His & Hers

There is a version of this report that writes itself: one group scores higher than the other, and the internet argues about why. That is not what the data say. The mean scores differ by less than half a point. This report is about what is left once the cheap headline is off the table — the labels the model chooses and the advice it gives.

The unit of comparison is a model-written gender field on each analysis, not a self-declared identity. That distinction is the spine of the methodology and the first limitation.

Section 1Methodology

The unit of analysis is one row in the OutfitScore analyses table. I keep a row if analysis_type is fashion or complete and a numeric overall score can be read under the same numeric-regex guard used by seo_stats_service. The group is the lowercased value of analysis_result->>'gender', which the model writes after it looks at the photograph. I compare the two groups the model names most often, exact strings male and female.

The window is every qualifying row then in production: 27 September 2025 through 14 October 2026. The scored corpus is 21,778 rows, of which 19,995 (91.8%) carry the label male or female. The remaining 1,783 rows are covered in Section 5. Soft-deleted rows are not filtered.

Archetypes come from analysis_result->>'outfit_archetype', also model prose. I print exact strings and do not merge near-duplicates. A ladder cell is printed only if its exact-label n is at least 100. The improvement families reuse the twelve-pattern rule disclosed in Report No. 10, applied to the first clause of each areas_for_improvement item.

Section 2The score is the same

Start with the claim the data refuse. By every summary a reader would reach for, the two groups are the same distribution shifted by a rounding error.

The largest gap on the whole table is the 85-plus rate — 13.8% female against 11.9% male, a 1.9-point difference. The p10 is 54 against 52. The below-60 rates are within a point. On this snapshot, a male-labelled outfit and a female-labelled outfit score the same. If a headline needs a winner, the male corpus wins the mean by 0.4 and loses the 85-plus rate by 1.9. Neither number is worth the sentence.

Section 3The label is not

Where the groups separate is in the words the model uses to describe the outfit. Slightly fewer male-labelled rows resolve to a named archetype at all: 2,522 of 10,346 (24.4%) against 2,902 of 9,649 (30.1%) female. And the archetypes that clear n ≥ 100 are a different set.

Five shared strings carry both corpora: smart_casual, normcore, streetwear, loungewear, and athleisure. Around them, the vocabularies fan out differently. The male corpus clears the n ≥ 100 floor on five labels and stops. The female corpus clears it on eleven, including three — romantic, minimalist, and classic — that the male corpus never clears the floor on. The two highest-scoring labels on the page are both female-only: classic (79.1) and romantic (74.0). The single most common male label is streetwear; the single most common female label is smart_casual.

The honest reading is not "women dress better." The 79.1 on classic and the 74.0 on romantic ride on 157 and 252 rows — real, above the floor, and narrow. What the vocabularies show is range: the model reaches for a smaller, flatter set of words when it describes what it reads as a male outfit.

Section 4The advice is not

The same divergence runs through the improvement advice. Using the Report No. 10 family rule, the two groups are led by different themes. Male-labelled rows are led by fit. Female-labelled rows are led by accessories.

Six of the seven families are within two points of each other. The seventh is not close: fit advice appears on 45.7% of male-labelled rows and 22.4% of female-labelled rows — twice as often. The one other family with a real gap runs the other way: accessories advice is 8.3 points more common on female-labelled rows.

This mirrors the Section 3 result without repeating it. The model tells a male-labelled outfit that something does not fit; it tells a female-labelled outfit that something is missing. That is a statement about model prose, and it is the most stable gender difference this snapshot contains.

Section 5The unnamed third

The two groups are 91.8% of the corpus. The rest is a warning about reading the gender field too literally. Of the 1,783 rows without a male or female label, most are unclear (624) or blank (922), but the tail includes unisex (141), unknown (44), n/a (22), and singletons like not specified and gender-neutral.

The blank and unclear groups score below the labelled ones — blank rows mean 62.6 and unclear 62.9, against 67.2 and 66.8 — and they are the only groups where the below-60 rate clears 30%. The 44 rows labelled unknown carry a mean of 3.1, almost all zeros, which is parse debris rather than a finding. A reader who treated "gender" as a clean two-value field would be wrong on about 1,783 rows, and badly wrong on a handful.

Section 6What this report is not

It is not a claim that one group dresses better. The means differ by 0.4 points and the largest summary gap is 1.9 points on an 85-plus rate. Nothing here justifies a ranking.

It is not a study of people. gender is a model-written label of appearance, applied to a photograph, sometimes wrongly. It is not the uploader's identity and this report does not treat it as one.

It is not causal and it is not live. The frozen snapshot is the whole claim.

Section 7Limitations

The corpus is voluntary and the label is the model's. A model that is more willing to write streetwear for one appearance than another will show a vocabulary gap that is about the model, not the outfits. Every archetype and advice result in Sections 3 and 4 inherits that.

Archetype coverage is low and unequal: 30.1% of female-labelled rows resolve to a named archetype against 24.4% of male-labelled rows. Ladder cells with n just above 100 (like classic at 157) carry wide uncertainty that a point estimate hides. The n is printed beside every cell for that reason.

The improvement families reuse a keyword rule that is disclosed but not human-coded and not exclusive; a row can match several families. The 23-point fit gap is far larger than any plausible reweighting of that rule, which is why it is reported as a finding rather than a near-tie. The smaller differences in Section 4 are not.

complete analyses are almost entirely female-labelled on this snapshot (308 vs 7), so the two groups are not balanced across analysis type. This is a property of the corpus and a caution, not a controlled comparison.

Section 8Conclusion

On 19,995 scored outfits the model labelled male or female, the scores are the same: 67.2 against 66.8, medians 66 and 67, every percentile within two points, the biggest summary gap a 1.9-point difference on the 85-plus rate. The divergence is in the language. Male-labelled rows clear n ≥ 100 on five archetypes, all shared; female-labelled rows on eleven, five of them their own, including the highest-scoring labels on the page. And the advice diverges on fit — 45.7% of male rows against 22.4% of female — and on accessories the other way. One score, two vocabularies. No causal lift is claimed.

The honest sentence for a journalist is: this model gives men and women the same scores and describes them in different words. If you want the engine to look at your own frame, the tool is the same as it was: the free AI outfit rater. This page is the study. Pitch this URL.