The Weakness Taxonomy

OutfitScore does two things after it scores an outfit: it writes a short list of strengths and weaknesses, and it pins a few labelled callouts onto the image. This report is about the first channel — the prose advice — and a shorter look at the second.

Everyone assumes the advice is about colour. Report No. 08 already showed colour is not where score is lost. This report asks a different question: when the model writes a sentence about what to improve, what is the sentence about? The answer, on 21,286 outfits, is not colour.

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, a numeric overall score can be read under the same numeric-regex guard used by seo_stats_service, and areas_for_improvement is a non-empty array. Makeup, body-type, color-season, roast, and accessory-only analyses are excluded. Soft-deleted rows are not filtered. The window is every qualifying row then in production: 27 September 2025 through 14 October 2026.

The scored corpus is 21,778 rows. Of those, 21,314 carry the areas_for_improvement key and 21,286 hold at least one item. That is 97.7% of the scored corpus, which makes this the most complete field I have published on. Rows without a single improvement item are real but rare (492 rows, 2.3%).

How a theme family is defined

Each item in areas_for_improvement is a short sentence the model wrote, not a closed enum. To count themes without hand-coding 60,000 sentences, I take the first clause of each item (the text before the first :, ,, or -), lowercase it, and test it against twelve keyword patterns. A row belongs to a family if any of its items matches that family's pattern. The full rule:

footwear    = footwear | shoe | sneaker | boot | heel | trainer | loafer | sandal | slipper
accessories = accessor | bag | belt | jewel | watch | scarf | hat | cap | sunglass
styling     = styl | cohesion | cohesive | coordinat | visual interest
fit         = fit | fitting | tailor | sleeve | trouser length | jacket
layering    = layer
colour      = color | colour | palette
proportion  = proportion | balance
texture     = textur | fabric | material
silhouette  = silhouette | shape | volum | boxy | oversized
occasion    = occasion | dress code | formality | appropriate
waist       = waist | tuck
hem         = hem | pooling | cuff | break

Two disclosures matter. First, the families are not exclusive: one sentence like "the styling is flat and the accessories are missing" matches two families, so a row can be in several and the shares sum past 100%. Second, the rule is mine, not the model's. It is designed to be transparent and reproducible, not to be the only defensible reading. Section 2 prints the raw strings too, so a reader can re-bucket if they disagree.

Section 2The taxonomy

Here are the twelve families across all 21,286 rows, ordered by how many rows carry them. The right-hand column is the mean overall score of the rows in that family — a first hint that the advice is not the same at the top and the bottom of the score range.

Three families are close to half the corpus each. Accessories leads. That is the single most counter-intuitive result in this report: the model's most common suggestion, by a margin, is to add something, not to change something. Styling and footwear follow. Colour, the thing every outfit-advice column leads with, lands eighth at 15.0% — the same position colour harmony took among the five scored pillars in Report No. 08.

The mean-score column separates the families into two groups. Finishing themes — accessories (69.8), hem (69.5), waist (69.1), footwear (69.0) — appear on higher-scoring rows. Structural themes — occasion (58.1) and proportion (61.8) — appear on lower-scoring rows. The advice a row receives is already telling you which end of the range it lives on.

Section 3The fragmentation problem

The family rollup is a convenience. Underneath it, the model writes the same idea many different ways, and the raw strings are worth printing because they are the actual output. Accessories alone arrives as four distinct strings; footwear as at least three; proportion as three.

Reading the raw strings changes one conclusion. Colour's 15.0% is spread across color coordination (560), color palette (253), and a long tail that never reaches the top of the list. It is not that colour advice is rare because the model dislikes the word; it is that colour advice is genuinely a small slice of what the model says. The "not the lever" finding in Report No. 08 shows up a second time, in a different field, written a different way.

It also warns a reader off a naive word count. A dashboard that greps for "accessories" once would miss lack of accessories and accessorizing; one that merges them would invent a family the model never named. Both the raw table and the disclosed rule are printed so neither mistake is forced.

Section 4Three tips, almost always

Before the bands, the shape of the list itself. The model does not flood a user with advice. The mode is three items, and 87.8% of rows carry exactly two, three, or four.

A three-item list is the product design, and it sets up the frequency tables: a family can appear on at most the share of rows that have a slot to put it in. Three families crossing 40% is therefore not a small effect — an accessory or styling note is present on close to half of every three-line critique the model writes.

Section 5The advice escalates with the score

Split the rows by current score and the family mix changes order. Below 60, the leading theme is styling (53.8%) and occasion is at its relative high (23.1%). At 85-plus, accessories leads (54.8%) and styling falls to 34.4%, while fit drops from 35.5% to 23.2%.

The direction is consistent with the mean-score column in Section 2. A low-scoring outfit is told about cohesion and occasion fit — structural problems. A high-scoring outfit is told to add accessories and reconsider footwear — finishing touches. The model is not giving everyone the same advice; it is grading the advice to the score it just gave.

Section 6The second channel: pins

Prose advice is not the only feedback the model emits. On some analyses it also pins labelled callouts to the photograph: strengths, weaknesses, and "boosters." This channel is far sparser. Only 2,619 of 21,778 scored rows (12.0%) carry any pins, so it is a niche, not a parallel product. Where it fires, footwear dominates.

Across the three footwear labels (footwear, footwear choice, missing footwear) the pin channel names footwear 578 times — 15.3% of all weakness pins, more than the next three labels combined. For context, the same pins file holds 3,048 strength pins and 1,000 boosters, so strengths are pinned more often than weaknesses. The model pins what it likes slightly more than what it does not.

The pin taxonomy agrees with the prose taxonomy at the top. Footwear is the most common prose family after accessories and styling, and the most common pinned weakness by a wide margin. Two channels, one recurring subject: the shoes.

Section 7What this report is not

It is not a style guide. This report counts how often the model raises each theme. It does not say that adding accessories raises a score, and it does not rank the themes by how much they would help. Counting mentions and measuring effect are different jobs; the effect question is Report No. 08's, and the answer there was "not colour" for the scored pillars.

It is not a claim about what users are told by a human stylist, nor a claim that the model's emphasis is correct. It is a description of one model's output on one window of uploads.

It is not live. The frozen snapshot is the whole claim.

Section 8Limitations

The corpus is voluntary. People who upload to an AI rater are not a random sample of dressers, and the 2.3% of scored rows with no improvement item at all are excluded from every family table. If those rows differ systematically, the shares are biased by an amount this snapshot cannot measure.

The family rule is a keyword map, not human coding. It is disclosed in full precisely because a different reasonable mapping — merging accessorizing differently, splitting styling from cohesion — would move a share by a point or two. No inter-rater kappa was computed. The raw string table is printed so the mapping can be checked.

Families overlap by construction. "Accessories 45.2%" and "styling 44.2%" are not disjoint groups; they are two ways of counting, over the same three-item lists, and they sum well past 100%. Read each share on its own, not as a slice of a pie.

The band table reports only the families that clear the publication floor in every band it names. Occasion appears in the below-60 column and is left blank where its share in a higher band fell below the threshold or out of the top row; the blank is an omission, not a zero.

Pins are a small, self-selected channel. The 12.0% coverage means the pin tables describe the kinds of uploads that trigger pins, not the corpus. They are reported as a supporting signal, not a headline.

Section 9Conclusion

On 21,286 scored outfits carrying improvement advice, the model's most common theme is accessories (45.2%), then styling (44.2%) and footwear (39.3%). Colour is eighth at 15.0%. The mode is three tips per outfit (59.6%). The critique escalates with the score: below-60 rows are led by styling (53.8%) and occasion (23.1%), 85-plus rows by accessories (54.8%) and footwear (39.7%). The sparse pin channel (12.0% of rows) is footwear-led (578 of 3,767 weakness pins). Advice is not about colour, and the advice you get depends on the score you already have. No causal lift is claimed.

The honest sentence for a journalist is: asked what to fix, this model talks about accessories, styling, and shoes — and it talks about colour last. If you want the engine to critique your own frame, the tool is the same as it was: the free AI outfit rater. This page is the study. Pitch this URL.