Catalog facts · Image measurements · Open questions
What can we trust in a preview?
Rechoice checks narrow properties of generated images and keeps them separate from the store's product facts. This page reruns the fidelity gate and the other checks on every published render, including two registered gate studies and the rejected bag probe.
The fidelity gate: does the preview show what was asked for?
For top previews with the required inputs, the gate makes three local measurements, with no model call: garment colour against the catalog (ΔE ≤ 10.0), print busyness against the top bought on the same photo (quieter at ≤ 0.5× its edge detail), and the share of the image outside a fixed garment band that changed (stable at ≤ 3.0%). A result outside these thresholds is flagged, and the shopper must choose it knowingly. These measurements do not establish overall fidelity.
Each render was meant to show something. Grouped by that, for the renders made after the thresholds were fixed (the two registered gate studies, on 6 people):
15 / 26plain-top alternatives passed: all three measurements agree with the request
11plain tops held back, each with its reason on the card: 7 for colour, 4 for print, 4 for changes outside the garment (some for more than one)
4 / 8patterned tops registered to fail were held back
6 / 6repeat renders of earlier colour flags were flagged again
flaggedthe rejected bag probe, which redrew the whole scene
What the held-back tops have in common: every print flag is the utility shirt, whose pockets the print measure reads as pattern; every outside-band flag is on the one full-length photo; and the blind AI rater judged all ten colour-flagged renders it saw a match, once it allowed for the photo's light. Independent human labels are registered as an addendum (assets/gate_labels_addendum.json) but have not been collected. Across all 53 published alternatives, including 13 retrospectively assessed renders and the repeats: 25 pass, 6 to check yourself, 22 held back; 36 of 46 print measurements agree with the catalog's quiet-or-busy label.
The print bound was chosen on these development renders: plain tops scored 0.11 to 0.42 of the bought blouse's edge detail and patterned tops 0.50 to 0.70. The stability box and pixel threshold are the held-out study's frozen values. This report is therefore not an unseen test of the gate.
Registered study v2: four new people, an attempt to make the gate fail, and repeats
Registered before any render (report, protocol assets/gate_study_v2.json): each of 4 people the gate had never measured got the floral blouse, five plain long-sleeved tops and two patterned sweaters registered as tops that should not pass. The three v1 colour flags were rendered twice more on the same photos. Thresholds were not changed after the results; a blind AI rater labelled every render.
4 / 8patterned sweaters the gate did not pass (catch rate)
16 / 20plain tops measured quieter than the blouse
16 / 20plain-top colours within ΔE 10.0 of the catalog
25 / 32renders stable outside the garment band
12 / 20plain tops where all three measurements agree with the request
What held. Each repeat reproduced its v1 colour flag within 0.2 ΔE, across all six repeats of three flagged cases; these retries did not remove the colour flags. On the three half-body photos all 24 renders stayed within the outside-band threshold.
What did not. The four plain tops not measured quieter were all the utility shirt, whose pockets, buttons and seams the print measure reads as pattern. The sweaters were weak negatives: a yoke band and broad stripes are less busy than the dense floral, the blind rater judged all eight quieter, and the gate passed six of them on print. On the full-length photo the stability band flagged seven of eight renders at 3.0 to 4.5 percent; the rater confirmed redrawn trousers or sleeves in three. The rater also saw small changes inside the garment band, such as sunglasses hung on a shirt, that the fixed band cannot see. On colour the rater accepted all ten flagged renders, allowing for the photo's light; the colour check does not correct for it.
Repeated case
v1 render
Repeats
Flag in every repeat
model-b · CALM-04
differs, ΔE 13.6
differs, ΔE 13.6; differs, ΔE 13.6
yes
model-b · CALM-03
differs, ΔE 19.4
differs, ΔE 19.6; differs, ΔE 19.4
yes
model-c · CALM-07
differs, ΔE 11.1
differs, ΔE 10.9; differs, ΔE 11.1
yes
Blind rater
Print
Colour
Outside the garment
AI reviewer (claude-sonnet-5-5)
28 / 34 agree
16 / 26 agree
30 / 38 agree
Raters saw only the images: the photo, the blouse on the same photo, the product photo and the preview, in a fixed shuffled order, without SKUs, roles or gate output. The protocol also named the owner as a human rater; the owner chose not to label, so no human labels are reported yet. An addendum registers more human raters. Where the rater and the gate disagree, the reasons are in the label file and the limits below.
Development evidence, with every denominator
36 development outcomes. These pilots were already inspected or used while building the checks. This is a retrospective evaluation, not an independent accuracy benchmark.
6 / 7plain-top colours within ΔE 10.0 on the public files
1 / 2earring alternatives in proportion before correction; the reference pair is excluded
1 / 1out-of-proportion earrings corrected and remeasured within the tolerance
3patterned tops excluded from a single-colour score
12shade outcomes with exploratory measurements, no calibrated match verdict
0new assessments of whether the requested difference is visible; pending review
36 published final outcomes across tops, beauty, jewellery, hair, shoes, looks and the bag probe. Unpublished results are excluded from pixel measurements and described on the pilot-evidence page. No combined pass rate is calculated.
Headline measurements above use development cases only. The registered photo holdout has its own report and denominators.
Render
Split
Role
Busyness
Print ratio
Outside change
Colour
Verdict
pilot-01 · model-a · CALM-04
development
alternative
28.69
0.42
0.0%
close
shows
pilot-02 · model-a · ORIGINAL
development
reference
69.11
–
0.0%
pattern
reference
pilot-03 · model-a · CALM-07
development
alternative
14.42
0.21
0.0%
close
shows
pilot-04 · model-b · CALM-01
development
alternative
23.29
not measured
0.0%
close
check
pilot-05 · model-b · BOLD-01
development
alternative
48.67
not measured
0.0%
pattern
check
pilot-06 · model-b · OVER-01
development
alternative
9.81
not measured
0.0%
close
check
pilot-07 · model-c · BOLD-02
development
alternative
34.91
not measured
0.0%
pattern
check
pilot-08 · model-c · CALM-03
development
alternative
30.49
not measured
0.0%
close
check
pilot-09 · model-c · CALM-02
development
alternative
52.71
not measured
0.0%
close
check
pilot-10 · model-c · SHORT-03
development
alternative
23.04
not measured
0.0%
differs
concern
heldout-v1-01 · model-d · ORIGINAL
held_out
reference
68.58
–
0.0%
pattern
reference
heldout-v1-02 · model-d · CALM-04
held_out
alternative
26.24
0.38
0.0%
close
shows
heldout-v1-03 · model-d · CALM-07
held_out
alternative
7.86
0.11
0.0%
close
shows
heldout-v1-04 · model-e · ORIGINAL
held_out
reference
74.88
–
0.0%
pattern
reference
heldout-v1-05 · model-e · CALM-04
held_out
alternative
27.23
0.36
0.0%
close
shows
heldout-v1-06 · model-e · CALM-07
held_out
alternative
9.22
0.12
0.0%
close
shows
gate-study-v1-01 · model-b · ORIGINAL
live_study
reference
70.01
–
0.0%
pattern
reference
gate-study-v1-02 · model-b · CALM-04
live_study
alternative
28.11
0.4
0.0%
differs
concern
gate-study-v1-03 · model-b · CALM-07
live_study
alternative
9.82
0.14
0.0%
close
shows
gate-study-v1-04 · model-b · CALM-03
live_study
alternative
24.94
0.36
0.0%
differs
concern
gate-study-v1-05 · model-c · ORIGINAL
live_study
reference
72.17
–
0.0%
pattern
reference
gate-study-v1-06 · model-c · CALM-04
live_study
alternative
30.96
0.43
0.0%
close
shows
gate-study-v1-07 · model-c · CALM-07
live_study
alternative
17.16
0.24
0.0%
differs
concern
gate-study-v1-08 · model-c · CALM-03
live_study
alternative
30.43
0.42
0.0%
close
shows
gate-study-v2-01 · model-f · ORIGINAL
live_study_v2
reference
79.78
–
3.1%
pattern
reference
gate-study-v2-02 · model-f · CALM-01
live_study_v2
alternative
30.73
0.39
3.62%
close
concern
gate-study-v2-03 · model-f · CALM-02
live_study_v2
alternative
41.64
0.52
3.29%
differs
concern
gate-study-v2-04 · model-f · CALM-03
live_study_v2
alternative
38.18
0.48
3.03%
close
concern
gate-study-v2-05 · model-f · CALM-04
live_study_v2
alternative
33.87
0.42
2.96%
close
shows
gate-study-v2-06 · model-f · CALM-07
live_study_v2
alternative
10.36
0.13
4.5%
close
concern
gate-study-v2-07 · model-f · BOLD-01
live_study_v2
alternative
35.36
0.44
3.05%
pattern
concern
gate-study-v2-08 · model-f · BOLD-02
live_study_v2
alternative
34.53
0.43
3.41%
pattern
concern
gate-study-v2-09 · model-g · ORIGINAL
live_study_v2
reference
70.11
–
0.0%
pattern
reference
gate-study-v2-10 · model-g · CALM-01
live_study_v2
alternative
32.54
0.46
0.0%
close
shows
gate-study-v2-11 · model-g · CALM-02
live_study_v2
alternative
37.38
0.53
0.0%
close
concern
gate-study-v2-12 · model-g · CALM-03
live_study_v2
alternative
24.02
0.34
0.0%
close
shows
gate-study-v2-13 · model-g · CALM-04
live_study_v2
alternative
23.87
0.34
0.0%
close
shows
gate-study-v2-14 · model-g · CALM-07
live_study_v2
alternative
19.32
0.28
0.0%
close
shows
gate-study-v2-15 · model-g · BOLD-01
live_study_v2
alternative
31.94
0.46
0.0%
pattern
shows
gate-study-v2-16 · model-g · BOLD-02
live_study_v2
alternative
33.51
0.48
0.0%
pattern
shows
gate-study-v2-17 · model-h · ORIGINAL
live_study_v2
reference
61.58
–
0.0%
pattern
reference
gate-study-v2-18 · model-h · CALM-01
live_study_v2
alternative
18.26
0.3
0.0%
close
shows
gate-study-v2-19 · model-h · CALM-02
live_study_v2
alternative
31.68
0.51
0.05%
differs
concern
gate-study-v2-20 · model-h · CALM-03
live_study_v2
alternative
20.98
0.34
0.0%
differs
concern
gate-study-v2-21 · model-h · CALM-04
live_study_v2
alternative
23.78
0.39
0.0%
close
shows
gate-study-v2-22 · model-h · CALM-07
live_study_v2
alternative
6.58
0.11
0.0%
close
shows
gate-study-v2-23 · model-h · BOLD-01
live_study_v2
alternative
35.64
0.58
0.0%
pattern
concern
gate-study-v2-24 · model-h · BOLD-02
live_study_v2
alternative
27.91
0.45
0.0%
pattern
shows
gate-study-v2-25 · model-i · ORIGINAL
live_study_v2
reference
69.38
–
0.0%
pattern
reference
gate-study-v2-26 · model-i · CALM-01
live_study_v2
alternative
27.63
0.4
0.0%
close
shows
gate-study-v2-27 · model-i · CALM-02
live_study_v2
alternative
48.04
0.69
0.0%
differs
concern
gate-study-v2-28 · model-i · CALM-03
live_study_v2
alternative
25.08
0.36
0.0%
close
shows
gate-study-v2-29 · model-i · CALM-04
live_study_v2
alternative
29.34
0.42
0.0%
close
shows
gate-study-v2-30 · model-i · CALM-07
live_study_v2
alternative
9.75
0.14
0.0%
close
shows
gate-study-v2-31 · model-i · BOLD-01
live_study_v2
alternative
38.51
0.56
0.0%
pattern
concern
gate-study-v2-32 · model-i · BOLD-02
live_study_v2
alternative
34.55
0.5
0.0%
pattern
shows
gate-study-v2-33 · model-b · CALM-04
repeat
alternative
28.12
0.4
0.0%
differs
concern
gate-study-v2-34 · model-b · CALM-03
repeat
alternative
24.96
0.36
0.0%
differs
concern
gate-study-v2-35 · model-c · CALM-07
repeat
alternative
17.41
0.24
0.0%
differs
concern
gate-study-v2-36 · model-b · CALM-04
repeat
alternative
28.11
0.4
0.0%
differs
concern
gate-study-v2-37 · model-b · CALM-03
repeat
alternative
24.94
0.36
0.0%
differs
concern
gate-study-v2-38 · model-c · CALM-07
repeat
alternative
17.66
0.24
0.0%
differs
concern
bag-probe · model-a · BAG-RED-HOBO
development
probe
–
54.95%
–
concern
Print busyness is mean edge strength inside the garment area; it is not a judgement of taste or a pattern classifier.
The stability measurement uses a fixed garment band; hands, hair and shadows inside the band are not checked.
Sleeves and neckline are not measured; they remain catalog facts.
A verdict of shows means the three measurements agree with the request; it does not establish fit, fabric or exact colour.
Colour is compared with the catalog colour without correcting for the light in the photo. In gate-study-v2 the blind AI rater accepted all ten renders the colour check flagged, allowing for the photo's lighting.
Print busyness also counts construction detail. The utility shirt's pockets, buttons and seams read as not quieter on all four new photos, where the blind AI rater saw a plain shirt.
The two patterned sweaters registered as tops that should fail were weak negatives: a yoke band and broad stripes are less busy than the dense floral, and the blind AI rater judged all eight quieter. A live test against a top as busy as the blouse is still to be run.
On the full-length photo Clothes V4 also redrew trousers and waistlines in some renders; the stability band flagged seven of its eight renders, at 3.0 to 4.5 percent, including the blouse itself.
The protocol named the owner as a second, human rater; the owner chose not to label, so only the AI rater is reported.
Catalog eligibility checks stock, policy, budget and confirmed controls.
It does not prove that the image shows those facts faithfully.
Is the product detail recognizable?
prior clothes reviews, attributed to the AI assistant, plus narrow colour and relative-size measurements.
No new independent human review; pattern, construction and physical fit are not automatically verified.
Is the requested difference visible?
A separate rubric field for a reviewer to score.
rated so far. A pilot task is not evidence of a shopper's confirmed request.
Did unrelated parts stay unchanged?
Outside-region pixel change is reported for each comparable image.
Fixed approximate boxes and compression can affect it. This historical development measurement is exploratory. The gate above uses a separate 3% threshold; neither verifies identity, pose or background preservation.
Are misleading results handled?
The colour warning, earring correction and the rejected bag probe are visible below. The rejected necklace is described in pilot evidence.
These are known development examples, not a measured failure-detection sensitivity.
Reproduce it without an API key
Protocol 1.0, recorded 2026-10-29. The command reads committed public images, manifests and reference colours. It makes no provider calls and needs no private pilot files or seeded catalog.
Score product detail as recognizable, partly changed, misleading or not assessed; requested difference as visible, not visible, not applicable or not assessed; unrelated parts as preserved, changed or not assessed; and misleading result as yes, no or not assessed.
Fill in the reviewer name, identify human or AI review, and mark blinding truthfully. Null or not-assessed fields do not count as reviewed. Do not mark a requested difference visible when no comparison request was recorded. Keep the original and corrected render assessments distinct.
The six-case photo holdout was registered before generation and is reported separately. Its source photos were unused in the available prior project records; its catalog products were already used in development. Inspect all six first outputs and the frozen measurements.
No new rubric ratings have been supplied. Prior AI reviews remain separately attributed.
Inspect every published outcome
pilot-01 · development · tops
Sage Knit Top
Plain knit; long sleeves drawn over bare arms · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: close. Colour check: close to the product photo. ΔE 1.8.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Plain green knit, crew neck and long sleeves all kept; the heather texture survives.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-02 · development · tops
Midnight Floral Blouse
Busy floral print; does the pattern survive · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: pattern. Patterned top: the preview may simplify the pattern. Check the product photo.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
The floral on navy is broadly kept with smaller flowers; the shirt collar is drawn as a softer open neck.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-03 · development · tops
Ribbed Turtleneck
Turtleneck detail; source is on a dress form · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: close. Colour check: close to the product photo. ΔE 5.5.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Ribbed red turtleneck kept; the cuffs gain a small ruffle that the source does not have.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-04 · development · tops
Indigo Denim Shirt
Collar, buttons and chest pockets · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: close. Colour check: close to the product photo. ΔE 6.4.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Collar, buttons, both flap pockets and the yoke are kept; the sleeves are drawn rolled up.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-05 · development · tops
Fair Isle Knit Sweater
Fair Isle yoke; source includes mannequin legs · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: pattern. Patterned top: the preview may simplify the pattern. Check the product photo.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Fair Isle yoke and fuzzy body kept; a grey layer shows at the neck and cuffs, likely from the mannequin.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-06 · development · tops
Burgundy Knit Top
Plain dark knit on a dark-shirted model · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: close. Colour check: close to the product photo. ΔE 2.7.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Plain burgundy knit kept closely despite the dark shirt underneath.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-07 · development · tops
Sunset Stripe Sweater
Bold horizontal stripes · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: pattern. Patterned top: the preview may simplify the pattern. Check the product photo.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Navy body, red, yellow and teal bands, striped sleeves and ribbed cuffs all kept; the bottom band runs out of frame with the photo.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-08 · development · tops
Heather Mauve Knit
Plain light knit · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: close. Colour check: close to the product photo. ΔE 2.7.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Plain mauve knit, crew neck and long sleeves kept.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-09 · development · tops
Khaki Utility Shirt
Utility shirt with pockets; arm crosses the torso · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: close. Colour check: close to the product photo. ΔE 1.5.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Collar, pockets and buttons kept; it is styled open at the neck with rolled sleeves, which the product may not suit.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
pilot-10 · development · tops
Sky Blue Tee
Short-sleeve control case · 2026-10-19 · cloth-v4
Original demo photoCatalog product photoOriginal YouCam result
garment_colour: differs. Colour check: this preview looks lighter than the product. Trust the product photo for colour. ΔE 14.8.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
Prior AI review, not blinded
Short sleeves and crew neck kept; the blue comes out noticeably lighter, which the colour check flags.
Claude (AI assistant) at the owner's request, 2026-10-27. This rating predates this rubric and does not count as an independent evaluation.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-01 · development · lip
230 Terracotta
Lip probe: confirms the Makeup VTO lip payload. The purchased shade in the demo · 2026-10-27 · makeup-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 4.5.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-02 · development · lip
220 Rosewood
The demo's first cooler shade · 2026-10-27 · makeup-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 2.7.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-03 · development · lip
240 Mauve
The demo's second cooler shade; with pairs 1 and 2, a complete recorded comparison · 2026-10-27 · makeup-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 4.3.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-07 · development · foundation
220W Light Warm
Foundation probe: confirms the Makeup VTO foundation payload. The purchased shade in the demo · 2026-10-27 · makeup-vto-1.0
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 7.4.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-08 · development · foundation
210C Light Cool
The demo's first cooler shade · 2026-10-27 · makeup-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 6.3.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-09 · development · foundation
310C Tan Cool
The demo's second cooler shade; with pairs 7 and 8, a complete recorded comparison · 2026-10-27 · makeup-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-11 · development · nail
01 Coral Crush
Nail probe: confirms the Nail VTO payload and that the demo hand is accepted. The purchased shade · 2026-10-27 · nail-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 2.1.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-12 · development · nail
02 Cherry
The demo's first cooler shade · 2026-10-27 · nail-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 6.9.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
beauty-13 · development · nail
04 Lilac
The demo's second cooler shade; with pairs 11 and 12, a complete recorded comparison · 2026-10-27 · nail-vto-1.1
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 4.4.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
jewellery-01 · development · earrings
Baroque Pearl Drops
Earring probe: confirms the Earrings VTO payload and that Face A's ear is found. The pair bought in the demo · 2026-10-28 · earring-vto-1.0
Original demo photoCatalog product photoOriginal YouCam result
relative_size: size-reference. Size check: the pair bought, which the other sizes are measured against. Rendered 142 px.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
jewellery-02 · development · earrings
Blue Crystal Drops
The demo's first suggestion for smaller and silver: same style, shorter · 2026-10-28 · earring-vto-1.0
Original demo photoCatalog product photoOriginal YouCam resultRechoice correction, remeasured separately
relative_size: size-differs. Size check: renders 50% longer than its length implies. Check the length in the product details. Rendered 140 px. After correction: size-close, 92 px.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
jewellery-03 · development · earrings
Blue Topaz Studs
The demo's second suggestion: a silver stud; with pairs 1 and 2, a complete recorded comparison · 2026-10-28 · earring-vto-1.0
Original demo photoCatalog product photoOriginal YouCam result
relative_size: size-close. Size check: renders in proportion to the pair bought. Rendered 28 px.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
necklace-01 · development · necklace
Amethyst Y Chain
Necklace probe: confirms the Necklace VTO payload and that Model A's neck is found. The pair bought in the demo · 2026-10-28 · necklace-vto-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
necklace-03 · development · necklace
Bloodstone Bib
The demo's suggestion for shorter in silver; with pairs 1 and 2, a complete recorded comparison · 2026-10-28 · necklace-vto-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
necklace-04 · development · necklace
Lapis Collar
A collar at a middling length, to see a second collar form · 2026-10-28 · necklace-vto-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
hair-01 · development · hair
7.3 Golden Blonde
Hair colour probe: confirms the AI Hair Color payload on a half-body photo with the hair down. The shade bought in the demo · 2026-10-28 · hair-color-1.0
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 16.3.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
hair-02 · development · hair
7.1 Ash Blonde
The demo's first cooler shade for "too brassy" · 2026-10-28 · hair-color-1.0
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 31.0.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
hair-03 · development · hair
6.1 Dark Ash Blonde
The demo's second cooler shade; with pairs 1 and 2, a complete recorded comparison · 2026-10-28 · hair-color-1.0
Original demo photoOriginal YouCam result
applied_shade: uncalibrated. Exploratory colour reading only; no calibrated match or requested-change verdict. ΔE 11.2.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
accessories-01 · development · shoes
Red Heeled Ankle Boots
Shoes probe: confirms the Shoes V3 payload and that Model F's feet are found. The pair bought in the demo · 2026-10-28 · shoes-v3-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
accessories-02 · development · shoes
Cream Block-Heel Peep-Toes
The demo's suggestion for a lower heel in a neutral colour · 2026-10-28 · shoes-v3-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
accessories-03 · development · shoes
Metallic Platform Brogues
The flattest pair: a lower heel in a bright colour · 2026-10-28 · shoes-v3-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
accessories-04 · development · shoes
Nude Suede Pumps
A neutral pair with a heel as high as the boots · 2026-10-28 · shoes-v3-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
accessories-bag-probe · development · bag
Red Slouch Shoulder Bag
Bag probe: test whether only the bag changes. · 2026-10-28 · bag-v2-1.0
Original demo photoCatalog product photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: not comparable. Exploratory; not a preservation verdict.
Rejected pilot result. Bag V2 drew the red shoulder bag well, but it made a new picture around it: a gallery background, a green dress in place of the white tee and jeans, and a new pose. Its documentation describes preset styles, chosen at random by default, and a fixed output size. That cannot show a shopper's own bag on them, and each bag would come with a different outfit, so the comparison would not be fair. The other three bags did not run, and Rechoice compares bags by their catalog details only.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
looks-01 · development · look
Look 1
The judge path's first suggested top with the lipstick and the necklace the shopper keeps. The first two steps were made by the looks probe on 28 October 2026 · 2026-10-28 · cloth-v4, makeup-vto-1.1, necklace-vto-1.0
Original demo photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
looks-02 · development · look
Look 2
The judge path's second suggested top, with the same lipstick · 2026-10-28 · cloth-v4, makeup-vto-1.1, necklace-vto-1.0
Original demo photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
looks-03 · development · look
Look 3
The top bought, with the same lipstick: what the shopper has now, made the same way · 2026-10-28 · cloth-v4, makeup-vto-1.1, necklace-vto-1.0
Original demo photoOriginal YouCam result
image_fidelity: not_measured. No validated automated product-detail check for this category.
Outside-region pixel change: 0.0%. Exploratory; not a preservation verdict.
New rubric review pending: product detail, requested difference, unrelated changes and misleading result.
Historical development-report limits
These limits describe the 36-outcome development report below. The original photo holdout and later registered gate studies are reported separately above; the gate adds its own 3% stability threshold. Earlier photo-holdout images were used to develop the later gate and are retrospective for that gate.
All current outcomes are development cases already inspected or used while building the checks. There are zero held-out outcomes.
Measurements are recomputed from committed public JPEGs, so they can differ from earlier full-resolution private-pilot readings.
Garment colour and relative earring size are narrow checks, not an overall preview-accuracy score. The reference earring is excluded from the size denominator.
Applied-shade readings are uncalibrated. No shade accuracy or visible-requested-change success rate is claimed.
Outside-region pixel change uses fixed approximate boxes, a 256-pixel working image and threshold 40. It is exploratory, not a face-identity, pose or background-preservation verdict; no pass threshold is defined.
The ten existing clothes reviews were written by an AI assistant, were not blind, and are separate from any new rubric ratings. No independent human review is implied.
The rejected bag probe is included. The rejected necklace result and unpublished private pilot outputs are described in the evidence but cannot be remeasured from this public dataset.
These are synthetic demo people and a small catalog, not a shopper study or evidence of reduced returns. There is no combined pass rate.
For a future held-out study, select new cases and freeze this protocol before collecting outputs. Keep development and held-out results separate and report all exclusions.