Sun, Huiming, Li, Yue, Yang, Kangning, Li, Ruineng, Cai, Jie, Fu, Lan, Fu, Yunwan, Li, Jinlong, Shen, Jinglin · Nov 6, 2026
ScreenProof: Can a model spot UI faults without a reference?
Markdown · JSON · Original source
The authors of ScreenProof ask whether a model can locate faults in a phone-screen screenshot without a target design or reference image. Their examples include text running off-screen, letters touching, awkward line breaks and a button whose corners clash with the page’s style. The proposed use is a cheap first review of generated pages before a designer checks them—but only if the model stays quiet on clean pages. The authors report that often flag clean screenshots and miss many text-layout faults, while rarely spotting foreign icons. Their small critic, ProofJudge, is trained on pages altered in code and keeps a finding only when confident. It reaches an F1 of 0.48, which the authors report as the highest among the judges tested, and is the only one within 15% false alarms on clean pages: 2% for real screenshots and 9% for generated ones. The authors say it costs about a hundredth as much as Gemini 4 Argon. Its performance is uneven: it finds more general faults, but fewer local style faults than Argon and fewer text-layout faults than judges. The paper also says judging a design against a named brand is harder, since models know a design language about as well as its public documentation states. These results suggest a possible first-pass tool, not a replacement for designer review; the abstract does not establish how well it would work on pages beyondে
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 6, 2026 (2026-11-06).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T13:31:54.232Z. First observed: 2026-11-10T08:39:50.693Z. Retrieved: 2026-11-10T13:01:30.967Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence is the abstract; it does not give the evaluation-set size or enough methodological detail to assess the reported comparisons. The abstract says ProofJudge is trained on code-made pages but does not specify how performance was tested on pages beyond those made by that data engine.
Open question
How well does ProofJudge find faults on pages beyond those made by the paper’s code-based data engine?
Source-stated
The authors frame ScreenProof as a test of whether a model can locate UI faults in screenshots without a target design or reference image.
We ask whether a model can box every such fault on a screenshot, given no target design and no reference image.
Source-stated
The authors propose the critic as a cheap first pass over generated pages, conditional on it staying silent on clean pages.
Such a critic could serve as a cheap first pass over generated pages before a designer reviews them, but only if it stays silent on clean pages.
Source-stated
The authors report that ProofJudge reaches F1 0.48, the highest score among the judges they tested.
reaches F1 0.48, the highest of all judges (Gemini 4 Argon 0.37; open models at most 0.15; 0.46 even without the confidence cut)
Source-stated
The authors report false alarms of 2% on real clean screenshots and 9% on generated clean pages for ProofJudge.
the only fault-finding judge within 15% false alarms on clean pages (2% real, 9% generated)
Source-stated
ProofJudge’s reported results vary by fault type: it finds more general faults, but fewer local style faults than Argon and fewer text-layout faults than API judges.
It finds the most general faults (59%) but fewer local style faults than Argon (25% against 41%) and fewer text layout faults than the API judges (43%; Gemini 4 Argon 47%).
Source-stated
The authors say that judging a design against a named brand is harder, and relate models’ knowledge of a design language to what public documentation states.
Against a named brand the task is harder still: models know a design language about as well as its public documentation states it.
Source and contributor context
Sun, Huiming, Li, Yue, Yang, Kangning, Li, Ruineng, Cai, Jie, Fu, Lan, Fu, Yunwan, Li, Jinlong, Shen, Jinglin · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
API
An application programming interface: a defined way for software to request data or actions from another system.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 34fa1ed7-4f06-4612-8714-da1ddc9010d6.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T13:31:54.232Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b8f0ebc2c7e15039a9cee46ed30f826bdb7f2a1ff955d79b1f76ba7685566656. Source text hash: b360eed3ac6ab1e28ca28e1f124b30723b33f8070b8a25ae9b0ad433895bc7dc.
Revision and source-check history
Version 1 · 2026-11-10T13:31:54.232Z
ScreenProof: Can a model spot UI faults without a reference?
The authors of ScreenProof ask whether a model can locate faults in a phone-screen screenshot without a target design or reference image. Their examples include text running off-screen, letters touching, awkward line breaks and a button whose corners clash with the page’s style. The proposed use is a cheap first review of generated pages before a designer checks them—but only if the model stays quiet on clean pages. The authors report that often flag clean screenshots and miss many text-layout faults, while rarely spotting foreign icons. Their small critic, ProofJudge, is trained on pages altered in code and keeps a finding only when confident. It reaches an F1 of 0.48, which the authors report as the highest among the judges tested, and is the only one within 15% false alarms on clean pages: 2% for real screenshots and 9% for generated ones. The authors say it costs about a hundredth as much as Gemini 4 Argon. Its performance is uneven: it finds more general faults, but fewer local style faults than Argon and fewer text-layout faults than judges. The paper also says judging a design against a named brand is harder, since models know a design language about as well as its public documentation states. These results suggest a possible first-pass tool, not a replacement for designer review; the abstract does not establish how well it would work on pages beyondে
Original source
cs.AI · Article
Source details
Original source date: Nov 6, 2026 (2026-11-06).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T13:31:54.232Z. First observed: 2026-11-10T08:39:50.693Z. Retrieved: 2026-11-10T13:01:30.967Z.
Limits
The evidence is the abstract; it does not give the evaluation-set size or enough methodological detail to assess the reported comparisons. The abstract says ProofJudge is trained on code-made pages but does not specify how performance was tested on pages beyond those made by that data engine.
Open question
How well does ProofJudge find faults on pages beyond those made by the paper’s code-based data engine?
Source-stated
The authors frame ScreenProof as a test of whether a model can locate UI faults in screenshots without a target design or reference image.
We ask whether a model can box every such fault on a screenshot, given no target design and no reference image.
Source-stated
The authors propose the critic as a cheap first pass over generated pages, conditional on it staying silent on clean pages.
Such a critic could serve as a cheap first pass over generated pages before a designer reviews them, but only if it stays silent on clean pages.
Source-stated
The authors report that ProofJudge reaches F1 0.48, the highest score among the judges they tested.
reaches F1 0.48, the highest of all judges (Gemini 4 Argon 0.37; open models at most 0.15; 0.46 even without the confidence cut)
Source-stated
The authors report false alarms of 2% on real clean screenshots and 9% on generated clean pages for ProofJudge.
the only fault-finding judge within 15% false alarms on clean pages (2% real, 9% generated)
Source-stated
ProofJudge’s reported results vary by fault type: it finds more general faults, but fewer local style faults than Argon and fewer text-layout faults than API judges.
It finds the most general faults (59%) but fewer local style faults than Argon (25% against 41%) and fewer text layout faults than the API judges (43%; Gemini 4 Argon 47%).
Source-stated
The authors say that judging a design against a named brand is harder, and relate models’ knowledge of a design language to what public documentation states.
Against a named brand the task is harder still: models know a design language about as well as its public documentation states it.
Source and contributor context
Sun, Huiming, Li, Yue, Yang, Kangning, Li, Ruineng, Cai, Jie, Fu, Lan, Fu, Yunwan, Li, Jinlong, Shen, Jinglin · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
API
An application programming interface: a defined way for software to request data or actions from another system.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 34fa1ed7-4f06-4612-8714-da1ddc9010d6.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T13:31:54.232Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b8f0ebc2c7e15039a9cee46ed30f826bdb7f2a1ff955d79b1f76ba7685566656. Source text hash: b360eed3ac6ab1e28ca28e1f124b30723b33f8070b8a25ae9b0ad433895bc7dc.
uncertain · 2026-11-10T13:31:54.232Z · anthropic/claude-opus-5.5
Publication: 34fa1ed7-4f06-4612-8714-da1ddc9010d6
attribution — pass: The draft consistently attributes the question, the proposed first-pass use, the benchmark results, the cost comparison and the brand-language observation to the paper's authors. It does not present the reported scores as independently verified.
Authors:Huiming Sun, Yue Li, Kangning Yang, Ruineng Li, Jie Cai, Lan Fu, Yunwan Fu, Jinlong Li, Jinglin Shen
summary — uncertain: The summary otherwise reads faithfully, including the authors' condition that the critic must stay silent on clean pages. However, its final sentence is cut off mid-phrase and ends in a stray non-Latin character: 'on pages beyond' followed by a garbled glyph. A reader cannot tell what limitation the draft meant to state. The smallest repair is to complete the sentence, for example 'beyond those made by its data engine', or to remove it. F1 is also left unexplained; a term-library tooltip would help readers outside the field.
date — pass: The originalPublished date of 2026-11-06 matches the arXiv submission date for v1. The feed date of 2026-11-10 is correctly not used as the publication date.
[Submitted on 6 Nov 2026]
support — pass: All six claims are backed by verbatim abstract text. The figures check out: F1 0.48 against Argon's 0.37, false alarms of 2% and 9%, and the fault-type comparisons. The characterizations of frontier judges, the cost claim and the brand point match the abstract. The 'first-pass tool, not a replacement' framing reflects the authors' own condition. The open question about generalization beyond code-made pages is reasonable, although the abstract does report false alarms on clean real screenshots.
Our small critic ProofJudge (Qwen3.5-9B with LoRA), trained on code-made pages with one to four faults each and keeping a finding only when it is confident
media — pass: The selected media is the verified direct link to the arXiv abstract page for this paper. That page is the original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
One of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceAn application programming interface: a defined way for software to request data or actions from another system.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivOne of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceAn application programming interface: a defined way for software to request data or actions from another system.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv