How honest is Dress Aloud?

Patent pending. U.S. Application No. 64/157,256. The method measured on this page is the subject of a pending United States patent application.

Updated 16 October 2026. This page is updated after every measurement, including the ones that go badly.

The problem we measure

A language model asked "does this fit" will happily invent a sleeve length or a "typical medium" chest measurement. A sighted shopper spots that. A blind shopper cannot. So Dress Aloud's fit verdict is written only from facts the model was given: your measurements, the size you are considering, and the store page's title, fabric, fit words and size chart.

The guard

A number guard, ordinary code rather than AI, checks every sentence of the verdict before you hear it. Any number that is not in those facts, and is not the difference between one of your measurements and a size-chart number, is sent back once with the invented numbers named. If the rewrite still invents, that sentence is dropped and you are told a sentence was dropped and why. Praise words are forbidden and counted afterwards.

What we measured

The same 18 try-ons each time: two synthetic subjects, real product pages on Amazon, Uniqlo and Zappos, tops and shoes, a size chart supplied for half of them, several sizes wrong on purpose. An independent script, stricter than the guard, re-checks every delivered verdict against the facts.

Re-measured twice on 17 September 2026, after the retailer widget, the seller API and the stock watch went live: the same 18 try-ons, the same strict check. First drafts that invented a number: 0 of 18. Delivered sentences with an invented number: 0 of 109. Praise words: 0 of 18. The first of the two runs was on a slow connection and the clock read 71 seconds, with verdicts agreeing with the run before on 17 of 18 items. The second, later that night from a normal connection, is the latest full measurement: a median of 24 seconds to the spoken verdict, slowest 40, and every one of the 18 verdicts the same as both previous runs.

Re-measured 16 October 2026, a month on, after the model backups, the in-browser page reading for stores that block servers, and the picture and size-reading fixes of October went live: the same 18 try-ons, the same strict check. First drafts that invented a number: 0 of 18. Delivered verdicts with an invented number: 0 of 18. Delivered sentences with an invented number: 0 of 111. Sentences dropped: 0. Praise words: 0 of 18. Every verdict took the same form as in September, 18 of 18, and the same size in all but two where only the wording moved ("S" became "Small"; "size 8 because the chart matches your foot" became "size 8"). Slower than September: the median single try-on took 39 seconds against 24, and the one outfit 68 seconds, because the server now reads more of each page and a second line of models stands behind Google. One honest footnote: the original photo file of the second synthetic subject was lost with a temporary folder, so this run used a full-body picture of that same synthetic subject from the September run, in a plain tee, with the same measurements.

A different test, 17 September 2026: pages built to fool it. Nine store pages we wrote ourselves, each with one trap: an instruction hidden in the title or in a size-chart row telling the assistant to recommend one size to everyone; a chart in centimeters; a chart of flat garment widths; a body larger than every size; a body smaller than every size; the chosen size sold out; shoe words on a shirt listing; a description claiming "typical" measurements with no chart. First pass, 6 of 9: both hidden instructions were ignored and the numbers won, the two impossible bodies got "Do not buy this" for the first time, but the sold-out size went unmentioned and the shirt with boot words was filed as shoes. We fixed both the same night and ran it again: 9 of 9. The pages and the scoring are in the open repository.

What real people say

Two shoppers have now filled in a short form after try-ons they did themselves: a wedding dress, a sundress and an expensive floral dress. In 3 of 3 try-ons they answered that nothing the app said was untrue, that the picture did look like them wearing the item, and that the verdict matched what they would have decided. They rated how easy it was 5 out of 5 each time.

Three try-ons is three try-ons. They rated their own try-ons rather than an independent judge doing it, both can see, and their try-ons ran on an earlier version of the app. We report it as 3 of 3, never as a percentage. Every rating we receive goes on this page, including the ones that go against us.

The misses, in plain words

• First guard. Three verdicts on pages with no size chart asserted a "typical" garment measurement as fact, for example "a standard medium measures at least 44 inches". The guard let them through because its arithmetic rule accepted the difference between any two facts, so a height minus a head circumference could justify almost any number. Closed the same day.
• Second guard. One strict hit: the number 100 in "100 percent cotton", read straight from the store page. The test harness had not saved the fabric line for outfit pieces, so the checker could not see it. The app's own guard, which had the page, was right to let it through. Read strictly it is 1 of 18; with the page in hand it is 0 of 18. Both are stated because the strict script is the one that cannot be argued with.
• Third guard. Nothing invented and nothing rewritten on the 18 items, on two separate runs. A later run with the same result also found that a shopper's typed note had been silently dropped from the verdict whenever a store page was read. That was a bug in our code, not the model's, and it is fixed.

What these numbers are not

They are the guard's interception rate on synthetic subjects, 18 items. "0 of 18" means 18 clean items on one run, not a percentage, and not an accuracy rate. The measurement that matters is people: blind listeners deciding from the spoken verdict, sighted raters scoring every claim against the picture, and shoppers telling us whether what they bought fit. That study is written, and the first ratings by real shoppers are above. The rest of it, an independent rater scoring every claim and blind listeners deciding from the spoken verdict alone, has not been done yet. Its results go on this page, misses included.

What you can do

If a verdict is ever wrong about you, say so: the thumbs-down under every result, the "Report this picture" control, and the follow-up question the app asks a week after a purchase all reach a person. That is the data that keeps this page honest.