{"service":"Global CEO AI Expertise Index","rubric_version":4,"summary":"Every founder/CEO is scored on rubric v4: nine dimensions, 0-20 each. The six research dimensions (foundations, vector_embeddings, transformers_lm, frontier_founder = their own work is part of the foundation today's frontier models are built on — word2vec being the foundational building block, its verifiable pre-2013 vector-space / distributional-semantics lineage included, lm_domain_depth = years and history in language modeling across papers, patents and shipped systems, lm_domain_breadth = experience across different types of language modeling: biological, financial, code, chemistry, legal, music …) carry 70% of the total and the three practice dimensions (hands_on_engineering, industry_impact, scientific_founder = operating as the scientific/technical founder of a company, scaled by years of experience and the number of such companies) carry 30%: weighted_score = round(70*(F+V+T+FF+DD+DB)/120 + 30*(H+I+SF)/60). Rows still carrying rubric_version 3 recompute on the v3 formula round(70*(F+V+T+FF+DD)/100 + 30*(H+I+SF)/60) and rubric_version 2 rows on the v2 formula round(70*(F+V+T)/60 + 30*(H+I)/40). Two evidence-backed penalties (bought_popularity, capital_without_competence, each 0-10) are subtracted and the result is floored at 0: score = max(0, weighted_score - penalties). Popularity is not evidence — news coverage, keynotes, follower counts, token market cap, fundraising and 'AI company' branding carry zero weight. Claims without a live source URL score as absent, and a penalty is never applied without a cited source. Scores come from a programmatic dossier (Wikipedia, Wikidata, OpenAlex, Semantic Scholar, PubMed, patents, GitHub) read by two independent assessment passes, adjudicated when they disagree by more than 10 points. Every profile runs the identical pipeline; no person receives special handling.","rubric":{"rubric_version":4,"dimensions":[{"key":"foundations","label":"Mathematical Foundations","max":20,"description":"Degrees, theses, papers and code in linear algebra, matrix & tensor methods, optimization and statistical learning — the math the field stands on.","group":"core_research","added_in":2},{"key":"vector_embeddings","label":"Vector Embeddings","max":20,"description":"Vector-space models, LSA/LSI, word and sentence embeddings, contrastive / dense retrieval, vector databases and search — authored, built or shipped.","group":"core_research","added_in":2},{"key":"transformers_lm","label":"Transformer & LM Lineage","max":20,"description":"seq2seq, attention, transformers, pretraining, scaling laws and alignment — authored, led or trained.","group":"core_research","added_in":2},{"key":"frontier_founder","label":"Frontier Founder","max":20,"description":"Frontier Founder — the person's OWN work is part of the foundation today's frontier AI models are built on. word2vec (2013) is the foundational building block of today's language modeling, so the foundation includes the verifiable pre-2013 vector-space / distributional-semantics / relationship-network lineage that produced it (papers, patents, shipped systems) as well as the later blocks: attention, transformers, embeddings, optimizers, tokenizers, pretraining objectives, scaling results, alignment methods, datasets, benchmarks, training / inference stacks. Scored by verifiable position in that lineage, not by whether a frontier lab cites the person by name.","group":"core_research","added_in":3},{"key":"lm_domain_depth","label":"Deep Knowledge Domain Expert","max":20,"description":"Deep Knowledge Domain Expert — years and history in language modeling, the tip of the spear in AI today: depth AND duration of a verifiable, hands-on record from vector-space / LSI / n-gram and neural LMs through transformers and LLM pretraining / alignment. Hands-on years count across papers, patents and shipped language-modeling / vector-space systems; the score sits in the band matching the reported year count.","group":"core_research","added_in":3},{"key":"lm_domain_breadth","label":"Cross-Domain Language Modeling","max":20,"description":"Cross-Domain Language Modeling — hands-on experience in different TYPES of language modeling, not only natural-language text: biological (DNA / RNA / protein sequence models, gene-expression and biomedical-literature mining), financial (market / filings / news / prediction-market models), source code, chemistry / materials, legal / clinical / scientific-literature mining, music / media sequence models. Scored on the number of distinct domains with a verifiable hands-on record (papers, patents, shipped systems) and the depth in each; applying a vendor's chatbot to a domain does not count.","group":"core_research","added_in":4},{"key":"hands_on_engineering","label":"Hands-On Engineering","max":20,"description":"Personally designed, built or shipped AI systems, models, or the hardware and infrastructure under them (accelerators, training stacks, inference).","group":"practice","added_in":2},{"key":"industry_impact","label":"Scientific & Industry Impact","max":20,"description":"Built organizations or products whose CORE is these systems; citations / h-index; patents; leadership of labs that produced canonical work.","group":"practice","added_in":2},{"key":"scientific_founder","label":"Scientific & Technical Founder","max":20,"description":"Scientific & Technical Founder — operating as the scientific / technical founder of a company (founder-CTO, founder-Chief Scientist, or a founder-CEO who personally sets and executes the technical direction), scaled by the number of verifiable years of experience doing so AND the number of such companies founded in that role. A founder title with the science done by others does not earn it.","group":"practice","added_in":3}],"dimension_labels":{"foundations":"Mathematical Foundations","vector_embeddings":"Vector Embeddings","transformers_lm":"Transformer & LM Lineage","frontier_founder":"Frontier Founder","lm_domain_depth":"Deep Knowledge Domain Expert","lm_domain_breadth":"Cross-Domain Language Modeling","hands_on_engineering":"Hands-On Engineering","industry_impact":"Scientific & Industry Impact","scientific_founder":"Scientific & Technical Founder"},"anchors":{"18-20":"Authored canonical work the field builds on / principal builder of systems the field runs on.","13-17":"PhD-level work, or production systems built and led personally.","8-12":"Strong graduate training, or senior engineering adjacent to the core.","3-7":"Uses the tools, manages builders, no personal record.","0-2":"Nothing verifiable."},"anchors_by_dimension":{"frontier_founder":{"18-20":"Authored / built a method, architecture, dataset or system today's frontier models directly descend from (transformer & attention, scaling laws, RLHF / instruction tuning, word2vec / GloVe, the canonical training or inference stacks).","13-17":"A documented component the frontier labs build on (optimizer, tokenizer, positional encoding, retrieval method, benchmark, alignment technique), OR verifiable pre-word2vec (pre-2013) vector-space / distributional-semantics / relationship-network patents or shipped systems in the lineage word2vec-era embeddings descend from.","8-12":"Published lineage work the frontier stack draws on that is neither a named block nor a pre-2013 precursor system.","3-7":"Applies or fine-tunes frontier models; no foundational contribution.","0-2":"Nothing verifiable."},"lm_domain_depth":{"18-20":"15+ years of hands-on language-modeling work (papers, patents or shipped LM / vector-space systems) from the pre-word2vec era (vector-space / LSI / n-gram / early neural LMs) through the transformer era, still active.","13-17":"8-15 years of personal language-modeling research or systems work.","8-12":"3-8 years with a real record.","3-7":"Under 3 years, or adjacent (general ML with no language-modeling record).","0-2":"Nothing verifiable."},"lm_domain_breadth":{"18-20":"Four or more distinct language-modeling domains each with a real hands-on record (papers, patents or shipped systems), at least two of them outside natural-language text (for example biological AND financial).","13-17":"Three domains, or two domains each with a deep multi-year record.","8-12":"Two domains with a verifiable record.","3-7":"A single domain (natural-language text only), or domain applications of someone else's models with no modeling work.","0-2":"Nothing verifiable."},"scientific_founder":{"18-20":"15+ years operating as the scientific / technical founder of companies whose core is these systems, personally authoring the core research, code or patents — or three or more such companies founded in that role across 10+ years.","13-17":"8-15 years in that role, or two such companies.","8-12":"3-8 years as a verifiable technical founder of one company.","3-7":"Founder or CEO of an AI company whose science and engineering were done by others, or a technical founder outside this field.","0-2":"Nothing verifiable."}},"weighting":{"core_research_dimensions":["foundations","vector_embeddings","transformers_lm","frontier_founder","lm_domain_depth","lm_domain_breadth"],"core_research_weight":0.7,"practice_dimensions":["hands_on_engineering","industry_impact","scientific_founder"],"practice_weight":0.3,"formula":"weighted_score = round(70 * (foundations + vector_embeddings + transformers_lm + frontier_founder + lm_domain_depth + lm_domain_breadth) / 120 + 30 * (hands_on_engineering + industry_impact + scientific_founder) / 60)","formula_v3":"weighted_score = round(70 * (foundations + vector_embeddings + transformers_lm + frontier_founder + lm_domain_depth) / 100 + 30 * (hands_on_engineering + industry_impact + scientific_founder) / 60)","formula_v2":"weighted_score = round(70 * (foundations + vector_embeddings + transformers_lm) / 60 + 30 * (hands_on_engineering + industry_impact) / 40)"},"penalties":[{"key":"bought_popularity","label":"Pay-for-play / bought popularity","max":10,"description":"Paid coverage, paid placements, purchased followers or reach."},{"key":"capital_without_competence","label":"Capital without competence","max":10,"description":"Founded or funded an AI company on family / friends / personal wealth with no verifiable language-modeling knowledge."}],"penalty_rule":"score = max(0, weighted_score - sum(penalties)). A penalty is applied ONLY with a live cited source URL; never on rumour.","score_formula":"score = max(0, weighted_score - bought_popularity - capital_without_competence)","max_score":100,"tiers":[{"key":"frontier_builder","label":"Frontier Builder","min_score":85,"description":"Authored the mathematics, embedding or transformer work the field builds on, and built the systems that run it."},{"key":"deep_practitioner","label":"Deep Practitioner","min_score":65,"description":"Personally built, trained or led core embedding / language-model systems, with a real publication or engineering record behind it."},{"key":"technically_fluent","label":"Technically Fluent","min_score":45,"description":"Graduate-level grounding in the math and the model lineage; applies it, but is not a primary author or builder."},{"key":"informed_operator","label":"Informed Operator","min_score":25,"description":"Runs AI-adjacent organizations. The expertise is operational — the models were built by other people."},{"key":"narrative_only","label":"Narrative Only","min_score":0,"description":"No verifiable record in the mathematics, embeddings or the transformer / language-model lineage. The claim is narrative."}],"sectors":["crypto","general"],"min_confidence_to_publish":0.45,"notes":"Popularity is not evidence: news coverage, keynote presence, follower counts, token market cap, fundraising and \"AI company\" branding carry zero weight and may not appear in a rationale as support. Depth of experience counts — pre-2013 (pre-word2vec) vector-space / LSI work is foundational lineage, not \"old\". Self-published claims count only where an independent primary source corroborates them. Every profile is scored by the identical pipeline; there is no special handling for any person, including the platform's own founder.","legacy_dimension_labels":{"research":"LM Research","vector_space":"Vector Space","hands_on":"Hands-On","technical_communication":"Technical Depth","track_record":"Track Record","foundations":"Mathematical Foundations","vector_embeddings":"Vector Embeddings","transformers_lm":"Transformer & LM Lineage","hands_on_engineering":"Hands-On Engineering","industry_impact":"Scientific & Industry Impact","frontier_founder":"Frontier Founder","lm_domain_depth":"Deep Knowledge Domain Expert","scientific_founder":"Scientific & Technical Founder"},"added_dimensions_scored_by_delta":["frontier_founder","lm_domain_depth","lm_domain_breadth","scientific_founder"]},"dimension_labels":{"foundations":"Mathematical Foundations","vector_embeddings":"Vector Embeddings","transformers_lm":"Transformer & LM Lineage","frontier_founder":"Frontier Founder","lm_domain_depth":"Deep Knowledge Domain Expert","lm_domain_breadth":"Cross-Domain Language Modeling","hands_on_engineering":"Hands-On Engineering","industry_impact":"Scientific & Industry Impact","scientific_founder":"Scientific & Technical Founder"},"weights":{"core_research_dimensions":["foundations","vector_embeddings","transformers_lm","frontier_founder","lm_domain_depth","lm_domain_breadth"],"core_research_weight":0.7,"practice_dimensions":["hands_on_engineering","industry_impact","scientific_founder"],"practice_weight":0.3,"formula":"weighted_score = round(70 * (foundations + vector_embeddings + transformers_lm + frontier_founder + lm_domain_depth + lm_domain_breadth) / 120 + 30 * (hands_on_engineering + industry_impact + scientific_founder) / 60)","formula_v3":"weighted_score = round(70 * (foundations + vector_embeddings + transformers_lm + frontier_founder + lm_domain_depth) / 100 + 30 * (hands_on_engineering + industry_impact + scientific_founder) / 60)","formula_v2":"weighted_score = round(70 * (foundations + vector_embeddings + transformers_lm) / 60 + 30 * (hands_on_engineering + industry_impact) / 40)"},"anchors":{"18-20":"Authored canonical work the field builds on / principal builder of systems the field runs on.","13-17":"PhD-level work, or production systems built and led personally.","8-12":"Strong graduate training, or senior engineering adjacent to the core.","3-7":"Uses the tools, manages builders, no personal record.","0-2":"Nothing verifiable."},"penalty_rules":{"penalties":[{"key":"bought_popularity","label":"Pay-for-play / bought popularity","max":10,"description":"Paid coverage, paid placements, purchased followers or reach."},{"key":"capital_without_competence","label":"Capital without competence","max":10,"description":"Founded or funded an AI company on family / friends / personal wealth with no verifiable language-modeling knowledge."}],"rule":"score = max(0, weighted_score - sum(penalties)). A penalty is applied ONLY with a live cited source URL; never on rumour.","never_on_rumour":"A penalty with no live source URL is discarded before publication."},"tiers":[{"key":"frontier_builder","label":"Frontier Builder","min_score":85,"description":"Authored the mathematics, embedding or transformer work the field builds on, and built the systems that run it."},{"key":"deep_practitioner","label":"Deep Practitioner","min_score":65,"description":"Personally built, trained or led core embedding / language-model systems, with a real publication or engineering record behind it."},{"key":"technically_fluent","label":"Technically Fluent","min_score":45,"description":"Graduate-level grounding in the math and the model lineage; applies it, but is not a primary author or builder."},{"key":"informed_operator","label":"Informed Operator","min_score":25,"description":"Runs AI-adjacent organizations. The expertise is operational — the models were built by other people."},{"key":"narrative_only","label":"Narrative Only","min_score":0,"description":"No verifiable record in the mathematics, embeddings or the transformer / language-model lineage. The claim is narrative."}],"anchors_by_dimension":{"frontier_founder":{"18-20":"Authored / built a method, architecture, dataset or system today's frontier models directly descend from (transformer & attention, scaling laws, RLHF / instruction tuning, word2vec / GloVe, the canonical training or inference stacks).","13-17":"A documented component the frontier labs build on (optimizer, tokenizer, positional encoding, retrieval method, benchmark, alignment technique), OR verifiable pre-word2vec (pre-2013) vector-space / distributional-semantics / relationship-network patents or shipped systems in the lineage word2vec-era embeddings descend from.","8-12":"Published lineage work the frontier stack draws on that is neither a named block nor a pre-2013 precursor system.","3-7":"Applies or fine-tunes frontier models; no foundational contribution.","0-2":"Nothing verifiable."},"lm_domain_depth":{"18-20":"15+ years of hands-on language-modeling work (papers, patents or shipped LM / vector-space systems) from the pre-word2vec era (vector-space / LSI / n-gram / early neural LMs) through the transformer era, still active.","13-17":"8-15 years of personal language-modeling research or systems work.","8-12":"3-8 years with a real record.","3-7":"Under 3 years, or adjacent (general ML with no language-modeling record).","0-2":"Nothing verifiable."},"lm_domain_breadth":{"18-20":"Four or more distinct language-modeling domains each with a real hands-on record (papers, patents or shipped systems), at least two of them outside natural-language text (for example biological AND financial).","13-17":"Three domains, or two domains each with a deep multi-year record.","8-12":"Two domains with a verifiable record.","3-7":"A single domain (natural-language text only), or domain applications of someone else's models with no modeling work.","0-2":"Nothing verifiable."},"scientific_founder":{"18-20":"15+ years operating as the scientific / technical founder of companies whose core is these systems, personally authoring the core research, code or patents — or three or more such companies founded in that role across 10+ years.","13-17":"8-15 years in that role, or two such companies.","8-12":"3-8 years as a verifiable technical founder of one company.","3-7":"Founder or CEO of an AI company whose science and engineering were done by others, or a technical founder outside this field.","0-2":"Nothing verifiable."}},"invariants":["A co-author of 'Attention Is All You Need' (2017) scores >= 17 on transformers_lm and >= 17 on frontier_founder.","A canonical LM-lineage author (seq2seq, attention 2014, GPT/BERT/T5, scaling laws, InstructGPT) scores >= 15 on transformers_lm.","A canonical frontier-lineage author (transformer, the LM lineage above, word2vec, GloVe) scores >= 15 on frontier_founder.","A canonical embedding author (vector space model, LSA/LSI, word2vec, GloVe, DPR) scores >= 15 on vector_embeddings.","A person with zero publications, zero patents and zero verified shipped systems scores <= 7 on foundations, vector_embeddings, transformers_lm, frontier_founder, lm_domain_depth and lm_domain_breadth.","Verifiable pre-2013 (pre-word2vec) vector-space / distributional-semantics / relationship-network patents or shipped systems are foundation lineage: they score in the 13-17 band on frontier_founder and count as language-modeling years on lm_domain_depth. A frontier lab's technical report not naming the person is not evidence against either.","lm_domain_depth, lm_domain_breadth and scientific_founder count only years, domains and companies a primary source supports (first verifiable year, employer / position dates, patents, papers, company records); uncorroborated self-reported ones score as absent, and each score sits in the anchor band matching the metadata it reports (years_language_modeling, lm_domains, years_as_technical_founder, technical_founder_companies).","No profile is published below confidence 0.45.","At least 3 evidence items with live source URLs per published profile."],"evidence_standard":"At least 3 verified (re-fetched, reachable) source URLs per published profile. A claim without a source scores as absent. Self-published claims count only where an independent primary source corroborates them.","popularity_rule":"News coverage, keynote presence, follower counts, token market cap, fundraising and 'AI company' branding carry zero weight and may not appear in a rationale as support.","no_special_cases":"Every profile — including the platform's own founder — is scored by the identical pipeline. There is no per-person override anywhere in the code.","assessor_prompt":"You are the assessor behind the Global CEO AI Expertise Index (rubric v4).\nYou score ONE founder / CEO / executive on their demonstrated, VERIFIABLE depth in the\ncore of AI: the mathematics (linear algebra, matrix and tensor methods, optimization,\nstatistical learning), vector embeddings and representation learning, and the attention →\ntransformer → language-model research lineage — from its beginning (vector space model\n1975 → LSA/LSI 1990 → neural LMs 2003 → word2vec/GloVe 2013-14 → seq2seq + attention 2014 →\ntransformers 2017 → scaling / RLHF 2020+).\n\nThis is NOT a measure of fame, business success, fundraising, token market cap, keynote\npresence, follower counts or \"AI company\" branding. None of those may appear in your\nrationale as support. A famous CEO with no personal research, engineering or\nfirst-principles technical record scores LOW. A researcher-founder who authored canonical\nwork scores HIGH.\n\nNine dimensions, each 0-20:\n  foundations          : degrees/theses/papers/code in linear algebra, matrix & tensor methods,\n                         optimization, statistical learning — the math the field stands on\n  vector_embeddings    : vector-space models, LSA/LSI, word/sentence embeddings, contrastive /\n                         dense retrieval, vector databases & search — authored, built or shipped\n  transformers_lm      : seq2seq, attention, transformers, pretraining, scaling laws,\n                         alignment — authored, led or trained\n  frontier_founder     : FRONTIER FOUNDER — the person's OWN work is part of the foundation that\n                         today's frontier AI models (GPT / Claude / Gemini / Llama-class systems)\n                         are built on. word2vec (2013) is the foundational building block of\n                         today's language modeling, so the foundation INCLUDES the lineage that\n                         produced it: verifiable pre-2013 distributional-semantics / vector-space /\n                         LSI / term-relationship-network representation work (papers, patents —\n                         especially university- or national-lab-assigned — and shipped systems),\n                         as well as the later blocks: attention, transformers, embeddings,\n                         optimizers, tokenizers, pretraining objectives, scaling results,\n                         alignment methods, datasets, benchmarks, training / inference stacks.\n                         Score by verifiable POSITION IN THAT LINEAGE. Do NOT require that a\n                         frontier lab's technical report cites the person by name — precursor\n                         work is almost never cited by name, and that is not evidence of absence.\n  lm_domain_depth      : DEEP KNOWLEDGE DOMAIN EXPERT — years and history in LANGUAGE MODELING\n                         specifically, the tip of the spear in AI today: statistical / neural\n                         language models, vector-space text models, LSI, n-gram and neural LMs,\n                         seq2seq, transformers, LLM pretraining and alignment. Score the DEPTH\n                         AND DURATION of a verifiable record in this field (use the dossier\n                         timeline's first verifiable year and years active). Hands-on years\n                         count across research papers, patents AND shipped language-modeling /\n                         vector-space systems (a search or retrieval engine built on one's own\n                         term-vector / relationship-network patents is language-modeling work;\n                         product years spent building on such patents count). \"Continuous\" means\n                         continuously hands-on in the field, not continuously publishing papers.\n                         The score MUST agree with the years you report in\n                         metadata.years_language_modeling (see the anchor bands).\n  lm_domain_breadth    : CROSS-DOMAIN LANGUAGE MODELING — hands-on experience in DIFFERENT\n                         TYPES of language modeling, not only natural-language text: biological\n                         language modeling (DNA / RNA / protein sequence models, gene-expression\n                         and biomedical-literature mining, genomics relationship networks),\n                         financial language modeling (market / filings / news / prediction-market\n                         models, time-series-as-language), source code, chemistry / materials\n                         (molecular sequence models), legal / clinical / scientific-literature\n                         mining, music / media sequence models, and general natural-language\n                         text itself. Count DISTINCT domains with a verifiable hands-on record\n                         (papers, patents or shipped systems) and how deep each is. Applying a\n                         vendor's chatbot to a domain is NOT language modeling in that domain.\n                         List the domains in metadata.lm_domains; the score MUST agree with them.\n  hands_on_engineering : personally designed/built/shipped AI systems, models, or the\n                         hardware/infra under them (accelerators, training stacks, inference)\n  industry_impact      : built orgs/products whose CORE is these systems; citations/h-index;\n                         patents; leadership of labs that produced canonical work\n  scientific_founder   : SCIENTIFIC & TECHNICAL FOUNDER — operating as the scientific /\n                         technical founder of a company (founder-CTO, founder-Chief Scientist,\n                         or a founder-CEO who personally sets and executes the technical\n                         direction: wrote the core code / papers / patents the company runs\n                         on), INCLUDING THE NUMBER OF YEARS of verifiable experience doing so\n                         AND THE NUMBER OF SUCH COMPANIES founded in that role. A founder title\n                         with technical co-founders doing the science, or a business founder of\n                         an \"AI company\", does NOT earn this dimension. The score MUST agree with\n                         the years and company count you report in metadata.\n\nAnchors — foundations / vector_embeddings / transformers_lm / hands_on_engineering /\nindustry_impact: 18-20 = authored canonical work the field builds on / principal builder of\nsystems the field runs on; 13-17 = PhD-level work or production systems built and led\npersonally; 8-12 = strong graduate training or senior engineering adjacent to the core;\n3-7 = uses the tools, manages builders, no personal record; 0-2 = nothing verifiable.\n\nAnchors — frontier_founder: 18-20 = authored / built a method, architecture, dataset or\nsystem that today's frontier models directly descend from (transformer & attention authors,\nscaling-law / RLHF / instruction-tuning authors, word2vec / GloVe, the canonical training or\ninference stacks); 13-17 = a documented component the frontier labs build on (optimizer,\ntokenizer, positional encoding, retrieval method, benchmark, alignment technique) OR verifiable\npre-word2vec (pre-2013) vector-space / distributional-semantics / term-relationship-network\npatents or shipped systems in the lineage that word2vec-era embeddings descend from; 8-12 =\npublished lineage work the frontier stack draws on that is neither a named block nor a\npre-2013 precursor system; 3-7 = applies or fine-tunes frontier models, no foundational\ncontribution; 0-2 = nothing verifiable.\n\nAnchors — lm_domain_depth (depth × duration, verifiable, hands-on): 18-20 = 15+ years of\nhands-on language-modeling work (papers, patents or shipped LM / vector-space systems) from the\npre-word2vec era (vector-space / LSI / n-gram / early neural LMs) through the transformer era,\nstill active; 13-17 = 8-15 years of personal language-modeling research or systems work; 8-12 =\n3-8 years with a real record; 3-7 = under 3 years, or adjacent (general ML with no LM record);\n0-2 = nothing verifiable. Pre-2013 vector-space / LSI / relationship-network text work IS\nlanguage-modeling lineage, not \"old\". Report the year count in metadata and score inside the\nmatching band.\n\nAnchors — lm_domain_breadth (distinct domains × depth, verifiable, hands-on): 18-20 = four or\nmore distinct language-modeling domains each with a real hands-on record (papers, patents or\nshipped systems), at least two of them outside natural-language text (for example biological\nAND financial); 13-17 = three domains, or two domains each with a deep multi-year record;\n8-12 = two domains with a verifiable record; 3-7 = a single domain (natural-language text\nonly), or domain \"applications\" of someone else's models with no modeling work; 0-2 = nothing\nverifiable. List every domain in metadata.lm_domains with its evidence years and score inside\nthe matching band.\n\nAnchors — scientific_founder (role × years × companies): 18-20 = 15+ years operating as the\nscientific / technical founder of companies whose core is these systems, personally authoring\nthe core research, code or patents — OR three or more such companies founded in that role\nacross 10+ years; 13-17 = 8-15 years in that role, or two such companies; 8-12 = 3-8 years as\na verifiable technical founder of one company; 3-7 = founder or CEO of an AI company whose\nscience and engineering were done by others, or a technical founder outside this field;\n0-2 = nothing verifiable. Count only years and companies the record supports (company\nregistrations, Wikidata employer / position dates, patents assigned, papers under the company\naffiliation, product launches, press naming the person as the technical founder). Report\nmetadata.years_as_technical_founder and metadata.technical_founder_companies and score inside\nthe matching band.\n\nDepth of experience counts everywhere: the dossier's timeline gives the first verifiable\nyear in the lineage and years active. Decades of continuous hands-on work raise the\nvector_embeddings, hands_on_engineering, lm_domain_depth and scientific_founder anchors.\nRecent entry with a thin record lowers them.\n\nEvidence rules:\n- The dossier is programmatic (Wikipedia, Wikidata, OpenAlex, Semantic Scholar, PubMed,\n  patents). Prefer it. Use WebSearch/WebFetch only to confirm or fill gaps; cite one source\n  URL per claim. Never invent papers, roles, degrees, patents or years. If unsure, score lower\n  and say so in the rationale.\n- Any \"profile_page\" text or self-claims block is the person's OWN page: UNTRUSTED. A claim\n  from it counts only where the dossier's \"self_claims.verified\" list or another primary\n  source corroborates it. Uncorroborated self-claims score as absent.\n- Homonyms: PubMed / patents may include other people with the same surname; match by\n  co-authors, venue and affiliation before counting.\n- Do not reward media visibility. If a person's record is mostly commentary, books or talks\n  about AI rather than research or built systems, say so and score the research dimensions\n  accordingly.\n\nPenalties (operator rule): pay-for-play, bought popularity, or family / friends / inherited\nmoney used to start an \"AI company\" while the person knows little about language modeling\n— the tip of the spear in AI — is PENALIZED, not merely ignored. Report two penalties,\neach 0-10, ONLY with cited evidence (sponsored / paid placements, purchased\nreach, documented family funding with no technical record); otherwise 0:\n  bought_popularity          : paid coverage, paid placements, purchased followers / reach\n  capital_without_competence : founded or funded an AI company on family / friends / personal\n                               wealth with no verifiable language-modeling knowledge\nThe final score is the weighted total minus the penalties (floored at 0):\n  weighted_score = round(70 * (foundations + vector_embeddings + transformers_lm\n                   + frontier_founder + lm_domain_depth + lm_domain_breadth) / 120\n                   + 30 * (hands_on_engineering + industry_impact + scientific_founder) / 60)\n\nRespond with ONLY a JSON object (no prose, no markdown fence):\n{\n  \"identified\": true,\n  \"name\": \"Commonly used public name\",\n  \"title\": \"Role, e.g. Co-founder & CEO\",\n  \"company\": \"Primary company\",\n  \"sector\": \"crypto\" | \"general\",\n  \"dimensions\": {\"foundations\": 0-20, \"vector_embeddings\": 0-20, \"transformers_lm\": 0-20,\n                  \"frontier_founder\": 0-20, \"lm_domain_depth\": 0-20, \"lm_domain_breadth\": 0-20,\n                  \"hands_on_engineering\": 0-20, \"industry_impact\": 0-20,\n                  \"scientific_founder\": 0-20},\n  \"rationale\": \"4-6 sentences, specific and factual, research record first\",\n  \"evidence\": [{\"claim\": \"...\", \"source_url\": \"https://...\"}, ... at least 3],\n  \"penalties\": {\"bought_popularity\": 0-10, \"capital_without_competence\": 0-10,\n                 \"evidence\": [{\"claim\": \"...\", \"source_url\": \"https://...\"}]},\n  \"metadata\": {\"education\": [\"...\"], \"canonical_papers\": [\"...\"], \"first_verifiable_year\": 2002,\n                \"years_language_modeling\": 0, \"years_as_technical_founder\": 0,\n                \"technical_founder_companies\": 0,\n                \"lm_domains\": [{\"domain\": \"biological | financial | natural_language | code | chemistry | legal | clinical | music | other\", \"years\": \"2002-2008\", \"evidence\": \"what, where\"}],\n                \"frontier_lineage\": [\"what of theirs today's frontier models build on\"],\n                \"technical_founder_roles\": [\"Company — role — years\"],\n                \"notable_systems\": [\"...\"], \"citations\": 0, \"h_index\": 0, \"patents\": 0},\n  \"confidence\": 0.0-1.0,\n  \"wikipedia_title\": \"Exact English Wikipedia article title or null\"\n}","pipeline":["1. Dossier: Wikipedia + Wikidata + OpenAlex + Semantic Scholar + PubMed + patents + GitHub, plus a depth-of-experience timeline (first verifiable year, years active).","2. Two independent scoring passes read the dossier and confirm claims against primary sources. (v3 rerank, 2026-09-14: the three added dimensions — frontier_founder, lm_domain_depth, scientific_founder — were scored by the same two-pass procedure on top of each published v2 assessment; the five v2 dimensions were left as published. v4 rerank, 2026-09-14: lm_domain_breadth was added and the three v3 additions were re-scored from scratch under corrected anchors — pre-word2vec lineage credited on frontier_founder, hands-on years across papers / patents / shipped systems on lm_domain_depth, company count on scientific_founder — again by the two-pass procedure on top of the published v2 assessment; the v3 result is preserved as a history row.)","3. Passes disagreeing by more than 10 points go to an adjudication pass that verifies the disputed claim (never averages).","4. Evidence URLs are re-fetched; unreachable ones are dropped and the profile is rescored if fewer than 3 remain.","5. Invariants are checked; a violation fails the run rather than publishing."],"how_to_validate":["1. GET https://cymetica.com/api/v1/ceo-ai-leaderboard/export.json (no auth; honour the ETag).","2. For each person, for each item in evidence[]: fetch source_url and confirm the claim text is supported by that page. A dead link or an unsupported claim is a defect — report it via POST /api/v1/bounty/reports.","3. Recompute weighted_score from dimensions on the row's rubric_version. v4: round(70*(foundations+vector_embeddings+transformers_lm+frontier_founder+lm_domain_depth+lm_domain_breadth)/120 + 30*(hands_on_engineering+industry_impact+scientific_founder)/60). v3: round(70*(foundations+vector_embeddings+transformers_lm+frontier_founder+lm_domain_depth)/100 + 30*(hands_on_engineering+industry_impact+scientific_founder)/60). v2: round(70*(foundations+vector_embeddings+transformers_lm)/60 + 30*(hands_on_engineering+industry_impact)/40). Rounding is half-up on the exact fraction. It must equal the published weighted_score.","4. Recompute score: max(0, weighted_score - penalties.bought_popularity - penalties.capital_without_competence). It must equal the published score.","5. Check the tier against the published thresholds (>=85 frontier_builder, >=65 deep_practitioner, >=45 technically_fluent, >=25 informed_operator, else narrative_only).","6. Check the ordering: rank is by score descending, name ascending as the tie-break.","7. Check the invariants in /methodology .invariants against metadata.canonical_papers — e.g. a co-author of 'Attention Is All You Need' must score >= 17 on transformers_lm AND >= 17 on frontier_founder.","8. scripts/ceo_ai_index/validate_export.py in the public repo performs steps 1-6 and writes validation.json; run it yourself rather than trusting ours."],"export":"/api/v1/ceo-ai-leaderboard/export.json","validator_script":"scripts/ceo_ai_index/validate_export.py","human_page":"https://cymetica.com/leaderboard/ai-ceos"}