1claim_id,claim,claim_type,evidence_type,supporting_keys,contradicting_keys,corpus_artifact,confidence,verification_status,notes2G-01,"The automatically selected landmark candidates of the theory and software_pl areas (150 works) are dominated by non-research records: only 29 are in-scope research contributions, while 33 are misclassified, 27 are theses, 24 are re-deposited classics with a recent record date and 22 are textbooks or chapters.",established,metadata_only,,,literature_synthesis/agent_G/contributions.csv (column verdict); manuscript/tables/tab_theory_landmark_audit.tex,0.9,human-verified,"Verdicts assigned by Agent G from title/venue/type metadata and abstracts; the verdict scheme is documented in make_contributions.py. Human-verified here means verified by the specialist agent, not by the PI (checkpoint H-3 pending)."3G-02,"Seven candidate records carry OpenAlex identifiers that, resolved live on 2026-11-03, return a different and far more cited work (e.g. the record titled 'Exploiting Generative AI to Scale up Intelligent Tutoring Systems', DOI 10.4230/lipics.itp.2023.19, resolves to 'Attention Is All You Need'), and their c36 values match the other work's citation count.",established,metadata_only,,,literature_synthesis/agent_G/verification_lookups_raw.txt; literature_synthesis/agent_G/contributions.csv (column id_resolves_to_other_work),0.95,machine-verified,"Live OpenAlex lookups of W4385245566, W2964015378, W4293846201, W4311415873, W4294558607, W3177813494, W4297781990. Mechanism (identifier merge or reuse between snapshot and live database) is inferred, not observed."4G-03,"The computational_complexity category is the least precise of the six: 15 of its 25 candidates are works from psychology, management, education or metrology that carry the phrase 'complexity' or 'complexity theory'.",established,abstract,,,literature_synthesis/agent_G/contributions.csv (primary_category = computational_complexity),0.9,human-verified,Consistent with the lexical signals of taxonomy/ontology.yaml; per-category precision from the validation sample (taxonomy/validation) was not available to Agent G.5G-04,"Metaphor-based metaheuristic optimisers (arithmetic, aquila, lemming, snake, langur, starfish) account for six of the 25 algorithms candidates and for most of the post-2022 highly cited algorithm 'articles'; the genre is contested in the operational-research literature.",established,abstract,abualigah2021aoa;abualigah2021aquila;sorensen2013metaphor;aranha2021metaphor;camachovillalon2022exposing,,literature_synthesis/agent_G/contributions.csv (verdict = metaheuristic),0.85,human-verified,"The critique papers argue that many such algorithms re-describe existing methods; this review reports the critique, it does not re-evaluate the algorithms."6G-05,"Across the six categories the multi-label count grew faster in 2023-2025 than in 2018-2022 (e.g. software engineering 25.5% vs 0.9% CAGR, computational complexity 22.7% vs 1.9%), but the corpus-wide multi-label CAGR also rose from 6.1% to 26.0%, so the category-level acceleration is largely the corpus-wide acceleration.",established,corpus_statistic,,,results/graph/category_counts_yearly.csv via macros \ThyCagrMultilabel*; tab_theory_summary,0.85,machine-verified,Post-period CAGR covers two years (2023 to 2025) and is sensitive to indexing lag (Section sec:m-truncation).7G-06,"Measured by primary-label share, all six categories lost ground between 2018 and 2025 (algorithms 5.83% to 3.79%, software engineering 5.81% to 4.07%, formal verification 1.06% to 0.52%, programming languages 0.88% to 0.47%).",established,corpus_statistic,,,results/graph/category_counts_yearly.csv via \ThySharePrimary*Pct,0.9,machine-verified,"Relative decline reflects faster growth elsewhere (AI categories), not absolute contraction."8G-07,"The results/bibliometrics/cagr_by_category.csv file reports counts for the six categories (e.g. algorithms n_2018 = 9,716) that are about three times smaller than the primary counts of results/graph/category_counts_yearly.csv (29,148) and therefore predates the current classification; its CAGR values should not be quoted for these categories until regenerated.",established,corpus_statistic,,,results/bibliometrics/cagr_by_category.csv vs results/graph/category_counts_yearly.csv (macros \ThyCagrfile* vs \ThyCount*),0.95,machine-verified,"bibliometrics/04_diversity_cagr.py reads category_counts_yearly.csv, so re-running it would resolve the discrepancy; flagged to the director."9G-08,"Interrupted-time-series slope changes at 2022-12 are positive and HAC-significant for all six categories (1.0% to 1.8% per month), but the placebo test rejects at the 5% level for none of them (placebo slope p from 0.062 to 0.431), so the post-2022 slope change is not distinguishable from pseudo-events in the baseline.",established,corpus_statistic,,,"results/econometrics/breaks.csv rows cat_algorithms, cat_computational_complexity, cat_programming_languages, cat_compilers, cat_formal_verification, cat_software_engineering (columns chatgpt_slope, chatgpt_slope_p, placebo_slope_p)",0.9,machine-verified,No causal reading. Algorithms and computational complexity are the closest to the 5% threshold (0.062).10G-09,"Software engineering is the only one of the six categories with a positive level change at 2022-12 (4.8%, p = 0.124) and a significant positive level change at the GPT-4 date (0.102, p = 0.003); the other five show negative or null level changes at 2022-12.",established,corpus_statistic,,,"results/econometrics/breaks.csv (chatgpt_level, gpt4_level, gpt4_level_p)",0.85,machine-verified,Level changes partly reflect seasonality and the dating of conference proceedings; the placebo level p for software engineering is 0.862.11G-10,"Co-classification with the LLM category rose between 2022 and 2025 in every theory/software category, most strongly in software engineering (4.0% to 21.2% of its works), programming languages (4.2% to 18.3%) and compilers (3.1% to 14.9%), and least in algorithms and computational complexity (1.3% to 5.3%).",established,corpus_statistic,,,"results/graph/category_cooccurrence_yearly.csv and category_counts_yearly.csv via \ThyCooc*LlmTwoZeroTwo{Two,Five}Pct; tab_theory_llm_cooccurrence",0.9,machine-verified,"Co-classification is a lexical/topic signal (Part IV), not evidence of method adoption."12G-11,"Reference flows from the six categories to the LLM category remain small in the post period (software engineering 2.5%, computational complexity 1.8%, compilers 1.2%, programming languages 0.7%, formal verification 0.4%, algorithms 0.3% of references), and within-category self-citation shares changed little (algorithms 65.1% to 65.0%, formal verification 49.9% to 47.0%).",established,corpus_statistic,,,results/graph/category_flows_yearly.csv aggregated by period via \ThyFlow*; \ThyInflow*,0.85,machine-verified,"Flows are computed on the categorised subset of references (Part VIII reference coverage); the LLM category is young, which depresses its share of cited works mechanically."13G-12,"Computational complexity shows the largest drop in self-citation share (30.0% to 20.6%) and the largest rise in references to HCI (3.0% to 6.7%), consistent with the category's contamination by social-science 'complexity' works rather than with a change in theoretical research.",contested,corpus_statistic,,,results/graph/category_flows_yearly.csv via \ThyFlowComputationalComplexityTo*; literature_synthesis/agent_G/contributions.csv,0.6,machine-verified,"Interpretation; the flow statistics are machine-verified, the attribution to contamination is the specialist's reading of the candidate list."14G-13,"Formal verification receives a growing share of its inbound citations from computer vision (6.7% baseline to 13.8% post), consistent with the neural-network verification literature (Reluplex, Marabou, DeepPoly, VNN-COMP) being cited by vision robustness work.",contested,corpus_statistic,katz2017reluplex;katz2019marabou;singh2019deeppoly;brix2023vnncomp,,"results/graph/category_flows_yearly.csv via \ThyInflowFormalVerificationFromComputerVision{Baseline,Post}Pct",0.6,machine-verified,The attribution to neural-network verification is not tested on the flow data (no paper-level breakdown).15G-14,"The major theoretical results of the window (almost-linear max-flow, negative-weight SSSP in near-linear time, Kelley-Meka 3-AP bounds, Williams's time-space simulation, the directed SSSP sorting barrier, optimal open-addressing bounds) do not appear among the field-normalised landmark candidates, which favour generic-title book chapters and re-deposited classics; the landmark selection is therefore not a valid instrument for identifying theoretical breakthroughs.",established,metadata_only,chen2022maxflow;bernstein2022negative;kelley2023threeap;williams2025sqrtspace;duan2025sorting;farachcolton2024openaddressing,,literature_synthesis/landmarks/theory_landmarks.csv; literature_synthesis/agent_G/contributions.csv,0.85,human-verified,"The listed results are known to the specialist from the verified bibliographic records (titles, venues); their content is not reviewed here beyond the abstracts."16G-15,Petri-net reachability was settled in the baseline period: a non-elementary lower bound (STOC 2019) followed by matching Ackermann-completeness (FOCS 2021); the STOC 2019 paper is one of only two in-scope research candidates in the computational_complexity list.,established,abstract,czerwinski2019petri;czerwinski2021ackermann;leroux2021notprimitive,,literature_synthesis/landmarks/theory_abstracts.jsonl (W2900179285),0.9,machine-verified,Abstract of the STOC paper reports the lower bound; the 2021 results are cited from their verified records.17G-16,"LLM-assisted program analysis and compiler testing entered the top venues of the programming-languages community in 2024 (LLift and WhiteFox at OOPSLA), and LLM-based compiler optimisation models were released in 2025 (LLM Compiler at CC); these works report found bugs or optimisation recovery rates but, per their abstracts, no formal guarantees.",established,abstract,li2024llift;yang2024whitefox;cummins2025llmcompiler;munley2024llm4vv;xia2024fuzz4all,lopes2021alive2;yang2011csmith,"literature_synthesis/landmarks/software_pl_abstracts.jsonl (W4396242417, W4403223258, W4407942537)",0.85,machine-verified,Contrast with translation validation (Alive2) and Csmith-style differential testing is the specialist's framing.18G-17,"Machine-learning-guided automated theorem proving reached about 60% of Mizar theorems in the hammer setting (ITP 2023), and LLM-based provers (Baldur, LeanDojo, DeepSeek-Prover) report whole-proof generation on Lean/Isabelle benchmarks; the LeanDojo and DeepSeek-Prover results are preprints and the benchmark-level figures cannot be compared across systems.",established,abstract,jakubuv2023mizar;first2023baldur;yang2023leandojo;xin2024deepseekprover;yang2024formalmath,,literature_synthesis/landmarks/software_pl_abstracts.jsonl (W4385245566),0.8,machine-verified,The MizAR abstract is attached in the corpus to a wrong title (claim G-02).19G-18,"Field evidence on AI coding assistants is conflicting: three randomised field experiments report a 26.08% (s.e. 10.3%) increase in completed tasks (Management Science 2026), while a 2025 randomised trial preprint with experienced open-source developers reports slower completion; the populations (enterprise developers vs expert maintainers of mature repositories) and outcomes (tasks completed vs time per issue) differ.",contested,abstract,cui2026genaiwork;peng2023copilot;ziegler2022copilot,becker2025metr,literature_synthesis/landmarks/software_pl_abstracts.jsonl (W7131867455),0.8,machine-verified,The METR study is a preprint; the reconciliation by population/outcome is the specialist's reading of the abstracts.20G-19,"Security evaluations of LLM-generated or LLM-classified code (Copilot vulnerability rates, user studies, CWE classification by open models) consistently report high detection or completion rates paired with poor fine-grained security reasoning.",established,abstract,pearce2022asleep;perry2023insecure;gasparlopes2025vuln,,literature_synthesis/landmarks/software_pl_abstracts.jsonl (W7110046239),0.75,machine-verified,Three studies with different designs; 'consistently' refers to these three only.21G-20,"The compilers category is small (0.09% of primary labels in 2025) and its candidate list contains ten misclassified works captured by the tokens 'just-in-time', 'compiler', 'compiling' and 'polyhedral' used outside computing.",established,abstract,,,literature_synthesis/agent_G/contributions.csv (primary_category = compilers); \ThySharePrimaryCompilersTwoZeroTwoFivePct,0.9,human-verified,Lexical mechanism inferred from titles; the ontology's keyword regexes were not re-run by Agent G.22G-21,Expressivity results for transformers (log-precision transformers lie in uniform TC0; survey of formal-language results) are the main post-2022 contribution of complexity theory to the study of language models visible in the verified bibliography; they bound what a single forward pass can compute and motivate chain-of-thought as a route to greater power.,speculative,metadata_only,merrill2023parallelism;strobl2024survey,,manuscript/references_part07a.bib,0.6,unverified,Not from the corpus landmark set; the relevance claim is the specialist's and the content is known only from verified titles and abstracts.23G-22,"Algorithms with predictions (learning-augmented algorithms) is a baseline-period research programme (ski rental and scheduling with ML predictions; competitive caching with advice; CACM overview 2022) that the corpus records through a 2024 arXiv re-posting of the 2018 paper, illustrating how re-depositions distort the dating of landmarks.",established,abstract,mitzenmacher2022predictions;lykouris2021caching,,literature_synthesis/landmarks/theory_abstracts.jsonl (W2891784792),0.85,machine-verified,The OpenAlex record W2891784792 carries publication year 2024 and DOI 10.48550/arxiv.2407.17712.24G-23,"AI-discovered algorithms (AlphaTensor matrix multiplication, AlphaDev sorting routines, FunSearch, AlphaEvolve) are reported in Nature or as preprints by their developers; the improvements are on specific instances (small matrix sizes, short sorting networks, cap-set constructions) and no independent replication appears in the verified bibliography.",contested,metadata_only,fawzi2022alphatensor;mankowitz2023alphadev;romeraparedes2024funsearch;novikov2025alphaevolve,,manuscript/references_part07a.bib; manuscript/references.bib,0.7,unverified,"AlphaEvolve is a preprint; replication status is asserted from the absence of such works in the bibliography, not from a search of the corpus."25G-24,"Verified-systems research continued in the window with Verus (SOSP 2024), a Rust-based verification framework, extending the CompCert/seL4/RustBelt lineage; the corpus candidate list records this lineage only through re-deposited UNSW theses dated 2022.",established,metadata_only,lattuada2024verus;leroy2009compcert;klein2009sel4;jung2017rustbelt;jung2021saferust,,"literature_synthesis/agent_G/contributions.csv (W2327217190, W2491926874, W630549275)",0.8,machine-verified,Verus content known from its verified title and venue only.26