Update README.md

README.md

@@ -89,22 +89,6 @@ Synthetic checkbox crops generated from PDF answer-sheet templates using the scr
 
 ---
 
-## Limitations & Known Issues
-### Gray-ink misclassification
-The training dataset uses synthetic marks with pixel grayscale values in the 10–80 range, simulating dark pen strokes. During real-world testing, lightly-marked checkboxes (pixel values ~128–200, typical of light gray ink or faint pencil strokes) fall outside this distribution and are frequently misclassified as empty.
-
-An interim workaround — applying Otsu's binarization to normalize gray marks to black before inference — has been deployed in the Plom integration (plom/plom!3702), but this is a pre-processing patch rather than a model fix.
-
-### Erased-mark ambiguity
-Erased marks and lightly-filled marks occupy the same grayscale range (~128–200). The model currently has no mechanism to distinguish between:
-
-A checkbox lightly filled with gray ink (should be filled)
-A checkbox that was marked and then erased (should be empty)
-This is an open problem: solving it likely requires either multi-threshold classification (e.g., a three-class filled / erased / empty output) or additional context such as smear patterns or residual graphite texture.
-
-### Future Work
-Retrain on a wider grayscale range: Expand the synthetic data generation to include filled marks across the full 10–200 pixel intensity range to handle gray ink natively without Otsu pre-processing (tracked in plom/plom#4277).
-Erased mark handling: Investigate whether texture features or a three-class head (filled / erased / empty) can separate light fills from erasure residue.
 
 ## Inference
 
@@ -178,3 +162,22 @@ Six answer-sheet layouts are supported. Each variant has a corresponding templat
 Designed for automated grading pipelines that pre-crop answer-sheet checkboxes (after scan alignment and template matching) and need a fast, CPU-friendly classifier to decide filled vs. empty. Intended to run as part of the [plom-digit-recognition-server](https://gitlab.com/plom/plom).
 
 **Out of scope**: general object detection, handwriting recognition, forms with non-checkbox mark types.
+
+
+## Limitations & Known Issues
+### Gray-ink misclassification
+The training dataset uses synthetic marks with pixel grayscale values in the 10–80 range, simulating dark pen strokes. During real-world testing, lightly-marked checkboxes (pixel values ~128–200, typical of light gray ink or faint pencil strokes) fall outside this distribution and are frequently misclassified as empty.
+
+An interim workaround — applying Otsu's binarization to normalize gray marks to black before inference — has been deployed in the Plom integration (plom/plom!3702), but this is a pre-processing patch rather than a model fix.
+
+### Erased-mark ambiguity
+Erased marks and lightly-filled marks occupy the same grayscale range (~128–200). The model currently has no mechanism to distinguish between:
+
+A checkbox lightly filled with gray ink (should be filled)
+A checkbox that was marked and then erased (should be empty)
+This is an open problem: solving it likely requires either multi-threshold classification (e.g., a three-class filled / erased / empty output) or additional context such as smear patterns or residual graphite texture.
+
+### Future Work
+Retrain on a wider grayscale range: Expand the synthetic data generation to include filled marks across the full 10–200 pixel intensity range to handle gray ink natively without Otsu pre-processing (tracked in plom/plom#4277).
+Erased mark handling: Investigate whether texture features or a three-class head (filled / erased / empty) can separate light fills from erasure residue.
+