Asking for data runs a model call, so it needs an access key. Unlock or request access.
26 datasets · 13 public · 13 behind an access key
Every published signal on the board: one tracked AI development per row, with its significance, audience lenses, touched claims, and the analyst brief (what happened, why it matters, what is contested) plus the counterpoint.
Every published signal in the same row shape as the external-scan dataset, for a downstream intake that already speaks that schema: discovery metadata, the full editorial writeup composed into full_text, and the signal-native fields appended.
Every public evidence row: which source or signal bears on which claim, in which direction, at what weight, with the quoted excerpt.
The touch matrix in long form: one row for every published signal that bears on every claim it touches, with the direction of the touch. The dataset for direction-balance questions.
Every node of the argument map in one flat table: the open questions, the candidate stances under them, the falsifiable claims with their tests, the organizing frames, and the bridge claims that link domains.
The wiring of the argument map: which claims support or contradict which stances, which claims feed the bridge claims, and which frames organize what.
The AI terminology dependency graph: every concept with its short definition, full explanation, settled or contested status, prerequisite concepts, and linked argument-map claims.
The bibliography behind the Atlas: every source that backs at least one evidence row or one published signal, with counts and a full-text availability flag.
The complete retained text of every published signal's underlying article: curated source text where it exists, otherwise the page text the discovery pipeline fetched.
The curated arXiv shortlist: papers the research pipeline kept, with abstracts, triage summaries, advisory claim links, and citation counts.
The firewall handoff of the reviewed arXiv shortlist: every tracked or noted paper with its extracted finding and retained full text.
The Startup Scout watchlist: young AI companies tracked as acquisition candidates, organized by vertical, with what each builds and the AI technology itself.
The event timelines of tracked Scout companies: funding rounds, launches, and notable news, one row per event.
One day of the automated external news scan: items across financial services and technology topics discovered from public press feeds and web search, with full text, a model summary, taxonomy tags, and named entities.
One day of the Intel Desk collection engine: documents discovered about tracked companies from public press feeds, web search, and SEC filings, with full text, extracted facts, and a model summary.
The Intel Desk registry: the companies under standing coverage, organized by tier, with their public identifiers and the machine's merged dossier summary.
Structured, provenance-carrying facts about tracked companies, extracted by enrichment from collected items: one fact per row, with its dimension, value, and as-of date.
LLM-free structured series for tracked companies, about a decade deep: the full FDIC call-report field set per bank, FR Y-9C holding-company consolidated items, curated SEC EDGAR XBRL concepts, and CFPB complaint counts: one metric-period value per row.
Everything public about the organization the Atlas is written for, since ChatGPT: its SEC filings and press releases, news, research papers, AI patents, comment letters and regulatory documents, plus the AI timeline and the cited profile built from them, in one row shape.
Everything public the Atlas holds about each tracked company, compiled for a model to read: a cited profile and timeline, public metrics, a peer table, extracted facts, recent developments and the public record, one row per section. The same content downloads as a base or brief markdown document per company.
The AI Tooling Monitor catalog: every product the market scan has kept, curated categories plus emergent feature tags, the agent's fit read, and the machine's merged dossier and deep dive.
The event timelines of every non-dismissed Tooling Monitor product: launches, funding, feature and pricing changes, partnerships, changelog entries, and notable news, one row per event.
The Tooling Monitor's feature matrix in long form: one row per (cataloged or parked product, normalized feature tag), for who-does-what comparisons across a category.
The public AI Tooling Monitor catalog: every cataloged product with its category, what it does, who it is for, and its features. No agent scores, dossier, or curation detail.
The standing-hypothesis scoreboard: every saved thesis report with its matched-signal counts, direction balance, one-sidedness and thin-coverage flags, and a link to the full public report.
The portal's own schema, as data: one row per column of every dataset, with its type and definition. The machine-readable answer to what is in the Atlas.
Public datasets carry only the published layer of the Atlas. The key-gated exports add retained article text and machine-extracted records; the argument map's personal layer (confidence values, rationales, source reliability priors) never enters any download, and the one reviewer-set value that does ship, the research export's rigor prior, rides only behind the access key. Retained article text is an internal working corpus; link to the original source when sharing outward.