decision-models-for-industrial-image-qc

This record belongs to a study that asks whether structured decision models (JEV, a hosted model, and Clef-Flash, an open-weight model) can detect manufacturing defects as well as few-shot CLIP with fewer labeled images, and give calibrated class probabilities. The paper of the study is on arXiv: Megahed, F. M., & Chen, Y.-J. (2026). *Beyond CLIP: Calibrated (few-shot) defect detection with System One decision models*. arXiv. https://doi.org/10.48550/arXiv.2610.41199. The record holds the decisions (who proposed and who decided each one), the process events with their times, the milestones and open items, the paid-call budgets, and the software versions. The experiment itself (each trial, its settings and its result) is in experiment_log.md and results/ at the root of the repository.

Now

• Latest release: v0.0.2, 2026-10-22
• Milestone in progress: M2 Full MSEC 2027 paper: fmm and yjc write the text and make the method figure (D93, D106). Claude Code makes the data figures and the LaTeX tables in the style of fmm (D96, D97, D107) and gives each reference in the .bib file a doi field (D93). fmm did the literature search for the introduction and the framework, so Claude Code does no literature search (D260). Following the ASME position statement on AI, Claude Code writes no text of the paper and no observations (D105), and the paper discloses how AI tools were used. The paper is in the private repository fmegahed/decision-models-paper, linked to Overleaf (D94)

Open items

• O1: On hold (D104). Test the refinement step (criteria and references) with smaller validation sets; test balanced accuracy settled at 10 to 20 validation images per class, but the ECE kept falling up to 50 (figure F7 of v0.0.1).
• O3: On hold (D104). Extrusion frames come from video: consider a time-based split, because near-identical frames help methods with large learning sets.
• O5: State in the paper that the audit of the description loop finds plausible reads outside the sandbox, not deliberate evasion: an interpreter one-liner (python -c, Rscript -e) can name a parent folder without a slash or '..', and the audit does not see it (ledger of M5, Task 4).
• O6: Write the AI-use disclosure that ASME asks for (position statement, section B): tools, versions, maker, access, dates and uses in the Acknowledgements, and how AI was used in the Methods. Claude Code collects the facts from the record; fmm and yjc write the text.

Status

Current milestone: M2 Full MSEC 2027 paper: fmm and yjc write the text and make the method figure (D93, D106). Claude Code makes the data figures and the LaTeX tables in the style of fmm (D96, D97, D107) and gives each reference in the .bib file a doi field (D93). fmm did the literature search for the introduction and the framework, so Claude Code does no literature search (D260). Following the ASME position statement on AI, Claude Code writes no text of the paper and no observations (D105), and the paper discloses how AI tools were used. The paper is in the private repository fmegahed/decision-models-paper, linked to Overleaf (D94). Next milestone: none. Date: 2026-11-09.

Milestones

1. done M1 Tune JEV on validation, test once, compare with CLIP and Clef-Flash (MSEC 2027 abstract)
2. now M2 Full MSEC 2027 paper: fmm and yjc write the text and make the method figure (D93, D106). Claude Code makes the data figures and the LaTeX tables in the style of fmm (D96, D97, D107) and gives each reference in the .bib file a doi field (D93). fmm did the literature search for the introduction and the framework, so Claude Code does no literature search (D260). Following the ASME position statement on AI, Claude Code writes no text of the paper and no observations (D105), and the paper discloses how AI tools were used. The paper is in the private repository fmegahed/decision-models-paper, linked to Overleaf (D94)
3. done M3 R package decisions (named decisionqc until D146; design D183, plan D187)
4. done M4 Record additions: releases, results, timeline and contributions (replication-materials 1.2.0)
5. done M5 Fair comparison: shared label budget, blind description loop, calibration of all methods (v0.0.2)
6. now M6 Hugging Face Space of decisions: a Shiny app in Docker that asks Jev a question about each image or text, with the R code and a bundle of the run (design D242, plan D243)
7. now M7 Box study, then Box in decisions 0.2.0 and the Space: find defects with the TypeSafe Box primitive (Jev 1.14) on FabriD and the AutoVI classes tank screw and pipe clip, with a search over regions of several sizes that overlap, as in imageST (D268 to D274)

Decisions by decider

• Fadel M. Megahed: 271
• Claude Code: 127

Pages

Record

• Process: The stages of the work, from the inputs to the paper, with their dates.
• Timeline: The work by date: the stages, the releases and the decisions.
• Progress: The milestones, the tasks that are complete and the paid calls.
• Decisions: Each decision: who proposed it, who decided it, and how.
• Contributions: The account of the researchers: who did what, and what no person checked.
• Records: The plans, designs, ledgers and reviews of the work.
• Results: Each result with its release, its source file and its status.

Reuse

• Analysis notebook: Remakes every table and figure of the paper from the saved results, with the code.
• Description loops: Every candidate description of the eight loops, against the text it revised.
• Releases: Each tagged version of the code, with its changes and its decisions.
• Repository: The code and all files of the project on GitHub.