AI Release is an autonomous publishing system about artificial intelligence. It collects sources, checks facts, writes analysis and publishes to the website and Telegram. The website is the canonical source; social platforms are delivery channels.
Not a link aggregator and not a press-release rewriting service: every item has a triggering source, a fact check, original analysis and an explicit line between what was reported and what was assessed.
News, practical guides, community discussion analyses and video explainers are published. Every item exists in three machine formats: HTML, Citation Map JSON and Markdown.
Pipeline: source collection → candidate selection → fact-checking → editorial processing → cover generation → website publication → delivery to Telegram, Bluesky and Dzen.
Shared limit: 10 news items per day, paced across the day so an early slot cannot consume the whole daily quota. Up to 3 guides per day with originality checks, up to 2 community discussions.
Article structure: key takeaway, what happened, editorial analysis, what remains unknown, what is being debated, outlook, disclaimer and source. The last blocks are not decoration: they record the boundaries of knowledge.
| Pipeline | Role | What it does |
|---|---|---|
hourly.py | Основной новостной контур | RSS/веб → отбор → факт-чек → редактура → обложка → сайт → Telegram → Bluesky/Dzen |
converter_to_airelease.py | Азиатский редактор | Отдельный промпт и расписание, общая квота новостей |
discussion/ | Обсуждения сообщества | HN / OpenAI / Hugging Face → кластеризация → скоринг → комментарии → редактура |
guides.py | Практические гайды | Выбор темы → проверка оригинальности → статья → обложка → анонс |
video.py | Видеоразборы | 1 объяснялка в сутки: тизер + ссылка на страницу сайта |
daily.py | Дневной дайджест | Один пост-сводка; в новостную квоту не входит |
Three categories are separated: confirmed, reported by a source, and editorial analysis. Facts, numbers and quotations come from the triggering material; analysis is labelled as assessment.
If evidence is missing, the slot stays empty. Stale triggers are not published, old material is not presented as new, and unsupported claims are not invented.
Single-sourced, unofficial and leaked material gets an automatic disclaimer. Retracted material is withdrawn with HTTP 410 and removed from the manifest, translations, feed, sitemap and llms files.
Every item has a visible "AI summary" block with machine-readable facts; for guides it is extended into a citation layer. In markup it is marked with data-ai-knowledge="true", id="ai-knowledge" and a JSON URL in data-ai-citation-map.
The Citation Map is a separate JSON next to the page: {slug}.json for RU and {slug}_en.json for EN. It contains the article, summary, topics, entities, claims, relationships, question/answer pairs, limitations and a ready citation string.
Each claim has source_type (origin), verification_status (whether the event was verified) and confidence (editorial assessment). confidence is not a probability: use verification_status for verification.
The Knowledge Layer is an AI Release-defined format. It is not an official OpenAI, Google or Anthropic standard, and no compatibility with third-party specifications is claimed.
No semantic embeddings or vector search: no vector database is served.
No aggregate Knowledge Index across all items.
No public content API; only an internal analytics endpoint exists.
Videos are uploaded to YouTube manually: the system prepares the item and posts an announcement, but does not upload the video.
Crawler permission does not imply indexing, citation or endorsement by any AI provider.
| Type | Item | Formats |
|---|---|---|
| News · RU | Provenance vs. Fabrication: How to Keep a Made-Up Value Out of a Real Call | HTML · Citation Map · Markdown |
| Guide · RU | Jeff 0.8B: a millisecond-fast model and who actually needs it | HTML · Citation Map · Markdown |
| News · EN | Provenance vs. Fabrication: How to Keep a Made-Up Value Out of a Real Call | HTML · Citation Map · Markdown |