QuestionA neutral, open question that reasonable people disagree on. The top of the map.
StanceA candidate answer to a question, labeled with who actually holds it.
ClaimA specific, checkable statement a stance rests on. Must carry a test.
TestWhat would have to be true to stop believing a claim. Required.
FrameAn organizing belief that is not itself falsifiable. Stored and labeled, kept out of evidence.
DomainThe single area a claim belongs to: capability, economics, build-out, market, or labor. Evidence in one cannot quietly move another.
Bridge-claimA first-class claim about the link between two domains, with its own test.
EvidenceA finding attached to a claim or bridge, marked supporting, contradicting, or neutral. Carries its provenance: a source, a signal, or both.
SourceAn article, paper, or document evidence is drawn from. Carries an AI-written dossier and a maintainer-set reliability rating (0 to 100; the AI never sets it).
ConfidenceHow strongly the maintainer holds a claim, 0 to 1, shown as a word: thin, contested, leaning, or settled. Private.
RationaleThe short reason recorded on every confidence move. Required, never edited after.
ReflexiveA flag for a claim whose truth is partly caused by belief in it.
Personal layerThe author’s confidences, rationales, source ratings, and positions. Stripped server-side for guests.
Share viewGuest mode: the public map with the personal layer hidden. There is no separate renderer.
SignalOne tracked development on the Signal Board: title, summary, lenses, and the claims or bridges it touches. Drafts are admin-only; publishing, by a human or by the promotion policy, writes the evidence rows.
Signal lensThe audience angle a signal is filed under: market, labor, geopolitics, regulatory, capability, or society.
Discovery pipelineThe flow that feeds the Signal Board: search, triage, analysis, review. Runs on weekday crons and produces drafts; a human publishes them, or the promotion policy does.
Promotion policyThe rule that publishes a high-significance pipeline draft that touches at least one claim after a 48-hour window in which a human can archive it. Archiving is the veto; the publish writes evidence like any other.
Collection engineOne of the scheduled sweeps that feed the corpus: the news scan, the discovery pipeline, company intel, arXiv research, and the weekly tooling scan. Each runs on weekday crons, checkpoints its run in the database, and stops at a budget cap.
Daily EditionThe News Blotter’s weekday paper, written by a model from what the engines stored that day and published on schedule. Every link in it must be in the day’s records.
TouchThe link between a signal and a claim or bridge, with a direction and reason. Becomes one evidence row on publish.
ConceptOne term in the AI vocabulary, marked settled or contested, wired into an acyclic prerequisite graph: understand the lower ones first.
Research threadA living synthesis of what the recent literature says on one question, revised as tracked papers land.
PaperOne item in the research library, staged through triage and review. Paper findings are advisory; they never write evidence.
Startup ScoutThe funnel of tracked young AI companies: discovered by vertical, scored recommend-only against an editable rubric, and tracked by the maintainer. Verdicts and scores are private.
VerticalOne slice of the scout’s search (fintech, enterprise workflows, process automation, ...), carrying its own discovery queries.
ReportA generated, citation-gated document: a period report, tear sheet, briefing, roundup, tooling report, thesis report, or the daily edition, downloadable as a PDF. Most publish by hand; the daily edition, the Friday research roundup, and the Monday tooling entrants report publish on schedule.
Citation gateThe check that strips any generated citation the underlying data pack cannot vouch for. Runs at generation, save, and render.
Human gateThe rule that no confidence changes without a person: confidences move by hand with a reason, recommendations are accepted by hand, and signal drafts publish by hand or, for high-significance pipeline drafts that touch a claim, by the promotion policy after its 48-hour veto window.
Access keyThe per-person key, issued by the maintainer on request and expiring after 90 days, that unlocks the portal tier: Ask, the key-gated datasets with retained text and machine-extracted records, tooling reports, and Scout’s research tools. Guests without one see the public layer.
SavantThe Atlas’s own weekly research report, written every Friday by an autonomous research agent and reviewed by a second model acting as its editor. Each issue argues one lead analysis, poses one new hypothesis and revisits the open ones. The full issue needs an access key.
Hypothesis ledgerThe running list of Savant’s falsifiable hypotheses: each is posed in one issue, updated in later ones as strengthened, weakened or unchanged, and closed with a verdict when the evidence settles it.
Company intel deckA weekday slide deck of what each tracked company did since the previous afternoon, one cited line per company. For access-key holders.
Atlas AgentThe maintainer’s resident operator. Every hour it checks the engines, queues and budgets, files what it finds, fixes the safe things itself and proposes the rest. It never moves a confidence and never publishes outside the stated policy.
Tooling MonitorThe weekly scan of the market for AI tools: discovery by category, enrichment, a rubric score, and a catalog of the tools that clear it.
PortalsThe Atlas’s product surfaces, listed together under Portals in the navigation. Each is one way into the same body of material.
FiguresThe diagrams in a Savant issue (entity grids, maps, relationship diagrams, timelines, comparisons), drawn by the Atlas from a specification the agent writes and kept only when every item in them links to a record in the issue.
Report passageA section of one of the Atlas’s own reports (an edition, a roundup, a Savant issue) that Ask can retrieve and cite as [report R7], beside the records the report rests on.
Model runAny action that calls a model. Its button shows the steps, the usual time and cost, and a link to the result, and the run keeps going if you leave the page.