The AI Atlas is a personal project. It runs on the maintainer’s own hosting accounts (Vercel for the application, Supabase for the Postgres database) and is not an employer system, an official product, or the view of any organization the maintainer works for or with. The site is public by URL: anything not behind the access key or the admin password can be read by anyone who has the address. The public surfaces are Signal Board, News Blotter, Savant, Claims & Theses, Report Portal, Data Portal, Research Portal, Startup Scout, and Tooling Monitor. The confidence levels on the argument map are the maintainer’s own judgments and nothing else.
Every collection engine reads public sources only: the engines hold no subscriptions and never log in, so what is stored is what a page serves to an anonymous reader, whether fetched directly or through the Jina reader. A page that refuses to be fetched is kept as a title and a link (plus the search result’s snippet, where there is one) rather than as text.
Text collected here is sent to outside services to be processed. Anthropic’s Claude models run the reasoning legs, from dossiers to Savant’s writer and editor. OpenRouter routes calls to open-weight models from several labs, run by whichever inference provider OpenRouter selects for the call. Tavily performs news search. Jina fetches pages the site cannot fetch itself. Resend sends the agent’s daily brief and access-request notices by email to the maintainer. Questions typed into Ask are also sent to an embedding model (text-embedding-3-small) through OpenRouter so the meaning search can run; see How Ask finds things. Vercel hosts the application and, through Vercel Web Analytics, counts page views: each view sends the page path, referrer, and coarse device and country data to Vercel, without cookies and without a per-visitor identifier that persists across days. Supabase hosts the database. On the Tooling Monitor your browser loads vendor logos directly from Google’s favicon service and from GitHub, so those two services see your request; nothing from that is stored here. Each of these providers applies its own retention and training terms to what is sent to it, and this site does not control those terms.
Article text and filings, for summaries, tags, and fact extraction. Extracted facts, when a dossier or a report is composed from them. An Ask question, the record excerpts retrieved for it, and the recent conversation: the question and the previous user turn go to the classifier (a Qwen model via OpenRouter); the question, the excerpts, and up to the last twelve turns of the conversation (about 8,000 characters, your questions and the earlier answers) go to the answering model (Claude Haiku via Anthropic). The text of a document uploaded on a Scout profile or the add-source form. The free-text context field on the tooling report console, which is written into the report pack. When web search is switched on in Ask, or when a fresh question the records do not cover switches it on automatically for that turn, the question also goes to Anthropic’s search tool.
The database keeps the full text of every collected article, filing, and paper; the text extracted in the browser from uploaded documents; the facts, tags, scores, and summaries the models produce; and the data packs behind every generated report, including the tooling report’s context field. The Atlas’s own reports (the Daily Edition, the research roundup, Savant) are cut into passages and indexed for Ask the same two ways as the records, so an answer can cite a report passage beside the records under it. The one-time history since November 2022 is stored apart from the daily collections and keyed on each article’s own date, so none of it appears as today’s news; its drafts reach the map only when the maintainer publishes them. Everything is kept indefinitely. There is no retention limit and no deletion job today. Archiving a signal or a draft hides it from the public surfaces; it does not delete the row. Every model run started from a button (a report, a diagnosis, an AI suggestion) leaves a small record: its steps, how long it took, what it cost, and who started it, the maintainer or an access key by its id, never a name; that record is what lets a run continue after its page is closed. Uploaded documents are never stored as files: the browser extracts the text, sends the text, and discards the file. The one exception is the feedback box, whose screenshots are stored in the database and are visible only to the maintainer. The Data Portal’s schema map (/datasets/schema) shows the shape of every table this stores, grouped by subsystem, with no row of data ever shown.
The feedback box stores the email address you enter (it is required), the page you were on, your browser’s user-agent string, and any screenshots you attach. Only the maintainer can read a ticket. Source dossiers and Scout company profiles carry facts about named authors, founders, and team members drawn from public sources; a dossier notes the basis of each claim, and a Scout profile lists the pages it consulted. Once the access-request form ships, a request will store a name and a work email address.
Do not enter confidential information into Ask, a document upload, or any form on this site. Everything you type may be stored here and sent to the processors above.
Three tiers. A guest, with no session at all, sees the public layer: the surfaces above with confidences stripped, published signals only, and the 13 public datasets. An access key (per-person, issued by the maintainer on request through the request form, expiring after 90 days unless renewed; a legacy shared key remains during the migration) unlocks Ask, the 13 key-gated datasets (retained article text, machine-extracted records including items no human has reviewed, agent scores, rigor priors, and dossier fields), tooling reports and drafts, and the Scout research tools and document upload. The admin, one password, sees the personal layer (confidences, rationales, reliability priors), every draft, and the operating consoles.
Most of what the models produce is a proposal a human then commits or discards. Eight things publish without a human step, and a ninth is handed to access-key holders unreviewed:
Field Reports are a separate case: research reports written on request. The person asking (the admin, or a holder of a personal access key) edits the research plan and starts the run; Claude models then research the Atlas records, search the web for gaps, and write the report, labeling every paragraph by whether it rests on Atlas records, web sources, both, or the model’s own analysis. A report stays visible to the person who ran it and the admin until the admin publishes it to other access-key holders. Field Reports are never public and are never added to Ask’s shared corpus. Each access key has a daily spending allowance for them.
The Daily Edition, the research roundup, the tooling reports, the generated sheets, the thesis reports, Savant, Field Reports, the Briefcase’s briefs, and the company intel deck pass a citation gate, so a link the records cannot vouch for is stripped. Period reports are edited by the admin before saving; a signal draft’s claim touches are checked against live claim codes but its text is not link-gated. A confidence on the argument map never moves without a human-written rationale.