How the Atlas is built.

1. The unit of analysis is the technique, not the paper

A single modern AI product deploys dozens of research ideas. Attributing “GPT = one paper” is reductive and wrong. The Atlas defines its unit as a technique: a named, bounded contribution with exactly one canonical origin paper and strong community consensus.

Techniques are curated by hand — there are about 50 that matter in the modern era. Variants and refinements are tracked as prior_art links between techniques, not as separate entries.

2. What counts as a “deployment”

A product deploys a technique when there is public, citable evidence that the technique is in the production system. Evidence must come from one of:

• The product's own model card or technical report
• The inventor's official blog describing a specific deployment
• Peer-reviewed analysis confirming the deployment
• Strong community consensus across multiple independent sources

Inference from benchmark behaviour (“model X probably uses technique Y because of pattern Z”) is not sufficient. Speculation is excluded.

3. Velocity calculation

velocity_days = deploy_date − technique.origin_date

• origin_date = arXiv v1 submission date (or journal publication where no arXiv exists)
• deploy_date = the product's first_seen in our knowledge graph, or the product's publicly-stated release date, whichever is earlier
• Self-invented deployments (Anthropic ships Constitutional AI → internal velocity) are computed but flagged separately; they answer a different question than external-adoption velocity

4. Confidence tiers

5. Data sources

• arXiv — authoritative paper metadata (submission date, authors, abstract)
• Papers With Code — technique taxonomy + benchmark associations (imported where applicable)
• Semantic Scholar — citation graphs and author affiliation history
• Model cards / technical reports — for the ~30 major commercial products tracked
• Official company blogs — Anthropic, OpenAI, Google DeepMind, Meta AI, Mistral, DeepSeek, Moonshot, xAI, Alibaba, Zhipu, NVIDIA
• Inference library release notes — vLLM, TGI, llama.cpp, Transformers — used for “commercially deployable” milestones where no single product shipped first

6. Known gaps

• Closed-source products under-report. When a major model publishes no technical report (e.g., some Gemini variants), we only track techniques that are independently confirmed.
• Chinese-language papers are under-represented. We rely primarily on arXiv and English-language tech reports.
• Incremental engineering improvements are not techniques. Better dataset filtering, prompt templating, or RLHF hyperparameter tuning are not tracked.
• Multiple origin candidates are collapsed. Where several papers introduce the same idea near-simultaneously, we cite the most-cited canonical entry and list alternates under prior_art. Challenge us if you think we picked wrong.

7. Open dataset

All technique, paper, and deployment records are freely downloadable under Creative Commons Attribution 4.0. API endpoints:

• /api/v1/atlas/techniques — all 49 techniques
• /api/v1/atlas/technique/{slug} — one technique + its deployments
• /api/v1/atlas/product/{slug} — one product's recipe
• /api/v1/atlas/velocity — global statistics
• /api/v1/atlas/graph — full graph for visualization
• /research-frontier/data.json — full dataset snapshot

Suggested citation: gentic.news Deployment Atlas (2026). Version 1.0. https://gentic.news/research-frontier

8. How to challenge a claim

If you think an attribution is wrong, a confidence tier is overconfident, or an important technique is missing: email corrections@gentic.news with the specific claim and your counter-evidence. We version the dataset and publish diffs.