Before changing a limit, collect recent real failures, older good images from the line, and recent good images spanning normal variability.
Factory samples are a weak baseline because they freeze the process at buyoff. Compare the piles in reverse milestone order: HMI, I/O, math, tools, fixture, image. Watch the custom-view beacon beside the image and note whether fails track light, position, part appearance, or cycle time. Then add temporary acquisition in a scratch copy: delay lines, charts, overlay graphics, on-screen values, and timers. Insert rows or use the right margin so the diagnostics do not collide with the job. Setpoint changes should stay inside ranges justified by a study. Retraining a pattern or character model needs a copy of the old model first. If the old model carried an angle, the new train resets that angle to zero, so the previous offset has to be restored as a constant or robot continuity breaks.