How the score works

Eight dimensions, each 0–20. It measures a person's own verifiable record in the core of AI — not their company's, and not their visibility.

• Mathematical Foundations — degrees, theses, papers and code in linear algebra, matrix & tensor methods, optimization and statistical learning: the math the field stands on.
• Vector Embeddings — vector-space models, LSA/LSI, word and sentence embeddings, contrastive and dense retrieval, vector databases and search — authored, built or shipped.
• Transformer & LM Lineage — seq2seq, attention, transformers, pretraining, scaling laws and alignment — authored, led or trained.
• Frontier Founder — the person's own work is part of the foundation today's frontier AI models are built on: the architecture, attention, embeddings, optimizers, tokenizers, pretraining objectives, scaling results, alignment methods, datasets, benchmarks or training / inference stacks those models descend from.
• Deep Knowledge Domain Expert — years and history in language modeling, the tip of the spear in AI today: the depth and duration of a verifiable, continuous record from vector-space / LSI / n-gram and neural LMs through transformers and LLM pretraining and alignment.
• Hands-On Engineering — personally designed, built or shipped AI systems, models, or the hardware and infrastructure under them.
• Scientific & Industry Impact — built organizations or products whose core is these systems; citations and h-index; patents; leadership of labs that produced canonical work.
• 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. A founder title with the science done by others does not earn it.

Research is overweighted. The five research dimensions carry 70% of the total and the three practice dimensions 30%: score = round(70 × (F+V+T+FF+DD)/100 + 30 × (H+I+SF)/60). Two evidence-backed penalties are then subtracted, each 0–10: bought_popularity (paid coverage, placements or purchased reach) and capital_without_competence (an AI company started on family, friends or personal wealth with no verifiable language-modeling knowledge). A penalty is never applied without a live cited source.

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. Anchors: 18–20 authored canonical work the field builds on · 13–17 PhD-level work or production systems built personally · 8–12 strong graduate training or senior engineering adjacent to the core · 3–7 uses the tools, manages builders · 0–2 nothing verifiable.

Years count, and only verifiable years. Deep Knowledge Domain Expert: 18–20 is 15+ years of hands-on language-modeling work from the pre-word2vec era through transformers · 13–17 is 8–15 years · 8–12 is 3–8 years · 3–7 under 3 years or adjacent. Scientific & Technical Founder: 18–20 is 15+ years as the scientific / technical founder of companies whose core is these systems · 13–17 is 8–15 years or multiple such companies · 8–12 is 3–8 years · 3–7 a founder whose science was done by others. Frontier Founder: 18–20 authored a building block today's frontier models directly descend from · 13–17 a component the frontier labs cite and build on · 8–12 lineage work the stack draws on · 3–7 applies frontier models only. Years are counted from primary sources (first verifiable year, employer and position dates, patents, company records) — self-reported years score as absent.

Tiers: Frontier Builder 85+   Deep Practitioner 65–84   Technically Fluent 45–64   Informed Operator 25–44   Narrative Only <25

Every profile is scored from a programmatic dossier (Wikipedia, Wikidata, OpenAlex, Semantic Scholar, PubMed, patents, GitHub) by two independent passes, adjudicated when they disagree by more than 10 points. Every cited source is re-fetched and unreachable ones are dropped. Low-confidence results are returned to you but not published to the board. There is no special handling for any person, including this platform's own founder.

Check it yourself. export.json carries every person's dimensions, weighted total, penalties, evidence URLs and metadata; methodology carries the assessor prompt, the anchors, the invariants and the step-by-step validation recipe. Fetch each source, confirm the claim, recompute the total. Find a defect and file it.