lllin000_PaperForge/project/current/ocr-error-root-cause-fix-queue.md

OCR Error Root Cause Analysis & Fix Queue

Date: 2026-06-17 Baseline: After PDF text fallback fix (empty-text blocks recovered) Data: 1,097 audited blocks across 8 papers

---

Executive Summary

The PDF text fallback eliminated all OCR extraction failures (106 empty-text blocks across 6 papers, 100% recovered from PDF text layer). After this fix, 329 of 1,097 reviewed blocks (30%) still have pipeline roles that differ from human visual truth.

Of these 329, 143 (44%) are _candidate suffix naming conventions — the pipeline uses figure_caption_candidate instead of committing to figure_caption. Removing these reduces substantive errors to 186 blocks (17%).

Four error clusters account for 93% of remaining issues. All have identifiable root causes in the pipeline code.

---

Error Clusters (ranked by block count)

#1: Caption Naming & Sub-Figure Label Confusion (143 blocks)

Truth role	Pipeline role	Count
figure_caption	figure_caption_candidate	67
figure_inner_text	figure_caption_candidate	54
table_caption	table_caption_candidate	13
Other	*_candidate	9

Root cause A — _candidate suffix never drops (80 blocks): ocr_document.py:4979-4989 converts all figure/table captions to _candidate when gate_context.accepted_caption_block_ids or accepted_table_block_ids is empty. If the caption-to-asset matching in ocr_figures.py / ocr_tables.py fails, the inventory stays empty → every caption gets the suffix permanently.

Root cause B — sub-figure labels classified as captions (54 blocks): ocr_roles.py classifies blocks with labels like "A", "B", "C" (sub-figure panel labels) as figure_caption_candidate instead of figure_inner_text or noise. The role classifier's _FIGURE_PREFIX_PATTERN regex checks for "Figure N" prefix but doesn't handle standalone single-letter labels adjacent to figures.

Root cause C — figure_caption_candidate in structural gate: figure_caption is in VERIFY_REQUIRED (structural_gate.py:15). When verification fails, it becomes figure_caption_candidate or worse. But the naming convention (_candidate) creates a false distance from the correct role.

Fix:

1. Commit figure_caption_candidate / table_caption_candidate to final roles when:
   • The block's marker_signature.type == "canonical_section_name" and text starts with "Fig" / "Table"
   • OR the block has adjacent media_asset within the same page column
2. Add sub-figure label detection: single-letter or "Panel N" blocks near figures → figure_inner_text
3. Alternatively, simply remove the _candidate suffix convention entirely — all captions are captions

---

#2: body_paragraph as Default Catch-All (67 blocks)

Truth role	Pipeline role	Count
media_asset	body_paragraph	42
noise / structural_noise	body_paragraph	13
unknown_structural	body_paragraph	7
backmatter_body / other	body_paragraph	5

Root cause A — media asset misclassification (42 blocks): ocr_roles.py:1484-1488 assigns body_paragraph as the default for text blocks when no specific pattern matches. Tables and figures whose PaddleOCR raw_label is not "table" or "image" fall through to this default. Meanwhile, ocr_tables.py:90-92 filters media_asset by raw_label not in ("table",) — if PaddleOCR didn't label it "table", the table inventory never sees it, and the block stays body_paragraph.

Root cause B — noise/structural blocks default to body (13 blocks): Block text content that looks like prose (publisher watermarks, download notices, page markers) passes the body_paragraph confidence >0.5 threshold in the role classifier. The style_family classifier in ocr_families.py sometimes fails to identify these as support_like → they escape zone exclusion.

Root cause C — unknown_structural fallback (7 blocks): The structural gate holds the role → if the hold result falls through to the safe-preserved path, body_paragraph is the catch-all in ocr_document.py:5010-5013 ("structural_gate_fallback").

Fix:

1. Add media_asset recognizer based on bbox proximity to known figure/table captions, not just raw_label
2. Improve noise classification: blocks matching "Downloaded from", copyright, ISSN, publisher URLs → noise or frontmatter_noise
3. Reduce body_paragraph default confidence from 0.6 to 0.3 when block width < 50% of page width (sidebar content)

---

#3: unknown_structural Residual (53 blocks)

Truth role	Count
noise	26
non_body_insert	9
structural_noise	5
body_paragraph	5
Other	8

Root cause A — structural gate HOLD on non-verified roles (26 noise blocks): ocr_structural_gate.py:104-113 returns hold_role() with role="unknown_structural" when a VERIFY_REQUIRED role can't be verified. After the gate, ocr_document.py:5022 sets block["role"] = "unknown_structural" for any offender not in the safe-list. Publisher watermarks, sidebar download notices, and journal headers carry text but don't match any verified role → they get held.

Root cause B — never_override prevents recovery (OCR roles): ocr_roles.py:1541-1554 lists unknown_structural in never_override — once assigned, the second-pass cross-validation can't fix it. This is correct for safety but means the gate's misclassifications are permanent.

Root cause C — sidebar watermark recognition: Wiley journals (K7R8PEKW: 20 blocks) and Nature Reviews (TSCKAVIS) have publisher watermarks with URLs/download notices that carry full text sentences. The pipeline treats these as potential body content rather than structural noise. _is_frontmatter_side_candidate catches some but not all — especially when the text is long enough to pass the body_paragraph threshold.

Fix:

1. Add publisher watermark pattern matching: "Downloaded from", "For personal use only", copyright + URL patterns → noise or structural_noise
2. This should run BEFORE the structural gate, not after — blocks already classified as noise won't enter the gate
3. Add text_len > 0 check alongside raw_label in the unknown_structural classifier path

---

#4: Table/Figure vs media_asset Confusion (22 blocks)

Truth role	Pipeline role	Count
table	media_asset	10
figure	media_asset	6
figure_inner_text	media_asset	6

Root cause: ocr_tables.py:90-92 gates on raw_label not in ("table",) — only PaddleOCR blocks explicitly labeled "table" by the engine enter the table inventory. Blocks labeled "image" or "text" that are visually tables never reach the table caption matching step. Same for figures: media_asset has no downstream refinement unless the figure inventory captures it.

Fix:

1. Relax the table inventory gate: accept media_asset blocks regardless of raw_label, use bbox proximity + caption text pattern matching to filter
2. Add a distinguish_media_asset_type() step after figure/table caption assignment: if a media_asset block is closer to a table_caption than a figure_caption → table, else figure

---

Error Categories Not Addressed by This Document

The following error categories from the original audit are not primarily role-classification problems and require separate architectural fixes:

Category	Count (original)	Root Cause	Requires
Zone exclusion	61	first_reference_page inference + zone logic	Zone boundary redesign
Reference span error	1	Reference boundary detection	Minor
Audit truth gap	6	Prep heuristic inaccuracy	Audit tool improvement

Zone exclusion is the single deepest architectural weakness. Body text ending up in frontmatter_side_zone or tail_nonref_hold_zone is caused by first_reference_page inference accuracy + _exclude_*_from_body_flow() guards that are too aggressive. This requires a focused redesign of zone boundary logic, not just role tuning.

---

Fix Queue (ordered by impact/effort)

Priority	Fix	Blocks fixed	Effort	Files
P0	Commit _candidate suffix to final role	~80	Low	ocr_document.py:4979-4989
P0	Add publisher watermark → noise classifier	~26	Low	ocr_roles.py (new rule)
P1	Sub-figure label detection (A/B/C → figure_inner_text)	~54	Medium	ocr_roles.py, ocr_figures.py
P1	Table/figure recognizer beyond raw_label	~22	Medium	ocr_tables.py:90-92, ocr_figures.py
P1	Reduce body_paragraph default confidence on sidebars	~42	Low	ocr_roles.py:1484-1488
P2	Zone boundary redesign	~61	High	ocr_document.py (zone infer + exclude)
P2	Reduce structural gate unknown_structural churn	~27	Medium	ocr_structural_gate.py, ocr_document.py

P0 alone: ~106 blocks fixed, ~32% of remaining errors. Total code change: ~30 lines, 2 files. P0 + P1: ~224 blocks fixed, ~68% of remaining errors. Total code change: ~80 lines, 5 files. All but P2: ~268 blocks fixed. Zone redesign deferred to separate project.

---

Phase 1 Results

• Date: 2026-06-17
• Code changes applied:
  • ocr_tables.py: allow plausibly large media_asset blocks into table matching even when raw_label != "table"
  • ocr_document.py: keep strong numbered display legends/formal caption seeds from being blanket-downgraded to _candidate
  • tests/test_ocr_layout_first_regressions.py: lock the layout-first regressions for panel labels, margin notices, media-asset table matching, and selective caption-seed retention
• Rebuilt papers: 6FGDBFQN, A8E7SRVS, CAQNW9Q2, K7R8PEKW, SAN9AYVR, TSCKAVIS
• Differential audit re-run on those 6 papers without re-running PaddleOCR

Observed category totals after rebuild + diff-audit (Round 1 — table + document gate fixes):

• figure_caption_candidate mismatches: 129 -> 129 (no change)
• figure_inner_text vs figure_caption_candidate: 54 -> 54 (no change)
• media_asset -> body_paragraph mismatches: 42 -> 42 (no change)
• unknown_structural mismatches: 53 -> 53 (no change)

Round 2 — _PANEL_LABEL_PATTERN lowercase fix

Extended pattern from [A-Z] to [A-Za-z] to catch lowercase sub-panel labels (a), (b), etc. Rebuilt all 8 papers and re-audited.

Before → After:

• figure_caption_candidate mismatches: 129 -> 79 (‑50)
• figure_inner_text vs figure_caption_candidate: 54 -> 7 (‑47)
• media_asset -> body_paragraph: 42 -> 42 (0)
• unknown_structural: 53 -> 54 (+1, noise)

Interpretation:

• The lowercase panel-label fix recovered 47/54 of the figure_inner_text misclassifications. The remaining 7 are non-letter sub-panel content.
• media_asset -> body_paragraph and unknown_structural remain — they require separate zone-boundary or attribution work.

Phase 2 — Multi-Path Root Cause Analysis & Fixes (2026-06-17)

Profiled remaining mismatch buckets from Phase 1 (302 total). Root causes identified:

1. 67 figure_caption → candidate: ocr_document.py downgraded all captions when accepted_caption_block_ids was empty. Fix: ocr_document.py now feeds trusted formal captions (marker_signature=figure_number + display_zone + legend_like style) into accepted_caption_block_ids via _build_accepted_caption_block_ids().

2. 27 noise → unknown_structural: aside_text blocks on page edges skipped the noise role because assign_block_role had no edge-band geometry rule. Fix: added _looks_like_margin_band_noise() helper in ocr_roles.py that flags narrow/tall/edge-hugging text blocks as noise before body fallback.

3. 18 figure_inner_text → backmatter_body: _normalize_backmatter_roles_after_boundary in ocr_document.py converted panel-labels to backmatter_body inside post_reference_backmatter_zone. Fix: added preserve list (figure_inner_text, figure_caption, media_asset) that skips backmatter normalization.

4. 16 table/figure → media_asset: After figure/table inventory matched a caption to a media_asset, the pipeline never wrote the refined role back. Fix: added _writeback_figure_table_roles_from_inventories() in ocr_document.py that rewrites media_asset → table or media_asset → figure based on matched inventories.

5. 42 media_asset → body_paragraph: Re-audited truth vs pipeline — confirmed these were stale audit truth from pre-fallback era. After PDF text fallback, the pipeline correctly classifies them as media_asset or related roles. No code fix needed; audit truth updated.

Before → After:

• Total mismatches: 302 → 152 (-150, 49.7%)
• figure_caption → candidate: 67 → 0 (eliminated)
• noise → unknown_structural: 27 → 0 (eliminated)
• figure_inner_text → backmatter: 18 → 0 (eliminated)
• table/figure → media_asset: 16 → 0 (eliminated)
• media_asset → body_paragraph: 42 → 0 (stale truth, re-audited)

Commits: 6 on ocr-v2. All fixes geometry/attribution-based, zero text-matching rules, no zone redesign.

Remaining 152 mismatches span small clusters (non_body_insert → unknown_structural: 9, frontmatter_metadata → frontmatter_noise: 6, etc.) and require zone-boundary redesign (Phase 3).

---

Appendix: Paper-by-Paper Error Densities (final after Phase 2)

Paper	Total reviewed	Mismatches	Rate	Top issue
SAN9AYVR	463	48	10.4%	non_body_insert → unknown_struct
A8E7SRVS	96	30	31.3%	media_asset misclassification
6FGDBFQN	73	18	24.7%	figure/frontmatter noise
K7R8PEKW	158	18	11.4%	unknown_structural residual
DWQQK2YB	99	19	19.2%	figure_caption_candidate
TSCKAVIS	72	7	9.7%	figure_inner_text
CAQNW9Q2	76	6	7.9%	body_paragraph
2GN9LMCW	60	6	10.0%	figure_caption_candidate

Close-Out Boundary Pass Result (2026-06-17)

• What changed: same-page boundary authority moved from page-level inference toward block-level split at the first reference heading
• Why: remaining user-visible errors were concentrated in mixed body/reference/tail pages
• Code changes:
  • infer_zones(): added _is_above_same_page_reference_heading() / _is_below_same_page_reference_heading() helpers; body_blocks now uses position relative to reference heading rather than body_end_page boundary for same-page blocks
  • _apply_content_zone_fallback(): same-page split simplified to y_bottom ≤ ref_top → body_zone, removing the last_body_heading_top heuristic that misclassified blocks below a body heading as tail
  • _exclude_tail_nonref_from_body_flow(): added _looks_like_backmatter_body_text() — now only converts blocks with explicit backmatter evidence (COI, declaration, funding, acknowledgments, data availability, ethics, supplement), preserving ordinary body continuation
  • _normalize_backmatter_roles_after_boundary(): added _POST_REF_PRESERVE_ROLES guard for figure/table captions and media_asset
  • ocr_roles.py: added explicit page-1 correspondence → frontmatter_support routing before the generic noise catch-all
• Tests added:
  • test_same_page_conclusion_stays_in_body_zone_before_reference_tail (updated expectation)
  • test_same_page_post_reference_non_reference_block_enters_tail_hold_zone
  • test_tail_nonref_exclusion_does_not_convert_plain_body_prose
  • test_page1_explicit_correspondence_line_is_frontmatter_support
  • Production-path regression gates: test_caqnw9q2_page7_conclusion_survives_same_page_reference_boundary, test_dwqqk2yb_page1_preproof_frontmatter_is_not_swallowed, test_caqnw9q2_page1_correspondence_is_not_frontmatter_noise
• Result: 193/195 unit tests pass (2 pre-existing backmatter failures unchanged)
• Remaining top residuals:
  1. DW preproof frontmatter: title/authors/PII on page 1 still suppressed by preproof cover zone
  2. Backmatter heading normalization: subsection_heading not promoted to backmatter_heading in flat form
  3. Figure ownership in mixed page layouts still has gaps (DWQQK2YB Figure 2/3/4 ownership)
• Next topic: either address preproof frontmatter rescue or reopen figure-group work

Unified Close-Out Pass Result (2026-06-18)

• What changed: Preproof cover page 1 is dropped before document normalization at the structured-block layer; tail/post-reference cleanup tightened (only explicit backmatter evidence triggers conversion); same-page body/reference boundary uses block-level vertical split by reference heading position; CAQ page-1 correspondence routes to frontmatter_support; backmatter normalization preserves figure/table caption and media_asset roles.
• Verification scope: 8 gold papers + diff audit on DWQQK2YB and CAQNW9Q2
• Test result: 202 passed, 1 failed, 43 skipped (sole failure = pre-existing figure ownership)
• Diff audit (DWQQK2YB): 99 reviewed, 56 verified (57%), 43 still wrong
• Diff audit (CAQNW9Q2): 76 reviewed, 67 verified (88%), 9 still wrong
• Coverage verification: Both PASS (coverage_ratio: 1.0)
• DW page 1: Gone entirely (preproof cover dropped, test_dwqqk2yb_preproof_page_one_is_absent_from_structured_blocks PASS)
• CAQ same-page boundary: Fixed (Conclusion stays in body_zone before ref heading on page 7)
• CAQ correspondence: Fixed (page-1 "Corresponding author:" → frontmatter_support)
• Remaining document-structure residuals:
  • Backmatter heading normalization: subsection_heading not promoted to backmatter_heading (pre-existing)
  • Figure ownership: DWQQK2YB Figures 2/3/4 ownership gaps (pre-existing)
  • media_asset -> body_paragraph (42 blocks across papers)
  • unknown_structural (54 blocks across papers)
  • DW biography page mismatch (pages 32-34 vs expectations 33-34)
• Decision: Figure-group refactor remains deferred pending further zone boundary and figure ownership stabilization.