Releases
v0.0.1
2026-10-20 23:32 UTC
Pre-alpha. JEV tuned on validation and tested once; comparison with CLIP and Clef-Flash; calibration; the MSEC 2027 abstract and the replication document.
| ID | Decision | Proposed by | Decided by |
|---|---|---|---|
| D1 | Keep a replication record with the replication-materials skill. | Fadel M. Megahed | Fadel M. Megahed |
| D2 | Make all manifest paths relative to the local imgs/ folder. | Fadel M. Megahed | Fadel M. Megahed |
| D3 | Pin the model to jev-1.13.0 in the runner, because jev-latest can change. | Claude Code | Claude Code |
| D4 | Run batch 1: a baseline and a repeat baseline for microstructure, textiles and extrusion. | Claude Code | Fadel M. Megahed |
| D5 | Freeze 50 validation images per class. | Claude Code | Fadel M. Megahed |
| D6 | Add an opt-in upscale argument (default FALSE) to the JEV client. | Claude Code | Fadel M. Megahed |
| D7 | Set up the replication record after the experiments. | Claude Code | Fadel M. Megahed |
| D8 | Make a panelled progress chart like the autoresearch chart. | Fadel M. Megahed | Fadel M. Megahed |
| D9 | Include center crop and downsizing as experiments to reduce tokens. | Fadel M. Megahed | Fadel M. Megahed |
| D10 | Run the rest of the queue for the three priority datasets, with a cap of 1,000 calls. | Claude Code | Fadel M. Megahed |
| D11 | Judge criteria, reference and size changes by AUC (+0.02) and tune the threshold last. | Claude Code | Fadel M. Megahed |
| D12 | Choose each threshold as the midpoint of the widest run of thresholds with the maximum validation BA. | Claude Code | Claude Code |
| D13 | Tune pipe_staple and pan before one final test of all datasets. | Claude Code | Fadel M. Megahed |
| D14 | Stop pan tuning while its validation images have copies in the test set. | Claude Code | Claude Code |
| D15 | Delete the duplicate pan images from the test folder. | Fadel M. Megahed | Fadel M. Megahed |
| D16 | Raise the paid-call cap to 5,000. | Fadel M. Megahed | Fadel M. Megahed |
| D17 | Delete all files with the word copy in their names. | Fadel M. Megahed | Fadel M. Megahed |
| D18 | Keep a smaller pan image size when it is no worse, because the aim is to save tokens. | Claude Code | Claude Code |
| D19 | Use the bright-clip wording for pipe_staple; at the AUC ceiling, compare BA at tuned thresholds. | Claude Code | Fadel M. Megahed |
| D20 | Match the test sizes of the prior CLIP study, not its few-shot learning sizes. | Fadel M. Megahed | Fadel M. Megahed |
| D21 | Delete pan test image 21.JPG, which the prior test folder does not have. | Claude Code | Fadel M. Megahed |
| D22 | Add the real textile images as a transfer test. | Claude Code | Fadel M. Megahed |
| D23 | Run the one-time final test for the frozen configurations. | Claude Code | Fadel M. Megahed |
| D24 | Run CLIP on the same partition, a zero-shot JEV ablation and a calibration check. | Fadel M. Megahed | Fadel M. Megahed |
| D25 | Give the zero-shot ablation its own validation threshold, with the same rule. | Claude Code | Claude Code |
| D26 | Raise the paid-call cap to 7,000 if needed. | Fadel M. Megahed | Fadel M. Megahed |
| D27 | Run CLIP in a separate, pinned conda environment. | Claude Code | Claude Code |
| D28 | Replicate CLIP with the learning-set size that the CLIP paper reports for each experiment. | Fadel M. Megahed | Fadel M. Megahed |
| D29 | Compare Platt scaling with isotonic regression. | Fadel M. Megahed | Fadel M. Megahed |
| D30 | Drop the real-textile transfer example. | Fadel M. Megahed | Fadel M. Megahed |
| D31 | Exclude copies of test images from the CLIP learning pool. | Claude Code | Claude Code |
| D32 | Wait with the repository, because the scope and name can change with Clef-Flash. | Fadel M. Megahed | Fadel M. Megahed |
| D33 | Back-fill the decisions of this session into the record. | Claude Code | Fadel M. Megahed |
| D34 | Run Clef-Flash locally with the same protocol. | Claude Code | Fadel M. Megahed |
| D35 | Delete only the extrusion pool frames that copy test frames. | Claude Code | Fadel M. Megahed |
| D36 | Send images to Clef-Flash as an ordered list with a text legend, because its SystemOne server ignores TypeSafe assets. | Claude Code | Claude Code |
| D37 | Submit the MSEC 2027 title and abstract in the rewrite of the owner. | Fadel M. Megahed | Fadel M. Megahed |
| D38 | Add Ying-Ju (Tessa) Chen as a co-author. | Fadel M. Megahed | Fadel M. Megahed |
| D39 | Name the R package decisionqc and the research repository decision-models-for-industrial-image-qc. | Claude Code | Fadel M. Megahed |
| D40 | Organize the repository before the first push. | Fadel M. Megahed | Fadel M. Megahed |
| D41 | Make the research repository public, so the record website works on the free plan. | Claude Code | Fadel M. Megahed |
| D42 | License the repository under MIT. | Claude Code | Fadel M. Megahed |
| D43 | Cite the CLIP study as the 2026 Journal of Quality Technology article, which is open access. | Fadel M. Megahed | Fadel M. Megahed |
| D44 | Set the JEV budget to about USD 5, the free credit for a new user. | Fadel M. Megahed | Fadel M. Megahed |
| D45 | Use APA 7 style for citations, with the published metadata of each work. | Fadel M. Megahed | Fadel M. Megahed |
| D46 | Use short, direct section headers in all documents. | Fadel M. Megahed | Fadel M. Megahed |
| D47 | Make one Quarto document that reproduces all paper figures and tables; it reads saved predictions by default and has switches to re-run the paid and GPU steps. | Claude Code | Fadel M. Megahed |
| D48 | Produce tables T1 to T7, figures F1 to F7 and the appendix table of all trials. | Claude Code | Fadel M. Megahed |
| D49 | Use theme_bw with Dark2, annotations and a polished font; save each figure as PNG, PDF and SVG in one-column and two-column sizes. | Claude Code | Fadel M. Megahed |
| D50 | Use Outfit for figure titles and Cabin for all other figure text. | Fadel M. Megahed | Fadel M. Megahed |
| D51 | Name groups in their chart color in the subtitle (ggtext) in place of a legend. | Fadel M. Megahed | Fadel M. Megahed |
| D52 | Round reported numbers half up (0.725 to 0.73), as in the submitted abstract, because floating point makes round() and sprintf() give 0.72. | Claude Code | Claude Code |
| D53 | Show all tables of the replication document as interactive DT tables; keep the CSV and LaTeX files for the paper. | Fadel M. Megahed | Fadel M. Megahed |
| D54 | In figures F3 and F4, give each method its own row or column, with larger points and no vertical nudging. | Fadel M. Megahed | Fadel M. Megahed |
| D55 | Publish figure F1 with its image sources on the figure; its AutoVI (pipe staple) tiles stay under CC BY-NC-SA 4.0. | Fadel M. Megahed | Fadel M. Megahed |
| D56 | Keep beyond_clip_analysis.html in the repository and show it on the record website as an extra page (replication-materials 1.1.0). | Fadel M. Megahed | Fadel M. Megahed |
| D57 | Put "(Few-Shot)" in parentheses in the title: Beyond CLIP: Calibrated (Few-Shot) Defect Detection With Structured Decision Models. | Fadel M. Megahed | Fadel M. Megahed |
| D58 | Also report CLIP with the learning sets of the CLIP paper at a threshold tuned like ours, next to the rule of that paper. | Fadel M. Megahed | Fadel M. Megahed |
| D59 | Tune that CLIP threshold by leave-one-out on the learning set (identical frames left out together), because the learning sets hold most of our validation images (all 16 for pans). | Claude Code | Claude Code |
| D60 | In the interactive T4, mark each JEV and Clef-Flash result against the best CLIP result on the same dataset: up, level or down at a margin of 0.02 balanced accuracy. | Fadel M. Megahed | Claude Code |
| D61 | Record the R packages in an renv lockfile (renv.lock) without activating renv in the project, because the project folder is in Dropbox. | Fadel M. Megahed | Claude Code |
| D62 | The benchmark is a CLIP baseline with 50 learning images per class (pans: 10), the low end of the abstract of the CLIP paper, with the defect images split evenly over the defect types; it replaces the best CLIP result as the reference of the glyphs, and the other columns stay. | Fadel M. Megahed | Fadel M. Megahed |
| D63 | Add a dot-chart column to the interactive T4: one row per method, balanced accuracy on a common axis, dashed line at the CLIP baseline (inline SVG, no script). | Fadel M. Megahed | Fadel M. Megahed |
| ID | Result | Value |
|---|---|---|
| H1 | JEV with 2 to 4 references: test balanced accuracy on the five datasets | 0.73 to 1.00 (pans 1.00, extrusion 0.73, textiles 0.86, pipe staple 0.98, microstructure 0.92) |
| H2 | JEV against CLIP with the same references: equal or better on all five datasets | microstructure equal (0.92 and 0.92); JEV better on the other four |
| H3 | JEV against CLIP with the learning sets of the CLIP paper (20 to 700 labeled images): JEV better on three datasets, CLIP slightly better on two | pipe staple 0.98 vs 0.57, pans 1.00 vs 0.91, microstructure 0.92 vs 0.85; textiles 0.86 vs 0.975, extrusion 0.73 vs 0.82 |
| H4 | Calibration of the JEV probabilities on the test sets | raw ECE 0.11 to 0.23; isotonic regression 0.01 to 0.04 on four of the five datasets |
| H5 | Clef-Flash with the frozen JEV configurations: test balanced accuracy | 0.68 to 0.99 |
| H6 | Benchmark of the analysis document: the best CLIP result on the test set for each dataset | pans 0.94, extrusion 0.83, textiles 0.98, pipe staple 0.61, microstructure 0.96 |
| H7 | JEV with references against the CLIP baseline (50 learning images per class, pans 10; rule of the CLIP paper): JEV higher on four datasets, lower on textiles | pans 1.00 vs 0.91, extrusion 0.73 vs 0.70, textiles 0.86 vs 0.98, pipe staple 0.98 vs 0.57, microstructure 0.92 vs 0.84 |
| H8 | Validation size (JEV with references): test balanced accuracy settles at about 10 to 20 validation images per class; the calibration error keeps falling up to 50 | mean test balanced accuracy at 20 and at 50 per class: extrusion 0.726 and 0.725, textiles 0.848 and 0.855, pipe staple 0.991 and 0.980, microstructure 0.920 and 0.920 (pans have 8 per class only) |