This is a real finding from one of our own design reviews — caught by Probe, our PCB design-review tool. Here is what it read.

The most dangerous number in a review is one that is exactly right.

We were reviewing a USB-C to ATX supply of ours, and came to its 12 V rail. The feedback divider works out to 12.000 V. Not approximately. The reference voltage times the divider ratio gives twelve, to three decimal places. Nothing about the rail invites a second look: the nominal is correct, it is well centred, and there is 400 mV of margin to the board’s own power-good supervisor, which releases at 11.6 V.

The nominal was never the question. The question is the bound.

Start with the converter’s reference alone. Its datasheet specifies the feedback reference to within one per cent, and it does so across the whole junction-temperature and input-voltage range the part is rated for, which is stated in the heading above the table rather than in the rows. Reference tolerance by itself puts the rail between 11.88 V and 12.12 V. Still comfortable.

Now add the divider, and the rail splits into two different products.

Built with one per cent resistors, the worst-case low is 11.66 V. That clears the supervisor by 60 mV, which is thin but real, and the board works.

Built with five per cent resistors, the worst-case low is 10.82 V. That is 776 mV below the trip point, which means power-good could never assert at all. The supply comes up, the rail is within its own specification, and the board reports itself dead.

Why it can’t be decided from the files

Which of those two boards gets built is not decidable from anything on file, because the bill of materials states no tolerance for any resistor on the board.

That is the finding, and it is worth being precise about what the review did with it. It recorded the nominal. It recorded the bound it could actually establish, which was the one the reference tolerance alone supports. It named the field that was empty. And it declined to assume one per cent.

That last decision is the one we would defend hardest, because assuming one per cent was available and would have felt like diligence. It computes cleanly, it clears the threshold, and it produces a tidy pass. It would also have been a guess wearing a result’s clothes, and it would have removed the only thing about this rail worth escalating.

This is the shape of defect that reaches production. Every individual number is fine. The nominal is exact, the supervisor threshold is sensible, the margin at nominal is generous, and the board works on the bench in front of you. It fails later, on a purchasing decision taken by somebody with no reason to believe that a resistor tolerance on a feedback divider was load-bearing, because nobody had written down that it was.

The most useful thing a review produced that day was not a defect. It was the name of an empty field, and the reason it mattered.