acceptance-gate — what "done" means

A result is done when it has survived a check that could have failed, run by a route that could not have known the answer. It is never done merely because the computation that produced it finished.

Why the gate exists: computation produces confident wrong answers at every stage, and none of them announce themselves. A fit converges to the wrong basin and reports small residuals. An integer-relation search "recognizes" numerical noise as a famous constant, because search enough constants and something always matches a short prefix. Two evaluators built on the same series expansion agree to great depth on the same wrong value. The agent that produced the answer — human or model — is the least qualified judge of it, because everything that produced the answer also produces reasons to believe it. So "done" is defined externally: agreement with an independent route, at points that entered no fit, to more digits than the fit could enforce, deepening when precision is raised, measured by machinery that has demonstrably failed on a wrong answer. Each clause of that sentence is there because its absence has, at some point, let a wrong result through.