any CS topic and publication date 2026-10-01 to 2026-11-02, by day and by primary field & 2026-11-03 onward, within the public budget & \texttt{api\_increment/}\allowbreak\texttt{ledger.json}: calls, cost, records and pages per slice; raw pages kept verbatim \\
++DBLP & \texttt{dblp.xml.gz} dated 2026-11-02 22:39 UTC, 1{,}116{,}270{,}104 bytes, with \texttt{dblp.dtd} and \texttt{dblp.xml.gz.md5} & 2026-11-02 & published MD5 \texttt{ff1c424c3b4cef36}\allowbreak\texttt{582f647c36df27c0} verified on the whole file before parsing; parse report with size, MD5 of the first GiB, counts by type and year \\
++arXiv & OAI-PMH, set \texttt{cs}, metadata prefix \texttt{arXiv}, datestamps 2018-01-01 to 2026-11-02 & 2026-11-02 to 03 & \texttt{harvest\_state.json} (pages, records, seconds per window); 758 raw XML pages kept; \texttt{arxiv\_parse\_}\allowbreak\texttt{report.json} \\
++ACL Anthology & \texttt{anthology+abstracts.bib.gz}, 46{,}723{,}382 bytes & 2026-11-02 & SHA-256 \texttt{44c3f7e907f221b740}\allowbreak\texttt{bea09069818a3d7c0d}\allowbreak\texttt{fd88b2da3b750d9f02}\allowbreak\texttt{955efc58b3}; \texttt{acl\_parse\_}\allowbreak\texttt{report.json} \\
++\bottomrule
++\end{tabularx}
++\end{table}
++
++\subsection{OpenAlex works snapshot}
++\label{sec:appB-openalex}
++
++\paragraph{Bucket, release and manifest.}
++The works entity of the OpenAlex snapshot \citep{priem2022openalex} is
++distributed as gzip-compressed JSON Lines files in the public S3 bucket
++\texttt{s3://openalex/} under the prefix \texttt{data/jsonl/works/}, partitioned by the date at which the
++records were last updated (\texttt{updated\_date=YYYY-MM-DD/}\allowbreak\texttt{part\_NNNN.gz}).
++The release read is the one whose \texttt{manifest.json} carries the date
++2026-10-14; the manifest lists 1{,}898 part files with, for each, its URL,
++its content length in bytes and its record count, together with the totals
++(\LinkSlimTotal{} works, 675{,}298{,}563{,}638 bytes). A copy of the manifest
++is kept as \texttt{data/raw/openalex/works\_manifest\_2026-10-14.json} and is
++the reference against which the ledger is checked. No AWS client was
++installed on the node; the bucket is public and each object was fetched over
++HTTPS from \url{https://openalex.s3.amazonaws.com/} with \texttt{curl}
++(\texttt{-L --retry 5 --retry-delay 3 --retry-all-errors -C -}, the last
++option resuming an interrupted transfer), up to five outer attempts per file
++with a linearly increasing pause, and the transfer was accepted only when the
++byte count on disk equalled the content length in the manifest. The measured
++single-stream throughput from the node was about 71~MB/s and the aggregate
++over eight streams about 110~MB/s, which fixed the worker count at eight and
++the wall time at one hour and forty-six minutes for the whole bucket
++(\texttt{ingestion/openalex\_snapshot.py --workers 8}).
++
++\paragraph{Filter applied on the stream (pass 1).}
++Each part file was decompressed in memory and parsed line by line
++(\texttt{orjson}); a line that failed to parse was counted as malformed and
++skipped, and the count of such lines (zero for every file) is in the ledger.
++For every parsed work $w$ the script evaluated the predicate
++\begin{equation}
++  \mathrm{CS}(w) \;=\; \Bigl[\exists\, \theta \in \mathrm{topics}(w) :
++  \mathrm{field}(\theta) = \text{\texttt{fields/17}}\Bigr]
++  \;\vee\; \Bigl[\mathrm{field}\bigl(\mathrm{primary\_topic}(w)\bigr) = \text{\texttt{fields/17}}\Bigr],
++  \label{eq:appB-csfilter}
++\end{equation}
++that is, whether any of the up to three topics attached to the record, or its
++primary topic, belongs to the OpenAlex field with identifier~17
++(\emph{Computer Science}; \citealp{openalex2024topics}), and kept the
++\emph{complete} JSON record when $\mathrm{CS}(w)$ held and
++\texttt{publication\_year} lay in $[2018, 2026]$. Records flagged
++\texttt{is\_xpac} (OpenAlex's expansion of coverage from non-Crossref
++sources) were kept at this stage with their flag, so that the exclusion could
++be applied later by the eligibility predicate rather than silently at
++ingestion. The retained records were re-compressed with Zstandard (level~6)
++into one file per part, \texttt{works/cs/}\allowbreak\texttt{updated\_date=X\_\_part\_NNNN.jsonl.zst},
++about 20~GB in total for
++\PrismaIdentifiedOpenalexCsTopicTwoZeroOneEightTwoZeroTwoSixPassOne{}
++records. No text search of any kind was performed: the scope of the OpenAlex
++layer is defined entirely by the topic classifier's field assignment, which is
++the search strategy appropriate to a field-level mapping and whose
++consequences are discussed in Part~III\@.
++
++\paragraph{Slim table of every work.}
++In the same pass the script wrote, for \emph{every} work in the snapshot
++regardless of field or year, one row of a slim Parquet table with the
++following columns: OpenAlex identifier (as an integer), DOI (lower-cased,
++prefix removed), publication year and date, type, \texttt{cited\_by\_count},
++primary topic, subfield, field and domain identifiers, the value of
++$\mathrm{CS}(w)$, the XPAC, open-access and retraction flags, language,
++primary source identifier, number of referenced works, number of authors,
++whether an abstract is present, and FWCI\@. The 1{,}898 per-part files were
++later concatenated into \texttt{compact/slim\_all\_works.parquet}
++(\LinkSlimTotal{} rows). The slim table is what allows a reference from a
++computer-science work to a work in another field to be typed and dated
++(Part~IV), and it is also what made the second pass of
++Section~\ref{sec:appB-pass2} possible without re-reading the bucket for a
++search.
++
++\paragraph{Integrity and provenance.}
++Before parsing, the SHA-256 digest of the gzip bytes of each part file was
++computed and written, with the file's URL, content length, manifest record
++count, parsed record count, retained record count, malformed-line count,
++download and total times, attempt count and timestamps, to a SQLite ledger
++(\texttt{works/ledger.sqlite}, table \texttt{files}); a second table
++(\texttt{runs}) records the manifest date and totals, the worker count, the
++host name and the command line of every invocation. The ledger makes the job
++resumable (a file with status \texttt{done} is never re-read) and is the
++provenance record of the raw layer. The raw \texttt{.gz} files were deleted
++after a successful parse (decision D6 of the plan): the snapshot is a public,
++versioned release, the manifest and the per-file digests permit an
++independent party to re-download and verify any part, and keeping 675~GB of
++compressed JSON on the node would have precluded the rest of the pipeline.
++Three checks were made at the end of the pass and are recorded in
++\texttt{data/raw/README.md}: all 1{,}898 files have status \texttt{done}; the
++sum of parsed records over files equals the manifest total; and the sum of
++malformed lines is zero. The OpenAlex snapshot is released under CC0~1.0
++(\texttt{LICENSE.txt} in the bucket, copied next to the manifest), which
++permits redistribution of derived record-level data.
++
++\paragraph{Entity snapshots.}
++The smaller entity snapshots were downloaded from the same bucket by
++\texttt{ingestion/}\allowbreak\texttt{openalex\_entities.py}, using each
++entity's own manifest, with the same length check and SHA-256 ledger
++(\texttt{entities\_ledger.json}). The manifests list 4{,}516 topics,
++252 subfields, 26 fields, 4 domains, 268{,}629 sources, 143{,}455 institutions,
++10{,}713 publishers, 41{,}799 funders and 65{,}026 concepts; the topics
++snapshot was flattened to \texttt{topics\_flat.csv} (identifier, name,
++subfield, field, domain, works count, keywords, description) and is the
++vocabulary over which the ontology of Appendix~A maps OpenAlex topics to
++categories. The authors entity was not downloaded, because the authorship
++blocks embedded in each work record already carry the author identifier,
++display name, ORCID, position and institutions; the keywords entity
++(1{,}881{,}723 entries in its manifest) was not downloaded either, the
++per-work keyword lists being embedded as well.
++
++\subsection{Second snapshot pass by identifier set}
++\label{sec:appB-pass2}
++
++The first pass retains what OpenAlex calls computer science. The record
++linkage of Section~\ref{sec:appB-linkage} showed that
++\LinkDblpUnlinkedWithDoiInSlim{} DBLP records of 2018--2026 have a DOI that
++exists in the slim table of all works but not in the retained subset, that is,
++works indexed by DBLP that OpenAlex labels with no computer-science topic.
++Rather than accept OpenAlex's boundary, the pipeline reads the bucket a
++second time in \emph{identifier-set mode}. The linkage script writes
++the list \texttt{data/metadata/needed\_oa\_ids.csv} of OpenAlex identifiers to
++recover (\LinkNeededOaIdsSecondPassDblpDoi{} reached through a DBLP DOI and
++\LinkNeededOaIdsSecondPassArxivDoi{} through an arXiv DOI), and
++\texttt{openalex\_snapshot.py --id-list} streams every part file again,
++keeping exactly the records whose identifier is in the set and writing no slim
++rows. The identifier set is loaded in each worker process (macOS uses the
++\texttt{spawn} start method, so module globals are rebuilt per worker), the
++same download, length and SHA-256 discipline applies, and the output lands in
++\texttt{works\_pass2/cs/} with its own ledger. The recovered records are
++normalised with the suffix \texttt{\_p2} and the linkage is rerun on the
++union of both passes (\texttt{ingestion/run\_pass2\_chain.sh}). The second
++pass costs the same bandwidth as the first---about one hour and fifty minutes
++of streaming---because an identifier lookup cannot be pushed to the bucket;
++the price was accepted because the alternative, fetching the identifiers one
++by one through the API, would have exceeded the public budget by two orders of
++magnitude. The PRISMA counts of Part~III were generated from the first pass
++(\PrismaIdentifiedOpenalexTotalRecordsLoaded{} records loaded); the second
++pass adds its records to $\mathcal{W}_0$ when it completes and the counts are
++regenerated by the same scripts, so the identification total in the final
++PRISMA figure is \pending{PrismaIdentifiedOpenalexTotalRecordsLoaded after the second pass}
++rather than the first-pass value quoted here.
++
++\subsection{OpenAlex API increment}
++\label{sec:appB-increment}
++
++The snapshot ends on 14~October 2026 and the observation window on
++2~November 2026. The gap was covered by the REST API
++(\texttt{https://api.openalex.org/works}) without a key. Two constraints of the
++unkeyed API shaped the harvest and are recorded here because they bound what
++the increment can contain. First, cursor paging requires a key, so the
++harvest used basic page numbering, which the API caps at 10{,}000 results per
++filter expression; the harvest is therefore \emph{sliced}, one request series
++per publication day $d$ and per primary-field class, with the filter
++\begin{quote}\ttfamily
++topics.field.id:17, from\_publication\_date:$d$, to\_publication\_date:$d$,
++primary\_topic.field.id:17 \quad (slice ``cs\_primary'')\\
++topics.field.id:17, from\_publication\_date:$d$, to\_publication\_date:$d$,
++primary\_topic.field.id:!17 \quad (slice ``cs\_secondary'')
++\end{quote}
++for every day from 2026-10-01 to 2026-11-02 (the start was placed two weeks
++before the snapshot date to absorb records that OpenAlex back-dates), 200
++results per page, a \texttt{select} list of the fields used by the
++normaliser (identifiers, dates, type, language, locations, open access,
++authorships, topics, keywords, referenced works, citation counts, abstract
++inverted index, flags, FWCI, funders, indexed-in, full-text flag, APC), the
++polite-pool \texttt{mailto} parameter and a descriptive user agent. A slice
++whose reported \texttt{count} exceeds 10{,}000 is marked \texttt{truncated} in
++the ledger; the two-way split by primary field was chosen because a single
++day of computer-science output in 2026 can exceed the cap while each half
++rarely does. Second, the unkeyed API is metered: list-and-filter calls are
++charged at about \$0.0001 each against a shared budget of \$0.10 per day,
++and the script stops at \$0.085 per day (a margin for the free single-record
++lookups used for reference verification), sleeps 0.12~s between calls, and
++halts on HTTP~429, 403 or any 5xx response so that the next day's run resumes
++from the ledger. Every page is stored verbatim as
++\texttt{api\_increment/pages/<day>\_<slice>/page\_NNNN.json}; the ledger
++(\texttt{api\_increment/ledger.json}) records per slice the pages, records,
++reported count, completion and truncation flags, and globally the number of
++calls and the cost. At the time of writing the harvest had used the budget
++of several consecutive days and was not complete; its totals
++(\pending{IncrementApiCalls}, \pending{IncrementRecordsRetrieved},
++\pending{IncrementTruncatedSlices}) will be reported from the ledger when the
++last slice closes. The conversion step (\texttt{ingest\_api\_increment.py})
++rewrites the pages into the same \texttt{.jsonl.zst} part format as the
++snapshot filter (\texttt{works\_pass3/}), skipping any work whose identifier
++is already in the pass-1 or pass-2 tables and writing an
++\texttt{increment\_report.json} (pages, new works, works already present); the
++number of works added by the increment is \pending{IncrementNewWorks}. Two
++limitations follow from the design and are not recoverable from the ledger: a
++truncated slice loses the records beyond the 10{,}000th in the API's default
++ordering, and a record created after 2~November 2026 with a publication date
++inside the window is absent whatever its date. Both affect only the last
++weeks of a window that is already treated as right-censored (Part~III), so no
++analysis depends on them.
++
++\subsection{DBLP}
++\label{sec:appB-dblp}
++
++\paragraph{Dump and version.}
++DBLP \citep{ley2002dblp, ley2009dblp} publishes its whole database as a
++single XML file. The file read is \texttt{dblp.xml.gz} as served on
++2~November 2026 (file timestamp 22:39 UTC), 1{,}116{,}270{,}104 bytes, together
++with the document type definition \texttt{dblp.dtd} and the published
++checksum file \texttt{dblp.xml.gz.md5}. The MD5 of the whole file,
++\texttt{ff1c424c3b4cef36582f647c36df27c0}, was computed and compared with the
++published value before parsing; the parse report additionally stores the file
++size and an MD5 of the first gibibyte as a quick identity check for later
++sessions. DBLP is released under the Open Data Commons Attribution licence
++(ODC-BY~1.0), which permits the redistribution of derived data with
++attribution. The download from \texttt{dblp.org} proceeded at about 12~MB/s,
++the slowest link of the whole acquisition.
++
++\paragraph{Parsing rule.}
++The DTD matters: DBLP titles use XML entities for accented characters, and a
++parser that does not resolve them against \texttt{dblp.dtd} corrupts titles
++and therefore the title-based linkage. The parser
++(\texttt{ingestion/dblp\_parse.py}) is \texttt{lxml.etree.iterparse} over the
++gzip stream with \texttt{load\_dtd=True}, \texttt{resolve\_entities=True} and
++\texttt{huge\_tree=True}, run from the directory containing the DTD because
++the dump references it relatively. It emits one row per record of the types
++\textit{article}, \textit{inproceedings}, \textit{proceedings}, \textit{book},
++\textit{incollection}, \textit{phdthesis}, \textit{mastersthesis} and
++\textit{data}; person records (\texttt{www}) are consumed and discarded. For
++each record it keeps the key, type, modification date, publication-type
++attribute, title (concatenated text of the title element, so that embedded
++markup such as subscripts is flattened), year, author names and DBLP person
++identifiers, editors, journal, booktitle, volume, number, pages, publisher,
++series, the DOI (taken from the first electronic-edition URL beginning with
++\texttt{https://doi.org/}, lower-cased), the full list of electronic-edition
++URLs, the URL, the cross-reference to the proceedings record, and the
++\emph{stream}, the part of the key before its last slash
++(\texttt{journals/tocs}, \texttt{conf/nips}), which is DBLP's venue identifier
++and the unit on which the ontology's venue signals are defined. All years are
++parsed; counts by type and by year for the entire archive are written to
++\texttt{dblp\_parse\_report.json}, and the records of 2018--2026
++(\LinkDblpTwoZeroOneEightTwoZeroTwoSix{} after the type restriction of the
++linkage step) are the ones that enter the corpus. The total number of
++publication records in the dump is \pending{DblpParsedRecords}. A second
++script (\texttt{dblp\_streams.py}) lists the streams with at least one hundred
++records of type article or inproceedings in 2018--2026, with their first and
++last years and a sample journal or booktitle name; this list
++(\texttt{taxonomy/dblp\_streams\_2018\_2026.csv}) is the input of the
++venue-to-category mapping and is not itself a corpus filter.
++
++\subsection{arXiv}
++\label{sec:appB-arxiv}
++
++\paragraph{Protocol and windows.}
++arXiv \citep{ginsparg2011arxiv} metadata were harvested through its OAI-PMH
++endpoint \texttt{https://oaipmh.arxiv.org/oai} with the verb
++\texttt{ListRecords}, the set \texttt{cs} and the metadata prefix
++\texttt{arXiv}, which returns arXiv's own schema (identifier, dates, title,
++abstract, authors with forenames and keynames, categories, comments, journal
++reference, DOI, licence, report number, MSC and ACM classes). The harvest was
++sliced into calendar-month windows on the OAI \texttt{from}/\texttt{until}
++parameters from 2018-01-01 to 2026-11-02; each window was walked with the
++resumption token until exhausted, and every response page was written
++verbatim to \texttt{raw/<window>/page\_NNNN.xml} before being parsed, so that
++the Parquet table can be rebuilt from the raw pages
++(\texttt{--parse-only}) without touching the network. The harvester waited
++3.5~s between requests, honoured the \texttt{Retry-After} header on HTTP~503
++(arXiv's flow-control response), waited thirty seconds on other errors, and
++gave up a window after twelve consecutive failures (none occurred). The
++state file \texttt{harvest\_state.json} records, per window, the number of
++pages, of records and of seconds, and the harvest of the 107 windows produced
++758 pages in about one hour.
++
++\paragraph{Datestamp semantics and deduplication.}
++A point about OAI-PMH that affects the counts must be stated. The
++\texttt{from}/\texttt{until} parameters select on the record's
++\emph{datestamp}, which is the date of its last modification, not on its
++submission date. A monthly window therefore contains every record touched in
++that month, including old e-prints whose metadata were edited, and a record
++edited in two months appears in two windows. The parser keeps, for each arXiv
++identifier, the occurrence with the latest datestamp, skips records whose
++header carries \texttt{status="deleted"}, normalises whitespace in titles and
++abstracts, lower-cases the DOI, takes the first listed category as the
++primary category, and records the set specifications. The result is
++\LinkArxivCsRecords{} distinct records, of which
++\LinkArxivCreatedTwoZeroOneEightTwoZeroTwoSix{} have a \texttt{created} date
++(first submission) in 2018--2026; the difference consists of older e-prints
++modified during the window, which are kept in the table for linkage but do
++not belong to the observation window. The \texttt{created} date is the one
++used as $d_{\mathrm{arx}}(w)$ in Part~IV\@. arXiv's metadata are offered for
++reuse under its metadata terms; the e-prints themselves were not downloaded
++in the acquisition described here.
++
++\subsection{ACL Anthology}
++\label{sec:appB-acl}
++
++The ACL Anthology \citep{radev2013aclanthology, acl2024anthology}
++distributes its complete bibliography as a BibTeX export. The variant with
++abstracts (\texttt{anthology+}\allowbreak\texttt{abstracts.bib.gz}) was downloaded on
++2~November 2026 (46{,}723{,}382 bytes; SHA-256 in
++Table~\ref{tab:appB-acq}). BibTeX is not a regular language, so the parser
++(\texttt{ingestion/acl\_parse.py}) walks the text by brace depth: it locates
++each \texttt{@type\{} header, finds the matching closing brace, splits the
++body at its first comma into the citation key (which is the Anthology
++identifier, for instance \texttt{P19-1001} or \texttt{2023.acl-long.1}) and
++the field list, and reads each field as a braced, quoted or bare value with
++nested braces allowed. Fifteen fields are kept (title, author, editor,
++booktitle, journal, month, year, address, publisher, URL, DOI, pages,
++abstract, volume, number); braces inside values are removed and whitespace
++normalised. Entries of type \texttt{comment}, \texttt{string} and
++\texttt{preamble} are skipped. The export is released under CC~BY~4.0. The
++number of entries parsed is \pending{AclParsedEntries}; the linkage step uses
++only the \textit{inproceedings} and \textit{article} entries
++(\LinkAclRecords{}), of which \LinkAclTwoZeroOneEightTwoZeroTwoSix{} fall in
++2018--2026. The Anthology's role is a coverage check for natural-language
++processing venues rather than a corpus source in its own right; an ACL entry
++with no OpenAlex counterpart is counted in the linkage report but is never a
++work of the corpus.
++
++\subsection{What was sought and not obtained}
++\label{sec:appB-notobtained}
++
++The mandate lists sources beyond the four used, and the integrity rules
++require that inaccessible data be recorded rather than worked around. The
++record is as follows.
++
++\begin{description}
++  \item[Semantic Scholar Academic Graph.] An anonymous request to the
++  endpoint \texttt{graph/v1/paper/} for \texttt{arXiv:2303.08774} returned HTTP~429 with the
++  instruction to apply for a key, and the datasets endpoint requires a key
++  outright \citep{kinney2023s2ag}. No key was requested because the project
++  rule is not to assume credentials and because applying would have made the
++  acquisition depend on an approval outside the project's control. The
++  consequence is the absence of the influential-citation and citation-intent
++  signals from every analysis, stated wherever it matters.
++  \item[OpenAlex API key.] A free key raises the daily budget from \$0.10 to
++  \$1 and enables cursor paging. Obtaining it requires the principal
++  investigator to confirm an e-mail; this was not done without instruction,
++  and the increment of Section~\ref{sec:appB-increment} was run under the
++  unkeyed budget. The snapshot route does not depend on the key.
++  \item[Full texts.] No full text was downloaded during corpus construction.
++  Abstracts come from the OpenAlex inverted index (and from arXiv and the
++  Anthology for their records). The plan reserves lawful full-text access
++  (arXiv e-prints at one request per three seconds, open-access locations)
++  for at most 10{,}000 high-priority works in a later phase; the
++  knowledge-graph extraction of Part~IV is abstract-based and labelled as
++  such.
++  \item[Crossref.] The REST API was reached (\texttt{works?filter=}\allowbreak\texttt{from-pub-date:2022-11-30\&rows=0}
++  returned a count for all fields) and is used for the verification of every
++  DOI cited in the manuscript and for sample cross-checks of types and
++  citation counts; it was not used as a corpus source, OpenAlex already
++  carrying its metadata.
++  \item[Other sources listed in the mandate.] Kaggle arXiv snapshot,
++  Microsoft Academic Graph, Papers with Code, institutional repositories: no Kaggle credentials were configured and
++  OAI-PMH made the snapshot unnecessary; the Microsoft Academic Graph was
++  discontinued in 2021; benchmark and code repositories are scoped to the
++  reproducibility analyses of a later phase and are not bibliographic
++  sources; institutional repositories enter only through OpenAlex's
++  indexing of them, which the generic-deposit rule of
++  Section~\ref{sec:appB-exclusions} then restricts.
++\end{description}
++
++\subsection{PRISMA-S search record}
++\label{sec:appB-prismas}
++
++PRISMA-S \citep{rethlefsen2021prismas} is written for keyword searches of
++bibliographic databases. A scope-filtered mapping has no search string, but
++its sixteen items can still be answered, and answering them makes explicit
++which items have no counterpart. Table~\ref{tab:appB-prismas} gives the
++record; the PRISMA~2020 flow itself \citep{page2021prisma, haddaway2022prismaflow}
++is Figure~\ref{fig:prisma} and Table~\ref{tab:prisma} of Part~III\@.
++
++\begin{table}[htbp]
++\centering\footnotesize
++\caption{PRISMA-S items answered for the WP44 corpus.}
++\label{tab:appB-prismas}
++\begin{tabularx}{\textwidth}{>{\raggedright\arraybackslash}p{0.5cm}>{\raggedright\arraybackslash}p{3.2cm}>{\raggedright\arraybackslash}X}
++\toprule
++No. & Item & This study \\
++\midrule
++1 & Database names & OpenAlex (works and entity snapshots), DBLP, arXiv, ACL Anthology; Crossref for verification only. \\
++2 & Multi-database searching & Each source read in full under its own scope filter; no interface or vendor platform (direct bucket, dump and OAI access). \\
++3 & Study registries & Not applicable. \\
++4 & Online resources and browsing & None. \\
++5 & Citation searching & None for identification; OpenAlex reference lists are analysed but do not add records to the corpus. \\
++6 & Contacts & None. \\
++7 & Other methods & None. \\
++8 & Full search strategies & OpenAlex: predicate~\eqref{eq:appB-csfilter} with publication year 2018--2026, then identifier-set recovery of DBLP- and arXiv-linked works (Section~\ref{sec:appB-pass2}), then the API filter of Section~\ref{sec:appB-increment}. DBLP: whole dump, record types listed in Section~\ref{sec:appB-dblp}, years 2018--2026 at linkage. arXiv: OAI set \texttt{cs}, datestamps 2018-01-01 to 2026-11-02, \texttt{created} in 2018--2026 at linkage. ACL: whole export, types inproceedings and article, years 2018--2026 at linkage. \\
++9 & Limits and restrictions & Year window only; no language, type or country restriction at search time (types are an eligibility criterion, Section~\ref{sec:appB-exclusions}). \\
++10 & Search filters & None beyond the scope predicates; the OpenAlex topic classifier is the instrument that defines ``computer science'' on that layer and its properties are discussed in Part~III\@. \\
++11 & Prior work & No search strategy was reused from a prior review. \\
++12 & Updates & One update by API increment to 2026-11-02; no further update is planned for this version. \\
++13 & Dates of searches & OpenAlex snapshot released 2026-10-14, read 2026-11-02; DBLP dump 2026-11-02; arXiv harvest 2026-11-02 to 03; ACL export 2026-11-02; API increment from 2026-11-03. \\
++14 & Peer review of the search & No librarian review; the strategy is preregistered in \texttt{protocol/PROTOCOL.md} (P3) and audited by the independent agents of Phase~8. \\
++15 & Total records & OpenAlex \PrismaIdentifiedOpenalexCsTopicTwoZeroOneEightTwoZeroTwoSixPassOne{} (pass 1); DBLP \PrismaIdentifiedDblpTwoZeroOneEightTwoZeroTwoSix{}; arXiv \PrismaIdentifiedArxivCsTwoZeroOneEightTwoZeroTwoSix{}; ACL \PrismaIdentifiedAclTwoZeroOneEightTwoZeroTwoSix{} (all 2018--2026). \\
++16 & Deduplication & Deterministic record linkage across sources (Section~\ref{sec:appB-linkage}) and preprint--publication merging within OpenAlex (Part~III, Part~IV); OpenAlex works are the counting unit; \PrismaExcludedEFiveDuplicate{} preprint members of merged pairs are flagged E5. \\
++\bottomrule
++\end{tabularx}
++\end{table}
++
++\subsection{Exclusion codes and their predicates}
++\label{sec:appB-exclusions}
++
++The eligibility stage (\texttt{systematic\_review/define\_corpus.py}) computes
++seven exclusion flags and two membership flags as columns of the table
++\texttt{corpus\_works}, each a SQL expression over the normalised OpenAlex
++record and the linkage tables. Table~\ref{tab:appB-codes} paraphrases the
++predicates. Three lists on which they depend are reproduced verbatim from the
++script, because their content is the substance of the type rule and of
++amendment A-002.
++
++\begin{itemize}
++  \item $T_{\mathrm{keep}}$ (types kept without further condition):
++  \texttt{article}, \texttt{conference-}\allowbreak\texttt{paper}, \texttt{preprint},
++  \texttt{book-}\allowbreak\texttt{chapter}, \texttt{book}, \texttt{dissertation},
++  \texttt{report}, \texttt{review}, \texttt{software-}\allowbreak\texttt{paper}.
++  \item $T_{\mathrm{hard}}$ (types excluded even for DBLP-linked records):
++  \texttt{editorial}, \texttt{erratum}, \texttt{retraction},
++  \texttt{peer-}\allowbreak\texttt{review}, \texttt{paratext}, \texttt{conference-}\allowbreak\texttt{abstract},
++  \texttt{reference-}\allowbreak\texttt{entry}, \texttt{dataset}, \texttt{software},
++  \texttt{supplementary-}\allowbreak\texttt{materials}, \texttt{libguides}, \texttt{letter},
++  \texttt{grant}, \texttt{standard}.
++  \item $\mathcal{G}$ (generic deposit platforms, matched on the primary
++  source's display name): ``Zenodo (CERN European Organization for Nuclear
++  Research)'' and ``Zenodo''; ``Figshare''; ``Institutional Repositories
++  DataBase (IRDB)''; ``LA Referencia (Red Federada de Repositorios
++  Institucionales de Publicaciones Cient\'{\i}ficas)''; ``Dialnet (Universidad de
++  la Rioja)''; ``DOAJ (DOAJ: Directory of Open Access Journals)''; ``OSTI OAI
++  (U.S. Department of Energy Office of Scientific and Technical
++  Information)''.
++\end{itemize}
++
++\begin{table}[htbp]
++\centering\footnotesize
++\caption{Exclusion and membership codes of \texttt{define\_corpus.py}, with their predicates paraphrased from SQL. $w$ is an OpenAlex record; \textit{in\_dblp}, \textit{in\_acl} and \textit{arxiv\_linked} are read from the link tables; $t_0$ = 2018-01-01, $t_1$ = 2026-11-02.}
++\label{tab:appB-codes}
++\begin{tabularx}{\textwidth}{>{\raggedright\arraybackslash}p{0.9cm}>{\raggedright\arraybackslash}p{3.4cm}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{1.9cm}}
++\toprule
++Code & Column & Predicate (true = flagged) & Marginal count \\
++\midrule
++E1 & \texttt{e1\_outside\_window} & \texttt{publication\_date} is null, or $< t_0$, or $> t_1$ & \PrismaExcludedEOneOutsideWindow{} \\
++E2 & \texttt{e2\_type} & $\tau(w) \notin T_{\mathrm{keep}}$ and not (\textit{in\_dblp} and $\tau(w) \notin T_{\mathrm{hard}}$); a null type counts as outside $T_{\mathrm{keep}}$ & \PrismaExcludedETwoType{} \\
++E3 & \texttt{e3\_xpac} & \texttt{is\_xpac} and not \textit{in\_dblp} & \PrismaExcludedEThreeXpac{} \\
++--- & \texttt{xpac\_kept\_dblp} & \texttt{is\_xpac} and \textit{in\_dblp} (kept, flagged) & \PrismaExcludedXpacKeptDblp{} \\
++E4 & \texttt{e4\_no\_title} & title null or empty after trimming & \PrismaExcludedEFourNoTitle{} \\
++E5 & \texttt{e5\_duplicate} & $w$ is the preprint member of a surviving preprint--published pair (Part~IV, eq.~\eqref{eq:duppair}) & \PrismaExcludedEFiveDuplicate{} \\
++E6 & \texttt{e6\_retracted} & \texttt{is\_retracted} (flag only; not part of eligibility) & \PrismaExcludedESixRetracted{} \\
++E7 & \texttt{e7\_generic\_deposit} & primary source name $\in \mathcal{G}$ and not \textit{in\_dblp} and no linked arXiv record & \pending{PrismaExcludedESevenGenericDeposit} \\
++core & \texttt{cs\_core} & \textit{in\_dblp} or \texttt{cs\_primary} or (\texttt{cs\_any} and linked arXiv primary category begins with \texttt{cs.}) & \PrismaCsCoreEligible{} eligible \\
++pen. & \texttt{cs\_penumbra} & \texttt{cs\_any} and not \texttt{cs\_core} & \PrismaCsPenumbraEligibleOtherwise{} otherwise eligible \\
++\midrule
++--- & \texttt{eligible} & $\neg$E1 $\wedge$ $\neg$E2 $\wedge$ $\neg$E3 $\wedge$ $\neg$E4 $\wedge$ $\neg$E5 $\wedge$ $\neg$E7 $\wedge$ core & \PrismaCsCoreEligible{} \\
++\bottomrule
++\end{tabularx}
++\end{table}
++
++Three further columns are derived in the same statement and are used
++throughout the manuscript. The first date, \texttt{first\_date}, is the least
++of the publication date, the merged preprint's date and the arXiv creation
++date, except that a publication date falling on 1~January is treated as
++year-only precision and is overridden by either of the other two when
++available; \texttt{date\_precision} is \texttt{year} when the date falls on
++1~January and no finer date exists, \texttt{month} otherwise; and
++\texttt{period} is \texttt{pre}, \texttt{baseline}, \texttt{post} or
++\texttt{future} according to whether the first date precedes $t_0$, lies in
++$[t_0, \text{2022-11-30})$, in $[\text{2022-11-30}, t_1]$, or after $t_1$.
++The PRISMA counts are computed twice, marginally (the column of
++Table~\ref{tab:appB-codes}) and sequentially in the order E1, E2, E3, E4, E5,
++E7, and the script asserts that the loaded total equals the sum of the
++sequential exclusions and the after-exclusion count
++(\PrismaAfterExclusions{}); the macro set of Part~III reports both. The E6 and
++penumbra rows are not exclusions and do not enter the sum. The numbering
++conflict between the protocol's E7 (penumbra) and the code's E7
++(generic deposit) is resolved by amendment A-005 in favour of the code.
++
++\subsection{Record-linkage cascade and thresholds}
++\label{sec:appB-linkage}
++
++Part~IV states the cascade in prose; Table~\ref{tab:appB-cascade} gives it as
++executed, step by step, with every threshold. The normalised title
++$\eta(\cdot)$ lower-cases and deletes every character outside $[a$--$z0$--$9]$;
++``unique'' means that the normalised title occurs exactly once on that side
++among the records still unlinked at that step (for OpenAlex in the arXiv and
++ACL cascades, unique among all OpenAlex records, the exclusion of already
++linked identifiers being applied to the pair instead); the length condition
++is on the raw title for DBLP and OpenAlex and on the normalised title for
++arXiv and ACL\@. The DBLP title step adds a first-author check that the other
++cascades do not have: the last whitespace-separated token of the first
++author's name on each side must coincide when both are present. Each link
++carries its method label, which is the stratum of the precision sample of
++Part~IV; the counts per method are in \texttt{data/metadata/linkage\_report.json}
++and Table~\ref{tab:linkage}.
++
++\begin{table}[htbp]
++\centering\footnotesize
++\caption{The deterministic linkage cascade of \texttt{ingestion/link\_sources.py}. Steps within a cascade apply only to records unlinked by the previous steps.}
++\label{tab:appB-cascade}
++\begin{tabularx}{\textwidth}{>{\raggedright\arraybackslash}p{1.6cm}>{\raggedright\arraybackslash}p{1.4cm}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{1.8cm}}
++\toprule
++Cascade & Step & Join condition and thresholds & Method label \\
++\midrule
++\multirow{3}{*}{DBLP--OA}
++ & 1 & equal lower-cased DOI; all matching pairs kept (many-to-many allowed) & \texttt{doi} \\
++ & 2 & equal arXiv identifier (DBLP: parsed from \texttt{ee} URLs by \texttt{arxiv.org/abs/([\^{} ]+)}; both sides stripped of \texttt{v[0-9]+\$}); both records unlinked & \texttt{arxiv} \\
++ & 3 & equal $\eta$, raw title $\ge 25$ characters, $\eta$ unique on both sides, $|y_{\mathrm{DBLP}} - y_{\mathrm{OA}}| \le 1$, first-author surnames equal when both present & \texttt{title\_year} \\
++\midrule
++\multirow{3}{*}{arXiv--OA}
++ & 1 & equal arXiv identifier (OpenAlex side: \texttt{ids.arxiv}, else DOI prefix \texttt{10.48550/arxiv.}, else \texttt{arxiv.org/abs|pdf/} in any location URL); all pairs kept & \texttt{arxiv\_id} \\
++ & 2 & equal lower-cased DOI; arXiv record unlinked & \texttt{doi} \\
++ & 3 & equal $\eta$, $|\eta| \ge 25$, unique on both sides, $|y_{\mathrm{created}} - y_{\mathrm{OA}}| \le 2$, OpenAlex record unlinked & \texttt{title\_year} \\
++\midrule
++\multirow{3}{*}{ACL--OA}
++ & 1 & equal lower-cased DOI & \texttt{doi} \\
++ & 2 & Anthology identifier equal to the segment after \texttt{aclanthology.org/} in any OpenAlex landing-page URL (with or without trailing slash); ACL record unlinked & \texttt{url} \\
++ & 3 & equal $\eta$, $|\eta| \ge 25$, unique on both sides, $|y_{\mathrm{ACL}} - y_{\mathrm{OA}}| \le 1$, OpenAlex record unlinked & \texttt{title\_year} \\
++\bottomrule
++\end{tabularx}
++\end{table}
++
++The record types admitted on the DBLP side are \textit{article},
++\textit{inproceedings}, \textit{incollection}, \textit{book},
++\textit{phdthesis} and \textit{data} with year 2018--2026; on the ACL side
++\textit{inproceedings} and \textit{article}; on the arXiv side every record
++of the harvest, the year condition being applied through the \texttt{created}
++year in the title step only. The linkage runs inside one DuckDB database
++with 24 threads and a 60~GB memory limit and completes in about ten seconds
++on the compacted tables, which is why it was possible to rerun it after each
++amendment rather than patch its output. The precision sample of 500 pairs
++(250 DOI, 200 title--year, 50 arXiv, the last stratum empty for the reason
++given in Part~IV) is drawn by \texttt{systematic\_review/linkage\_sample.py}
++with seed 44, and its automatic agreement checks (title similarity at least
++0.9, year difference at most one, first-author surname agreement) are the
++interim figure until the principal investigator's review (checkpoint H-1).
++
++\subsection{Reproducing the acquisition}
++\label{sec:appB-repro}
++
++The acquisition can be repeated from the runbook in six commands, in the
++order of Table~\ref{tab:appB-acq}, on a machine with Python~3.12 and the
++locked environment of \texttt{reproducibility/environment\_M3U96b.txt}:
++\begin{enumerate}
++  \item \texttt{openalex\_snapshot.py} with the manifest copy (then once more with \texttt{--id-list});
++  \item \texttt{openalex\_entities.py};
++  \item \texttt{dblp\_parse.py};
++  \item \texttt{arxiv\_oai.py};
++  \item \texttt{acl\_parse.py};
++  \item \texttt{openalex\_api\_increment.py}, followed by \texttt{ingest\_api\_increment.py}.
++\end{enumerate}
++Two of these are not exactly repeatable and the reader should know which.
++The OpenAlex bucket retains only recent releases, so a later reader must
++either locate the 2026-10-14 release or accept that a newer release will
++differ in record counts and in the nightly-recomputed fields (citation
++counts, FWCI); the per-file digests in the ledger allow a reader with the
++same release to prove identity, and a reader with a different release to
++quantify the difference by record count. The arXiv OAI feed is likewise
++live: a re-harvest of the same datestamp windows returns records as they now
++are, and the raw pages archived with the project are the only way to recover
++the state of 2--3~November 2026. The DBLP dump and the ACL export are dated
++files whose hashes are published here, and a copy of each is kept with the
++project's raw layer.

modified manuscript/appendices/appC_derivations.tex +419 −1

++% ============================================================================
++% Appendix C — Mathematical derivations
++% Written by Agent L (Appendices B/C instance), 2026-11-03. Derivations and
++% proofs referenced by Part IV (sections/part04_methodology.tex); notation
++% follows Part IV. Each derivation is checked against the implementation in
++% econometrics/structural_change.py, bibliometrics/03_citation_impact.py,
++% bibliometrics/04_diversity_cagr.py, knowledge_graph/02_category_flows.py,
++% taxonomy/classify.py, taxonomy/validate_sample.py and
++% experiments/emerging_detection.py; where the code and the usual textbook
++% convention differ, the code's convention is stated and the difference noted.
++% ============================================================================
++
++\section{Mathematical derivations}
++\label{sec:appC}
++
++Part~IV defines the statistics of this study and states their properties;
++this appendix supplies the derivations behind those statements, in the order
++in which Part~IV uses them. Each subsection fixes the object as implemented,
++derives the property that the main text relies on, and records any point at
++which the implementation departs from the convention of the cited source.
++The derivations are elementary and self-contained; their purpose is not
++novelty but the assurance that the quantities tabulated in Part~IX are what
++the text says they are. Throughout, $n$ is the length of a monthly series,
++$y_t = \log(N_t + 1)$ its logarithm, and $\mathbf{1}[\cdot]$ the indicator
++function.
++
++\subsection{The interrupted time series as a linear regression}
++\label{sec:appC-its}
++
++\paragraph{Design matrix.}
++Equation~\eqref{eq:its} of Part~IV is the linear model $y = X\beta +
++\varepsilon$ with $y \in \R^n$ and $X \in \R^{n \times 15}$ whose columns
++are, for $t = 0, \dots, n-1$ and event index $t_e$,
++\begin{equation}
++  x_t \;=\; \bigl(1,\; t,\; \mathbf{1}[t \ge t_e],\; (t - t_e)\,\mathbf{1}[t \ge t_e],\;
++  \mathbf{1}[\mathrm{month}(t) = 2],\; \dots,\; \mathbf{1}[\mathrm{month}(t) = 12]\bigr)^{\top},
++  \label{eq:appC-design}
++\end{equation}
++exactly the data frame built by \texttt{\_design()} in
++\texttt{econometrics/structural\_change.py} (columns \texttt{const},
++\texttt{t}, \texttt{post}, \texttt{t\_post}, \texttt{m2}, \dots,
++\texttt{m12}). The post-event trend is centred at $t_e$ rather than at zero
++so that $\beta_2$ is the level change \emph{at} the event month and not the
++extrapolated intercept difference at $t = 0$; with an uncentred post-trend
++the two parametrisations are related by $\beta_2^{\mathrm{unc}} = \beta_2 -
++\beta_3 t_e$, and only the centred one has the interpretation used in the
++text. The eleven month dummies are linearly independent of the constant
++because January is omitted; the full-column-rank condition requires at least
++one observation in every month of the year, satisfied by any series of
++twelve months or more, and at least one post-event month with $t > t_e$,
++guaranteed by the driver's rule that twelve months follow every event.
++The ordinary least-squares estimator is $\hat\beta = (X^{\top}X)^{-1}
++X^{\top}y$ and the residuals $\hat\varepsilon = y - X\hat\beta$.
++
++\paragraph{The HAC covariance estimator.}
++Under the assumption that $\varepsilon_t$ is a stationary, serially
++correlated and possibly heteroskedastic sequence with
++$\E[x_t \varepsilon_t] = 0$, the asymptotic covariance of $\sqrt{n}(\hat\beta
++- \beta)$ is $Q^{-1}\Omega Q^{-1}$ with $Q = \lim n^{-1}X^{\top}X$ and
++$\Omega = \sum_{\ell=-\infty}^{\infty} \Gamma_\ell$, where
++$\Gamma_\ell = \E[x_t \varepsilon_t \varepsilon_{t-\ell} x_{t-\ell}^{\top}]$
++is the autocovariance of the score at lag $\ell$. The estimator of
++\citet{newey1987hac} replaces the infinite sum by a kernel-weighted finite
++one,
++\begin{equation}
++  \hat\Omega_L \;=\; \hat\Gamma_0 + \sum_{\ell=1}^{L} w_\ell\,\bigl(\hat\Gamma_\ell + \hat\Gamma_\ell^{\top}\bigr),
++  \qquad
++  \hat\Gamma_\ell = \frac{1}{n}\sum_{t=\ell}^{n-1} \hat\varepsilon_t \hat\varepsilon_{t-\ell}\, x_t x_{t-\ell}^{\top},
++  \qquad
++  w_\ell = 1 - \frac{\ell}{L+1},
++  \label{eq:appC-nw}
++\end{equation}
++and the variance estimate is $\hat V = n\,(X^{\top}X)^{-1} \hat\Omega_L
++(X^{\top}X)^{-1}$, which is the expression of Part~IV once the factors of
++$n$ are absorbed into the un-normalised $\hat\Gamma_\ell$. The Bartlett
++weights $w_\ell$ are what make $\hat\Omega_L$ positive semi-definite for every
++sample: $\hat\Omega_L$ can be written as $\frac{1}{n(L+1)}\sum_{t}
++\bigl(\sum_{s=t-L}^{t} \hat u_s\bigr)\bigl(\sum_{s=t-L}^{t} \hat u_s\bigr)^{\top}$
++up to end effects, with $\hat u_s = x_s \hat\varepsilon_s$, a sum of outer
++products, whereas the truncated sum with $w_\ell \equiv 1$ is not guaranteed
++to be positive semi-definite. The implementation calls
++\texttt{statsmodels} with \texttt{cov\_type="HAC"} and \texttt{maxlags=6}, so
++$L = 6$ and $w_\ell = 1 - \ell/7$ as stated in Part~IV, without the
++small-sample degrees-of-freedom correction and with $p$-values from the
++normal approximation to the $t$-ratio, which is the library's default for
++robust covariances. With $n$ of the order of one hundred and fourteen
++regressors, the difference between the normal and the $t_{n-15}$ reference
++distribution is immaterial at the magnitudes of the $t$-statistics reported;
++what matters is the placebo rank of the next paragraph, which does not
++depend on any reference distribution.
++
++\paragraph{The placebo $p$-value as a rank statistic.}
++Let $\mathcal{T}' = \{t' : 24 \le t' \le n - 13\}$ be the candidate event
++indices (the implementation's \texttt{min\_pre = 24} and
++\texttt{min\_post = 12}) and, for each $t' \in \mathcal{T}'$, let
++$t^{(2)}_{t'} = \hat\beta_2(t') / \widehat{\mathrm{se}}(\hat\beta_2(t'))$ be
++the HAC $t$-ratio of the level change when the model
++\eqref{eq:appC-design} is fitted with $t_e = t'$. The placebo value of the
++true event $t_e \in \mathcal{T}'$ is
++\begin{equation}
++  p^{\mathrm{plac}} \;=\; \frac{1}{|\mathcal{T}'|}\sum_{t' \in \mathcal{T}'}
++  \mathbf{1}\bigl[\,|t^{(2)}_{t'}| \ge |t^{(2)}_{t_e}|\,\bigr]
++  \;=\; \frac{\mathrm{rank}\bigl(|t^{(2)}_{t_e}|\bigr)}{|\mathcal{T}'|},
++\end{equation}
++where $\mathrm{rank}$ is the descending rank of the true event's absolute
++$t$-ratio among the $|\mathcal{T}'|$ values, ties counted as not better
++(\texttt{>=}). Three properties follow directly. Because $t_e \in
++\mathcal{T}'$, the sum contains the term $t' = t_e$, so $p^{\mathrm{plac}}
++\ge 1/|\mathcal{T}'|$, which for the default window ($n = 101$, hence
++$|\mathcal{T}'| = 65$) is about $0.015$: a placebo rank cannot be smaller,
++however large the true statistic, and a reported value of $1/|\mathcal{T}'|$
++means ``largest of all candidate dates'', not ``significant at that level''.
++Because the statistic is a rank, it is invariant to any monotone
++transformation of the $|t|$-ratios and in particular does not depend on the
++HAC bandwidth through anything but the ordering it induces. And if the
++candidate dates were exchangeable under a null hypothesis of no break
++anywhere, $p^{\mathrm{plac}}$ would be a valid permutation-type $p$-value
++with $P(p^{\mathrm{plac}} \le \alpha) \le \alpha$; they are not exchangeable,
++since all $|\mathcal{T}'|$ regressions share the same $n$ observations and
++adjacent candidates give nearly identical fits, so the value is reported as
++a rank and read as such. The same construction gives the placebo rank of
++the slope change from $t^{(3)}_{t'}$. The driver additionally computes the
++rank restricted to the baseline candidates $\{t' < t_e\}$, which answers the
++protocol's original question (is the reference date special relative to the
++pre-period?) and is reported next to the full-range rank.
++
++\subsection{Bai--Perron multiple breaks by dynamic programming}
++\label{sec:appC-bp}
++
++\paragraph{Segment costs and additivity.}
++For a segment $[i, j)$ of the series with $j - i \ge h$, define
++\begin{equation}
++  S(i, j) \;=\; \min_{a, b \in \R}\ \sum_{t=i}^{j-1} (y_t - a - b\,t)^2,
++  \label{eq:appC-segcost}
++\end{equation}
++the residual sum of squares of a linear trend fitted to the segment alone
++(the implementation solves the two-parameter least-squares problem for each
++segment with \texttt{numpy.linalg.lstsq}; with \texttt{trend=False} the
++regressor $t$ is dropped and $S(i,j)$ is the centred sum of squares). The
++model with $m$ breaks $0 < \tau_1 < \dots < \tau_m < n$ assigns its own
++$(a, b)$ to each of the $m+1$ segments $[\tau_{k-1}, \tau_k)$ with
++$\tau_0 = 0$ and $\tau_{m+1} = n$. Because the parameters of different
++segments are distinct and each observation enters exactly one segment, the
++residual sum of squares of the whole model separates,
++\begin{equation}
++  \mathrm{SSR}(\tau_1, \dots, \tau_m) \;=\; \sum_{k=1}^{m+1} S(\tau_{k-1}, \tau_k),
++  \label{eq:appC-additive}
++\end{equation}
++and the global minimisation over $\tau$ decomposes: the optimal partition of
++the first $j$ observations into $m+1$ segments contains, as its first $m$
++segments, an optimal partition of the first $\tau_m$ observations into $m$
++segments. This is the principle of optimality, and it is the only property of
++the cost that the dynamic programme uses; it holds for any cost that is a
++sum over segments of a function of the segment alone, which is why the same
++recursion serves the piecewise-constant model of Section~\ref{sec:appC-pelt}.
++It fails for a model with parameters shared across segments (a common slope,
++or the seasonal dummies of the ITS), which is why the segmented regression
++carries no seasonal terms and the series is seasonally adjusted beforehand
++by subtracting the baseline-period month means
++(\texttt{run\_structural\_change.py}).
++
++\paragraph{The recursion.}
++Let $C_m(j)$ be the minimum of \eqref{eq:appC-additive} over partitions of
++observations $0, \dots, j-1$ into $m+1$ admissible segments. Then
++\begin{equation}
++  C_0(j) = S(0, j), \qquad
++  C_m(j) = \min_{\,mh \le i \le j-h} \bigl\{ C_{m-1}(i) + S(i, j) \bigr\},
++  \qquad j \ge (m+1)h,
++  \label{eq:appC-dp}
++\end{equation}
++the lower limit of $i$ ensuring that $m$ segments of length at least $h$
++fit before the last break and the upper limit that the last segment has
++length at least $h$. The argument minimiser $i^{\ast}_m(j)$ is stored, and
++the breaks of the $m$-break solution for the full series are recovered
++backward: $\tau_m = i^{\ast}_m(n)$, $\tau_{m-1} = i^{\ast}_{m-1}(\tau_m)$,
++and so on, which is what the dictionary \texttt{B[m][j]} of the
++implementation accumulates forward instead. This is the dynamic programme
++that \citet{bai2003computation} propose for the global minimiser of
++\citet{bai1998breaks}; the earlier literature used it for the
++piecewise-constant case, and its correctness requires nothing beyond
++\eqref{eq:appC-additive} and the induction on $m$: if $C_{m-1}(i)$ is the
++true minimum for every $i$, then any partition of $[0, j)$ into $m+1$
++segments has a last segment $[i, j)$ and a prefix partition of cost at least
++$C_{m-1}(i)$, so its cost is at least the right-hand side of
++\eqref{eq:appC-dp}, which is attained.
++
++\paragraph{Complexity.}
++The table of segment costs has one entry per admissible pair $(i, j)$, that
++is $\binom{n-h+1}{2} + O(n)$ entries, of order $n^2/2$. The recursion
++\eqref{eq:appC-dp} evaluates, for each $m \le M$ and each $j$, a minimum over
++at most $n$ candidates, so its cost is $O(n^2 M)$ operations; the
++back-tracking costs $O(M)$. The implementation fills the cost table by
++solving each segment's least-squares problem independently, at a cost
++proportional to the segment length, so the table costs $O(n^3)$ rather than
++the $O(n^2)$ obtainable by updating running sums as in
++\citet{bai2003computation}; for $n$ of the order of one hundred, as here,
++the difference is a fraction of a second and the simpler code was preferred
++for its transparency. Trimming enters only through the admissibility
++condition $j - i \ge h$; with $h = \max\{8, \lfloor 0.15\,n \rfloor\}$ and
++$M = 3$, the default window of $n = 101$ months gives $h = 15$ and a
++feasible region $(m+1)h \le n$ for every $m \le 3$, so all four candidate
++models exist.
++
++\paragraph{Information criteria as implemented.}
++For $m = 0, \dots, M$ let $\mathrm{SSR}_m = C_m(n)$ and let $p_m = k(m+1) +
++m$ be the parameter count, with $k = 2$ for the trend model ($k = 1$ for the
++mean model): one intercept and one slope per segment, plus the $m$ break
++positions, counted as parameters by convention. The two criteria computed by
++the code are
++\begin{align}
++  \mathrm{BIC}(m) &= n \log\frac{\mathrm{SSR}_m}{n} + p_m \log n, \label{eq:appC-bic}\\
++  \mathrm{LWZ}(m) &= n \log\frac{\mathrm{SSR}_m}{n - p_m} + p_m\, c_0\, (\log n)^{2.1}, \qquad c_0 = 0.299, \label{eq:appC-lwz}
++\end{align}
++and the selected number of breaks is $\hat m = \arg\min_m$ of each. The BIC
++form \eqref{eq:appC-bic} is the Gaussian log-likelihood, maximised over the
++error variance, in which $-2\log\hat L = n\log(\mathrm{SSR}/n) + \text{const}$,
++plus the Schwarz penalty $p \log n$ \citep{schwarz1978bic}; the two
++differences from a textbook BIC are that the break positions are counted
++among the parameters and that the constant and the segment variances are
++treated as common, both as in \citet{bai2003computation}. The LWZ form
++\eqref{eq:appC-lwz} is the modified Schwarz criterion of
++\citet{liu1997segmented}, whose penalty grows as $(\log n)^{2+\delta_0}$ with
++$\delta_0 = 0.1$ and whose variance estimate divides by $n - p_m$; the
++constant $c_0 = 0.299$ is the value they recommend and the one used in the
++simulations of \citet{bai2003computation}, which find that LWZ selects fewer
++spurious breaks than BIC under serial correlation at the cost of some
++power. The comparison across $m$ is legitimate because every $\mathrm{SSR}_m$
++is a global minimum over the same observations and the same admissible
++region; a sequential procedure that fixed earlier breaks before adding the
++next would not have this property and would need the sup-$F$ machinery that
++is not implemented. Neither criterion yields a confidence set for the break
++dates; the text reports the agreement of the two selections and the
++sensitivity to the truncation month instead.
++
++\subsection{PELT: cost, penalty and pruning}
++\label{sec:appC-pelt}
++
++\paragraph{Objective.}
++With the $\ell_2$ cost $\mathcal{C}(y_{i:j}) = \sum_{t=i}^{j-1} (y_t -
++\bar y_{i:j})^2$ for a segment $[i,j)$ and a penalty $\lambda > 0$ per
++changepoint, the problem of equation~\eqref{eq:pelt} of Part~IV is
++\begin{equation}
++  F(n) \;=\; \min_{m,\ \tau_1 < \dots < \tau_m}\ \Bigl\{ \sum_{k=1}^{m+1} \mathcal{C}(y_{\tau_{k-1}:\tau_k}) + \lambda m \Bigr\},
++  \label{eq:appC-peltobj}
++\end{equation}
++and, by the same additivity argument as in Section~\ref{sec:appC-bp}, it
++satisfies the optimal-partitioning recursion of \citet{jackson2005optimal},
++\begin{equation}
++  F(0) = -\lambda, \qquad F(j) \;=\; \min_{0 \le i < j}\ \bigl\{ F(i) + \mathcal{C}(y_{i:j}) + \lambda \bigr\},
++  \label{eq:appC-op}
++\end{equation}
++where the number of breaks is not fixed in advance but chosen by the
++penalty; the convention $F(0) = -\lambda$ cancels the penalty charged to the
++first segment. Solving \eqref{eq:appC-op} for all $j$ costs $O(n^2)$ cost
++evaluations.
++
++\paragraph{The pruning condition.}
++\citet{killick2012pelt} observe that if there exists a constant $K$ such that
++for all $i < j < l$
++\begin{equation}
++  \mathcal{C}(y_{i:j}) + \mathcal{C}(y_{j:l}) + K \;\le\; \mathcal{C}(y_{i:l}),
++  \label{eq:appC-Kcond}
++\end{equation}
++then any candidate $i$ for which
++\begin{equation}
++  F(i) + \mathcal{C}(y_{i:j}) + K \;\ge\; F(j)
++  \label{eq:appC-prune}
++\end{equation}
++can never be the last changepoint before any later $l > j$ and may be
++discarded from the candidate set. The proof is two lines. For $l > j$, by
++\eqref{eq:appC-Kcond} and \eqref{eq:appC-prune},
++$F(i) + \mathcal{C}(y_{i:l}) + \lambda \ge F(i) + \mathcal{C}(y_{i:j}) +
++\mathcal{C}(y_{j:l}) + K + \lambda \ge F(j) + \mathcal{C}(y_{j:l}) +
++\lambda$, so the partition that places a changepoint at $j$ after the
++optimal prefix of $[0,j)$ is at least as good as any that ends its last
++segment at $i$; hence $i$ never needs to be examined again. For the
++$\ell_2$ cost, \eqref{eq:appC-Kcond} holds with $K = 0$, because adding a
++split point to a segment can only reduce its residual sum of squares (the
++split model nests the unsplit one). The recursion with pruning examines, at
++each $j$, only the candidates not yet discarded; under the regularity
++conditions of \citeauthor{killick2012pelt} (changepoints whose number grows
++linearly with $n$) the expected number of surviving candidates is bounded and
++the total cost is linear in $n$, whence the name. For a short series with
++few changes, as here, the pruning saves little and the method is simply an
++exact solution of \eqref{eq:appC-peltobj}; the point of using it is
++exactness at a given penalty, not speed. The implementation
++(\texttt{ruptures.Pelt}, \citealp{truong2020ruptures}) is called with
++\texttt{model="l2"}, \texttt{min\_size=6} and \texttt{jump=1}, the last two
++restricting segments to at least six months and allowing a changepoint at
++every index; the library returns the segment end points including $n$, which
++the wrapper drops.
++
++\paragraph{The penalty $2\sigma^2 \log n$.}
++Suppose the observations are independent Gaussian with piecewise-constant
++mean and a common variance $\sigma^2$. Twice the negative log-likelihood of
++a partition, up to a constant, is $\sigma^{-2}\sum_k \mathcal{C}(y_{\tau_{k-1}:\tau_k})$.
++Each additional changepoint adds two parameters, a segment mean and a
++position, so the Schwarz criterion in this scale adds $2\log n$ per
++changepoint. Multiplying through by $\sigma^2$ to express the criterion in
++the units of the cost gives
++\begin{equation}
++  \sum_k \mathcal{C}(y_{\tau_{k-1}:\tau_k}) + 2\sigma^2 \log n \cdot m,
++\end{equation}
++which is \eqref{eq:appC-peltobj} with $\lambda = 2\sigma^2\log n$. The
++implementation replaces $\sigma^2$ by the sample variance of the whole
++series, $\widehat{\mathrm{Var}}(y) = n^{-1}\sum_t (y_t - \bar y)^2$
++(\texttt{numpy.var}, population form). This is the conservative choice
++discussed in Part~IV: in a series with a trend or with genuine level
++shifts, the overall variance exceeds the within-segment variance, so the
++penalty is larger than the within-segment Schwarz penalty and fewer
++changepoints are declared. \citet{yao1988schwarz} proves that the Schwarz
++criterion, with a known or consistently estimated $\sigma^2$, selects the
++true number of changes with probability tending to one in the
++piecewise-constant Gaussian model; the result does not extend to a series
++whose mean is a trend, for which a piecewise-constant fit must spend
++changepoints to track the slope, and the PELT dates of this study are read
++with that in mind.
++
++\subsection{Bayesian online changepoint detection}
++\label{sec:appC-bocpd}
++
++\paragraph{Conjugate updates.}
++Within a run the observations are modelled as
++$y \mid \mu, \sigma^2 \sim \mathcal{N}(\mu, \sigma^2)$ with the
++Normal--Inverse-Gamma prior
++$\mu \mid \sigma^2 \sim \mathcal{N}(\mu_0, \sigma^2/\kappa_0)$,
++$\sigma^2 \sim \mathrm{IG}(\alpha_0, \beta_0)$, whose joint density is
++proportional to
++\begin{equation}
++  (\sigma^2)^{-\alpha_0 - 3/2} \exp\Bigl\{ -\frac{\beta_0}{\sigma^2} - \frac{\kappa_0 (\mu - \mu_0)^2}{2\sigma^2} \Bigr\}.
++\end{equation}
++Multiplying by the likelihood of one observation $y$, proportional to
++$(\sigma^2)^{-1/2}\exp\{-(y-\mu)^2/(2\sigma^2)\}$, and completing the square
++in $\mu$,
++\begin{equation}
++  \kappa_0 (\mu - \mu_0)^2 + (y - \mu)^2
++  \;=\; (\kappa_0 + 1)\Bigl(\mu - \frac{\kappa_0\mu_0 + y}{\kappa_0 + 1}\Bigr)^2 + \frac{\kappa_0 (y - \mu_0)^2}{\kappa_0 + 1},
++\end{equation}
++shows that the posterior is again Normal--Inverse-Gamma with
++\begin{equation}
++  \mu_1 = \frac{\kappa_0 \mu_0 + y}{\kappa_0 + 1}, \quad
++  \kappa_1 = \kappa_0 + 1, \quad
++  \alpha_1 = \alpha_0 + \tfrac12, \quad
++  \beta_1 = \beta_0 + \frac{\kappa_0 (y - \mu_0)^2}{2(\kappa_0 + 1)},
++  \label{eq:appC-nigupdate}
++\end{equation}
++which are the four one-step updates of the implementation
++(\texttt{mu\_new}, \texttt{kappa\_new}, \texttt{alpha\_new},
++\texttt{beta\_new}). Applying \eqref{eq:appC-nigupdate} $r$ times and
++collecting terms gives the batch form quoted in Part~IV, with
++$\kappa_r = \kappa_0 + r$, $\mu_r = (\kappa_0\mu_0 + r\bar y)/\kappa_r$,
++$\alpha_r = \alpha_0 + r/2$ and
++$\beta_r = \beta_0 + \tfrac12\sum_{i \le r}(y_i - \bar y)^2 + \kappa_0 r (\bar y - \mu_0)^2 / (2\kappa_r)$;
++the equality of the two forms is a matter of algebra (induction on $r$) and
++is the reason the online recursion can be trusted to reproduce the batch
++posterior.
++
++\paragraph{Student-$t$ predictive.}
++The predictive density of the next observation integrates the Gaussian
++likelihood against the current Normal--Inverse-Gamma posterior. Integrating
++$\mu$ first gives $y \mid \sigma^2 \sim \mathcal{N}(\mu_r, \sigma^2 (1 +
++1/\kappa_r))$; integrating $\sigma^2 \sim \mathrm{IG}(\alpha_r, \beta_r)$
++against this scale mixture of normals gives
++\begin{equation}
++  p(y \mid r) \;=\; \int \mathcal{N}\Bigl(y;\ \mu_r,\ \sigma^2\frac{\kappa_r + 1}{\kappa_r}\Bigr)\,\mathrm{IG}(\sigma^2; \alpha_r, \beta_r)\, d\sigma^2
++  \;=\; t_{2\alpha_r}\Bigl( y;\ \mu_r,\ \frac{\beta_r(\kappa_r + 1)}{\alpha_r \kappa_r} \Bigr),
++  \label{eq:appC-student}
++\end{equation}
++a Student-$t$ density with $2\alpha_r$ degrees of freedom, location $\mu_r$
++and squared scale $\beta_r(\kappa_r+1)/(\alpha_r\kappa_r)$: the standard
++result that a normal with inverse-gamma mixing on the variance is a
++$t$ with degrees of freedom twice the shape and scale equal to the
++rate over the shape times the normal's variance factor. The code evaluates
++\eqref{eq:appC-student} through \texttt{scipy.stats.t.pdf} with
++\texttt{df = 2*alpha} and \texttt{scale = sqrt(beta*(kappa+1)/(alpha*kappa))},
++vectorised over the run lengths.
++
++\paragraph{Run-length recursion and the mass on $r = 0$.}
++Let $R_t(r) = P(r_t = r \mid y_{1:t})$ be the filtered run-length posterior
++after $t$ observations, with $R_0(0) = 1$, and let $\pi_t(r) = p(y_{t+1} \mid
++r_t = r, y_{1:t})$ be the predictive \eqref{eq:appC-student} evaluated with
++the hyperparameters of the run of length $r$ ending at $t$. With a constant
++hazard $H$ (a geometric prior on run lengths), the joint recursion of
++\citet{adams2007bocpd} is, for the growth and the change moves,
++\begin{equation}
++  \tilde R_{t+1}(r+1) = R_t(r)\,\pi_t(r)\,(1 - H), \qquad
++  \tilde R_{t+1}(0) = \sum_{r=0}^{t} R_t(r)\,\pi_t(r)\, H,
++  \label{eq:appC-rl}
++\end{equation}
++followed by normalisation, $R_{t+1}(r) = \tilde R_{t+1}(r) / \sum_{r'}
++\tilde R_{t+1}(r')$. Write $Z_t = \sum_{r} R_t(r)\pi_t(r)$ for the marginal
++predictive likelihood of $y_{t+1}$. Then $\sum_{r \ge 1}\tilde R_{t+1}(r) =
++(1-H) Z_t$ and $\tilde R_{t+1}(0) = H Z_t$, so the normaliser is $Z_t$ and
++\begin{equation}
++  R_{t+1}(0) \;=\; \frac{H Z_t}{H Z_t + (1-H) Z_t} \;=\; H \qquad \text{for every } t .
++  \label{eq:appC-r0}
++\end{equation}
++The filtered probability that a changepoint occurs \emph{at} the current
++step therefore equals the prior hazard identically, whatever the data: the
++run-length posterior carries information about where the last change was
++only through the relative masses at $r \ge 1$, which are reweighted by the
++predictive likelihoods, and not through the mass at $r = 0$, which the
++constant hazard fixes. This is why the detection statistics of Part~IV are
++not $R_t(0)$. The first, $P(r_t \le 3 \mid y_{1:t}) = \sum_{r=0}^{3} R_t(r)$,
++is $H$ plus the mass on runs of length one to three, that is, on changes
++that occurred one to three months earlier and have since accumulated
++evidence; its value exceeds $H$ only when the recent observations are
++better explained by a fresh run than by the continuation of the old one. The
++second is the maximum-a-posteriori run length $\hat r_t = \arg\max_r R_t(r)$,
++whose resets from a value above twelve to a value at most two are the
++online detector's changepoints. The implementation keeps the full
++$(n+1)\times(n+1)$ matrix $R$, adds $10^{-300}$ to the normaliser against
++underflow, and initialises $\mu_0$ and $\beta_0$ from the first six
++observations with $\kappa_0 = \alpha_0 = 1$; the prior mean run length
++$1/H = 60$ months is the only tuning constant. The recursion is the filtered
++one; the smoothed posterior of \citet{fearnhead2006exact} would require a
++backward pass and answers the retrospective question that the offline
++methods of Sections~\ref{sec:appC-bp} and~\ref{sec:appC-pelt} already
++address.

modified manuscript/appendices/appE_data_dictionary.tex +814 −1

Large diff (815 lines) is not loaded by default to keep the page fast.Load diff

modified manuscript/appendices/appH_reproducibility.tex +315 −1

Large diff (316 lines) is not loaded by default to keep the page fast.Load diff

modified manuscript/merge_bibs.py +1 −1

Large diff (8 lines) is not loaded by default to keep the page fast.Load diff

added manuscript/references_appAEH.bib +5 −0

Large diff (5 lines) is not loaded by default to keep the page fast.Load diff

added manuscript/references_appBC.bib +59 −0

Large diff (59 lines) is not loaded by default to keep the page fast.Load diff

added manuscript/references_verification_appAEH.csv +6 −0

Large diff (6 lines) is not loaded by default to keep the page fast.Load diff

added manuscript/references_verification_appBC.csv +6 −0

Large diff (7 lines) is not loaded by default to keep the page fast.Load diff

spb/wp44_uqo

WP44: Appendices A (taxonomy), E (data dictionary), H (reproducibility) by Agent L

Theme