part of: the tour · chapter 5 of 12
People, events and documents
The hubs: who appears where, what happened when, and what each document actually said.
The entity page: names, and other names
The entity’s name with a type badge (person, organisation, government, asset, place), the other spellings it has appeared under — also written as — and the separate records that were resolved into this one.
Why Real coverage spells things several ways, and the merges the model made are decisions that could be wrong. Listing them means a wrong merge is visible rather than buried in an id.
Connects Entities are the site’s busiest hubs: nearly every claim names at least one, and each of those names is a link here.
See it live An entity and its names
“Possibly the same as” — doubt as public content
Where the model suspects two records are one thing but has not decided, it says so on the page: “The model has not decided whether these are one thing or two. That doubt is shown rather than resolved silently — a wrongly merged cast would quietly rewrite every claim on this page.”
Why Identity resolution is where a system like this can do the most damage invisibly: merge two people and every claim about either becomes a claim about a person who does not exist. Publishing the uncertainty makes that failure mode auditable.
Connects It is the same instinct as the residual bar — the unresolved case gets a place on the page instead of being rounded away.
See it live An entity page
Involved in, and Statements attributed to it
Two distinct lists: the claims this entity appears in, with the standing of each, and the claims this entity is on record as having made — each marked independent or restatement · zero weight.
Why Being talked about and doing the talking are different relationships, and collapsing them is how an actor’s own press becomes evidence about them. Keeping the two lists apart is what makes the second one safe to publish.
Connects A third list — Questions it figures in — takes you from a name straight to the open uncertainties it bears on. The statements list is the reverse of the Voices section on every claim page, and its “zero weight” marks are the repetition rule showing its work.
See it live Both lists on one entity
Publishers, and the number that is not there
An entity that published documents lists them — and where a track record would go, the page says “Track record: not yet computed.”
Why A reliability score beside a publisher’s name is the single most quotable number a site like this could invent, and inventing it is exactly what it must not do. Naming the gap is more honest than filling it, and more useful than hiding the section.
Connects It is the same empty-state discipline as “narrative ranking pending” and “no independent evidence yet”: state the mechanism, never apologise, never fabricate.
See it live A publisher’s record
Structural facts, in both directions
Membership, ownership, parthood and location — rendered on both entities involved, so the subsidiary lists its owner and the owner lists its subsidiary.
Why A one-directional structural fact is a fact you can only find if you already knew where to look. Rendering both ends is the linkability contract applied to the boring relationships as strictly as to the interesting ones.
Connects Events the entity takes part in are listed the same way, in time order, and lead to the timeline.
See it live Structural facts
The event page
What happened, when — rendered as the record states it, open bounds and all — who and what took part with their roles, where it happened, and the claims made about it with their standing.
Why Events are the anchors that let a story be read chronologically rather than as a pile of assertions. Keeping the time exactly as stated, including its vagueness, keeps that chronology honest.
Connects An event’s ordering section (what precedes what) leads into the timeline view, and its anchoring Questions lead back to the splits it belongs to.
See it live An event
The source page: two clocks on the document itself
Publisher, publication date, and the date this system read it — published 2023-03-07 · read 2026-07-19 — plus a link out to the original.
Why The gap between publication and reading is part of the fog: a document from March that was only ingested in July was not informing anything in between. The two clocks apply to documents as much as to claims.
Connects Reading time is knowledge time, which is what the ?asof scrub replays.
See it live A document’s two dates
“A source is testimony, not proof”
That sentence sits at the top of every source page, with the rest of it: citing a source is a provenance fact — it records who said what, where — and never a verification that the claim is true.
Why It is the single most load-bearing sentence on the site, because “it has a citation” is how most people decide something is established. Putting it on the document’s own page is putting it where the temptation is.
Connects The same wording governs the cited evidence chip on every claim page.
See it live The sentence in place
What it claims — the document’s own contribution
Every claim this document produced, each shown as the canonical claim beside the verbatim span that produced it, and marked for how it landed: independent, converging with an existing claim, or contradicting one.
Why It is the view that answers “what did this outlet actually add” — as distinct from what it repeated. That is the question syndication makes hard to answer and this page makes trivial.
Connects Each canonical claim links to its own page, where this document appears in the Sources list. Both directions, always.
See it live What a document claims
The text itself is never republished
Quoted spans and a link out — never the body. The page says so, and records a content fingerprint of the text it read instead.
Why Two reasons, and both matter: it is other people’s work, and a copy here would silently become the version of record while the original changed. Quote and link is the honest form.
Connects The fingerprint is what lets revisions of the same document be recognised as revisions.
See it live A document, quoted not copied
Syndication, in both directions
Which documents this one restates, which restate it, and the ones the model suspects but has not confirmed — all linked, both ways.
Why This is where the repetition rule becomes inspectable. You can start at any of forty stories and walk back to the one document doing the work.
Connects Restatement is why those forty count once in every number on the site, and zero in the ranking.
See it live A document’s syndication
How it was read
The extraction provenance: which model read the document, when, whether the run completed, how many claims it produced — and, where an earlier reading was superseded, that it was.
Why Everything on this site downstream of a document depends on an extraction that could have been wrong. Naming the run makes the dependency visible and the failure traceable.
Connects The same provenance appears as the Origin line at the foot of every node the run produced.
See it live An extraction run
Browsing everyone and everything
Two plain tables: every entity with how many claims it appears in, how many statements are attributed to it and how many documents it published; every document with its publisher, dates, how many claims it produced and how many other documents restate it.
Why They are the site’s countable view — how much of the record each thing touches. The pages are careful to say that degree is a fact about the graph and never a judgement about the thing, because a large number beside a publisher’s name would otherwise read as authority.
Connects It is the same restraint as “track record: not yet computed”: publish the count, refuse the implication.
See it live Every entity · Every document
When a node genuinely has no neighbours
A few records — usually publishers minted at ingest whose documents have since gone — have nothing attached. Their pages say exactly that, and are marked as such for the site’s own link checker.
Why A page that hides its neighbours and a page that has none are different things, and only the first is a bug. Declaring the second keeps the checker strict without pretending the corpus is tidier than it is.
Connects The dead-end test reports both lists every run, so nothing is silently exempted — the next chapter is about that check.
See it live The dead-end test · Browse entities