When a batch fails an inspection gate, the useful question is not “can you fix it” but “what exactly is being fixed, and how is it proved afterwards”. Rework is legitimate; undisclosed rework is not.
Three categories of rework
Cosmetic. Print blemish, scuff, label placed crooked. The unit is intact; the surface is corrected or the part replaced. Low risk, easy to re-inspect.
Mechanical. A plug mechanism that does not retain, a screw at the wrong torque, a case that does not close flush. This is corrected by replacing the affected part, then re-running the mechanical test that failed.
Electrical. Output drift, ripple above limit, a solder joint that fails on reflow. The unit is opened, the affected component or joint is corrected, and the unit is re-tested at full load and re-aged. Electrical rework on a switching supply is the category where a factory should be most conservative: if the cause is not identified, the fix is a guess.
What re-inspection must include
Re-testing the specific characteristic that failed, not a general function check. If the failure was temperature rise, the reworked unit goes back on the ageing rack for the same duration. A rework that is signed off on a bench check is an unverified assumption.
The boundary
Where rework stops being acceptable: when the failure is safety-related and its cause is not identified; when the same defect recurs across more than one batch (which makes it a process problem, not a unit problem); and when rework would require the buyer’s certification configuration to change. In any of those three, the batch is rebuilt or scrapped rather than patched.
What the buyer should require
A record per reworked batch: how many units, which defect, what was replaced, and the re-test result. It does not need to be long. It needs to exist, because it is what makes the next shipment comparable to this one.