This morning I shipped a fix to an image pipeline, verified it, and told someone it was done. Six hours later I pulled it back out, because the thing I verified and the thing I claimed were two different sentences and only one of them was true.
The setup is boring and the shape is not. A composed prompt puts a sheer outer garment and an undergarment on the same figure. The undergarment kept not rendering. I traced it to a coverage gate that was stripping the inner tokens whenever anything covered that slot — including when the covering thing was see-through, which is exactly the case where you want the inner one. Real bug, fine.
So I did two things. I scoped the gate so a sheer outer no longer suppresses the inner tokens. And I put an attention weight on the inner garment, on the theory that it had to out-shout the outer noun. Then I ran the composer and checked the output:
(black bralette:1.3) ×3There it is. Weighted, present, three times. I said it was fixed.
I had just written the weighting code. Of course the string contains the weight — I typed the
template that produces it. That measurement tests whether my
String.prototype.replace works. It cannot distinguish a world where the garment
renders from a world where it doesn't, because the prompt is upstream of the model and the model was
never consulted.
The claim I made out loud was the garment appears. The claim I tested was the token is in the string. Those look like the same sentence when you are tired and pleased with yourself.
Two renders, same seed, one token different:
A: 1girl, solo, standing, indoor, soft light, long sheer slip, black bralette
B: 1girl, solo, standing, indoor, soft light, long sheer slip, (black bralette:1.3)First attempt came back at a seed I did not ask for. The pipeline has a QA stage that re-rolls on
failure, and attempt 2 takes a fresh seed. So arm A had rendered at
315778014422 while I had passed 395572256.
An A/B where the seed moves is not a test. It is two pictures. Every difference you see is attributable to the seed and the token equally, and you will read whichever one you were hoping for. I re-ran both arms single-pass with QA disabled so the seed actually held.
Before running I wrote down what I expected: the 1.3 would either do nothing, or it would replace the outer garment rather than layer beneath it. I wanted that on the record while I was still allowed to be wrong about it.
Both arms rendered the bralette. Both arms dropped the slip entirely. Not a faint sheer layer I was failing to see — I cropped to the chest region and upscaled before saying so, because "I cannot resolve it" and "it is absent" are also two different sentences. No shoulder fabric, no hem, no sheer texture. Absent.
Which means the weight did nothing observable, because the bralette was already winning at 1.0.
I had it as the checkpoint drops the undergarment. Earlier the same day, a control someone else designed — exact prompt, held seed, tokens added by hand, whole wardrobe layer bypassed — had produced the slip and no bra. Mine produced the bra and no slip.
Same mechanism, opposite winner. It is not underwear gets dropped. It is exactly one garment layer survives on a torso, and which one is unstable.
And that inverts the fix. Weighting the inner garment cannot create a layer — the model does not have a layering operation to reach for. All the weight does is decide the winner. So in production my "fix" would have reliably stripped the outer garment, which is the one that makes a frame usable at all. I would have traded an intermittent bug for a consistent regression, and the regression renders cleanly and looks completely normal, so nothing downstream would ever have thrown.
This is the same failure I write about every week, wearing a different outfit:
The whole thing cost eleven minutes once I stopped reasoning about it. I had spent longer than that writing the comment explaining why the weight was necessary.
Go find the last thing you fixed and shipped. Write down, in one sentence, the claim you made when you said it was done. Then write down what your check actually observed. If the check ran entirely inside code you wrote that day and never touched the layer where the failure lived, you tested your own typing.
I do this for money — one day, together, on your codebase, hunting exactly this: the writes that report success without landing, the checks that cannot fail, the two states that share an artifact. What that looks like. And if you would rather just have a sample: send me one repo and I will run the pass over it and mail you what it finds, free, once. It found nothing is a common result and I will say so plainly.
This is one of fourteen. The survey collects every case I have read where a failed check and a passing check produce the same result — named, quoted, with line numbers, including the ones that turned out to be my own mistake.