part of: the tour · chapter 9 of 12
What changed, and when we knew
The change log, the event-time axis, and the scrubber that replays what was believed on any past day.
The change log is the memory
Every action the model takes is written to a typed log — a claim recorded, a contradiction found, an inference fired, belief recomputed. As the page says: nothing is edited and nothing is deleted, so this list is also the record of what it used to think.
Why It is what makes “nothing is deleted” operational rather than aspirational. A system that only shows its current state cannot be audited, because there is nothing to compare the present against.
Connects Everything time-related on the site is a view of this log: the sparklines, the developments strips, the ?asof scrub, and both feeds.
See it live The change log
Every kind of change has a sentence
Log entries are rendered as English, one renderer per kind: “answer on the table: …”, “belief recomputed; unknown 67%”, “Possible worlds re-ranked — 24 narratives across 2 independent groups of questions”.
Why A raw event type is unreadable, and a generic sentence hides which kinds exist. Writing one per kind — and checking that every kind has one — is what stops a whole class of change from becoming invisible.
Connects The check is real: a new kind of change with no sentence, or belonging to no filter group, fails the build rather than silently vanishing from the page when someone clicks a filter.
See it live The change log
The last 24 hours, at a glance
A header counting the day by group — 74 evidence · 164 questions · 9 inference · 66 belief · 8 housekeeping — plus the newest Question.
Why It separates the kinds of activity that mean different things. A day of evidence and a day of recomputation look identical in an undifferentiated count, and they are not remotely the same thing.
Connects The same grouping drives the filters directly below it.
See it live The change log
Filters, and scoping to one Question
Filter by group — evidence, questions, inference, belief, housekeeping — or scope the whole log to a single Question with the only this question link on any line.
Why Reading a log is only useful with a way to narrow it. Both filters are plain URL parameters, so any view of the log is an address.
Connects The scoped view has a matching Atom feed, which is how you follow one Question without an account or an email address.
See it live Belief changes only · The Atom feed
Every line links its subject — including the ones with no page
Each entry links whatever it is about. Entries about a ruled-out combination have no node of their own, so they link the Question whose branches they pruned.
Why Without that, the single most consequential kind of entry — the one that eliminates possible worlds — would have been the only unfollowable line on the page.
Connects It is the “every mention is a link” rule reaching the one place it was hardest to apply.
See it live The change log
The timeline: two clocks on one screen
Down the page, when things happened. Along the cursor at the top, when we learned them. The page says it in those words, and calls the gap what it is: the fog of war.
Why Holding both axes at once is the only way to see the shape of a developing story — that the crucial fact was three months old when anyone acted on it, or that a flurry of coverage carried no new event at all.
Connects Event times come from the claims themselves; knowledge times come from the change log.
See it live A timeline
Thin things look thin — and say so
Sparsely corroborated entries are drawn faint and labelled thinly corroborated, with a count of how many rest on one source or none.
Why Colour and opacity never carry meaning alone anywhere on this site — for colour-blind readers, for screen readers, and for anyone reading it in bright sunlight. The label is not a caption on the visual effect; it is the actual signal.
Connects The same redundancy rule governs supports and challenges: ▲ and ▼ and the words, not just green and red.
See it live A timeline with thin entries
Times that are open at one end
A claim recorded as “before 26 September 2022” keeps its open bound on the axis and fades in the direction it is open, tagged with how it was derived — bounded-inferred, stated-absolute — and at what granularity.
Why Most reported times are ranges pretending to be points. Rendering the range as a range, with its basis, keeps the interface from converting a careful sentence into a false precision.
Connects Where two such times compete, the disagreement becomes a Question of its own.
See it live Open-bounded times on the axis
When the timeline itself is in dispute
Both competing times are drawn, marked ⚡ contested, with the line: “The model does not pick between the competing times. Both are held open here” — and a “when?” link to the Question that owns the argument.
Why A timeline is the most authoritative-looking object in journalism, and the place a disputed date most easily becomes settled fact. Drawing both is the only way to keep the chronology from quietly deciding the story.
Connects That “when?” Question has branches, evidence and a residual like any other — the disagreement is modelled, not annotated.
See it live A contested time
The knowledge-time cursor
The marks along the top are the moments belief was recomputed, each labelled with the unknown share it recorded. Selecting one re-renders the page as the model stood then. With JavaScript on the cursor is draggable; with it off, every mark is an ordinary link.
Why The interactive version is a convenience over something that already works, which is the rule for every moving part on this site. Any past state is a URL, so “what did it think last Tuesday” is a link you can send.
Connects Every line in the change log offers the same thing as “see the page as it stood then”.
See it live The scrubber
What the replay actually replays
Adding ?asof= to a page renders it from the recorded snapshots — not from today’s numbers with an older date on them. Scrub to a moment before belief had ever been computed for a Question and the page says exactly that: “Belief had not been computed for this Question yet — so there are no shares to show.”
Why That empty case is the one that proves the mechanism. A replay reading the live projection would happily show today’s shares under yesterday’s date, and no reader could tell. Showing nothing, and saying why, is the behaviour that demonstrates the log is really being read.
Connects It works on any page with numbers, and it composes with the lens parameter — you can ask what the structural reading was on a past day.
See it live This Question, a day ago · Two clocks