From c974ba74801566aeddfbbfd631e34600fe6a9af9 Mon Sep 17 00:00:00 2001
From: Research Assistant <research@example.com>
Date: Sat, 6 Jun 2026 13:28:46 +0800
Subject: [PATCH] fix: remove open-ended authors text heuristics, anchor to
 page-1 zone only

---
 _check_rebuild.py                             |  50 ++
 ...r-author-anchor-and-nonbody-insert-plan.md | 439 ++++++++++++++++
 ...06-ocr-structure-truth-convergence-plan.md | 480 ++++++++++++++++++
 paperforge/worker/ocr_roles.py                |  20 -
 tests/test_ocr_roles.py                       |  25 +
 5 files changed, 994 insertions(+), 20 deletions(-)
 create mode 100644 _check_rebuild.py
 create mode 100644 docs/superpowers/plans/2026-06-06-ocr-author-anchor-and-nonbody-insert-plan.md
 create mode 100644 docs/superpowers/plans/2026-06-06-ocr-structure-truth-convergence-plan.md

diff --git a/_check_rebuild.py b/_check_rebuild.py
new file mode 100644
index 00000000..5ff13a5a
--- /dev/null
+++ b/_check_rebuild.py
@@ -0,0 +1,50 @@
+from pathlib import Path
+import json
+
+vault = Path(r"D:\L\OB\Literature-hub")
+structured_path = vault / "System" / "PaperForge" / "ocr" / "SAN9AYVR" / "blocks.structured.jsonl"
+profiles_path = vault / "System" / "PaperForge" / "ocr" / "SAN9AYVR" / "role_span_profiles.json"
+
+if structured_path.exists():
+    print(f"structured blocks: {structured_path.stat().st_size} bytes")
+    with open(structured_path, encoding="utf-8") as f:
+        line = f.readline()
+        row = json.loads(line)
+        has_span = "span_metadata" in row
+        has_role = row.get("role")
+        has_conf = row.get("role_confidence")
+        print(f"  first block role={has_role}, conf={has_conf}, has_span_metadata={has_span}")
+
+    total = 0
+    with_span = 0
+    with open(structured_path, encoding="utf-8") as f:
+        for line in f:
+            total += 1
+            row = json.loads(line)
+            if row.get("span_metadata"):
+                with_span += 1
+    print(f"  total blocks: {total}, with span_metadata: {with_span} ({100*with_span/total:.1f}%)")
+else:
+    print("structured blocks not found")
+
+if profiles_path.exists():
+    profiles = json.loads(profiles_path.read_text(encoding="utf-8"))
+    print(f"\nrole_span_profiles.json: {len(profiles)} roles")
+    for role, p in sorted(profiles.items()):
+        print(f"  {role}: count={p['block_count']}, size={p['mean_size']}, quality={p['quality']}")
+else:
+    print("role_span_profiles.json not found")
+
+second_pass_total = 0
+span_alt_total = 0
+with open(structured_path, encoding="utf-8") as f:
+    for line in f:
+        row = json.loads(line)
+        ev = row.get("evidence", [])
+        for e in ev:
+            if "second_pass" in e:
+                second_pass_total += 1
+            if "span_alternatives" in e:
+                span_alt_total += 1
+print(f"\nSecond-pass applied: {second_pass_total} blocks")
+print(f"Span alternatives flagged: {span_alt_total} blocks")
diff --git a/docs/superpowers/plans/2026-06-06-ocr-author-anchor-and-nonbody-insert-plan.md b/docs/superpowers/plans/2026-06-06-ocr-author-anchor-and-nonbody-insert-plan.md
new file mode 100644
index 00000000..4d4dd01e
--- /dev/null
+++ b/docs/superpowers/plans/2026-06-06-ocr-author-anchor-and-nonbody-insert-plan.md
@@ -0,0 +1,439 @@
+# OCR Author Anchor And Non-Body Insert Plan
+
+Date: 2026-06-06  
+Scope: `ocr_roles.py`, `ocr_document.py`, `ocr_metadata.py`, `ocr_blocks.py`, related tests, real-paper validation on `SAN9AYVR`
+
+## Goal
+
+收掉当前最不稳的一条链：
+
+1. 开放式 `authors` 文本识别会误伤正文
+2. early-page 作者简介 / profile card / side insert 会被误救回正文
+3. `page` 号和人名信号都不该作为主判定依据
+
+目标是把这条线改成：
+
+- `authors` 只走 `frontmatter zone + Zotero/source metadata anchor`
+- `author bio` / `profile card` / `early-page side insert` 统一走 `non_body_insert` 路径
+- 先识别正文主 spine，再识别“不属于正文”的插入 cluster
+- 人名/履历句式只做增强信号，不做主判定
+
+## Current Root Cause
+
+`SAN9AYVR` 暴露了三类错误：
+
+### 1. 正文被误判成 `authors`
+
+例如 page 3 的：
+
+- `In Section 5, the focus is on ES based bioelectronics ... $^{8,49}$`
+
+被打成 `authors`。
+
+根因：
+
+- 当前 `ocr_roles.py` 对 author-like text 仍然开放式识别
+- superscript / comma / name-like token 在综述正文里很容易误触发
+
+### 2. page-1 / page-2 的作者简介块被当正文
+
+例如：
+
+- `Dr Ya Huang is currently ...`
+- `Zhenlin Chen is currently ...`
+- `Xinge Yu is the Associate Director ...`
+
+这些在 PDF 上是明显的 early-page 插入 profile blocks，不属于正文主链。  
+但现在要么直接被判 `body_paragraph`，要么先成 `frontmatter_noise` 再被 rescue 成 `body_paragraph`。
+
+根因：
+
+- 系统还缺少 `non_body_insert` 这一类结构类型
+- rescue 只看“像不像 body”，没先问“它是否属于 body spine”
+
+### 3. `authors` 和 `author bio` 的职责混在一起
+
+现在 `authors` 被拿来承担两种事情：
+
+- metadata/frontmatter 作者块
+- 各种 name-like non-body blocks
+
+这是不对的。两者必须拆开：
+
+- `authors`：frontmatter metadata anchor
+- `author_bio` / `profile insert`：non-body insert
+
+## Design Principles
+
+### 1. `authors` must be anchored, not guessed
+
+`authors` 只允许来自：
+
+- page-1 `frontmatter zone`
+- 与 Zotero/source metadata authors list 有足够匹配的 OCR block
+
+禁止：
+
+- 从正文任意 block 开放式猜 `authors`
+- 仅因有人名样式/上标/逗号就升成 `authors`
+
+### 2. Determine body spine first
+
+不是先识别 “sidebar”。
+
+而是：
+
+1. 先识别正文主 spine
+2. 再找不属于 body spine 的 early-page insert cluster
+3. 最后再用人名/履历语义做弱确认
+
+### 3. Name-like text is only a confidence booster
+
+人名信号只能在以下前提后使用：
+
+- 该 block 已被判定为 non-body cluster 成员
+
+用途：
+
+- 增强 `profile_insert` / `author_bio_insert` 置信度
+
+不能：
+
+- 直接产出 `authors`
+- 直接产出 `sidebar`
+
+### 4. Avoid absolute page heuristics
+
+不要写：
+
+- `page <= 3 -> maybe sidebar`
+- `page == 2 -> maybe author bio`
+
+可以使用的只有弱相对先验：
+
+- early document region
+- close to frontmatter/body transition
+- detached from body spine
+
+但 page 号不能单独决定角色。
+
+## Target End State
+
+After this plan:
+
+- metadata authors come from Zotero/source metadata truth, optionally aligned to a real OCR block
+- body text no longer gets promoted to `authors`
+- early-page author biography/profile cards become `non_body_insert`-family blocks
+- rescue no longer converts non-body insert blocks back into `body_paragraph`
+
+## Implementation Plan
+
+### Task 1: Remove open-ended `authors` text heuristics
+
+Files:
+
+- `paperforge/worker/ocr_roles.py`
+- tests in `tests/test_ocr_roles.py`
+
+#### 1.1 Remove generic author promotion from `text` blocks
+
+Find and delete or strongly demote logic that promotes arbitrary text to `authors` based on:
+
+- superscript affiliation markers
+- commas
+- name-like shapes
+- lack of year parens
+
+This logic is currently too broad and causes正文误识别.
+
+#### 1.2 Keep `authors` only in frontmatter zone
+
+Allow `authors` assignment only when:
+
+- page-1 zone detector returns `author_zone`
+- or a future anchored author-candidate check explicitly confirms it
+
+All other blocks should fall through to:
+
+- `body_paragraph`
+- `unknown_structural`
+- or later `non_body_insert`
+
+Acceptance:
+
+- body blocks can no longer become `authors` purely from text shape
+
+### Task 2: Add Zotero-anchored author candidate resolution
+
+Files:
+
+- `paperforge/worker/ocr_metadata.py`
+- optionally `ocr_roles.py`
+- tests in `tests/test_ocr_metadata.py`
+
+#### 2.1 Treat source metadata authors as truth
+
+For resolved metadata:
+
+- if Zotero/source metadata authors exist, they are canonical
+- OCR authors should only serve as alignment/traceability support
+
+#### 2.2 Add OCR block alignment, not OCR author guessing
+
+Implement a helper along the lines of:
+
+- `_match_author_block_to_source_authors(blocks, source_authors)`
+
+Use:
+
+- page-1 candidate blocks only
+- normalized name matching
+- token overlap / fuzzy similarity
+
+If matched:
+
+- keep block role as `authors`
+- store raw OCR authors evidence if useful
+
+If not matched:
+
+- metadata still uses source authors
+- do not invent OCR `authors`
+
+Acceptance:
+
+- wrong OCR `authors` block does not pollute `resolved_metadata.authors`
+
+### Task 3: Introduce `non_body_insert` / `profile_insert` regime
+
+Files:
+
+- `paperforge/worker/ocr_document.py`
+- `paperforge/worker/ocr_blocks.py`
+- `ocr_render.py` if render suppression is needed
+- tests in `tests/test_ocr_document.py`, `tests/test_ocr_render_stabilization.py`
+
+#### 3.1 Add a document-level detector for early-page non-body insert clusters
+
+New helper(s), for example:
+
+- `_detect_body_spine(blocks)`
+- `_detect_non_body_insert_clusters(blocks, body_spine, ...)`
+
+The detector should rely primarily on:
+
+- geometry
+- block continuity
+- cluster layout
+- style coherence
+
+Not on fixed text phrases.
+
+#### 3.2 What counts as a candidate cluster
+
+Signals:
+
+- block is not on the main body spine
+- block belongs to an early-page region near the body/frontmatter transition
+- multiple similar blocks appear in a local grid / side cluster
+- cluster width / line height / family differ from body spine
+- there is a visible vertical separation from body paragraphs
+
+Optional weak boost:
+
+- name-like lead
+- biography-like tense or CV-style sentence forms
+
+#### 3.3 Role strategy
+
+Do not overload `authors`.
+
+Either:
+
+- introduce explicit role `non_body_insert`
+
+or:
+
+- add `_non_body_insert = true` regime marker while keeping `body_paragraph` temporarily
+
+Recommendation:
+
+- use explicit role `non_body_insert`
+
+This keeps downstream logic simpler.
+
+Acceptance:
+
+- page 1/2 author bio blocks in `SAN9AYVR` no longer belong to body spine
+
+### Task 4: Block rescue from pulling insert clusters back into body
+
+Files:
+
+- `paperforge/worker/ocr_document.py`
+- tests in `tests/test_ocr_document.py`
+
+#### 4.1 Add rescue guard
+
+In `rescue_roles_with_document_context()`:
+
+- before rescuing `frontmatter_noise -> body_paragraph`
+- or similar body-promoting logic
+
+check:
+
+- whether the block belongs to a detected `non_body_insert` cluster
+
+If yes:
+
+- do not rescue to body
+
+#### 4.2 Body-like font is not enough
+
+This is the key fix for `SAN9AYVR`:
+
+- page-1/2 author bio blocks have body-like font size
+- but they are still not body
+
+So rescue must require:
+
+- body family match
+- and body-spine compatibility
+
+not just font match.
+
+Acceptance:
+
+- author bio/profile insert blocks stay out of正文 even when their font matches body
+
+### Task 5: Use style/geometry more explicitly for insert detection
+
+Files:
+
+- `paperforge/worker/ocr_profiles.py`
+- `paperforge/worker/ocr_document.py`
+
+#### 5.1 Add body-spine comparison helpers
+
+Useful block-level features:
+
+- block width
+- line height distribution
+- mean font size
+- font family
+- italic/bold ratios
+- x-column placement
+- gap above/below
+
+Use these to compare candidate blocks to the learned body spine.
+
+#### 5.2 Detection logic should be relative, not absolute
+
+No:
+
+- fixed page number
+- fixed width thresholds without page-relative normalization
+
+Yes:
+
+- narrower/wider than body family on the same paper
+- detached from body chain on the same page/spread
+- repeated cluster pattern on the same paper
+
+Acceptance:
+
+- insert detection generalizes across journals better than phrase rules
+
+### Task 6: Downstream behavior
+
+Files:
+
+- `ocr_metadata.py`
+- `ocr_render.py`
+- maybe `ocr_index.py`
+
+#### 6.1 Metadata
+
+- `non_body_insert` must not be used as author candidate
+- only anchored `authors` blocks may feed OCR frontmatter candidates
+
+#### 6.2 Render
+
+Default:
+
+- `non_body_insert` should not render in `fulltext.md`
+
+Optional:
+
+- if needed later, preserve them in separate object/appendix notes
+- but do not include in正文
+
+#### 6.3 Index
+
+- keep them out of main body index buckets
+- optionally add a separate bucket later if useful
+
+## Testing Plan
+
+### Unit / integration tests
+
+1. `tests/test_ocr_roles.py`
+- body paragraph with superscript citations must not become `authors`
+- page-1 author zone still yields `authors`
+
+2. `tests/test_ocr_metadata.py`
+- source/Zotero authors override bad OCR author candidates
+- OCR alignment works only for page-1 candidates
+
+3. `tests/test_ocr_document.py`
+- early-page insert cluster is detected as non-body
+- non-body insert is not rescued to `body_paragraph`
+
+4. `tests/test_ocr_render_stabilization.py`
+- `non_body_insert` does not appear in rendered正文
+
+### Real-paper validation
+
+Primary:
+
+- `SAN9AYVR`
+
+Must verify:
+
+- `resolved_metadata.authors` is correct
+- author bio blocks on pages 1-2 are absent from正文
+- page 3 `In Section 5 ...` is body, not authors
+
+Secondary regression:
+
+- `2GN9LMCW`
+- `7C8829BD`
+
+Must verify:
+
+- no regression in frontmatter authors
+- no regression in tail/backmatter structure
+
+## Acceptance Criteria
+
+This plan is complete when:
+
+1. arbitrary正文 text can no longer be promoted to `authors`
+2. metadata authors are anchored to source/Zotero truth
+3. early-page author biographies/profile cards are classified as non-body insert blocks
+4. rescue no longer drags these insert blocks back into正文
+5. `SAN9AYVR/fulltext.md` no longer contains author biography prose
+
+## Risks
+
+1. Overfitting insert detection to this one paper
+- avoid “is currently / received degree” as primary rules
+
+2. Hiding legitimate boxed正文 content
+- keep detection cluster-based and relative to body spine
+- validate on at least one paper with real正文 side notes later
+
+3. Confusing affiliation/footnote blocks with author bios
+- author bios are paragraph-like insert clusters
+- affiliation blocks remain page-1 frontmatter candidates
+
diff --git a/docs/superpowers/plans/2026-06-06-ocr-structure-truth-convergence-plan.md b/docs/superpowers/plans/2026-06-06-ocr-structure-truth-convergence-plan.md
new file mode 100644
index 00000000..147e60f3
--- /dev/null
+++ b/docs/superpowers/plans/2026-06-06-ocr-structure-truth-convergence-plan.md
@@ -0,0 +1,480 @@
+# OCR Structure Truth Convergence Plan
+
+Date: 2026-06-06  
+Scope: `paperforge/worker/ocr_pdf_spans.py`, `ocr_blocks.py`, `ocr_document.py`, `ocr_roles.py`, `ocr_profiles.py`, `ocr_render.py`, `ocr_rebuild.py`, downstream tests
+
+## Goal
+
+收掉当前 structured OCR pipeline 里最危险的三条分叉，让全文结构、span 样式、最终 `fulltext.md` 使用同一套真相：
+
+1. `span_metadata` 必须基于正确坐标系提取
+2. `DocumentStructure` 必须成为唯一的全文结构真相
+3. `role/profile` 的构建顺序必须和结构归一化顺序一致
+
+这份计划不重做 pipeline。它从当前代码状态出发，收敛现有实现。
+
+## Current Root Causes
+
+### 1. Span extraction likely uses the wrong coordinate system
+
+Current behavior:
+
+- `ocr_pdf_spans.extract_pdf_spans_for_block()` directly uses OCR block bbox as `page.get_text("rawdict", clip=rect)` PDF coordinates.
+- OCR block bbox is derived from rendered page image coordinates, not guaranteed PDF coordinates.
+- We already hit the same coordinate mismatch in figure/table cropping earlier.
+
+Impact:
+
+- `span_metadata` may be extracted from the wrong PDF region
+- style-based role inference becomes noisy or silently empty
+- all later span-profile logic rests on an unstable base
+
+### 2. Document structure is analyzed twice, in two different layers
+
+Current behavior:
+
+- `ocr_blocks.build_structured_blocks()`:
+  - seeds roles
+  - calls `analyze_document_structure()`
+  - calls `rescue_roles_with_document_context()`
+- `ocr_render._order_tail_blocks()`:
+  - calls `analyze_document_structure()` again
+  - does additional `_promote_tail_body_candidates()`
+  - does additional `_assign_tail_spread_ownership()`
+
+Impact:
+
+- structure layer and render layer can diverge
+- `fulltext.md` may not reflect the same truth used by metadata/index/health/inventory
+- debugging remains difficult because render still mutates structural semantics
+
+### 3. Profile construction happens before structural normalization settles
+
+Current behavior:
+
+- `ocr_blocks.build_structured_blocks()`:
+  - first-pass roles
+  - builds `role_profiles`
+  - only then runs `analyze_document_structure()`
+- `analyze_document_structure()` currently normalizes some roles in-place:
+  - boundary/container backmatter normalization
+  - body/frontmatter noise to backmatter body in boundary zone
+
+Impact:
+
+- `role_span_profiles.json` represents pre-normalized roles
+- rescue logic compares blocks against profiles that are partially stale
+- backmatter families are especially under-modeled
+
+## Non-Goals
+
+- Do not redesign figure/table pipeline here
+- Do not add new journal-specific text patches
+- Do not introduce absolute page gates or absolute font-size thresholds as primary logic
+- Do not rebuild the pipeline from scratch
+
+## Target End State
+
+The final pipeline should be:
+
+1. raw blocks
+2. span extraction on correct coordinates
+3. seed roles only
+4. document structure analysis and structural normalization
+5. profile building on normalized structure
+6. section-aware rescue
+7. downstream consumers render/index/health/inventory consume the same structure without re-inferring it
+
+In other words:
+
+- `ocr_roles.py` owns seed-role assignment only
+- `ocr_document.py` owns document segmentation and structural normalization
+- `ocr_profiles.py` owns profile construction and profile comparison
+- `ocr_render.py` renders; it does not invent structure
+
+## Implementation Plan
+
+### Task 1: Fix span extraction coordinate baseline
+
+Files:
+
+- `paperforge/worker/ocr_pdf_spans.py`
+- `paperforge/worker/ocr.py`
+- `paperforge/worker/ocr_rebuild.py`
+- tests in `tests/test_ocr_pdf_spans.py`
+
+#### 1.1 Add an explicit OCR-to-PDF coordinate mapping path
+
+Update `extract_pdf_spans_for_block()` and/or introduce a helper such as:
+
+- `_map_ocr_bbox_to_pdf_rect(page, bbox, page_width, page_height)`
+
+Use:
+
+- OCR page dimensions from `raw_blocks.page_width/page_height`
+- PDF page rect dimensions from `fitz.Page.rect`
+
+Expected mapping:
+
+- if OCR bbox is in rendered image coordinates, convert it back into PDF coordinates using scale ratios
+- avoid direct `fitz.Rect(*bbox)` on OCR image coordinates
+
+#### 1.2 Carry the required dimensions through the API
+
+Update call sites so span extraction receives:
+
+- `page_width`
+- `page_height`
+
+Likely signature change:
+
+- `extract_pdf_spans_for_block(pdf_doc, page_num, bbox, page_width=None, page_height=None)`
+
+#### 1.3 Preserve graceful degradation
+
+If mapping inputs are absent or invalid:
+
+- return `None`
+- do not crash rebuild/ocr
+
+#### 1.4 Add tests for coordinate mapping
+
+Add tests that verify:
+
+- OCR-space bbox is converted before clip extraction
+- no-span fallback remains stable when PDF is missing or bbox invalid
+
+Acceptance:
+
+- `span_metadata` extraction no longer relies on raw OCR bbox as PDF clip coords
+- tests cover mapped extraction path
+
+### Task 2: Make `ocr_roles.py` a pure seed-role pass
+
+Files:
+
+- `paperforge/worker/ocr_roles.py`
+- tests in `tests/test_ocr_roles.py`
+
+#### 2.1 Remove residual structure-owning logic from role assignment
+
+`assign_block_role()` should keep:
+
+- raw label priors
+- page-1 frontmatter zone heuristics
+- figure/table formal prefix detection
+- obvious abstract/reference/backmatter heading seeds
+- conservative body fallback
+
+It should stop owning:
+
+- late tail spread ownership
+- cross-page tail continuation logic
+- render-like backmatter reordering assumptions
+
+#### 2.2 Demote high-risk hard rules
+
+Specifically remove or weaken as primary rules:
+
+- `_is_backmatter_boundary_heading()` relative page gate:
+  - `if total_pages > 0 and (page_num / total_pages) < 0.5: return False`
+- visual heading shortcuts like:
+  - `font_size >= 14`
+  - `bold && font_size >= 11`
+
+Replace with:
+
+- heading-like style as one weak signal
+- let `ocr_document.py` and later profile matching validate boundary families
+
+#### 2.3 Keep boundary seeds conservative
+
+`backmatter_boundary_heading` should remain seedable, but only when multiple signals agree:
+
+- heading-like shape
+- candidate container wording or structure
+- not a known backmatter child heading
+- not references
+
+No absolute page number or hard half-document gate.
+
+Acceptance:
+
+- `assign_block_role()` becomes easier to reason about as a seed pass
+- tests no longer assert page-ratio gating behavior
+
+### Task 3: Move all document-level structural mutation into `ocr_document.py`
+
+Files:
+
+- `paperforge/worker/ocr_document.py`
+- `paperforge/worker/ocr_render.py`
+- tests in `tests/test_ocr_document.py`, `tests/test_ocr_render_stabilization.py`
+
+#### 3.1 Define `ocr_document.py` as the only owner of structural normalization
+
+`ocr_document.py` should own:
+
+- body/backmatter/references boundary detection
+- tail spread reconciliation
+- backmatter form classification
+- backmatter role normalization after boundary
+- tail body candidate promotion
+- tail spread ownership assignment
+
+Currently some of this logic still lives in `ocr_render.py`:
+
+- `_promote_tail_body_candidates()`
+- `_assign_tail_spread_ownership()`
+
+Move these into `ocr_document.py` or make `ocr_render.py` call normalized results only.
+
+#### 3.2 Make `analyze_document_structure()` return normalized blocks explicitly
+
+Current issue:
+
+- it mutates blocks in-place indirectly and returns only `DocumentStructure`
+
+Refactor to something like:
+
+- `analyze_document_structure(blocks) -> tuple[DocumentStructure, list[dict]]`
+
+or:
+
+- `normalize_document_structure(blocks) -> NormalizedDocument`
+
+where the returned artifact includes:
+
+- `document_structure`
+- `normalized_blocks`
+
+This avoids hidden in-place mutation and makes downstream ordering explicit.
+
+#### 3.3 Stop `ocr_render.py` from re-owning structure
+
+`render_fulltext_markdown()` should consume already-normalized blocks and a precomputed `DocumentStructure`.
+
+It should not:
+
+- call `analyze_document_structure()` again
+- promote tail bodies
+- assign tail ownership
+
+It may still do presentation-only ordering inside already-normalized page groups, but not semantic role mutation.
+
+Acceptance:
+
+- only one module owns structure mutation
+- render consumes structure, it does not invent it
+
+### Task 4: Reorder profile construction after structural normalization
+
+Files:
+
+- `paperforge/worker/ocr_blocks.py`
+- `paperforge/worker/ocr_profiles.py`
+- `paperforge/worker/ocr_document.py`
+- tests in `tests/test_ocr_document.py`, `tests/test_ocr_profiles.py` if present
+
+#### 4.1 Change `build_structured_blocks()` order
+
+Current order:
+
+1. seed roles
+2. build role profiles
+3. analyze document structure
+4. rescue
+
+Target order:
+
+1. seed roles
+2. analyze document structure + normalize roles
+3. build profiles on normalized roles
+4. rescue with normalized roles and normalized profiles
+
+This is the most important sequencing fix.
+
+#### 4.2 Expand profile semantics from role-only toward family-aware
+
+Do not throw away `build_role_span_profiles()`.
+Extend it so later rescue can reason about:
+
+- body family
+- heading family
+- backmatter heading family
+- reference family
+- caption family
+
+Minimal first step:
+
+- keep role profiles for compatibility
+- add helper(s) to derive family views from normalized roles
+
+Do not overdesign this phase; the main goal here is correct sequencing, not a full new taxonomy.
+
+Acceptance:
+
+- `role_span_profiles.json` is built from normalized structure, not pre-normalized seeds
+- backmatter/reference/body families are less stale
+
+### Task 5: Upgrade rescue to use normalized structure as context
+
+Files:
+
+- `paperforge/worker/ocr_document.py`
+- `paperforge/worker/ocr_blocks.py`
+- tests in `tests/test_ocr_document.py`
+
+#### 5.1 Keep rescue narrow but make it structurally consistent
+
+Current rescue is acceptable in scope, but must operate on:
+
+- normalized blocks
+- normalized profiles
+- final section boundaries
+
+It should continue to own:
+
+- frontmatter noise rescued back to body
+- body promoted to reference in references zone
+- weak heading demoted to body
+
+#### 5.2 Prevent rescue from fighting stale roles
+
+Once Task 4 is done, rescue no longer compares against pre-boundary profiles.
+
+That should especially improve:
+
+- backmatter child stability
+- reference consistency
+- weak heading/body disambiguation
+
+Acceptance:
+
+- rescue uses normalized context only
+- no rescue step depends on pre-normalization role families
+
+### Task 6: Make downstream products consume the same structure truth
+
+Files:
+
+- `paperforge/worker/ocr_rebuild.py`
+- `paperforge/worker/ocr.py`
+- `paperforge/worker/ocr_render.py`
+- possibly `ocr_health.py`, `ocr_index.py`
+
+#### 6.1 Thread document-structure output through downstream calls
+
+Where possible, pass:
+
+- normalized structured blocks
+- document structure
+- normalized profiles
+
+into downstream stages instead of forcing each layer to reconstruct assumptions.
+
+Minimal requirement:
+
+- render must not recompute structure
+
+Optional but desirable:
+
+- health/index may later write structure summaries for debugging
+
+#### 6.2 Consider persisting document structure artifact
+
+Add a derived artifact such as:
+
+- `structure/document_structure.json`
+
+Contents may include:
+
+- `body_end_page`
+- `backmatter_start`
+- `references_start`
+- `spread_start`
+- `spread_end`
+- `backmatter_form`
+
+This is useful for:
+
+- debugging
+- reproducibility
+- making render/index/health easier to compare
+
+Acceptance:
+
+- same structural truth is visible and reusable downstream
+
+## Testing Plan
+
+### Unit / integration targets
+
+1. `tests/test_ocr_pdf_spans.py`
+- mapped OCR-to-PDF coordinate extraction
+- graceful fallback
+
+2. `tests/test_ocr_roles.py`
+- remove expectations that depend on hard page gates or hard font thresholds
+- keep seed-role expectations conservative
+
+3. `tests/test_ocr_document.py`
+- boundary detection still works
+- backmatter normalization still works
+- rescue uses normalized roles
+
+4. `tests/test_ocr_render_stabilization.py`
+- render no longer depends on its own structural mutation
+- tail ordering remains correct when given normalized blocks
+
+5. if needed, add a new test around `build_structured_blocks()` ordering
+- asserts profile build happens after document normalization
+
+### Real-paper regression set
+
+Must re-run on:
+
+- `7C8829BD`
+  - Frontiers-style mixed tail spread
+- `2GN9LMCW`
+  - PeerJ-style declarations container + references + unnumbered headings
+
+Verification focus:
+
+- structure layer and rendered markdown agree on backmatter order
+- references remain stable
+- no reintroduced page marker mismatch
+- span metadata is present where expected and not obviously empty/misaligned
+
+## Acceptance Criteria
+
+This plan is complete when all of the following are true:
+
+1. `span_metadata` extraction no longer assumes OCR bbox == PDF coordinates
+2. `ocr_render.py` does not recompute or mutate document structure semantics
+3. `build_structured_blocks()` builds profiles after structural normalization
+4. rescue operates on normalized roles and normalized profiles
+5. `7C8829BD` and `2GN9LMCW` remain correct after rebuild
+6. tests cover the new sequencing and coordinate-baseline assumptions
+
+## Risks To Watch
+
+1. Over-coupling render to a new structure artifact too early
+- keep the first change minimal: pass normalized blocks first, persist structure later if useful
+
+2. Breaking old tests that implicitly relied on hard threshold behavior
+- update tests toward seed-role behavior, not old heuristic internals
+
+3. Span extraction still returning sparse data on scan-heavy PDFs
+- that is acceptable if graceful degradation remains intact
+- the important fix is to stop reading obviously wrong regions
+
+## Recommended Execution Order
+
+1. Task 1 — span coordinate fix
+2. Task 3 — make `ocr_document.py` the structure owner
+3. Task 4 — reorder profile construction
+4. Task 5 — rescue on normalized context
+5. Task 6 — downstream truth convergence
+6. finalize tests and real-paper rebuild verification
+
diff --git a/paperforge/worker/ocr_roles.py b/paperforge/worker/ocr_roles.py
index d93ebf90..47d50ca3 100644
--- a/paperforge/worker/ocr_roles.py
+++ b/paperforge/worker/ocr_roles.py
@@ -591,26 +591,6 @@ def assign_block_role(
                 evidence=[f"citation line pattern: {text[:60]}"],
             )
 
-        # Author list with superscript affiliation markers → authors, not noise
-        # Distinguish from citation lines (which have year in parens) and
-        # affiliation blocks (which have institutional keywords)
-        has_year_parens = bool(re.search(r"\(\d{4}[a-z]?\)", text))
-        has_inst_keyword = any(
-            kw in lower_txt for kw in ["department", "university", "institute", "college", "school of"]
-        )
-        if (
-            _AUTHOR_AFFILIATION_MARKER.search(text)
-            and "," in text
-            and not has_year_parens
-            and not has_inst_keyword
-            and len(text) < 500
-        ):
-            return RoleAssignment(
-                role="authors",
-                confidence=0.8,
-                evidence=[f"author list with affiliation markers: {text[:60]}"],
-            )
-
         # Affiliation block starting with superscript (page 1 only)
         if (
             still_frontmatter
diff --git a/tests/test_ocr_roles.py b/tests/test_ocr_roles.py
index f79d50d5..3627a61f 100644
--- a/tests/test_ocr_roles.py
+++ b/tests/test_ocr_roles.py
@@ -343,6 +343,31 @@ def test_heading_level_from_profile_match() -> None:
     pass  # Placeholder — will be updated after refactor
 
 
+def test_body_citation_not_authors() -> None:
+    from paperforge.worker.ocr_roles import assign_block_role
+    block = {
+        "block_label": "text",
+        "block_content": "In Section 5, the focus is on ES based bioelectronics $^{8,49}$",
+        "block_bbox": [100, 500, 800, 540],
+        "page": 3,
+    }
+    result = assign_block_role(block, page_blocks=[block], page_width=1200, page_height=1600)
+    assert result.role == "body_paragraph", f"Expected body_paragraph, got {result.role}"
+    assert result.confidence >= 0.5
+
+
+def test_frontmatter_author_zone_still_works() -> None:
+    from paperforge.worker.ocr_roles import assign_block_role
+    block = {
+        "block_label": "text",
+        "block_content": "Alice Smith, Bob Jones, Charlie Brown",
+        "block_bbox": [100, 300, 800, 330],
+        "page": 1,
+    }
+    result = assign_block_role(block, page_blocks=[block], page_width=1200, page_height=1600)
+    assert result.role == "authors", f"Expected authors, got {result.role}"
+
+
 def test_backmatter_boundary_detects_on_early_page() -> None:
     """Backmatter boundary should be detectable on papers with fewer
     than 8 pages, without a hard page gate."""