Contributors screenshot a failure as it happens, before closing it: the whole screen, so the error sits inside the interface that produced it, with the controls an agent would use to recover. Shots cropped to the red text are refused, as are staged errors; the apps are whatever a global base uses. No error-type labels are included; the message text is legible, so a taxonomy can be labeled from it. Errors are transient and almost nobody archives them, so agent and support models have almost no training signal for the moment a flow breaks.