# Agent E — Disciplinary review of the AI and machine-learning area (WP44, Phase 5)

Date: 2026-11-03. Inputs: `literature_synthesis/landmarks/ai_ml_landmarks.csv` (234 records) and `ai_ml_abstracts.jsonl` (190 non-empty abstracts); `results/bibliometrics/cagr_by_category.csv`; `results/graph/category_counts_yearly.csv`, `category_flows_yearly.csv`; `results/econometrics/breaks.csv`, `its_results.json`; `results/emerging/known_set_lead_times.csv`; `manuscript/references.bib` (verified) and `manuscript/references_part05.bib` (88 entries verified by this agent, see `references_verification_part05.csv`). Claim identifiers [E-nn] refer to `claims.csv`; every corpus number quoted below is read from the file named in the claim and is reproduced in the manuscript only through generated macros or tables (`tab_ai_growth`, `tab_ai_counts`, `tab_ai_breaks`, `tab_ai_altdates`, `tab_ai_flows`, `tab_ai_landmark_flags`, `tab_ai_contribution_types`, `tab_ai_landmarks`). Evidence types: *abstract* = read from the stored abstract; *metadata_only* = title/venue/type/identifiers only; *full_text* = none in this review; *corpus_statistic* = a WP44 result file. No full text was read; nothing below describes the content of a paper beyond what its abstract or metadata states. Preprints are named as preprints. The November 2022 boundary is a reference date, not a cause.

## 0. The evidence base and its defects

**The landmark set is not usable as delivered.** The automatic selection (`select_landmarks.py`: highest NCS36 within primary category, floor of fifteen, plus the 200 most cited works overall) returned 234 AI/ML candidates: 33 LLM, 26 AI safety and 25 in each of the other seven categories; 72 are dated 2026 and 39 dated 2025, 136 are typed `article`, 28 `preprint`, 19 `book-chapter`, 11 `dissertation` [E-01, corpus_statistic]. Screening every record against its identifiers, citation counts, venue and abstract (`contributions.csv`, column `data_quality_flag`) flags 44 of 234 (18.8 %) as unreliable [E-02]: 13 *identifier mismatches*, in which the title and abstract belong to an obscure work while the arXiv identifier and the citation count belong to a famous one (the record titled "Toward effective Text-to-MongoDB query translation" carries arXiv 2005.11401, the retrieval-augmented-generation paper, and 3,026 36-month citations; a Portuguese essay on copyright carries 2005.14165, GPT-3; an essay on "theatrical compliance" carries 2201.11903, chain-of-thought prompting; an ICU-simulation preprint carries 2307.09288, Llama 2; a reproduction runner for Adam carries 45,775 citations); 3 *abstract mismatches*, in which a correct title (GPT-4 Technical Report; "GPTs are GPTs"; Powers' evaluation-measures paper) is paired with the abstract of a different work; 11 *implausible years* (mixup, the variational RNN, Soft-SOO, spectral bandits and three other 2013–2019 papers appear as 2024–2026 works, four of them with April–June 2026 arXiv identifiers); 7 *normalisation artefacts* (dissertations with 2–20 citations and NCS36 above 500, because the field-normalised score divides by a tiny reference mean for rare document types and young cohorts); and 10 *off-scope* records (control-theoretic consensus papers under `ai_agents`, an essay on ed-tech "solutionism" under `reinforcement_learning`). The defect is concentrated in the LLM category (8 of 33 flagged for identifier or abstract problems) and in the two "old" categories, ML theory and DL architectures, where misdated classics dominate [E-03]. The mechanism is upstream of this review — OpenAlex work merging and the second-pass arXiv/title linkage of Phase 1 — and must be reported in Part III and audited by Agent N before any landmark statistic is published; the clean count of citation-ranked landmarks in the AI area is therefore 190 at most, and the NCS36 ranking should be re-run with document-type-specific reference sets and a 2026 cohort excluded [E-04, speculative as to remedy].

**What the set does contain.** Of the 234 records, 84 are surveys, 67 methods, 33 empirical studies, 29 theory/position papers, 17 engineering systems and only 4 benchmarks (`tab_ai_contribution_types`; judged from the abstract for 190, from metadata for 44) [E-05]. Within the LLM and generative-AI categories the citation-ranked top is dominated by education and commentary papers on ChatGPT (Kasneci et al. 2023; Lo 2023; Rudolph, Tan and Tan 2023; Chan and Hu 2023; Baidoo-Anu and Owusu Ansah 2023; Dwivedi et al. 2023; van Dis et al. 2023; Lim et al. 2023), not by the technical papers that introduced the models [E-06]. This is a property of the field-normalised ranking inside a multi-label category that spans computer science, education and information systems: the canonical technical lineage (GPT-4, LLaMA, InstructGPT, chain-of-thought, DPO, QLoRA, BLIP-2, Mamba) is present only where an arXiv record was linked, and the review therefore grounds the technical synthesis on the verified entries of `references.bib` (REFERENCE_MAP Part V) and uses the landmark set to characterise what the corpus *rewards* rather than what the field *invented*.

## 1. The corpus picture of the nine categories

*Volume.* The LLM category grows from a primary-category count of a few thousand works in 2018–2022 to tens of thousands in 2025, with a post-period CAGR of 1.085 against 0.189 in the baseline (`cagr_by_category.csv`); generative AI (0.870 vs 0.296), multimodal AI (0.869 vs 0.284), foundation models (0.802 vs 0.661) and AI safety (0.647 vs 0.294) follow; reinforcement learning (0.252 vs 0.208) and ML theory (0.197 vs 0.104) accelerate modestly; DL architectures decelerate slightly (0.155 vs 0.167); AI agents reverse from a negative baseline CAGR (−0.045) to 0.329 [E-07, corpus_statistic]. The LLM share of eligible works rises from 0.33 % (2018) to 5.06 % (2025) [E-08]. Multi-label counts (`tab_ai_counts`) are three to ten times larger than primary counts because the AI categories overlap heavily; the ordering of growth rates is the same under both definitions.

*Structural change.* The interrupted time series at 2022-12 (`breaks.csv`) shows, for the LLM series, a level change of 0.106 log points (HAC p = 0.075) and a slope change of 0.034 log points per month (p ≈ 10⁻⁶⁰), with a placebo p-value of 0.938 for the level and 0.200 for the slope; the Bai–Perron procedure selects a single break at 2022-07 and PELT a change point at 2023-10 [E-09]. The same pattern — no significant level jump at the reference month, a strongly significant slope change, placebo p-values that do not single out December 2022, data-driven breaks before or well after it — holds for multimodal (slope 0.024; break 2022-04) and generative AI (slope 0.019; break 2021-11). AI agents (level −0.210, p = 0.028; slope 0.040), reinforcement learning (level −0.164, p = 0.013) and AI safety (level −0.206, p = 0.002) show *negative* level changes with positive slope changes, i.e. a dip followed by acceleration; DL architectures is the only AI category with no significant slope change (−0.002, p = 0.136) [E-10]. At the alternative event dates the picture is monotone: level effects at the GPT-3 (2020-06) and DALL-E 2 (2022-05) dates are negative for every AI category, while those at the GPT-4 (2023-04) and o1 (2024-10) dates are positive and significant for LLM, multimodal and generative AI (`tab_ai_altdates`) [E-11]. The conservative reading is that the AI literature's acceleration is a trend change that was already under way in mid-2022 and steepened through 2023–2024; the monthly series cannot attribute it to a single product release, and the manuscript must not do so [E-12, established as a statement about the estimates].

*Knowledge flows.* In-corpus references of LLM works went to NLP for 69.6 % in 2019 and 68.6 % in 2022, but only 34.2 % in 2025, when 28.4 % go to the LLM category itself, 4.3 % to HCI and 3.4 % to software engineering (`category_flows_yearly.csv`) [E-13]. Generative-AI works cited computer vision for 53.6 % (2019) and 57.9 % (2022) of their in-corpus references and 33.8 % in 2025, when 13.7 % go to LLM and 6.6 % to HCI; AI-safety works cite LLM for 10.0 % and HCI for 11.5 % in 2025 against 0.1 % and 9.1 % in 2019 [E-14]. Reinforcement learning moves the other way — its share of references to robotics rises from 14.4 % to 23.1 % and to LLM stays below 1 % — and ML theory and DL architectures are essentially unchanged [E-15]. The LLM category has thus become a citation hub for generative AI, safety and agents while remaining peripheral to RL, theory and architectures; this is the corpus-level answer to RQ5 at the level of categories, and it is consistent with but does not establish the "foundation model" convergence thesis (Bommasani et al. 2021) [E-16, contested].

*Emergence lead times.* In the preregistered known-emergence set (`known_set_lead_times.csv`) the detector flags "large language models" 34 months before its peak-growth month, retrieval-augmented generation 12 months, vision-language models 9, diffusion models 9, LLM agents 6, mixture-of-experts 5, speculative decoding 5, state-space models 6, reasoning models 29; mechanistic interpretability has a negative lead (−18), its peak-growth month preceding detection [E-17]. The median lead over the set is 7 months. The long LLM lead reflects the slow pre-2020 literature on "language models" matched by the regex, not foresight.

## 2. Large language and foundation models

*Established.* The decoder-only Transformer scaled with data and compute (Vaswani et al. 2017; Brown et al. 2020; Chowdhery et al. 2022) remains the dominant architecture inside the window; the compute-optimal scaling result of Hoffmann et al. (2022) — more tokens per parameter than Kaplan et al. (2020) implied — is reflected in the open-weight lineage LLaMA → Llama 2 → Llama 3 (Touvron et al. 2023a,b; Dubey et al. 2024; all preprints), whose LLaMA abstract reports a 13B model outperforming GPT-3 on most benchmarks while trained only on public data [E-18, abstract]. Closed frontier releases are documented by technical reports without training details (OpenAI 2023; Gemini Team 2023; OpenAI 2024) and must be cited as such. Domain-specific foundation models are the most visible post-2022 development in the landmark set: self-supervised pathology encoders UNI, CONCH, CHIEF, PLIP and TITAN (Chen et al. 2024; Lu et al. 2024; Wang et al. 2024; Huang et al. 2023; Ding et al. 2025), radiology (Wu et al. 2025), dermatology (Yan et al. 2025), brain MRI (Tak et al. 2026), the Earth system (Bodnar et al. 2025), tabular data (Hollmann et al. 2025), time series (Jin et al. 2023 preprint; Liu et al. 2024) and human cognition (Binz et al. 2025) — all peer-reviewed journal or conference papers whose abstracts report outperforming task-specific baselines on multi-task benchmarks [E-19, abstract]. The adaptation toolkit — LoRA, QLoRA, LLaMA-Adapter, PEFT surveys (Hu et al. 2021; Dettmers et al. 2023; Zhang et al. 2023 preprint; Han et al. 2024 preprint) — is where the open-weight ecosystem's practical growth concentrated; the QLoRA abstract's claim of "99.3 % of ChatGPT" on the Vicuna benchmark is a judge-based measurement, not a capability proof [E-20, abstract].

*Contested.* Emergent abilities: Wei et al. (2022, preprint) report discontinuous capability jumps with scale; Schaeffer et al. (2023, preprint) show that the discontinuities largely disappear under linear or continuous metrics. Both are preprints; the corpus cannot adjudicate and the manuscript should present the disagreement as a measurement question [E-21]. "Sparks of AGI" (Bubeck et al. 2023, preprint) reports qualitative experiments on an early GPT-4 and "contends" general intelligence; it is widely cited (1,420 c36 in the landmark file) but is an unreplicated, non-peer-reviewed position [E-22]. Pure-RL reasoning: the DeepSeek-R1 Nature paper (Guo et al. 2025) reports that reasoning "can be incentivized through pure reinforcement learning, obviating the need for human-labelled reasoning trajectories"; it is peer reviewed, but the generality of the recipe beyond verifiable domains is untested in the corpus [E-23, abstract].

*Speculative.* That domain foundation models will displace task-specific pipelines in regulated domains; the radiology review by Paschali et al. (2025) and the data-centric survey of Zhang et al. (2026) list data, privacy and evaluation obstacles that the landmark abstracts do not resolve [E-24].

## 3. Alignment and post-training

The RLHF recipe (Christiano et al. 2017; Ouyang et al. 2022; Bai et al. 2022a,b) and its closed-form successor DPO (Rafailov et al. 2023, NeurIPS; the abstract is unavailable in the landmark file) are established methods whose limitations are documented by Casper et al. (2023, preprint) [E-25]. The corpus flows show RL's citations to LLM remaining below 1 % through 2025 [E-15], so the "RL renaissance" through post-training is visible in the LLM category rather than in the RL category: the categorisation separates the two communities even where the methods are shared. Jailbreak and adversarial-prompt work — universal suffixes (Zou et al. 2023, preprint), the "jailbroken" analysis (Wei et al. 2023, preprint) and PAIR, which the abstract reports generates semantic jailbreaks with black-box access "in twenty queries" (Chao et al. 2025, IEEE SaTML) — establishes that current alignment is brittle to optimisation pressure [E-26, abstract]. Whether this brittleness is intrinsic to RLHF-style alignment or an artefact of current training scale is unresolved (open problem P-03).

## 4. Reasoning, retrieval and tools

Chain-of-thought prompting (Wei et al. 2022; Kojima et al. 2022), tree-of-thought (Yao et al. 2023a) and Reflexion (Shinn et al. 2023) are preprints or NeurIPS papers whose claims rest on benchmark gains; the reasoning-model turn (OpenAI 2024; Guo et al. 2025) moves the computation to inference time. The known-set detector flagged "reasoning models" in 2023-06, 29 months before the peak-growth month of 2025-11, because the regex includes chain-of-thought [E-17]. Retrieval augmentation (Lewis et al. 2020; Karpukhin et al. 2020; Gao et al. 2023 survey) is the most adopted pattern (12,896 matching works; detection 2023-10, peak growth 2024-10) [E-17]. Tool use — ReAct, Toolformer, ToolLLM (Yao et al. 2023b; Schick et al. 2023; Qin et al. 2023) — is the bridge to agents. Hallucination remains an unsolved measurement problem (Ji et al. 2023; Huang et al. 2025); Singhal et al. (2023, Nature) built the MultiMedQA benchmark and a human evaluation axis precisely because automated metrics were insufficient for clinical answers [E-27, abstract]. Code generation (Chen et al. 2021; Li et al. 2022; Rozière et al. 2023; Li et al. 2023) and its productivity evidence (Peng et al. 2023; Ziegler et al. 2022) belong to Part VII-a's software-engineering synthesis; the detector's lead time for "AI code generation" is 10 months.

## 5. Multimodal and generative AI

*Established.* Contrastive image–text pre-training (Radford et al. 2021) and vision Transformers (Dosovitskiy et al. 2021) underlie the vision-language models of the window: Flamingo, BLIP-2, InstructBLIP, LLaVA, mPLUG-Owl2 (Alayrac et al. 2022; Li et al. 2023 preprint; Dai et al. 2023 preprint; Liu et al. 2023 preprint; Ye et al. 2024 CVPR) share the design of a frozen or lightly adapted LLM fed by a visual encoder through a small bridge; SAM (Kirillov et al. 2023) is the promptable-segmentation analogue [E-28]. LAION-5B (Schuhmann et al. 2022) made open CLIP-scale training possible. Diffusion models (Ho et al. 2020; Rombach et al. 2022; Saharia et al. 2022; Ramesh et al. 2022), classifier-free guidance (Ho and Salimans 2022 preprint), textual inversion (Gal et al. 2022 preprint), T2I-Adapter (Mou et al. 2024, AAAI), DiT (Peebles and Xie 2023) and SD3 (Esser et al. 2024 preprint) form the generative lineage, with MatterGen (Zeni et al. 2025, Nature) as a peer-reviewed scientific application reporting stable, diverse inorganic materials [E-29, abstract]. Speech follows the same self-supervised path (Baevski et al. 2020; Hsu et al. 2021; Radford et al. 2023).

*Contested.* Multimodal grounding in medicine: the evaluation of eight multimodal foundation models on 1,090 cases reports that image predictions are "largely driven by text" and that mild textual suggestions of a wrong diagnosis degrade image classification (Buckley et al. 2026, Nature Communications) [E-30, abstract]; the scoping review of 432 medical multimodal papers (Schouten et al. 2025) and the low-quality-data survey (Zhang et al. 2026) describe missing-modality and noise robustness as open. Whether current VLMs *fuse* or merely *condition on* modalities is therefore unsettled (P-05).

*Corpus.* Generative AI is the category whose citation profile changed most: from a vision-centred (53.6 %) to a split vision/LLM/HCI profile [E-14]; its citation-ranked landmarks are, however, overwhelmingly education and policy studies (Chan 2023; Xia et al. 2024; Bittle and El-Gayar 2025; McDonald et al. 2025; the OECD 2026 Outlook; Bastani et al. 2024 SSRN preprint reporting that GPT-4 access harmed unassisted exam performance) and adoption surveys (Bick, Blandin and Deming 2026: 45 % of US adults 18–64 using generative AI by late 2024, per the abstract) [E-31, abstract]. Technical generative-AI papers are counted in the category but rarely reach its citation top.

## 6. AI agents

*Established.* LLM-based agents are surveyed by Wang et al. (2024), Xi et al. (2023 preprint), Guo et al. (2024) and Cheng et al. (2026); memory mechanisms by Zhang et al. (2025); architectures include generative agents (Park et al. 2023), AutoGen (Wu et al. 2023 preprint), MemGPT (Packer et al. 2023 preprint) and "LLM as OS" proposals (Mei et al. 2024; Ge et al. 2023, preprints). Benchmarks — SWE-bench, WebArena, AgentBench (Jimenez et al. 2024; Zhou et al. 2023; Liu et al. 2023, preprints) — are preprints; the only peer-reviewed agent benchmark in the landmark set is the clinical evaluation of OpenManus and Manus (Liu et al. 2026, npj Digital Medicine) [E-32]. Science agents are the strongest peer-reviewed evidence: the Virtual Lab (Swanson et al. 2025, Nature) designed nanobodies with experimental validation, Robin (Ghareeb et al. 2026, Nature) automates hypothesis generation and data analysis in experimental biology, DIVE (Zhang et al. 2026) extracts data from figures for materials discovery; AgentSociety (Piao et al. 2026) simulates LLM-driven populations [E-33, abstract]. The AI-scientist preprints (Lu et al. 2024; Si et al. 2024) remain hypotheses.

*Contested.* The "agentic AI" definitional literature — Acharya et al. (2025), Sapkota et al. (2026), Bandi et al. (2025), Abou Ali et al. (2025), Allmendinger et al. (2026), Collaço et al. (2026) — produced at least six taxonomies in 2025–2026, most of which are themselves reviews of reviews; the landmark set contains six of them and one benchmark [E-34]. Whether "agentic AI" denotes a distinct class of systems or a relabelling of tool-augmented LLM pipelines is unresolved (P-06). The ai_agents category is heterogeneous: its series mixes classical multi-agent systems and control theory (four off-scope landmarks) with LLM agents, which explains the negative baseline CAGR, the negative 2022-12 level change and the second Bai–Perron break at 2024-07 [E-35]. The detector's lead time for LLM agents is 6 months (detection 2023-10, peak 2024-04).

*Speculative.* Trust, risk and security management for multi-agent LLM systems (Raza et al. 2026) is a framework without empirical validation.

## 7. Reinforcement learning

The deep-RL canon of the baseline period is intact in the landmark set — Rainbow, SAC, TD3, AlphaZero, AlphaStar, MAPPO (Hessel et al. 2018; Haarnoja et al. 2018 preprint; Fujimoto et al. 2018; Silver et al. 2018; Vinyals et al. 2019; Yu et al. 2022), the reproducibility warning of Henderson et al. (2018) and the curiosity study of Burda et al. (2019) — and the single post-2022 method landmark is DreamerV3 (Hafner et al. 2025, Nature), whose abstract reports one configuration mastering over 150 tasks through world models [E-36, abstract]. The MPC–RL survey (Reiter et al. 2026) and the MARL overview (Zhang et al. 2021) frame the field's control-theoretic side. RL's post-2022 slope change (0.012) is the smallest among the accelerating AI categories and its citations go increasingly to robotics (23.1 % in 2025) [E-15]. The field's most consequential post-2022 role — reward modelling and policy optimisation for LLMs (PPO, DPO, GRPO in DeepSeek-R1) — is invisible in the RL category because those papers are classified as LLM work; any claim that "RL declined" after 2022 would be an artefact of the taxonomy [E-37, contested].

## 8. Machine-learning theory and deep-learning architectures

The ML-theory landmark set contains no learning-theory result from the window: it is textbooks (Alpaydin; Zaki and Meira), evaluation metrics (Chicco and Jurman 2020), interpretability (Rudin 2019; Lipton 2018; Lundberg et al. 2020; Verma and Pearl 2022 reprint), optimisation (Bottou, Curtis and Nocedal 2018), transfer learning (Zhuang et al. 2021) and misdated 2013–2019 bandit papers [E-38]. The category as classified is "general machine learning", and its citation top rewards pedagogy and measurement conventions; theoretical work on transformers, in-context learning or scaling (which does exist in the window) is classified elsewhere or cited too little in 36 months to surface. The DL-architectures set is the mirror image: GNN surveys (Wu et al. 2021; Zhou et al. 2020), time-series Transformers and their refutation — Zeng et al. (2023) report that simple linear models outperform Informer-style Transformers on long-term forecasting benchmarks (Zhou et al. 2021; Wen et al. 2023) [E-39, abstract, contested within the field]; mixup (Zhang et al. 2018); Mamba (Gu and Dao 2023, preprint) and KAN (Liu et al. 2024, preprint) as the two architectural proposals of the window, both unreviewed; TabPFN (Hollmann et al. 2025, Nature) as the peer-reviewed one. Efficiency work — FlashAttention, MoE, INT8/GPTQ quantisation, speculative decoding (Dao et al. 2022; Fedus et al. 2022; Dettmers et al. 2022; Frantar et al. 2023; Leviathan et al. 2023; Gholami et al. 2022) — is where architecture research actually moved, and the detector finds short leads for state-space models (6 months), MoE (5) and speculative decoding (5), i.e. these topics were detected only months before their growth peaked [E-40]. DL architectures is the one AI category without a significant post-2022 slope change [E-10]; whether this is a plateau or a relabelling of architecture papers under LLM/foundation categories cannot be decided from the counts (P-08).

## 9. Evaluation

Static benchmarks (GLUE, SuperGLUE, MMLU, BIG-bench: Wang et al. 2018, 2019; Hendrycks et al. 2021; Srivastava et al. 2022), holistic evaluation (Liang et al. 2022), arena-style human preference (Chiang et al. 2024 preprint) and LLM-as-judge (Zheng et al. 2023 preprint) coexist with documented contamination (Sainz et al. 2023), benchmark saturation and the "benchmark lottery" (Dehghani et al. 2021; Ott et al. 2022; Raji et al. 2021; Kiela et al. 2021; Recht et al. 2019). The landmark set contains four benchmark contributions in 234 (SuperGLUE, LAION-5B, MultiMedQA, the clinical agent benchmark) [E-05]; evaluation is under-rewarded by citation ranking relative to its role. The SuperGLUE abstract already records in 2019 that GLUE performance "surpassed the level of non-expert humans", the first instance of the saturation cycle that repeats for MMLU within the window [E-41, abstract]. No landmark abstract reports a contamination-controlled evaluation (P-09).

## 10. Safety, alignment, fairness and societal evidence

*Established.* Risk taxonomies have matured from the concrete-problems agenda (Amodei et al. 2016; Hendrycks et al. 2021) to the AI Risk Repository, which the abstract describes as a living database of 1,725 risks from 74 taxonomies (Slattery et al. 2026, Patterns), the International AI Safety Report (Bengio et al. 2025) and the Science perspective on extreme risks (Bengio et al. 2024) [E-42]. Fairness and explainability surveys (Mehrabi et al. 2021; Barredo Arrieta et al. 2020; Ali et al. 2023; Miller 2019) dominate the AI-safety landmark top, i.e. the category's citation top is pre-LLM "responsible AI". Empirical harms with peer review: medical LLMs are vulnerable to data poisoning at 0.001 % of training tokens (Alber et al. 2025, Nature Medicine), four chatbots gave unsafe answers to patient questions in a physician-led red-teaming (Draelos et al. 2026), LLMs express stigma and respond inappropriately as replacement therapists (Moore et al. 2025, FAccT) [E-43, abstract]. Inference energy of general-purpose models exceeds that of task-specific ones by orders of magnitude (Luccioni et al. 2024; de Vries 2023) [E-44]. Labour and adoption: Eloundou et al. (2024), Dell'Acqua et al. (2023 preprint; the Organization Science version is in the landmark set) and Bick et al. (2026) are the peer-reviewed or preregistered evidence [E-45].

*Contested.* Catastrophic-risk framing (Hendrycks et al. 2023; Shevlane et al. 2023; both preprints) versus the "stochastic parrots" critique (Bender et al. 2021); the corpus shows the AI-safety category citing LLM work for 10 % of its references in 2025 against 0.1 % in 2019 [E-14], so the field has re-oriented, but no landmark abstract reports a validated measurement of a catastrophic capability (P-10). Governance: Papagiannidis et al. (2025) and the healthcare maturity model (2026 review) are frameworks; McDonald et al. (2025) document 116 university policies — descriptive, not evaluative.

## 11. Reconciliations and what the corpus cannot say

1. *Growth vs discontinuity.* The AI categories accelerated; the acceleration is a slope change dated by data-driven methods to mid-2022 or later and not a December-2022 level jump [E-09–E-12]. Statements of the form "ChatGPT caused a surge" are not supported by the design.
2. *Impact vs invention.* Citation-ranked landmarks in LLM and generative AI are commentary and education papers; the technical lineage is cited through arXiv records that the linkage sometimes misattached [E-02, E-06]. Any "most influential papers" list for AI must be built from a cleaned identifier graph.
3. *RL and architectures.* Their apparent stagnation is partly taxonomic [E-37, P-08].
4. *Agents.* Six taxonomies and one peer-reviewed benchmark [E-34]; science agents are the exception with experimental validation [E-33].
5. *Evidence tiers.* 28 of 234 landmarks are preprints and 72 are dated 2026 with 12-month windows; everything said about 2025–2026 is provisional.

Open problems feeding Part X are listed in `open_problems.csv` (P-01 to P-14).