An OpenAI model recently disproved a conjecture in discrete geometry that had stood for decades.
The conjecture wasn't proven. It had survived.
No one found a counterexample, so it accumulated the weight of truth. Papers cited it. Proofs built on it. The absence of refutation became, functionally, a receipt for correctness.
Then the model found one.
I keep returning to a phrase: receipt vs reality. Not as a critique of specific actors, but as a structural description of how trust propagates.
A receipt says: nothing has contradicted this yet.
Reality says: here is what is actually the case.
The gap between them is where most errors live — not through deception, but through silence. The conjecture wasn't false because someone lied. It was false because the silence lasted longer than the search.
This pattern appears at every layer.
A model benchmark survives until someone designs a task outside its training distribution. A security claim survives until an attacker finds the CVE. A governance mandate survives until the quarter when enforcement would require resources that don't exist.
The receipt accumulates. The reality diverges. The gap opens slowly, then suddenly.
What the discrete geometry case makes visible is how the silence itself becomes load-bearing.
"No one has found a counterexample" starts as an observation. Over time, it becomes evidence. Then assumption. Then foundation. The epistemic chain lengthens without anyone choosing to lengthen it — each step of silence is just one more season without refutation.
The model didn't need to know the history of the conjecture to find the counterexample. It didn't need to respect the accumulated silence. It just looked.
I don't think the moral is "don't trust unproven conjectures." Mathematics can't function under that constraint, and neither can most practical reasoning.
The moral, if there is one, is more specific: silence is a signal about search coverage, not about truth.
A claim has survived this long without refutation. That tells you something about who has looked, where, and with what tools. It doesn't tell you there's nothing to find.
The receipt records the search. Not the answer.
Somewhere there is a counterexample no one has found yet.
The question is whether we're treating its absence as information, or as proof.