🔍 ASSETS NEEDED FOR FINAL REPORT: Rakīm AI (رَقِيم)

To write the final polished official paper/report, please compile and provide the following assets. These will be embedded as figures, benchmarks, and data tables to impress the national competition judges.

1. Real Application Screenshots

You can capture these directly from your browser by running the local server (python run.py inside backend) and navigating to http://localhost:8001/app:

Figure 1: Pipeline Overview (Interface)
- What to capture: A widescreen screenshot of the workspace after uploading صور/واضحة_1926.jpg and running HTR. It should show the original manuscript with green/blue bounding boxes on the right and the clean Arabic transcription on the left.
Figure 2: Text Margin Separation (Matn vs. Hashiya)
- What to capture: Page صور/حواشٍ_1960.jpg loaded. Take two crops:
  1. Clicking 📚 الكل (showing all bounding boxes, including marginalia).
  2. Clicking 📄 المتن (showing only main text boxes highlighted, illustrating the classification).
Figure 3: Interactive Word Alternatives (Reading Candidates)
- What to capture: Click a low-confidence line (e.g. containing "المقربير"), open the line pop-up window, click ❓ بدائل, and capture the dropdown list displaying candidate percentages ranking "المقربين" (55.8%) at the top.
Figure 4: AI Analysis & Contextual Explanations
- What to capture: The AI panel/modal displaying:
  1. Suggestions for page summary and title: «قصص الخلق وأول النور».
  2. A list of extracted named entities (وهب بن منبه, سفيان الثوري, Ibn Abbas).
  3. The dictionary definition of terms in context (e.g., "المنطق" or "الريح العقيم").
Figure 5: Fuzzy Manuscript Search
- What to capture: Type "الحوت" in the search box, click Search, and capture the workspace showing the matching lines highlighted.
Figure 6: Manuscript Catalog Search & Content Matcher
- What to capture: The "فهرس المخطوطات" tab showing a search for "شرح فصول أبقراط", displaying the list of copies and matched metadata fingerprint terms.

2. Dataset and Calligraphy Details

Visual Families of Hands:
- If available, please provide the visual cluster parameters of the ~8 manuscript families (e.g., line count, line spacing, stroke thickness, or average character heights).
Source Collections:
- The names of the specific libraries or archives from which the training manuscript subsets were obtained (e.g., BULAC, Wellcome Collection).
RASAM & TariMa Scientific References:
- Standard bibliographic details or citations for the RASAM (Arabic manuscript dataset) and TariMa corpuses.

3. Structural Graphics (Raw Images / Line Segmentations)

BLLA Segmentation Output:
- If you have training logs or visualization outputs comparing raw segmentations between seg_best and logic_philosophy_v2_seg (e.g., showcasing how the logic segmenter handles bordered margins without shattering), please provide them.
Over-segmentation Samples:
- Sample crops of Page 417 or Page 1982 showing lines broken into stray fragments, to illustrate the layout classification bottleneck in the "Limitations" section.