Predictions
A prediction is only interesting if there is a moment at which it becomes wrong. Most published forecasting never reaches that moment. These rules exist to make sure this one does.
What counts as a prediction
Four things, all of them required. Drop any one and the claim is discarded at extraction rather than stored as a maybe.
-
A direction
Up, down, above, below, happens, does not happen. Not "pressure is building".
-
A size
A threshold or a magnitude. "Higher" is not a claim; it is satisfied by any move at all.
-
A referent
A specific measurable thing — a closing price, a statutory release, the settlement of an open question on a public forecasting market — with an agreed place to look it up.
-
A deadline
A date by which it must have happened. Open-ended predictions are never wrong, which is exactly why they are worthless.
How it resolves
The resolution criteria are fixed when the question is registered, before the outcome is known, and they are not touched afterwards. Once the deadline passes, the grade is mechanical.
Inconclusive is recorded and counted, not deleted. The rate of inconclusive outcomes is itself a quality measurement — a rising one means questions are being registered sloppily, and that shows up before it can quietly inflate the win rate.
The three grades
Mechanical- Confirmed — the threshold was met by the deadline.
- Falsified — it was not.
- Inconclusive — the referent could not be established: the data was not published, the feed had a hole, the question turned out ambiguous in a way registration did not catch.
What gets published
Falsified predictions get published alongside confirmed ones, in the same table, with the same prominence. There is no editorial pass between the resolver and the record. This is not modesty; it is the only arrangement under which a published track record carries any information at all.
Each resolved prediction carries what it was, when it was made, what the public market price of the same outcome was at that moment, when it resolved, and how it graded. The market price matters: a call that agreed with a price you could already have traded at is not a contribution, however right it turns out to be.
The honest state of the record
Published track record
None yetThere is not yet enough resolved history here to publish a meaningful track record, and none is being claimed.
The reason is in the funnel figures: of 1,173 claims extracted in a real run, five reached a grade. The bottleneck is not the volume of material coming in — it is how few claims in the wild are stated precisely enough to ever be settled. Fixing that is the current work, and it means changing what the system asks for, not how much it reads.
When the resolved set is large enough to say something that would survive scrutiny, it will appear on this page, with the losses included.
Why publish the rules before the record
Because the order matters. Anyone can publish a track record; the interesting question is always whether the criteria were set before or after the results came in. Publishing the rules first, in public, with the scoring criteria hash-pinned in advance, is the only version of this that can be checked by someone who does not trust you.