A sealed prediction records its text, criterion, timestamp or commit, and cryptographic digest before the relevant result is observed. The original stays immutable. Results are graded against the original criterion; failed and ill-posed predictions remain visible.
Sealing reduces hindsight editing. It does not prove three other things:
- that the hypotheses were independent of each other;
- that the prediction measured the quantity that matters;
- that no information leaked.
The second one bites. In EP-0004 the pre-registered retraction trigger did not fire, because the quantity chosen at sealing time was a poor proxy for the question it was meant to settle. Sealing is a procedure, not a guarantee of correctness.
Each note states what was known before sealing.