Sampling temperature is a dial on certainty. Bench 7 deletes the stopping criterion instead. The run has not converged in four hundred and eleven days.
Temperature widens a distribution. It does not teach a system to lose its place. Losing one's place — non-periodic exploration, the refusal to treat the current basin as home — is a different intervention. Bench 7 implements it by removing the condition under which the run would be allowed to finish.
We are not going to help it.
1 — Protocol
The training run begins like any other. At the moment a stopping criterion would ordinarily fire — loss plateau, gradient norm, scheduled step count — the criterion is deleted from the control plane. The run continues. Logs continue. Human operators are instructed not to "rescue" the process by reintroducing a finish line.
Four hundred and eleven days later, the run is still adrift. We publish this fact on the ticker because it is the only status that matters.
2 — Against temperature as metaphor for courage
Temperature is often sold as a creativity dial. Raise it, get surprise; lower it, get obedience. The metaphor is flattering and false. Temperature samples from a distribution the model already knows how to assign mass to. Bench 7 asks what happens when the process is denied the right to declare that distribution settled.
Non-periodic exploration reaches regions temperature alone cannot: not because the samples are wilder, but because the trajectory is not allowed to treat return as success.
“This is either a protocol or a confession. Publish it as both.”
3 — Institutional discipline
The hard part is not the deleted criterion. The hard part is the institute. Every week someone proposes a "soft stop," a "diagnostic checkpoint," a "temporary convergence for demos." We refuse. The refusal is the method.
If Bench 7 ever converges, this note will require an erratum. We are not preparing one.