What a seal is not · why that is deliberate
An EvalSeal signature proves less than you would like. It does not tell you a model is good, that the evaluation was well designed, that the tasks were the right tasks, or that whoever ran it knew what they were doing. It proves that these bytes are the bytes that existed at that moment, and that nobody has touched them since. That is a small claim. Every larger one on offer in this field is unfounded.
Each of these would be better than a receipt, if it worked. None of them fail because nobody has tried hard enough — they fail at the premise.
Which does not make it ninety-seven percent useful. It makes it unusable as evidence, because the party a detector accuses is the party with the motive and the budget to litigate that three percent. One publicised false accusation ends the detector, not the accused.
Model cards and methodology sections are prose, written afterwards, by the party being described, about an experiment nobody else can re-enter. Nothing in the document is bound to what actually executed. It reads as transparency and functions as a press release.
The classical answer, and it genuinely works — after forty years and a building full of people. It also carries a trap: it is the sentence trust us, spoken by a different mouth. A new institution cannot climb out. Its stamp is worth nothing until it is trusted, and it is not trusted until the stamp has been worth something for a long time.
Aggregate enough claims and the noise cancels. What is left is an unverifiable number with a tighter error bar and more authority than it earned. Aggregation improves precision. It does nothing for provenance.
A receipt makes no claim about quality. It claims occurrence — and its worth does not depend on who issued it. That single property is the one nothing else on the list has.
Every property here is deliberately weak. Weak claims are the only ones a new party is entitled to make, and the only ones that survive being checked.
A receipt does not make anyone honest. Nothing does.
What it changes is where dishonesty has to live. An unverifiable claim can be wrong quietly and indefinitely, because there is no moment at which it meets anything that could contradict it. A sealed record moves the same claim into the open, where it has an address, a timestamp, a signing key and a name beside it. Somebody can point at it. That is not a moral improvement — it is a structural one, and it is worth more than a promise.
Which is why a seal is issued expecting attack. A record nobody has contested is not stronger than one that has survived contest. It is only untested.
A record can be perfectly sealed and completely worthless. A badly designed task set, signed, is a badly designed task set with a signature on it. Nothing in the cryptography improves an evaluation, and nothing in it makes a number mean more than the run behind it earned.
A seal is a floor, not a ceiling. It rules out one specific failure — the record changing after the fact — and leaves every other failure exactly where it was, in the open, for someone to find.
Evidence that depends on trusting its author is not evidence. Remove that dependency and what remains is very small — and it is the only part worth publishing.