lllin000_PaperForge/docs/audit/6-paper-audit.md

OCR Truth Audit — 6-Paper Full Audit

Date: 2026-07-05 Auditor: Automated pipeline analysis + visual verification

---

Summary

Paper	Status	Fig Matched	Fig Captions	Unmatched	Issue Category
SKXTCE6M	🟡 Yellow	14/17	17	6	Table→figure routing
SRNJDAA2	🟡 Yellow	23/32	32	9	Body text → legend false positive
3FDT9652	🟡 Yellow	6/8	8	1	Caption→asset matching gap
6QNRHRKX	🟡 Yellow	0/7	7	7	Figure detection failure
SWDN9RHF	🔴 Red	1/1	1	0	Structural: no abstract
SAN9AYVR	🟢 Green	31/34	34	0	Clean (34 orphans normal)

---

1. SKXTCE6M — Table→Figure Routing ⚠️

Paper: Hashemi-Afzal et al., 2025 — Advancements in hydrogel design for articular cartilage

Finding: 6 table titles classified as figure_caption_candidate because raw_label=figure_title was assigned by OCR to bold/centered text above HTML <table> blocks. The descriptive text ("A summary of hydrogels...") becomes unmatched legend, disappears from rendered fulltext.

Output impact: 6 blocks of descriptive table content LOST from markdown. 4 tables rendered as images instead of HTML on pages 11/15/19/23 due to media_asset role.

Root cause: Raw label figure_title doesn't distinguish tables from figures. Adjacency check to <table> block missing.

Fix: In role classification: if raw_label=figure_title and next sibling is table_html or media_asset → assign table_caption.

---

2. SRNJDAA2 — Body Text → Legend False Positive ⚠️

Paper: (Book chapter on piezoelectric biomaterials, ~1970s)

Finding: 9 unmatched legends, all role=body_paragraph. They contain "Figure 11-10 shows the x-ray picture..." — inline references to figures in the running text. The pipeline picks up any block starting with "Figure X-Y " as a potential figure legend, even when it's clearly body text.

Output impact: MINIMAL. Text is correctly rendered as body paragraphs. The false positives only inflate unmatched_legends count, yellowing the health status.

Root cause: Legend detection is too aggressive. "Figure N-M shows..." should be distinguished from "Figure N-M. (caption)".

Fix: Filter body_paragraph blocks from legend candidates. If a block's role (after classification) is body_paragraph, it cannot simultaneously be a figure legend.

---

3. 6QNRHRKX — Complete Figure Detection Failure 🔴

Paper: Weiner & Macnab, 1970 — Superior Migration of the Humeral Head (radiology)

Finding: 7 figure captions (Fig. 2 through Fig. 7) identified but ZERO figure assets matched. Captions are very short ("Fig. 2" through "Fig. 7") — just labels, no descriptive text. Figures appear to be embedded in two-column text layout. Orphans are sub-panels (~12KB each) that couldn't be matched.

Output impact: SEVERE. Figures appear in markdown as bare > Fig. 2 blockquotes with no image. One descriptive caption text lost entirely.

Root cause: Short label-only captions lack the text overlap that the matching algorithm uses to pair captions with figure regions. Possibly combined with figures not detected as image regions by page layout analysis.

Fix: Needs investigation of the actual PDF layout — short captions in radiology papers may need a different matching heuristic.

---

4. 3FDT9652 — Single Caption Gap ⚠️

Paper: (Rotator cuff / shoulder tendon paper)

Finding: 6/8 figures matched. One unmatched legend: "A, Full-thickness tear of supraspinatus tendon..." — this is a partial caption starting with a sub-panel label. 2 orphans exist.

Output impact: 1 figure caption lost from rendered output.

---

5. SWDN9RHF — Structural Red, Figure Quality Good ✅

Paper: Roemer et al. — MRI-based semiquantitative assessment of subchondral bone marrow lesions

Finding: RED status is structural (abstract_found: False, structural_blockers: 2), not OCR quality. The single figure (Fig. 1) is correctly matched and cropped. The paper has no abstract section in the traditional sense.

Output impact: None for figure matching. Red status is a false alarm driven by paper formatting.

---

6. SAN9AYVR — Green but High Orphan Count ✅

Paper: (Large review, 81 pages, 31 matched figures)

Finding: All 31 figures correctly matched. 34 orphans exist but are tiny (2-7KB) — likely small icons, section dividers, plot markers. This is normal behavior for a 81-page paper with dense graphics.

Output impact: None.

---

Categorized Findings

By Fixability

Priority	Finding	Effort	Impact
P0	Table→figure routing (SKXTCE6M style)	~5 lines + test	6 lost content blocks per affected paper
P0	Body text → legend false positive (SRNJDAA2 style)	~3 lines + test	Noisy health status
P1	Short caption matching (6QNRHRKX style)	Needs layout research	Severe for affected papers
P2	Structural red for no-abstract papers (SWDN9RHF)	~2 lines	Wrong health signal

Cross-cutting Theme

The health status (yellow/red) is not a reliable indicator of figure matching quality. Three different mechanisms can trigger yellow:

1. Table→figure routing (real content loss) ← SKXTCE6M
2. Body text→legend false positives (noise only) ← SRNJDAA2
3. Unmatched legends from detection gaps (real) ← 6QNRHRKX

Without quality indicators (build_quality_indicators) the health report conflates these into one undifferentiated "yellow".