The Archetype Ladder

Every score on this site is written by a multimodal model, and the same pass that writes the score also writes a single word for the kind of outfit it just read. That word — the archetype — is a useful thing to rank, because it is the model telling us, in its own vocabulary, how it filed the outfit. The question this report asks is simple: when you line those words up by the score the model gave, what order do they fall into?

The answer is not a claim about dressing. It is a map of the model's own taste, and the map is worth seeing precisely because the model wrote both columns.

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 archetype is analysis_result.outfit_archetype, a single lower-case token the model writes. A row with an empty or absent archetype is treated as "unlabelled". Soft-deleted rows are not filtered.

One field needs care. undefined is not the same as unlabelled. It is a literal value the model writes when it can read the outfit but cannot place it in a style. The report keeps it as its own row and never merges it with the empty field.

Section 2Why this report starts in May 2026

The archetype field is younger than the corpus. Before May 2026 the model did not write it at all, so the earliest entries are blank by construction, not by choice. Coverage by month makes the shift plain.

Because of the cliff, comparing "labelled" to "unlabelled" across the whole corpus would really be comparing recent uploads to older ones. To keep the comparison honest, every figure in this report is restricted to the window where the field is populated: 1 May to 14 October 2026, in which 7,618 outfits score. Within it, 6,036 carry an archetype (79.2%) and 1,582 do not.

Section 3The ladder

Ranked by mean score, the twelve labels that clear the floor look like this. The span from top to bottom is 22.7 points — larger than the gap between any two score bands the site uses to describe a single outfit.

Two things stand out before any interpretation. First, the two most common labels — smart_casual (1,640) and streetwear (1,176) — sit in the middle of the ladder, near the corpus mean, which is what you would expect from the buckets that absorb the most ordinary outfits. Second, the ends are small. classic and romantic, the top two, are 271 and 258 rows; loungewear and undefined, the bottom two, are 446 and 220. The n is printed beside every cell because a handful of rows moves a small label's mean by a point.

Section 4The top of the ladder

classic leads at 77.7 and is the only label that puts a third of its rows at 85-plus (32.8%). romantic is second at 73.7 with 18.2% at 85-plus. Both are small — 271 and 258 rows — and both describe a deliberately composed, structured outfit: tailoring, coherent line, nothing arbitrary. The model's highest grades go to the outfits it files under the most structured words.

The next tier — edgy_grunge (69.7), minimalist (67.9), smart_casual (67.0), bohemian (66.8), streetwear (66.1), y2k (65.3) — clusters within 4.4 points and sits around the corpus mean. These are the labels that describe the bulk of what people upload, and they grade like it: nothing extreme, nothing embarrassing, centred on the middle.

Section 5The bottom of the ladder, and the word "undefined"

At the bottom, two named labels and one the model uses for failure. athleisure (62.0), normcore (59.3) and loungewear (55.0) describe deliberately low-effort, soft or plain dressing, and they grade lowest of the named labels. normcore has not one row at 85-plus; loungewear sends 63.5% of its rows below 60.

Then there is undefined. This is not the empty field; it is a word the model writes when it can read an outfit but cannot place it. Those rows average 41.4, and 90.5% of them fall below 60. The label behaves like an explicit admission of failure, and the scores beside it are the lowest in the report by a wide margin.

Section 6The unlabelled remainder

Within the window, 1,582 rows score but carry no archetype. They average 60.1 — below the labelled mean of 65.0 — and 42.9% of them fall below 60. Whatever makes the model withhold a word, those rows sit toward the bottom of the distribution.

The direction here is the opposite of Report No. 12, where not naming an occasion was associated with a higher score. There the missing field was a user choice; here it is a model omission, and it tracks with lower grades. The two are not the same kind of blank, which is exactly why each report prints its own population.

Section 7What this report is not

It is not a style ranking of real people, and it is not advice to dress classic. The archetype and the score come from the same model in the same pass, so the ladder is partly definitional: a model that already associates "classic" with clean tailoring will tend to grade the outfits it files there well. The report measures that tendency, not a property of the clothing.

It is not a causal claim. Nothing here says that changing your outfit's archetype would change its score, because no outfit is measured twice under two labels.

It is not the whole corpus. Every figure is restricted to 1 May – 14 October 2026, the window where the label exists. Read it as a snapshot of the model's vocabulary in that window, not of the platform's full history.

Section 8Limitations

The corpus is voluntary and the archetype is one model's free choice of word, not a human taxonomy. Twenty-five labels exist and only twelve clear the floor; the long tail (business_casual at 99, casual at 52, avant_garde at 42, and so on) is withheld rather than guessed at.

The top and bottom cells are small. classic (271) and romantic (258) lead on modest n, as do y2k (152) and bohemian (138) in the middle. A handful of rows would move their means by a point, which is why the n is printed beside every cell.

The circularity in Section 7 is the central caveat, not a footnote. The label and the score share an author, and this report can describe the association and name the mechanism as a hypothesis; it cannot isolate it.

Finally, the window is short. The archetype field has existed for roughly five months, so seasonal and product changes cannot be separated from genuine differences between labels.

Section 9Conclusion

On 7,618 scored outfits in the window where the model writes a style label, 6,036 carry one. Ranked by the score the same model gave, twelve labels span 22.7 points: classic leads (77.7; 32.8% at 85-plus), romantic follows (73.7), and normcore (59.3) and loungewear (55.0) trail. The model's failure label, undefined, is lowest of all (41.4; 90.5% below 60), and the unlabelled remainder averages 60.1. Because the label and the score share an author, the honest reading is the model's own taste made visible — a map of its priors, not a ranking of styles. Frozen snapshot. No causal lift is claimed.

The useful sentence is: line up the words the model uses for outfits by the grades it gives them, and the most structured labels sit at the top and the softest at the bottom. If you want it to file and grade your own frame, the tool is the same as it was: the free AI outfit rater. This page is the study. Pitch this URL.