The Breakout Was Real. The Recovery Wasn't.

By Zeeshan Mehdi · 18 September 2026 · 9 min read · 10 real scans, 63 photographs, 119 model calls, $3.27

Part two. The first post is The Number We Couldn’t Show You; this one assumes you’ve read it.

Two weeks ago I wrote that our Skin Score could see a spot appear but wasn’t allowed to say so. We shipped the fixes. Then I broke out for real, scanned every day, and checked every verdict against the pixels. The app was right on every day but one — and that one day taught us more than the rest.

01

What we said we’d do, and did

The first post ended with a list. Average three days of scans into one reading so a real breakout can clear the noise band. Guide the capture so consecutive photographs look alike. Stop marking people down for wearing sunscreen. Don’t loosen the threshold.

All of it shipped in the fortnight since, plus one thing the list didn’t have: the app now compares your scan to yesterday’s as well as to your very first one, because “what changed since yesterday” turned out to be a question the product literally could not answer.

02

Seven days of real scans, checked against the photographs

The breakout started on 4 September: one pimple on my cheek, which grew a head over two days, then two more appeared lower on the same cheek and one on my forehead. I scanned through the app every day or two, and kept the photographs, so every verdict could be checked against what was actually on my face — at full resolution, with the orientation matched by looking at which way the nose points rather than trusting the order the photos came in. (I got that wrong once in part one. Not again.)

Read down the score column against the middle one. The lowest score of the whole week is the morning the pustule peaked. Every day the lesions were there, both models called it: the comparison model reported an inflammatory change on five out of five scans across the peak, and the Skin Score’s own baseline comparison said texture — worse on every one of them.

One line deserves quoting. On the 10th the app wrote: “a few new mild papules present on forehead and right cheek.” There was a new forehead pustule, and there were new cheek papules, and none of them had been a papule four days earlier. That is not a generic acne sentence. It read a face that had changed in two places and named both.

The sunscreen fix held, too. The evening of the 10th was the shiniest capture in the whole set, and instead of “oiliness — worse” the app wrote “surface sheen slightly more visible under brighter overhead lighting”, flagged it as not confident, and moved on. Three scans after the fix, three times it attributed shine to the light rather than to my skin.

And the 18-second contradiction on the 7th — two guided captures, one “worse”, one “unchanged” — is the 33% false-positive rate from part one showing up in my own history. Still there. Still the same number.

03

The morning it said everything was fine

Then the last row. On the evening of the 10th the app said new papules, 76. Fifteen hours later, same face, it said everything was the same as my baseline, 82 — a six-point jump, exactly the size of the band. Had the lesions cleared overnight? I hadn’t touched them. So I measured.

For each lesion I took its redness relative to the skin immediately around it — a red-minus-green colour measure in a small disc on the spot, minus the same measure in a ring around it — which cancels out the overall exposure and colour of the room. Then the brightest point on each lesion, which is where a white head or a shiny cap shows up. Then the same numbers on plain skin, as controls.

Four of the five lesions were exactly as red the next morning. Nothing had cleared. What had changed was the shine.

Look at the plain-skin row. On the evening of the 10th, ordinary skin with nothing on it had bright points 16 to 38 units above its own median — the whole face was glossy, end-of-day oil or the evening’s moisturiser. Every raised lesion wore a specular cap, and a bright cap on a red dome reads as a white head. By morning the skin was matte, the caps were gone, and the same red discs were left with no relief to show. (One lesion, at the beard line, kept a genuine head both days. That one was a real pustule. The model saw it both times.)

So the “recovery” from 76 to 82 was not my skin. And, uncomfortably, the “new papules” of the evening before were partly the gloss too: the redness was real, the heads were light. The model’s read of a lesion depends on the surface finish of the skin around it.

04

The two things we were most afraid of

Before believing that, we had to rule out the two explanations that would have been much worse for the product. The morning scan was also the brightest capture of the week, and the one with the smallest face in the frame — a live guided capture puts the face at about 65% of the frame width, where a photo picked from the camera roll puts it at 85%. Either of those would mean the app misses breakouts as a matter of course, for everyone, on every ordinary scan.

So: one photograph, with a known breakout on it, manipulated one variable at a time, checked against a clean baseline three times per step.

Thirty out of thirty, down to a face at 55% of the frame — under a third of the pixels per lesion it had at full size. Then the same for light: brightness from 0.35 to 0.60 (real scans span 0.43 to 0.55), and two colour temperatures. Score range across all eight conditions: two points, against a repeat noise of one. The darkest version scored highest. And all 24 comparison judgements on relit identical skin came back zero.

The biggest fear didn't happen. The app does not miss a breakout because you held the phone further away or scanned in a dim room.

Six capture variables have now been manipulated directly on a single photograph — magnification up and down, framing, head angle, exposure, colour — and none of them moves either model. That leaves the one thing we measured in section 03 and could not manipulate: how shiny the skin was.

05

What we’re doing about the shine

The sunscreen fix from part one stopped the model from naming shine as oiliness. It does nothing about shine changing what a lesion looks like — that’s optics, not vocabulary, and no prompt fixes optics. The honest route is the same one that produced the sunscreen finding: a pair of photographs of my face minutes apart, once straight after washing, once after moisturiser, same room and light and distance. We haven’t taken it yet, and we’ve decided to ship first.

What ships now is two lines in the camera’s Tips sheet, because they’re the two things a user controls that the instrument can’t see past:

• Scan before moisturiser or SPF — shine changes what the scan sees.
• Same light and time of day as your last scan.

Not a fix. A boundary drawn around the instrument at the place we now know it is.

06

The fix that never fired

Part one had a section about three measurements that returned confident wrong answers in a single afternoon. Here are two more, and they were in the fixes we shipped.

When you average three days of scans, you must not count the same photographs twice — re-scanning yesterday’s photos is one measurement, not two, and counting it twice claims a tighter band than you’ve earned. We shipped that safeguard on 6 September, keyed on the storage IDs of the photos, on the basis that a re-scan carries the same IDs. The commit message said so.

A week later I pulled the actual database rows. A re-scan re-uploads the photos and gets fresh IDs every time. The safeguard had matched nothing. It shipped, ran on every scan, and did precisely nothing — while the capture measurements sitting in the next column over were byte-for-byte identical for every re-scan and would have caught all of them. It’s now keyed on those. Second one, same commit: the framing threshold was tuned against “four real captures”. They were two sets of photographs, each scanned twice.

A claim about database rows is made by querying the rows. Not from the memory of having seen them.

Neither would have been visible to a user. Both would have quietly made the averaged reading claim more precision than it had — which is the exact failure the averaging exists to prevent. The same measurement discipline that catches the model catches the people fixing it, and it only works if you actually run it on yourself.

07

Where this leaves the number

On direction, the Skin Score and the comparison behind it were right on every day the photographs could be checked. The week’s low point was the pustule’s peak. The one wrong day was wrong because of a thin film of moisturiser, and it would have been wrong in the same way for a dermatologist looking at the same two photographs.

On any single day, the number still can’t be trusted to the point — the 18-second contradiction on the 7th proves it, and the averaging that exists to fix that has only just started actually working. We’re not loosening the threshold. We said that in part one and the evidence since has made the case stronger, not weaker: the instrument has a real blind spot, we’ve found where it is, and we’ve drawn a line around it rather than pretending it isn’t there.

One person, one skin tone, one phone, one breakout. Same caveats as before. But this time the breakout was live, the scans were the real product, and every verdict was checked against the pixels. That’s the test that decides whether any of this deserves to be on your phone, and it passed the part we were most afraid it would fail.