spb/wp44_uqo

1# Agent F — Systems and Computing Infrastructure: disciplinary synthesis (WP44, Phase 5)23Scope: database systems, information retrieval, distributed computing, cloud computing, edge computing, computer networks4(area `data_distributed`); computer architecture, accelerators and AI hardware, operating systems, quantum computing,5sustainable computing (area `hardware_emerging`). Inputs: `literature_synthesis/landmarks/{data_distributed,hardware_emerging}_landmarks.csv`6and `_abstracts.jsonl`, `results/graph/category_counts_yearly.csv`, `results/graph/category_flows_yearly.csv`,7`results/bibliometrics/cagr_by_category.csv`, `results/econometrics/breaks.csv`, `manuscript/references.bib`.8Every claim below is indexed in `claims.csv` (F-01 … F-43); every landmark discussed is typed in `contributions.csv`;9unresolved questions are in `open_problems.csv` (OP-F01 … OP-F13). Evidence is metadata and abstracts only; no full text10was read. Preprints are named as such. Nothing in this review attributes a change to ChatGPT.1112## 1. What the evidence base is, and what it is not1314### 1.1 The candidate lists1516The two landmark lists contain 152 and 125 records, that is 25 per category after the floor rule of17`select_landmarks.py` (top works by NCS36 within primary category; for 2025–2026, c12 relative to the year mean). Between1810 (distributed computing, operating systems) and 20 (AI accelerators, sustainable computing) records per category fall in19the post period. Abstracts are available for 14 of 25 accelerator records and for 19–24 of 25 elsewhere (F-10).2021Two properties of the scoring rule shape what follows. First, because 2025–2026 records are ranked by c12 relative to a22year mean that is itself small, several 2026 "landmarks" have nine to thirty citations in total (F-11); they are23provisional and are treated here as indications of what is being cited now, not as established landmarks. Second,24NCS36 rewards records that OpenAlex has re-dated: the distributed-computing list opens with MapReduce, Lamport's25"Time, clocks" and "The Byzantine generals problem" recorded as 2018–2019 re-publications, an OpenMP `taskwait`26paper carries c36 = 12,891, an eBPF preprint dated 2020 carries the DOI of a 2026 OASIcs volume, and NumPy27("Array programming with NumPy", c36 = 5,512) is the top baseline record of computer architecture (F-36). These28anomalies are documented in `contributions.csv` and `claims.csv` and should be fed to the data-quality discussion of29Parts III and XI; they are not discussed as scientific contributions of the period.3031### 1.2 Misclassified records (excluded from the synthesis, listed for Agent K)3233The rule-based classifier assigns to the systems categories a number of records that are not systems research. The34clearest cases are: `cloud_computing` 110250037 (a dissertation on teaching Chinese as a second language), 436437573035(social health protection during COVID), 4385584867 and 4386827074 (energy-management book chapters);36`database_systems` 3080729669 and 3032576892 (student dissertations), 3005599391 (an ethnography of mobile phones),373037795158 (digital technology in pandemic response, *Lancet Digital Health*), 4324129606 (data quality of online38human-subjects panels), 4402762189 and 4408925021 (e-learning), 7161586974 (Paleobiology Database user guide);39`operating_systems` 3092732687 and 3134501457 (student dissertations), 4366587887 (educational technology),407130515337 (nursing students' AI literacy), 7162190207 (AI mental-health therapy), 7131624127 (feature selection),417127083763 (regulatory sandboxes); `quantum_computing` 4386285839 (strategic transformation of higher education),424407091026 (mammography screening); `sustainable_computing` 4403458785, 4407614581, 7122842330, 7128511066, 713163845243(higher education, banking, globalisation, product passports); `distributed_computing` 3182706339 (a statistics paper44on R-squared), 7130586201, 7123363719, 7118998287 (control theory); `computer_architecture` 3099878876 (NumPy),453046751266 and 3016932526 (molecular-dynamics and DFT codes), 2952332632 (autonomous-vehicle architecture);46`edge_computing` 3017623177, 3023398824, 3006456063, 3023617420 (COVID-era application reviews), 440180668647(generic AI overview); `information_retrieval` 4395483946 (generative AI in higher education). Roughly one record in48five in the two lists is of this kind. This is a finding about the taxonomy (OP-F12), and it means that the category-level49statistics of §2 carry classification noise whose magnitude is not estimated here.5051A second group is in scope but is not "research of the period": IETF RFCs (Multipath TCP 2020, SCTP 2022, BPF ISA522024), textbooks and handbooks (*Principles of Distributed Database Systems*, *Wireless Communications*, *Spatial Data53Science*), a Linked-Data chapter and a Semantic-Web reprint. They record what the field cites, not what it produced.5455### 1.3 How the review reads the rest5657After these exclusions, 74 records are discussed and typed in `contributions.csv`. Every statement about a paper's58content is taken from its abstract; where the abstract is missing (Nature 2025 photonic accelerator; the 2025 barren-plateau59review; the LLM-for-recommendation survey; the processing-in-memory primer) the statement is from title and venue60only, and `claims.csv` says so.6162## 2. The eleven categories in the corpus6364### 2.1 Volume and growth6566`cagr_by_category.csv` (eligible works by primary category; 2026 excluded as incomplete) separates the eleven67categories into two groups (F-04, F-05). Seven classical categories — computer architecture, operating systems,68distributed computing, cloud computing, computer networks, database systems and information retrieval — *declined*69between 2018 and 2022 (baseline CAGR between −0.018 for IR and −0.087 for operating systems) and *grew* between702023 and 2025 (post CAGR between 0.071 for networks and 0.198 for cloud). Four categories grew in both periods:71edge computing (0.133 then 0.126), quantum computing (0.103 then 0.287), sustainable computing (0.092 then 0.233) and72AI accelerators (0.148 then 0.092). The shares of the classical categories in total primary output fell throughout:73computer networks, the largest, from 0.118 in 2018 to 0.067 in 2025; architecture from 0.039 to 0.023; databases from740.033 to 0.019. Quantum computing rose from 0.0125 to 0.0196. The relative decline of the classical systems fields is75a pre-2022 phenomenon that continued after 2022 even as their absolute output recovered.7677### 2.2 Structural change7879In the interrupted time-series model of Part IV (`breaks.csv`, monthly log-volume, event 2022-12), all eleven80categories show a positive and significant change in trend (F-01): from 0.004 per month for edge computing81(p = 0.006) to 0.023 for information retrieval, with 0.020 for distributed computing, 0.019 for cloud, 0.018 for82accelerators and quantum, 0.016 for architecture, operating systems and sustainable computing, 0.014 for databases83and 0.009 for networks. This uniformity is itself informative: the trend steepening is not confined to the categories84that have anything to do with language models. The placebo test confirms the point (F-02): for every category the85share of placebo event dates whose slope statistic is at least as extreme is 0.062 or more (architecture 0.062,86distributed and edge 0.077, databases and IR 0.092, accelerators, quantum and sustainable 0.169, networks 0.200, cloud870.231, operating systems 0.477). Level changes at 2022-12 are heterogeneous in sign (F-03): negative and significant88for IR (−0.094), accelerators (−0.113) and quantum (−0.087); positive and significant for networks (+0.043) and89sustainable computing (+0.049); indistinguishable from zero for the others — and placebo level p-values are all above900.47. Bai–Perron break dates selected by BIC are scattered (F-06): 2022-01 and 2025-04 for quantum, 2022-02 for91networks, 2022-07 for sustainable computing, 2022-07 and 2024-07 for distributed computing and IR, 2022-11 and922025-01 for cloud (the only 2022-11 break), 2023-05 for architecture, 2023-09 for operating systems, 2023-11 for93databases, 2024-11 for accelerators, 2025-01 for edge computing. The reading that the corpus supports is a94field-wide acceleration of output from 2022–2023 onwards that the November 2022 date does not single out, superimposed95on a prior decline of the classical categories, so that mean reversion competes with any infrastructure-demand96explanation (F-39, OP-F13).9798### 2.3 Citation flows99100Aggregating `category_flows_yearly.csv` over citing years 2018–2022 and 2023–2026 (calendar years, not the 30 November101boundary), the systems categories remain strongly self-referential (F-07): in the post years the self-share is 84.9 %102for quantum computing, 72.4 % for networks, 59.2 % for IR, 57.7 % for architecture, 40.6 % for databases, 34.7 % for103accelerators, 33.1 % for cloud, 29.8 % for distributed computing, 24.4 % for edge computing, 16.0 % for operating104systems and 11.3 % for sustainable computing. Shares moved by less than four points between periods (F-08). Three105details matter for the synthesis (F-09): the `llm` category enters the seven largest destinations of IR references only106in the post years (3.0 %); operating systems now sends more references to computer architecture (18.8 %) than to107itself; edge computing and sustainable computing cite computer networks more than their own category in both periods,108which is as much a statement about where the classifier draws boundaries as about intellectual dependence.109110## 3. Computer architecture, accelerators and operating systems111112### 3.1 Architecture: the most cited work is machine-learning systems113114Once scientific software and misclassified records are removed, the post-2022 architecture landmarks are a coherent115set (F-37): *PyTorch 2* (ASPLOS 2024), whose abstract reports that TorchDynamo captures graphs by rewriting Python116bytecode and that TorchInductor yields geometric-mean speedups of 2.27× for inference and 1.41× for training on the117authors' model suite (F-16; a vendor-authored evaluation); *Splitwise* (ISCA 2024), which characterises LLM inference118as a compute-bound prompt phase and a memory-bound token-generation phase and places them on different machine119pools (F-14); *AWQ* (the GetMobile reprint of the MLSys 2024 paper), an activation-aware weight-quantisation method120with an on-device inference framework; the *AI and Memory Wall* perspective (IEEE Micro 2024), which states that peak121server FLOPS scaled at 3.0× per two years over twenty years against 1.6× for DRAM bandwidth and 1.4× for interconnect122bandwidth, and argues that decoder inference is bandwidth-bound (F-15); and a 2025 ACM Computing Surveys review of123deep-learning accelerators for heterogeneous HPC platforms (GPU, TPU, FPGA, ASIC, RISC-V, processing-in-memory,124neuromorphic). The baseline list adds the processing-in-memory primer and Ben-Nun and Hoefler's concurrency analysis125of distributed deep learning. The lineage is the one the bibliography already documents — the end of Dennard scaling126and the "golden age" of domain-specific architecture, TPU v1 and v4, GPU evolution — and the post-2022 landmarks are127its continuation under one workload. What is established is that the architecture community's most cited output of1282023–2025 concerns LLM training and serving; what is not established, from abstracts, is whether the memory-wall129diagnosis holds as the installed base changes (OP-F02).130131### 3.2 Accelerators: photonic and neuromorphic demonstrations, and a classification artefact132133The accelerator list is the most distorted by classification (F-12): at least ten of its twenty post-period records are134memristive or chaotic Hopfield-network circuit papers in *Integration*, *Neural Networks*, *IEEE TCYB* and *Chaos,135Solitons & Fractals*, a nonlinear-dynamics literature that the topic classifier maps to AI hardware. Digital136accelerators (TPU v4, GPUs) are absent from the list though present in the bibliography (F-43); they appear to be137classified under architecture. What remains is a photonic and neuromorphic thread: Loihi (IEEE Micro 2018), the138neuromorphic-computing perspective of *Nature Computational Science* (2022), the *Taichi* photonic chiplet139(*Science* 2024; the abstract reports 160 TOPS/W and on-chip 1,000-category classification), *Universal photonic140artificial intelligence acceleration* (*Nature* 2025; abstract unavailable), an optical-neural-network review141(*Light: Science & Applications* 2024), and 2026 *Nature Communications* papers on an electro-optic analog memory142(the abstract reports a >26× power saving in analysis relative to SRAM–DAC designs) and a hyper-multiplexed143matrix–matrix optical processor. These are device-level demonstrations with device-level efficiency metrics; the144abstracts do not report end-to-end energy including conversion and memory, and the review treats their system-level145impact as speculative (F-17, OP-F03). The category's level drop at 2022-12 (−0.113, p = 0.027) and late BIC break146(2024-11) do not suggest a language-model-driven surge in this literature.147148### 3.3 Operating systems: GPU resource management as the OS problem of the period149150Operating systems is the smallest of the eleven categories (1,363 primary works in 2025) and the one that declined151fastest before 2022 (−0.087 per year), recovering at 0.078 per year afterwards (F-35). Its post-period landmarks are152*PowerInfer* (SOSP 2024), a GPU–CPU hybrid engine that keeps "hot" neurons on a consumer GPU and computes "cold" ones153on the CPU, exploiting a power-law activation distribution; *Orion* (EuroSys 2024), which intercepts kernel launches154from several clients and schedules at operator granularity to limit interference; and *vAttention* (ASPLOS 2025),155which argues that PagedAttention's non-contiguous KV-cache layout imposes programming and performance overheads and156restores contiguity with CUDA virtual memory. Read together with vLLM (SOSP 2023, classified under distributed157computing), this is a live design controversy over memory abstractions for LLM serving (F-13, OP-F01), evaluated by158each system's own authors on particular hardware. The baseline landmarks — *Keystone* (EuroSys 2020, customisable TEEs159on RISC-V), *Nightcore* (ASPLOS 2021, sub-millisecond serverless), the Berkeley serverless view (preprint), persistent-160memory key-value stores — show the field's pre-2022 concerns (serverless, TEEs, persistent memory) that the landmark161lists no longer register after 2023. The BPF instruction-set RFC (2024) and an in-kernel aggregation preprint point to162kernel-extension technology as an infrastructure theme, on thin evidence (F-42). Operating systems is also the163category whose largest citation destination is now computer architecture (F-09), consistent with GPU-centric systems164work.165166## 4. Distributed, cloud and edge computing, and networks167168### 4.1 Distributed computing169170After the re-publication records and the control-theory papers are set aside, the post-period distributed-computing171landmarks are vLLM (SOSP 2023; the abstract reports 2–4× throughput at equal latency from near-zero KV-cache waste),172LOOM (a live race-bypassing system, recorded as a 2026 Columbia repository item) and OrbitalBrain (in-space ML173training on nanosatellites, OASIcs 2026). The baseline list, beyond the classics, contains a blockchain survey174duplicated under two OpenAlex identifiers, *Majority is not enough* (selfish mining) and a Korean consensus175pre-processing paper — blockchain was the distributed-computing growth topic of 2018–2020, and the post-period flows176still send 4.3 % of distributed-computing references to the blockchain category. The category's trend change at1772022-12 is among the largest (0.020) and its BIC breaks are 2022-07 and 2024-07. Nothing in the landmark set178supports a statement about consensus, transactions or fault tolerance in 2023–2026; the distributed-systems canon179of the bibliography (Paxos, PBFT, CAP, Spanner, Borg, Kubernetes, serverless) is cited but not renewed in the180most-cited slice.181182### 4.2 Cloud computing183184The post-period cloud landmarks split into three kinds (F-27, F-28). (i) LLMs for operations: *RCACopilot* (EuroSys1852024) matches incidents to handlers, aggregates diagnostics and predicts the root-cause category; the abstract reports186accuracy up to 0.766 on a year of Microsoft incidents, a proprietary evaluation (OP-F11). (ii) Reviews: deep187reinforcement learning for cloud scheduling (*Artificial Intelligence Review* 2024) and an MDPI systematic review of18830 studies on federated learning for cloud/edge security whose aggregate percentages (25 % privacy-risk reduction,18940 % better threat detection) are not interpretable without the primary studies (F-41). (iii) The 2026 slice, which190consists almost entirely of LLM-inference scheduling at smaller IEEE conferences: token-level GPU pooling,191contextual-bandit elastic scheduling for serverless multi-tenant inference, and carbon-aware prompt scheduling (CAPS,192trace-driven simulation, 26.8 % lower carbon per thousand tokens than round-robin as reported). These are provisional193landmarks (F-11) and single-testbed or simulation studies, but they show the direction of citation in 2026. The194baseline cloud landmarks were VM and task scheduling with metaheuristics, energy efficiency in data centres and195IoT–cloud integration; the continuity is scheduling, the discontinuity is the workload. Cloud computing's ITS level196change is zero (p = 0.936) while its trend change is 0.019 and its BIC breaks are 2022-11 and 2025-01 (F-29); it is197the only systems category with a selected break in the ChatGPT month, which the placebo result (0.231) does not allow198to be read as specific.199200### 4.3 Edge computing201202Edge computing is the category whose landmark composition changed most visibly and whose time series changed least203(F-30). Its baseline landmarks are blockchain–IoT integration papers (2018–2019), the *Edge Intelligence* survey204(*Proceedings of the IEEE* 2019) and COVID-era IoT application reviews; its post-period landmarks are surveys and205visions on language models at the edge: *Mobile Edge Intelligence for LLMs* (COMST 2025), *Toward Edge General206Intelligence with Agentic AI* (COMST 2026), *Edge–Cloud Collaborative Computing* (COMST 2026) and *AI Flow*207(2026), which proposes a device–edge–cloud framework for large models. The abstracts describe architectures,208taxonomies and challenges; none reports a measured deployment. Edge computing's trend change at 2022-12 is the209smallest of the eleven (0.004, p = 0.006) after a baseline CAGR of 0.133 that had already doubled the category's210share; its only BIC break is 2025-01. The LLM-at-the-edge literature is therefore, on this evidence, a change in211what the field writes surveys about rather than a measurable change in its output. Where inference should sit on the212device–edge–cloud continuum is an open empirical question (OP-F09).213214### 4.4 Computer networks215216Networks is the largest systems category and the one whose share of output fell most (0.118 to 0.067). Its most cited217post-2022 works are the 6G roadmap (COMST 2023) and the ISAC overview (JSAC 2022), both pre-dating or coinciding with218the window's start, the *Wireless Communications* textbook, the mobile-edge-LLM survey, and in 2026 tutorials on219large AI models and agentic AI for communications (JSAC 2026), a massive-IoT review and a pinching-antenna tutorial220(F-31). The positive level change at 2022-12 (+0.043, p = 0.021) has placebo p = 0.538 and the single BIC break is2212022-02; the series was already rising before ChatGPT. The deep-learning-for-wireless surveys of 2018–2019 show that222machine learning entered networking research well before the window; what the 2026 tutorials add is the vocabulary of223large models and agents, whose empirical content cannot be assessed from abstracts.224225## 5. Database systems and information retrieval226227### 5.1 Databases: data quality and vector search228229The database landmark list is the most contaminated by misclassification (§1.2). What remains is instructive (F-23,230F-40): the only engine-level item is vector search — the GPU similarity-search paper (IEEE TBD 2019) and *The Faiss231Library* (IEEE TBD 2025/26), whose abstract frames Faiss as the core of vector databases and describes its indexing,232compression and benchmarking; the other post-period landmarks are the *Data-centric AI* survey (CSUR 2025) and an233empirical study of six data-quality dimensions across nineteen ML algorithms (*Information Systems* 2025). The234baseline list contains BLOCKBENCH (TKDE 2018), *Datasheets for datasets* (CACM 2021) and a persistent B+-tree235dissertation. The bibliography's vector-DBMS surveys and the database community's position papers on LLMs236(*Proceedings of the VLDB Endowment* 2023; *SIGMOD Record* 2024) are not in the landmark set. The honest summary is237that the most cited database work of the period, as the corpus records it, is about data for machine learning and238about embedding search, not about transaction processing or query optimisation; whether vector search belongs in239dedicated systems or in extended relational engines is contested and untested at workload level (OP-F08). The240category's trend change (0.014) and BIC break (2023-11) are unremarkable among the eleven.241242### 5.2 Information retrieval: from graph recommendation to retrieval augmentation243244IR shows the clearest shift in what is cited (F-24). The 2018–2022 landmarks are graph-neural-network recommendation245— NGCF, KGAT, GraphRec, LightGCN (SIGIR 2020, which reports that feature transformation and non-linear activation246contribute little to collaborative filtering), self-supervised graph learning — and the GNN-recommendation surveys of247CSUR and ACM TORS. The 2023–2026 landmarks are RAG surveys (KDD 2024; CSUR 2026, which organises RAG into248pre-retrieval, retrieval, post-retrieval and generation; *Data Science and Engineering* 2026 on RAG for generated249content), surveys of LLMs for recommendation (TKDE 2024; *World Wide Web* 2024), and *C-Pack* (SIGIR 2024), which250releases training data, a Chinese embedding benchmark and the BGE embedding models. The methodological lineage is251continuous with 2020 (DPR, RAG, ColBERT); the landmark set contains no new retrieval architecture of 2023–2026, which252may reflect the citation lag of recent model papers as much as the field's output (F-25). In the corpus, IR has the253largest trend change at 2022-12 (0.023), a negative level change (−0.094, p = 0.019), BIC breaks at 2022-07 and2542024-07 (F-26), and is the only systems category whose post-period references reach the `llm` category in measurable255proportion (3.0 %). RAG evaluation — separating retrieval quality from parametric memory and from generation — has no256agreed protocol (OP-F07).257258## 6. Quantum computing259260Quantum computing is the category least connected to the rest (self-share 84.9 %), the fastest-growing after 2022261(CAGR 0.287) and the one whose level fell most clearly at 2022-12 (−0.087, p = 0.039) (F-22). Its landmarks are a262coherent experimental sequence that the window happens to contain (F-18): the NISQ framing (2018) and supremacy263experiment (2019) as baseline; then surface-code scaling from distance 3 to 5 on superconducting qubits (*Nature*2642023), the 280-physical-qubit neutral-atom logical processor (*Nature* 2024), the below-threshold distance-7265memories with logical error suppression factor Λ = 2.14 per distance step (*Nature* 2024), two-module trapped-ion266distributed computation by gate teleportation over an optical link (*Nature* 2025) and the 35-chip photonic *Aurora*267scale model (*Nature* 2025), which its authors describe as sub-performant (F-21). The 127-qubit "utility before fault268tolerance" experiment (*Nature* 2023) is contested: its abstract restricts the claim to circuit volumes beyond269brute-force classical simulation, and the classical-simulation responses that followed are not in the landmark set270and are not cited here because they are not in the verified bibliography (F-19, OP-F04). On the algorithmic side,271the barren-plateau review (*Nature Reviews Physics* 2025; abstract unavailable) and quantum-selected configuration272interaction (*Physical Review Research* 2026), which diagonalises classically in a subspace sampled from a noisy273device, mark the two responses to the trainability problem of variational algorithms (F-20, OP-F05). Qiskit's design274paper is a preprint; the IBM retrospective (*Journal of Supercomputing* 2025) is a narrative of one vendor's roadmap.275None of this literature has any connection with language models, and the corpus statistics for the category should276be read as the signature of a field growing on its own momentum.277278## 7. Sustainable computing279280Sustainable computing is small (1,639 primary works in 2025), grew at 0.233 per year after 2022, has a positive level281change at 2022-12 (+0.049, p = 0.048; placebo p = 0.754) and cites computer networks more than itself (F-34). Its282landmarks form a measurement literature on the footprint of machine learning (F-32): carbon of large-model training283(Patterson et al. 2021, preprint; BLOOM's 24.7–50.5 t CO2e, preprint), the carbon intensity of cloud instances284(FAccT 2022), generative-AI inference carbon today and in 2035 (HotCarbon 2023 workshop), the water footprint of AI285(preprint 2023, CACM 2025), green-AI and green-federated-learning reviews (2024–2025), rebound effects and Jevons'286paradox (FAccT 2025), data-centre carbon and water disclosure gaps (*Patterns* 2025/26), grid impacts of AI data287centres (*Nexus* 2026) and *TokenPowerBench* (AAAI 2026), a benchmark attributing inference energy to prefill and288decode. Several of the most cited items are preprints or workshop papers, and the quantitative statements they289carry (including industry figures quoted in abstracts, such as inference exceeding 90 % of LLM power) are not290independently verifiable from the corpus (F-33). The field's own diagnosis — that operators do not disclose AI-specific291energy and water, so fleet-scale footprints are modelled rather than measured — is the central open problem (OP-F10).292293## 8. Cross-cutting synthesis294295Three statements are established by the corpus. First, the systems categories accelerated after 2022 as a group,296and the acceleration is not specific to the ChatGPT date (F-01, F-02, F-06). Second, the classical systems categories297had been declining in absolute and relative terms since 2018 and recovered in absolute terms only; their share of298computer science continued to fall (F-04, F-05). Third, citation flows among systems categories were stable; the299only measurable new destination is IR's references to the `llm` category (F-07–F-09).300301Two statements are established by the metadata of the most cited works. The most cited post-2022 architecture,302operating-systems, distributed and cloud papers are LLM training and serving systems — compilation, phase splitting,303KV-cache management, GPU sharing, consumer-GPU inference, inference scheduling — and the most cited IR, edge and304networking papers are surveys and tutorials about language models (F-24, F-30, F-31, F-37, F-38). Quantum computing305and sustainable computing follow their own trajectories, the former experimental, the latter a measurement306literature on AI's footprint (F-22, F-32).307308What cannot be established from this evidence: that LLM-infrastructure research explains the volume recovery (its share309of each category's output is not measured; OP-F13); that any of the systems designs discussed is superior to its310competitors (each is evaluated by its authors; OP-F01, OP-F02); that photonic or neuromorphic efficiency figures hold at311system level (OP-F03); that pre-fault-tolerant quantum processors deliver utility (OP-F04). Finally, the landmark312lists themselves carry a classification-noise rate of roughly one record in five (OP-F12), which bounds the precision313of every category-level statement above and should be reported alongside the results of Part IX.314