How a claim gets on this ledger, and how it gets a verdict

How predictions are found

We watch each appearance through its automatic subtitles, clean them into readable text with timestamps, and use a language model to find candidate claims. Which appearances we watch is set per forecaster by the intake rules published below, and the channels we watch are listed at the foot of every page. Nothing the model produces is published automatically.

The falsifiability bar

A claim only qualifies if a neutral third party could, at some future date, judge it true or false against public evidence. Vague directional statements, opinions, and commentary about the past do not qualify. Conditional claims (“if X, then Y”) are marked as such.

The review gate

This section said, until 3 October 2026, that every candidate is reviewed by a person before it appears here and that the gate is permanent. That is no longer true, and the change is described here rather than quietly made, because a scoreboard that revises its own method without saying so is worth nothing.

From 3 October 2026 a claim can reach this ledger without a person approving it. How many actually have is counted below, from the record, rather than described here. Before any claim is judged at all, a set of checks written in ordinary code has to pass, and each one of them exists because of a specific failure:

• The quote must be found in the transcript of the appearance, in a stretch the speaker himself is talking. A guest’s words were once published under a forecaster’s name, and were withdrawn.
• The quote must be his own words, 25 at most, and long enough to carry the claim it is evidence for.
• The deadline must come from a time the speaker named himself, and those words must appear in his quote. If he named no time, the claim is published with no deadline and says so, and is never scored against a date. An audit found 165 of 218 queued candidates carrying a deadline nobody had said; the rule that catches them is a piece of plain text matching, not a judgement.
• The paraphrase must be free of advice-style wording.

Only what passes all of that is judged at all, and the judgement is narrow: is the claim falsifiable, and does the quote really contain what the paraphrase says it does. A claim that passes is published. A claim the reviewer is unsure about is left for the editor, and so is anything the checks above flagged.

Nothing is ever set aside by machine.A wrong publication is visible on a public page and can be withdrawn, and the history of it stays. A wrong rejection would be invisible: nobody, including us, would ever learn that a real claim had been thrown away. So only the generous direction is automated. Refusing a candidate is still a person’s decision, and the editor can withdraw anything the machine published.

How scoring works

A prediction stays pending until there is public evidence to resolve it. It is then marked confirmed, partial, wrong, or unverifiable. Every resolved verdict carries at least one public evidence link, and the reasoning is stated plainly. The headline accuracy figure is computed from the resolved calls only; it is never adjusted by hand.

Since 7 September 2026 a verdict can be issued by machine rather than by a person: the research is done automatically, the verdict is drafted from it, and it is published without anyone reading it first. Of the 47 verdicts standing today, 10 were issued that way and 37by the editor. Counted from the status history, not typed by hand. Each verdict says which it was on the prediction’s own page, along with the model that issued it. A machine will not score a claim against a deadline the speaker never gave, will not call a claim wrong before its deadline, will not issue the first verdict against a newly tracked forecaster, and will not re-settle a claim whose earlier draft the editor rejected.

This is stated here because the site asks readers to hold a forecaster to his record, and it would be a poor thing to do that while being vague about who keeps the record. The change was made on the measured agreement between the machine’s drafts and the editor’s decisions across 48 earlier verdicts. Those samples are small, so the honest form of them is a range rather than a figure: agreement on confirmed calls was 86% and could lie between 65% and 95%; on wrong calls it was 38% and could lie between 21% and 59%; on partial calls the machine and the editor agreed every time, but only six such calls have ever been judged, which is consistent with a true agreement rate as low as 61%. The editor can override any verdict, and every change stays in the public history.

What settles a pending claim

A pending prediction says in one sentence what would settle it. Where the claim has moving parts, those parts are listed as named conditions, and each one is marked observed, not observed, or contested. A condition marked observed always carries the public sources it was read from, so the reading can be checked rather than trusted, and the count is shown with its denominator: three of five named conditions observed, never a percentage.

Condition readings are drafted by an evidence monitor that watches the public record for events touching a pending claim, and they are published only after the owner approves them. The monitor cannot publish anything itself. A full set of observed conditions is still not a verdict: it means the claim looks ready to judge, and a person judges it.

Some claims have nothing that will settle them. A claim comes due when its end date arrives, and a claim can have no end date at all, because the speaker named no time for it. Those are marked no end date wherever they are counted: he named no end date. He named no time for these, so no date ever arrives, and they are never scored. They stay on the record as unjudged, and nothing here is getting round to them.

That is how the machinery works rather than a queue we are behind on. The research step asks for the claims whose end date has passed, and a claim with no end date is in no such list: nothing picks it up, so no verdict is ever drafted for it, and the rule stated under scoring above stops a machine scoring it in any case. The editor can still judge one by hand, and that is the only way one of these ever carries a verdict.

This is the cost of refusing to invent a date, and it is stated here rather than left for a reader to discover: a claim with no end date can stand unjudged for as long as the ledger does, and no amount of waiting will change that. We would rather pay it. The alternative is to pick a deadline the speaker never gave and score him against it, and that is worse, because it scores our own invention and publishes it as his record.

Who we track, and what counts

We track named people who make specific, checkable claims about the future on camera. Adding someone is a decision we publish here, dated, before any of their claims appear on the ledger, so a reader can hold us to our own rule.

Prof Jiang Xueqin, from 8 May 2024. Predictive History is the channel we watch Prof Jiang Xueqin on, and every lecture on it is watched in full. Codes begin JL.

Prof Robert Pape, rule published 30 July 2026, not yet active. There is no channel of Prof Pape’s own to watch: the on-camera claims appear in guest appearances, chiefly on Mario Nawfal’s channel. So the rule has to be narrower, and it binds us as much as it binds Prof Pape:

• Only appearances where Prof Pape speaks are watched, identified by the name Pape in the video title. Codes begin RP.
• We score the words as captioned, never a video title. Titles are written by the channel hosting the appearance, not by the speaker, and they routinely sharpen what was actually said. A claim we cannot anchor to Prof Pape’s own captioned words is not tracked at all.
• At most five appearances by Prof Pape enter the queue in any rolling seven days. Where there are more than five, the excess is not silently dropped: each declined appearance is recorded with its date and title, and the count is published below, so the gap is visible rather than hidden.
• Claims about what Prof Pape thinks, believes, or is likely to mean are not tracked. Only claims about events that either happen or do not.

Nobody named on this page is affiliated with this site, and nobody has been asked to take part. Everyone named is welcome to reply to anything here; the corrections route is on the disclaimer page and it is open to the subject of a prediction before anyone else.

The evidence standard

A verdict rests on public sources anyone can check: primary reporting from established news organisations, official statements and documents, and direct records of the event itself. Research is machine-assisted, but every source behind a verdict was opened and read before it was cited, and the links are published with the verdict. One source is the floor, not the target; contested claims get corroboration from outlets that do not share ownership or wire copy.

A verdict is our judgement of that record, stated as opinion, decided by a person. The machine drafts; it never decides. Where the evidence genuinely cuts both ways, the verdict is partial and the reasoning says what is debatable, rather than rounding to a cleaner answer. A headline hit rate appears only once twenty predictions have resolved, because a percentage built on fewer misleads.

Ongoing events: the settlement rule

Some predictions concern events still unfolding, like a war that may yet escalate. A claim with a deadline is judged against the world as it stood at that deadline. What happens afterwards does not change the verdict; it is appended to the prediction’s page as a dated development, so the record grows without being rewritten.

A published verdict is revised only if it was wrong on the evidence available at settlement, and the history of every status change stays visible on the page. Nothing is silently edited.

When a source changes: corrections

Sources move after publication. A story is rewritten, a page is pulled, an outlet corrects itself. When a source cited behind a verdict changes in a way that matters, we replace that citation with primary documents and independent reporting carrying the same fact, and we append a dated correction note to that verdict’s public history saying what changed and why.

Nothing already published is rewritten. The original note stays where it is, the correction sits below it with its own date, and the status history keeps every step. The record is append-only, so a reader can see what we said, when we said it, and what later changed our reading of it.

Anyone can ask for a correction, and the subject of a prediction is first in the queue. The route is on the disclaimer page.

Honest caveats

• The resolved denominator is small and will stay small for a long time, so early accuracy numbers carry wide uncertainty. We state the denominator openly rather than hiding it.
• Predictions often cluster around the same underlying event, so a single correct or incorrect call can move several verdicts at once. Treat the count as a record, not an independent sample.
• What we find depends on subtitle quality and on judgement about what counts as a claim. We show the exact quote and a link to the source so you can check every call yourself.