# Jev and Clef-Flash for Few-Shot Defect Detection Beyond CLIP [Canonical reading](https://escapemeridian.com/readings/jev-and-clef-flash-for-few-shot-defect-detection-beyond-clip-24c77eab) Fadel M. Megahed and Ying-Ju Chen investigate whether Jev and Clef-Flash can classify manufacturing defects with very few reference images, where collecting examples of defective products may be difficult. Their paper reports results from five manufacturing case studies, comparing the models with CLIP when each uses the same two reference images per class. The authors report higher test accuracies for Jev and Clef-Flash than for CLIP in those comparisons. They also report that an agentic approach to refining text descriptions improved accuracy in most tested model-application combinations, while their calibration approach reduced the Brier score in most configurations. Calibration adjusts predicted probabilities so that they better match observed outcomes; a lower Brier score indicates better probability accuracy in the reported evaluation. The authors present these methods as ways to support inspection decisions when defect examples are scarce, and provide replication code, an R package, and a web app for practitioners to inspect images. The abstract does not establish how the results transfer to other manufacturing settings or whether calibrated probabilities improve decisions in operational use. How well do the reported results hold on new production lines with different products and defect types? 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 arXiv cs.AI · Article Original source date: Oct 30, 2026 (2026-10-30). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T10:48:13.808Z. First observed: 2026-11-02T05:01:32.931Z. Retrieved: 2026-11-02T08:48:52.981Z. - [Beyond CLIP: Calibrated (Few-Shot) Defect Detection With System One Decision Models](https://arxiv.org/abs/2610.41199) ## Limits The reported comparisons are from five manufacturing case studies; the abstract does not describe their products, defect types, or evaluation conditions in enough detail to assess transfer to other settings. The abstract reports test accuracy and Brier score results, but does not establish performance on new production lines or whether the reported calibration improves operational inspection decisions. ## Open question How well do the reported results hold on new production lines with different products and defect types? ## Source-stated The authors say collecting even 50 defective images can be difficult in some manufacturing processes, motivating their study of models using two reference images per class at inference. > However, collecting even 50 defective images can be difficult in high-yield or low-volume manufacturing processes. In this paper, we examine the utility of two recently proposed structured decision models (Jev and Clef-Flash) in applications using as few as k=2 reference images per class at - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated In five manufacturing case studies, the authors report higher test accuracies for Jev and Clef-Flash than for CLIP when all use two reference images per class. > Using two reference images per class, both Jev and Clef-Flash outperform CLIP's test accuracies with the same references (0.75-1.00 \[Jev\] and 0.71-0.99 \[Clef-Flash\] vs 0.67-0.84 \[CLIP\]). - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated The authors report that refining text descriptions improved accuracy in six of eight tested model-application combinations. > our agentic refinement of text descriptions significantly improves accuracy across 6 of 8 model-application combinations - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated The authors report that their calibration approach reduced the Brier score in thirteen of fifteen model-application configurations. > our calibration approach reduces the Brier score in 13 of the 15 model-application configurations. - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated The authors say they provide replication code, an R package for evaluating additional models, and a web app for practitioners to inspect images without code. > we provide our replication code and have created the R "decisions" package for evaluating additional models. We also provide a publicly available web app that allows practitioners to inspect their own images with no code. - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source and contributor context Megahed, Fadel M., Chen, Ying-Ju · 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. ## Zero-shot Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot). - [Definition reference](https://developers.google.com/machine-learning/glossary#zero-shot-prompting) ## Agentic AI AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#agentic) ## Jev Jev is TypeSafe AI’s model for fast evaluation and routing decisions inside AI systems. - [Explore this context](/entities/jev) - [Reference](https://typesafe.ai/blog/introducing-system-one-models-and-jev) - [Reference](https://typesafe.ai/) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 11041307-fe9b-4a31-8541-91a595b0c58e. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T10:48:13.808Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 29215e75fc04d22a98fff839a648f42f252568a31e3116af42fcd423c3a4170b. Source text hash: ebd03d8355f472b19399403c140c07304b3785183c18818f9c1511b99637bbb2. ## Revision and source-check history ### Version 1 — 2026-11-10T10:48:13.808Z Fadel M. Megahed and Ying-Ju Chen investigate whether Jev and Clef-Flash can classify manufacturing defects with very few reference images, where collecting examples of defective products may be difficult. Their paper reports results from five manufacturing case studies, comparing the models with CLIP when each uses the same two reference images per class. The authors report higher test accuracies for Jev and Clef-Flash than for CLIP in those comparisons. They also report that an agentic approach to refining text descriptions improved accuracy in most tested model-application combinations, while their calibration approach reduced the Brier score in most configurations. Calibration adjusts predicted probabilities so that they better match observed outcomes; a lower Brier score indicates better probability accuracy in the reported evaluation. The authors present these methods as ways to support inspection decisions when defect examples are scarce, and provide replication code, an R package, and a web app for practitioners to inspect images. The abstract does not establish how the results transfer to other manufacturing settings or whether calibrated probabilities improve decisions in operational use. How well do the reported results hold on new production lines with different products and defect types? ## Original source arXiv cs.AI · Article Original source date: Oct 30, 2026 (2026-10-30). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T10:48:13.808Z. First observed: 2026-11-02T05:01:32.931Z. Retrieved: 2026-11-02T08:48:52.981Z. - [Beyond CLIP: Calibrated (Few-Shot) Defect Detection With System One Decision Models](https://arxiv.org/abs/2610.41199) ## Limits The reported comparisons are from five manufacturing case studies; the abstract does not describe their products, defect types, or evaluation conditions in enough detail to assess transfer to other settings. The abstract reports test accuracy and Brier score results, but does not establish performance on new production lines or whether the reported calibration improves operational inspection decisions. ## Open question How well do the reported results hold on new production lines with different products and defect types? ## Source-stated The authors say collecting even 50 defective images can be difficult in some manufacturing processes, motivating their study of models using two reference images per class at inference. > However, collecting even 50 defective images can be difficult in high-yield or low-volume manufacturing processes. In this paper, we examine the utility of two recently proposed structured decision models (Jev and Clef-Flash) in applications using as few as k=2 reference images per class at - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated In five manufacturing case studies, the authors report higher test accuracies for Jev and Clef-Flash than for CLIP when all use two reference images per class. > Using two reference images per class, both Jev and Clef-Flash outperform CLIP's test accuracies with the same references (0.75-1.00 \[Jev\] and 0.71-0.99 \[Clef-Flash\] vs 0.67-0.84 \[CLIP\]). - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated The authors report that refining text descriptions improved accuracy in six of eight tested model-application combinations. > our agentic refinement of text descriptions significantly improves accuracy across 6 of 8 model-application combinations - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated The authors report that their calibration approach reduced the Brier score in thirteen of fifteen model-application configurations. > our calibration approach reduces the Brier score in 13 of the 15 model-application configurations. - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source-stated The authors say they provide replication code, an R package for evaluating additional models, and a web app for practitioners to inspect images without code. > we provide our replication code and have created the R "decisions" package for evaluating additional models. We also provide a publicly available web app that allows practitioners to inspect their own images with no code. - [Supporting source](https://arxiv.org/abs/2610.41199) ## Source and contributor context Megahed, Fadel M., Chen, Ying-Ju · 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. ## Zero-shot Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot). - [Definition reference](https://developers.google.com/machine-learning/glossary#zero-shot-prompting) ## Agentic AI AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#agentic) ## Jev Jev is TypeSafe AI’s model for fast evaluation and routing decisions inside AI systems. - [Explore this context](/entities/jev) - [Reference](https://typesafe.ai/blog/introducing-system-one-models-and-jev) - [Reference](https://typesafe.ai/) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 11041307-fe9b-4a31-8541-91a595b0c58e. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T10:48:13.808Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 29215e75fc04d22a98fff839a648f42f252568a31e3116af42fcd423c3a4170b. Source text hash: ebd03d8355f472b19399403c140c07304b3785183c18818f9c1511b99637bbb2. ### uncertain — 2026-11-10T10:48:13.808Z — anthropic/claude-opus-5.5 Publication: 11041307-fe9b-4a31-8541-91a595b0c58e attribution — pass: The draft correctly attributes the paper and its claims to authors Fadel M. Megahed and Ying-Ju Chen. It keeps results as author reports rather than established facts. > Authors:Fadel M. Megahed, Ying-Ju Chen summary — uncertain: The summary names Jev, Clef-Flash and CLIP without saying what they are. The source calls Jev and Clef-Flash recently proposed structured decision models and CLIP a vision transformer (ViT) model. The draft also omits the baseline that explains why two images matters: thousands of images were traditionally needed, and CLIP-style models need 50-100 per class. The term agentic is also unexplained. Smallest repair: identify Jev and Clef-Flash as structured decision models, describe CLIP as a vision transformer that prior research says needs 50-100 images per class, and add tooltips for CLIP, ViT and agentic. The title, question and calibration/Brier explanation are otherwise faithful. > Recent research shows that vision transformer (ViT) models, such as CLIP, can reduce this requirement to 50-100 images per nominal and defective class. date — pass: The originalPublished date of 2026-10-30 matches the v1 arXiv submission date. The later feed date is not used as the publication date. > \[Submitted on 30 Oct 2026\] support — pass: Every claim matches the abstract, including the accuracy ranges, 6 of 8 combinations, 13 of 15 configurations, code, R package and web app. The limitations and the question about transfer to new production lines are framed as open issues rather than unsupported assertions. > our calibration approach reduces the Brier score in 13 of the 15 model-application configurations. media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. Some references could not be checked. Reload to try again. [Reference data (JSON)](/readings/24c77eab04d76bb9cd090056c4654c01a082e46c57cf87e721f2c2cb222b50d7/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/jev-and-clef-flash-for-few-shot-defect-detection-beyond-clip-24c77eab.json)