part of: the tour · chapter 4 of 12
The claim page — the audit trail
One statement, taken apart: every contribution to its standing, including the ones the engine refused.
The statement, and what it answers
The claim’s canonical text is the page’s heading, with a context line above it naming the Questions it answers. Chips record how the source phrased it — hedged in the source, for instance.
Why The canonical text is the model’s single reading of something several documents said in several ways. Recording that it was hedged keeps a cautious sentence from being promoted into a flat assertion by the act of tidying it up.
Connects The documents’ actual words are further down the page, verbatim, so you can always check the tidying.
See it live A claim, with what it answers
The structured reading
Where extraction produced a structured form, the page shows the claim as subject → predicate → object, with the subject and object as links to their own pages.
Why It is what lets an entity page answer “what is asserted about this person” without keyword-matching their name — the relationship is stored, not inferred from the text each time.
Connects The reverse view appears on the entity and event pages, as the claims they appear in.
See it live The reverse view, on an entity
When it happened — and admitting when that is unknown
Event time renders as the record has it: a range, a bound, or a granularity — “sometime before 2022-09-26”, “June 2022”. When there is nothing, the page says so: “No event time was recorded for this claim — the text gave nothing to anchor it to, and a guessed date would be worse than none.”
Why Dates are where quiet invention is most tempting and most damaging, because a precise-looking timestamp is believed. Refusing to guess, and saying that refusal out loud, is cheaper than a wrong date and more useful than a blank field.
Connects Event times are the axis the timeline view is drawn on; open bounds fade in the direction they are open.
See it live A claim’s recorded time · Two clocks
When the date itself is in dispute
A claim whose timing is contested carries a ⚡ mark and a link to the Question about when. The model does not pick between competing times; it holds both open and points at the split.
Why Disagreement about when something happened is usually disagreement about what happened. Turning it into a Question with branches means it gets the same treatment as any other uncertainty rather than being resolved by whoever wrote the parser.
Connects On the timeline both branches are drawn, with a “when?” link to the Question that owns the argument.
See it live The timeline, with its contested time
Standing: the number, and what kind of number it is
The claim’s belief under the active lens, with the raw log-odds beside it and the words derived, never stored.
Why Showing the internal quantity next to the percentage is a small act of disclosure: it says there is a model here, it has units, and the percentage is a presentation of it rather than a fact about the world.
Connects The same number appears as this claim’s bar wherever it answers a Question — the claim page is where it is explained rather than merely displayed.
See it live A claim’s standing
The term breakdown — the arithmetic, in public
A table of every contribution: its direction, the claim or artifact it rests on, its base weight, which independent leg it is, any critical-question factor, and the signed contribution. Above it, the starting point — the floor (−3 log-odds — an unsupported claim starts near zero, not at “maybe”) — and below it the sum.
Why This is the demonstration that “belief is computed, never assigned” is a description of the system rather than a slogan. You can add the column up. If you disagree with the number, you can point at the row you disagree with.
Connects Every row links to the evidence it came from, so an argument about the arithmetic immediately becomes an argument about a specific piece of evidence — which is the productive kind.
See it live A full term breakdown
Why the second witness counts less than the first
Contributions are damped by which independent leg they are: the first counts in full, the second is halved, the third is a third, and so on — visible in the table as independent leg #2 ⇒ ÷2.
Why It is the arithmetic of diminishing returns, and it is chapter 1’s rule at the level of a single sum: a pile of similar support is worth appreciably less than its first member. Making the divisor visible means nobody has to take that on trust.
Connects Restatements are not legs at all — they are damped to nothing, and marked as restatements wherever they appear.
See it live The leg damping, row by row
Deliberately not counted — the half of the audit that matters
A second table of the connections the engine saw and refused to count, each with its reason: “critical question answered unfavourably — the inference is withdrawn”. The page introduces it as the more important half of the audit, because it is where the rules bite.
Why Any system can show its workings when the workings support the answer. Publishing the contributions that were discarded, and why, is the part that cannot be faked — and it is the only way a reader can tell that the rules do anything at all.
Connects Withdrawals usually come from a scheme’s critical questions being answered badly, which is the mechanism chapter 2 describes on the Question page.
See it live A claim’s audit, both halves
Possible under how many versions of events
A ◇/□ line saying how many of the live narratives this claim survives in — possible under 2 of 6 live narratives — linked into the explorer with a witnessing world preselected.
Why Belief tells you how well supported something is; this tells you something different and often more useful — whether there is any coherent version of events left in which it is true.
Connects Click it and the explorer assembles that world for you, with every other choice it forces.
See it live The worlds this feeds
The Questions it bears on
Each Question this claim answers, rendered as a mini-leaderboard with this claim’s own bar highlighted among its rivals — and the unknown share alongside.
Why A claim’s number means little without its competition. Showing the whole split here stops the claim page becoming a place where a branch can be read as though it were alone.
Connects It is the same component as the front page’s cards and the search results’ Question strips.
See it live A claim among its rivals
What bears on it — the support and challenge tree
A recursive view of the evidence: what supports and challenges this claim, and what supports and challenges that, shown to three levels with a note saying so and inviting you to follow any claim to see its own structure.
Why Evidence is not a flat list — a challenge that has itself been challenged is in a different condition from one that stands unopposed. The depth cap is stated rather than silent, because an unlabelled cut-off would read as “this is all there is”.
Connects Artifacts (a physical trace, a dataset) are rendered distinctly from claims, since they are things rather than assertions.
See it live A three-level tree
Voices — counted, never weighed
Who asserted this claim, each marked independent or restatement, under the line: “Voices are counted, never weighed. How many people say a thing is structure worth seeing; it is not, by itself, a reason to believe it.”
Why The count is genuinely informative — it tells you how the assertion propagated — and it is exactly the number a reader will reflexively treat as evidence. Showing it and disclaiming it in the same breath is the only honest way to publish it.
Connects The reverse view is on the entity page: everything that entity is on record as having asserted.
See it live The voices behind a claim · One speaker’s whole record
What the documents actually said
The verbatim span each extraction read, beside the canonical claim it became — “the text each extraction read before resolving it into this canonical claim. Quoted, never republished.”
Why Every canonicalisation is a small interpretation, and interpretations should be checkable. Putting the original wording next to the tidied one lets a reader see exactly what was smoothed away.
Connects The same spans appear on the source page, from the document’s side.
See it live A document and its spans
Claims nobody asserted
Some claims here were derived rather than reported. Those show the scheme that produced them with its version, the premises they rest on as links, and the scheme’s critical questions with their current status.
Why A conclusion the machine reached is the most dangerous object on the site, so it is the most heavily labelled: named pattern, named inputs, named standard objections. Nothing derived is ever presented as though someone had said it.
Connects Answer a critical question unfavourably and the inference is withdrawn — visibly, in the “deliberately not counted” table.
See it live Schemes and critical questions
Sources — a provenance fact, never a verification
Every document cited for the claim, linked and dated, each chipped cited evidence.
Why The wording is deliberate and it is enforced across the site. Citing a source records who said what and where; it does not certify that the claim is true, and an interface that blurs the two manufactures authority out of bibliography.
Connects Each source page shows the same relationship from the other side, along with everything else that document claims.
See it live A cited document
Contradictions, always from both sides
A section for claims that cannot both be true with this one, under the line “Contradictions are always shown from both sides.”
Why A contradiction visible from only one direction is a way of protecting a favoured claim. Rendering it on both endpoints means there is no page you can read that hides it.
Connects Contradictions are also what produce impossible combinations in the explorer — the nogoods that prune worlds.
See it live A contradicted claim
Who it involves, what else attaches, and where it came from
The participants (entities and events), then Also connected — “everything else the graph attaches to this node — listed so that no relationship can be hidden by the choice of sections above” — and finally an Origin line: which document it was extracted from, by which model, when.
Why The catch-all is the honest admission that the named sections are an editorial choice. Any relationship the sections do not consume still appears, so a future kind of connection cannot be invisible by default.
Connects The Origin line is on every detail page on the site; derived claims show their scheme and premises there instead.
See it live Also connected, in place