1claim_id,claim,claim_type,evidence_type,supporting_keys,contradicting_keys,corpus_artifact,confidence,verification_status,notes2E-01,"The automatic AI/ML landmark selection returned 234 candidates (33 LLM, 26 AI safety, 25 in each other category), 72 dated 2026, 136 typed article, 28 preprint, 19 book-chapter, 11 dissertation.",established,corpus_statistic,,,"literature_synthesis/landmarks/ai_ml_landmarks.csv (columns primary_category, year, type); selection_report.json by_area.ai_ml",0.99,machine-verified,Counts recomputed by make_contributions.py.3E-02,"44 of the 234 AI/ML landmark records (18.8 %) are unreliable: 13 identifier mismatches, 3 abstract mismatches, 11 implausible years, 7 normalisation artefacts, 10 off-scope classifications.",established,metadata_only,,,literature_synthesis/agent_E/contributions.csv (column data_quality_flag); manuscript/tables/tab_ai_landmark_flags.tex,0.9,human-verified,"Judged record by record by Agent E from identifiers, citation counts, venue, title and abstract; needs Agent N audit of the Phase 1 linkage."4E-03,Identifier and abstract mismatches concentrate in the LLM category (8 of 33) and misdated classics in ML theory and DL architectures (4 each).,established,metadata_only,,,literature_synthesis/agent_E/contributions.csv; tab_ai_landmark_flags,0.9,human-verified,5E-04,The NCS36 ranking should be re-run with document-type-specific reference sets and the 2026 cohort excluded; the clean AI landmark count is at most 190.,speculative,corpus_statistic,,,literature_synthesis/agent_E/contributions.csv,0.6,unverified,Remedy proposed to the director; not implemented.6E-05,"Of the 234 landmark candidates, 84 are surveys, 67 methods, 33 empirical, 29 theory/position, 17 engineering and 4 benchmarks (judged from abstract for 190, metadata for 44).",established,abstract,,,literature_synthesis/agent_E/contributions.csv (contribution_type); tab_ai_contribution_types,0.85,human-verified,Contribution typing is a judgement; abstracts only.7E-06,The citation-ranked top of the LLM and generative-AI categories is dominated by education and commentary papers on ChatGPT rather than by the technical papers introducing the models.,established,metadata_only,kasneci2023chatgptgood;lo2023rapidreview;rudolph2023assessment;chan2023students;baidooanu2023chatgptedu;dwivedi2023sowhat;vandis2023priorities;lim2023ragnarok,,literature_synthesis/landmarks/ai_ml_landmarks.csv (rows with primary_category llm or generative_ai),0.9,human-verified,Property of field-normalised ranking inside a multi-label category spanning education venues.8E-07,Post-period CAGR (2023–2025) versus baseline (2018–2022): LLM 1.085 vs 0.189; generative AI 0.870 vs 0.296; multimodal 0.869 vs 0.284; foundation models 0.802 vs 0.661; AI safety 0.647 vs 0.294; RL 0.252 vs 0.208; ML theory 0.197 vs 0.104; DL architectures 0.155 vs 0.167; AI agents 0.329 vs −0.045.,established,corpus_statistic,,,"results/bibliometrics/cagr_by_category.csv (cagr_baseline, cagr_post); tab_ai_growth",0.95,machine-verified,File regenerated by Phase 4 on 2026-11-03 evening; values as of the table build.9E-08,The LLM category's share of eligible works rose from 0.33 % in 2018 to 5.06 % in 2025.,established,corpus_statistic,,,"results/bibliometrics/cagr_by_category.csv (share_2018, share_2025); tab_ai_growth",0.95,machine-verified,10E-09,For the LLM series the ITS at 2022-12 gives a level change of 0.106 (HAC p 0.075) and a slope change of 0.034 per month (p ~1e-60); placebo p 0.938 (level) and 0.200 (slope); Bai–Perron selects one break at 2022-07; PELT one change point at 2023-10.,established,corpus_statistic,,,results/econometrics/breaks.csv row cat_llm; its_results.json series.cat_llm; macros ItsSeriesCatLlm*; tab_ai_breaks,0.95,machine-verified,11E-10,"Multimodal, generative AI, agents, RL, ML theory and AI safety show significant positive slope changes at 2022-12 with no positive level jump (agents, RL and AI safety have negative level changes); DL architectures shows no significant slope change (−0.002, p 0.136).",established,corpus_statistic,,,results/econometrics/breaks.csv rows cat_*; tab_ai_breaks,0.95,machine-verified,12E-11,"Level effects at the GPT-3 (2020-06) and DALL-E 2 (2022-05) event dates are negative for all nine AI categories, while those at GPT-4 (2023-04) and o1 (2024-10) are positive and significant for LLM, multimodal and generative AI.",established,corpus_statistic,,,"results/econometrics/breaks.csv (gpt3_level, dalle2_level, gpt4_level, o1_level and p columns); tab_ai_altdates",0.9,machine-verified,Each alternative date is a separate ITS fit.13E-12,The post-2022 acceleration of AI publication is a trend change already under way in mid-2022 that steepened in 2023–2024; the monthly series cannot attribute it to a single product release.,established,corpus_statistic,,,"results/econometrics/breaks.csv; its_results.json (placebo, bai_perron, pelt)",0.85,machine-verified,Interpretive statement about the estimates; no causal claim.14E-13,"LLM works directed 69.6 % of in-corpus references to NLP in 2019 and 68.6 % in 2022, but 34.2 % in 2025, when 28.4 % went to the LLM category itself, 4.3 % to HCI and 3.4 % to software engineering.",established,corpus_statistic,,,results/graph/category_flows_yearly.csv (citing_cat llm); tab_ai_flows,0.95,machine-verified,15E-14,Generative-AI references to computer vision fell from 57.9 % (2022) to 33.8 % (2025) while references to LLM rose to 13.7 % and to HCI to 6.6 %; AI-safety references to LLM rose from 0.1 % (2019) to 10.0 % (2025) and to HCI from 9.1 % to 11.5 %.,established,corpus_statistic,,,results/graph/category_flows_yearly.csv; tab_ai_flows,0.95,machine-verified,16E-15,Reinforcement-learning references to robotics rose from 14.4 % (2019) to 23.1 % (2025) while references to LLM stayed below 1 %; ML theory and DL architectures flows are essentially unchanged.,established,corpus_statistic,,,results/graph/category_flows_yearly.csv; tab_ai_flows,0.95,machine-verified,17E-16,"The LLM category has become a citation hub for generative AI, safety and agents while remaining peripheral to RL, theory and architectures; this is consistent with but does not establish the foundation-model convergence thesis.",contested,corpus_statistic,bommasani2021foundation,,results/graph/category_flows_yearly.csv; tab_ai_flows,0.7,machine-verified,Category-level flows only; RQ5 answered at category granularity.18E-17,"Known-set lead times: large language models 34 months, reasoning models 29, RAG 12, vision-language models 9, diffusion models 9, LLM agents 6, state-space models 6, MoE 5, speculative decoding 5, AI code generation 10, mechanistic interpretability −18; median 7.",established,corpus_statistic,,,results/emerging/known_set_lead_times.csv (lead_months); macros Known*LeadMonths and KnownSetMedianLeadMonths,0.95,machine-verified,Long LLM lead reflects pre-2020 'language model' matches by the regex.19E-18,The LLaMA abstract reports that a 13B model trained only on publicly available data outperforms GPT-3 (175B) on most benchmarks; the paper is a preprint.,established,abstract,touvron2023llama;brown2020gpt3;hoffmann2022chinchilla;kaplan2020scaling,,ai_ml_abstracts.jsonl work 4322718191,0.9,machine-verified,Claim is the abstract's; not independently replicated here.20E-19,"Domain-specific foundation models (pathology UNI/CONCH/CHIEF/PLIP/TITAN, radiology RadFM, dermatology PanDerm, brain MRI BrainIAC, Earth system Aurora, tabular TabPFN, time series, cognition Centaur) are the most visible post-2022 development in the landmark set; their abstracts report outperforming task-specific baselines on multi-task benchmarks.",established,abstract,chen2024uni;lu2024conch;wang2024chief;huang2023plip;ding2025titan;wu2025radfm;yan2025panderm;tak2026brainiac;bodnar2025aurora;hollmann2025tabpfn;jin2023timellm;liu2024timesurl;binz2025centaur,,ai_ml_landmarks.csv rows with primary_category foundation_models / multimodal_ai,0.85,machine-verified,"Abstract-level claims by the authors; several records (UNI, CONCH, CHIEF, PLIP) have no abstract in the file."21E-20,"Parameter-efficient adaptation (LoRA, QLoRA, LLaMA-Adapter, PEFT surveys) is where the open-weight ecosystem's practical growth concentrated; QLoRA's '99.3 % of ChatGPT' is a judge-based measurement.",established,abstract,hu2021lora;dettmers2023qlora;zhang2023llamaadapter;han2024peftsurvey,,ai_ml_abstracts.jsonl works 7133233472 and 4361229539,0.8,machine-verified,22E-21,"Whether LLM capabilities are 'emergent' (discontinuous in scale) is contested: Wei et al. report discontinuities, Schaeffer et al. show they largely disappear under continuous metrics; both are preprints.",contested,metadata_only,wei2022emergent,schaeffer2023mirage,,0.9,machine-verified,Known disagreement; not adjudicable from the corpus.23E-22,"'Sparks of AGI' is a widely cited (c36 = 1,420 in the landmark file) but non-peer-reviewed, unreplicated position paper based on qualitative experiments on an early GPT-4.",established,abstract,bubeck2023sparks,bender2021parrots,ai_ml_landmarks.csv work 4360836968,0.9,machine-verified,24E-23,The DeepSeek-R1 Nature paper reports that reasoning can be incentivised through pure reinforcement learning without human-labelled reasoning trajectories; generality beyond verifiable domains is untested in the corpus.,established,abstract,deepseek2025r1nature;deepseek2025r1;openai2024o1,,ai_ml_abstracts.jsonl work 4414281281,0.85,machine-verified,Peer-reviewed version cited; the preprint is also in references.bib.25E-24,"Domain foundation models displacing task-specific pipelines in regulated domains is speculative; reviews list data, privacy and evaluation obstacles unresolved in the landmark abstracts.",speculative,abstract,paschali2025fmradiology;zhang2026datacentric,,ai_ml_abstracts.jsonl works 4407135816 and 4390690043,0.6,machine-verified,26E-25,RLHF and DPO are established post-training methods whose open problems are documented by Casper et al.,established,metadata_only,christiano2017rlhf;ouyang2022instructgpt;bai2022hh;bai2022constitutional;rafailov2023dpo;casper2023rlhfproblems,,,0.9,machine-verified,DPO record 7133208539 has no abstract in the landmark file.27E-26,"Current alignment is brittle to optimisation pressure: universal adversarial suffixes, jailbreak analyses and PAIR (semantic jailbreaks with black-box access in about twenty queries, per the abstract).",established,abstract,zou2023universal;wei2023jailbroken;chao2025pair,,ai_ml_abstracts.jsonl work 4410609100,0.9,machine-verified,Two of three are preprints; PAIR is peer reviewed (IEEE SaTML 2025).28E-27,Hallucination remains an unsolved measurement problem; Med-PaLM built MultiMedQA and a human evaluation framework because automated metrics were insufficient for clinical answers.,established,abstract,ji2023hallucination;huang2025hallucination;singhal2023clinical,,ai_ml_abstracts.jsonl work 4384071683,0.85,machine-verified,29E-28,"Vision-language models of the window (Flamingo, BLIP-2, InstructBLIP, LLaVA, mPLUG-Owl2) share the design of a frozen or lightly adapted LLM fed by a visual encoder through a small bridge module.",established,abstract,radford2021clip;dosovitskiy2021vit;alayrac2022flamingo;li2023blip2;dai2023instructblip;liu2023llava;ye2024mplugowl2;kirillov2023sam;schuhmann2022laion5b,,ai_ml_abstracts.jsonl works 4318718936 and 4402727669,0.85,machine-verified,"BLIP-2, InstructBLIP, LLaVA are preprint records."30E-29,"Diffusion generation (DDPM, latent diffusion, Imagen, DALL-E 2, classifier-free guidance, textual inversion, T2I-Adapter, DiT, SD3) forms the generative lineage; MatterGen is a peer-reviewed scientific application reporting stable, diverse inorganic materials.",established,abstract,ho2020ddpm;rombach2022ldm;saharia2022imagen;ramesh2022dalle2;ho2022cfg;gal2022textualinversion;mou2024t2iadapter;peebles2023dit;esser2024sd3;zeni2025mattergen,,ai_ml_abstracts.jsonl works 4288099666 4289785095 4393148714 4406472463,0.85,machine-verified,31E-30,"In an evaluation of eight multimodal foundation models on 1,090 medical cases, image predictions were largely driven by text and mild textual suggestions of a wrong diagnosis degraded image classification.",established,abstract,buckley2026textexploit,,ai_ml_abstracts.jsonl work 7164501397,0.85,machine-verified,Single study; abstract-level.32E-31,"The generative-AI category's citation-ranked landmarks are overwhelmingly education, policy and adoption studies (Bastani et al. preprint: GPT-4 access harmed unassisted exam performance; Bick et al.: 45 % of US adults 18–64 using generative AI by late 2024).",established,abstract,chan2023policyframework;xia2024assessmentreview;bittle2025integrity;mcdonald2025hepolicies;bastani2024harmlearning;bick2026rapidadoption,,ai_ml_landmarks.csv rows with primary_category generative_ai,0.9,machine-verified,"Bastani et al. is an SSRN preprint without abstract in the file; the exam-performance statement comes from the title and the published abstract elsewhere, not from the corpus file, and should be checked by Agent K."33E-32,"Agent benchmarks (SWE-bench, WebArena, AgentBench) are preprints; the only peer-reviewed agent benchmark in the landmark set is the clinical evaluation of OpenManus and Manus.",established,abstract,jimenez2024swebench;zhou2023webarena;liu2023agentbench;liu2026clinicalagents,,ai_ml_abstracts.jsonl work 7130481397,0.85,machine-verified,34E-33,"Science agents are the strongest peer-reviewed agent evidence: Virtual Lab (nanobodies with experimental validation), Robin (hypothesis generation and data analysis), DIVE (figure data extraction), AgentSociety (social simulation).",established,abstract,swanson2025virtuallab;ghareeb2026robin;zhang2026dive;piao2026agentsociety,lu2024aiscientist;si2024novelideas,ai_ml_abstracts.jsonl works 7161734976 4414991289 4407570917,0.8,machine-verified,Virtual Lab record has no abstract in the file (metadata only).35E-34,"The 'agentic AI' literature produced at least six taxonomies in 2025–2026, most of them reviews of reviews; the landmark set contains six of them and one benchmark.",established,abstract,acharya2025agenticsurvey;sapkota2026agentsvsagentic;bandi2025agenticreview;abouali2025agenticparadigms;allmendinger2026multiagentai;collaco2026agentichealthcare;cheng2026agentsdefinitions,,ai_ml_landmarks.csv rows with primary_category ai_agents,0.9,machine-verified,36E-35,"The ai_agents category mixes classical multi-agent systems and control theory with LLM agents, which explains its negative baseline CAGR, the negative 2022-12 level change and the second Bai–Perron break at 2024-07.",contested,corpus_statistic,,,contributions.csv off_scope rows; breaks.csv row cat_ai_agents; cagr_by_category.csv,0.7,machine-verified,Explanation is inferential.37E-36,"The deep-RL canon of the baseline period is intact in the landmark set; the single post-2022 method landmark is DreamerV3, whose abstract reports one configuration mastering over 150 tasks through world models.",established,abstract,hessel2018rainbow;haarnoja2018sac;fujimoto2018td3;silver2018alphazero;vinyals2019alphastar;yu2021mappo;henderson2018drlmatters;burda2018curiosity;hafner2025dreamerv3;zhang2021marloverview,,ai_ml_abstracts.jsonl work 4409147637,0.85,machine-verified,38E-37,RL's most consequential post-2022 role (reward modelling and policy optimisation for LLMs) is invisible in the RL category because those papers are classified as LLM work; a claim that RL declined would be a taxonomy artefact.,contested,corpus_statistic,schulman2017ppo;rafailov2023dpo;deepseek2025r1nature,,results/graph/category_flows_yearly.csv (reinforcement_learning -> llm < 1 %),0.7,machine-verified,39E-38,"The ML-theory landmark set contains no learning-theory result from the window; it consists of textbooks, evaluation metrics, interpretability, optimisation and transfer-learning surveys and misdated 2013–2019 bandit papers.",established,metadata_only,chicco2020mcc;rudin2019stopexplaining;lipton2018mythos;lundberg2020treeexplainer;bottou2018optimization;zhuang2021transfersurvey,,ai_ml_landmarks.csv rows with primary_category ml_theory; contributions.csv,0.9,human-verified,40E-39,Zeng et al. report that simple linear models outperform Transformer-based models (Informer and successors) on long-term time-series forecasting benchmarks; the finding is contested within the field.,contested,abstract,zeng2023transformerstsf,zhou2021informer;wen2023tstransformers,ai_ml_abstracts.jsonl work 4382203079,0.8,machine-verified,41E-40,"Architecture research moved to efficiency (FlashAttention, MoE, quantisation, speculative decoding, state-space models, KAN); the detector finds short leads for state-space models (6), MoE (5) and speculative decoding (5).",established,corpus_statistic,dao2022flashattention;fedus2022switch;dettmers2022int8;frantar2023gptq;leviathan2023speculative;gholami2022quantization;gu2023mamba;liu2024kan,,results/emerging/known_set_lead_times.csv,0.85,machine-verified,Mamba and KAN are preprints.42E-41,"The SuperGLUE abstract records in 2019 that GLUE performance surpassed non-expert humans, the first instance of the saturation cycle that repeats for later benchmarks; the landmark set contains four benchmark contributions in 234.",established,abstract,wang2019superglue;wang2018glue;hendrycks2021mmlu;srivastava2022bigbench;liang2022helm;sainz2023contamination;dehghani2021lottery,,ai_ml_abstracts.jsonl work 2990704537; contributions.csv,0.85,machine-verified,43E-42,"Risk taxonomies matured from the concrete-problems agenda to the AI Risk Repository (1,725 risks from 74 taxonomies per the abstract) and the International AI Safety Report.",established,abstract,amodei2016concrete;hendrycks2021unsolved;slattery2026riskrepository;bengio2025safetyreport;bengio2024extremerisks,,ai_ml_abstracts.jsonl work 4402698709,0.9,machine-verified,44E-43,Peer-reviewed empirical harms: medical LLMs vulnerable to data poisoning at 0.001 % of tokens; four chatbots gave unsafe answers in physician-led red-teaming; LLMs express stigma as replacement therapists.,established,abstract,alber2025poisoning;draelos2026unsafeanswers;moore2025stigma,,ai_ml_abstracts.jsonl works 4406152291 7128782475 4411542009,0.85,machine-verified,45E-44,Inference energy of general-purpose generative models exceeds that of task-specific models by orders of magnitude.,established,abstract,luccioni2024powerhungry;devries2023energy,,ai_ml_abstracts.jsonl work 4399450035,0.8,machine-verified,Magnitude statement is the authors'.46E-45,"Labour and adoption evidence rests on a preregistered field experiment (758 consultants), a task-exposure study and nationally representative adoption surveys.",established,abstract,dellacqua2023jagged;eloundou2024gpts;bick2026rapidadoption,,ai_ml_abstracts.jsonl works 7135096569 7124974472,0.85,machine-verified,'GPTs are GPTs' landmark record 4330336443 has a mismatched abstract; cite the verified Science article.47