auto-paper

Fully automatic paper writing + closed-loop self-improvement

Closed-loop architecture

Instructions (AGENTS.md) → Skill (auto-paper) → Writing (drafts/) → Hooks (audit) → update Skill/Instructions in return

Core concept (CONSTITUTION §22): auditable, decomposable, recomposable. See AGENTS.md for details.

Four audit dimensions:

Dimension	Check	Update target
Paper	Citations, word count, Anti-AI	`patch_draft` Revise paper
Methodology	Methodological reproducibility	Supplement methodological details
Compliance	Whether the Agent follows the process	`.memory/` Record deviations
Meta	Skill/Hook/Instruction	Update SKILL/Hook/AGENTS

Execution audit trail (Audit Trail)

Each execution produces in projects/{slug}/.audit/ :

File	Content
`pipeline-run-{ts}.md`	Phase summary + Hook statistics + Decision Log
`search-strategy.md`	Search strategy and results
`reference-selection.md`	Literature screening decisions
`concept-validation.md`	Novelty validation process
`hook-effectiveness.md`	Hook trigger rate/pass rate/false positive rate
`quality-scorecard.md`	Methodology + writing quality (0-10 score)
`checkpoint.json`	Breakpoint recovery:`last_completed_phase`, `phase_outputs`

Recovery logic: detect checkpoint.json → ask whether to continue from Phase N+1 or restart.

Dual Hook system

Copilot Hooks (A-D, real-time correction during writing) are defined in this file. Pre-Commit Hooks (P1-P8 + G1-G7) are defined in git-precommit/SKILL.md. The two are complementary: Copilot handles details, Pre-Commit handles the overall picture.

🚧 Hard Gate Enforcement (Code-Level, cannot be skipped)

SKILL.md is a soft constraint (the Agent may ignore it). The following MCP Tools are code-enforced hard limits。

Required MCP Tool call

Timing	MCP Tool	Description
After each Phase is completed	`pipeline_action(action="validate_phase", phase=...)`	Returns PASS/FAIL + missing artifacts; on FAIL, entering the next Phase is prohibited
At the start of each round of Phase 7	`pipeline_action(action="start_review")`	Starts the AutonomousAuditLoop state machine and returns the round context
At the end of each round of Phase 7	`pipeline_action(action="submit_review", scores=...)`	Submits the score and returns a verdict (CONTINUE/QUALITY_MET/MAX_ROUNDS/REWRITE_NEEDED)
Phase 7 needs a major rewrite	`pipeline_action(action="rewrite_section", sections=..., reason=...)`	Roll back to Phase 5 to rewrite the specified section, with at most 2 rollbacks
After each section of Phase 5 is completed	`pipeline_action(action="approve_section", section=..., decision=...)`	Autopilot: the Agent automatically approves after self-review. Manual: the user reviews → approve or revise. The Phase 5 gate requires all to be approved
When the user requests a pause	`pipeline_action(action="pause", reason=...)`	Pause the pipeline and record the draft hash to detect user edits
Resume after pause	`pipeline_action(action="resume")`	Resume the pipeline, detect user modifications, and suggest re-validation
At any point in the middle of the Pipeline	`pipeline_action(action="heartbeat")`	Returns the status of all Phases + remaining work items; the Agent cannot claim to be "done"
When the Agent needs a full check	`pipeline_action(action="doctor")`	Return 11-phase readiness, external MCP declaration/command availability, latest gate cache
After each Hook evaluation in Phase 5	`record_hook_event(hook_id, event_type)`	Record trigger/pass/fix/false_positive for A/B/C/E Hooks; the Phase 6 gate will verify that actual records exist
Before Phase 6 (audit stage)	`run_quality_checks(action="quality_audit", scores=...)`	Set a ≥4-dimension quality score + generate a scorecard/hook-effectiveness report; the Phase 6 gate verifies the score data
Before Phase 10 (self-improvement)	`run_quality_checks(action="meta_learning")`	Execute D1-D6 analysis + write meta-learning-audit.yaml / meta-learning-feedback.yaml; the Phase 10 gate verifies the analysis data

Enforce rules

Phase transition：Agent MUST call pipeline_action(action="validate_phase", phase=N) and receive PASS before Phase N+1 can begin
Phase 7 Review Loop：Agent MUST call pipeline_action(action="start_review") start,pipeline_action(action="submit_review", scores=...) end. Rounds cannot be skipped
Pipeline complete：Agent MUST call pipeline_action(action="heartbeat") confirm completion = 100% before claiming completion
All gate results are automatically recordedto .audit/gate-validations.jsonl, available for Phase 10 D-Hook analysis
Project structure validation：validate_project_structure() can be called independently of the pipeline to check the structural integrity of new/existing projects
Automatic prerequisite check for Phase > 1：pipeline_action(action="validate_phase", phase=...) automatically checks key artifacts of preceding Phases when Phase > 1 (WARNING level)
Audit data enforcement: The Phase 6 gate not only checks that report files exist, but also verifies quality-scorecard.yaml (≥4 dimensions, avg > 0) and hook-effectiveness.yaml (≥1 hook has event records)
Self-improvement data enforcement: The Phase 10 gate verifies that meta-learning-audit.yaml has complete analysis records (adjustments_count, lessons_count, suggestions_count), and produces meta-learning-feedback.yaml for the next round's Agent to directly continue from

🛡️ Anti-Compaction Recovery Protocol

Context compaction causes the Agent to lose pipeline progress and writing state. The following defenses automatically protect against this:

Layer 1: Automatic pipeline state writing

All gate tools (pipeline_action(action="validate_phase"), pipeline_action(action="heartbeat"), pipeline_action(action="start_review"), pipeline_action(action="submit_review")) automatically write to projects/{slug}/.mdpaper-state.json the pipeline_state field
No additional action by the Agent is required

Layer 2: Automatic writing progress writing (✍️ Writing Session Checkpoint)

write_draft() and patch_draft() automatically write to the writing_session field after success
Records: current section, filename, operation type, word count, list of all existing sections
No additional action by the Agent is required

Layer 3: Agent-initiated detailed checkpoint

Use workspace_state_action(action="checkpoint", section=..., plan=..., notes=..., references_in_use=...) to save reasoning state
Recommended timing: before starting each paragraph, before switching sections, when citing a large number of references
Save: writing outline, style decisions, reference keys currently in use, the Agent's thinking notes
This tool call is lightweight and does not affect the writing workflow

Layer 4: Automatic recovery

workspace_state_action(action="get") reads per-project state → get_recovery_summary() produces a banner
Pipeline banner：current phase, round, gate result, next action, failures
Writing session banner（✍️）：current section, file, word count, agent context, sections on disk
After recovery, use read_draft to read back the draft content and continue

Layer 5: Agent mandatory rules

Start of conversation(or first response after context compaction) → MUST call workspace_state_action(action="get")
If pipeline_state.is_active == true → continue directly from the Phase/Round indicated by the recovery summary
If writing_session.active == true → use read_draft to read back the draft and continue writing
Do not rerun from scratch any Phase that has already PASSED

11-Phase Pipeline (Phase 0-11, including Phase 6.5)

Phase 0: PRE-PLANNING（Source Materials + Journal Profile + Pipeline Config）🆕

Purpose: Before any writing begins, first register the raw materials provided by the user, then establish structured journal constraints + Pipeline behavior settings. Output：projects/{slug}/journal-profile.yaml + projects/{slug}/.audit/source-materials.yaml

── Step 0: 掃描用戶原始素材（不可略過）──
  1. 呼叫 `project_action(action="source_materials")`
     - 預設掃描 workspace root，排除 projects/src/tests/node_modules 等 generated/code trees
     - 偵測 `.docx`, `.xlsx`, `.xls`, `.csv`, `.tsv`, `.pdf`, `.rtf`, `.pptx`, `.txt`
     - 產出 `.audit/source-materials.yaml` 和 `.audit/source-materials.md`
  2. 若 manifest 有 `pending_asset_aware > 0`：
     - 對 `agent_next_steps.asset_aware_file_paths` 呼叫 asset-aware `ingest_documents`
     - 表格型 DOCX/XLSX 必須先轉成 context，不能只根據摘要腦補
     - 後續 Phase 4/5 的 asset plan / results drafting 必須以這些 primary_user_material 為優先證據
  3. 即使沒有找到素材，也必須保留 0-candidate manifest，證明 agent 已掃描過 workspace。

── Step 1: 資訊來源 ──
Agent 按優先順序取得期刊要求：
  a) 用戶提供 submission guide / for-authors PDF/URL
     → 解析文件 → 自動填入 YAML（需用戶確認）
  b) 用戶口頭說明（例：「BJA, original research, 3000字」）
     → Agent 先查 templates/journal-profiles/ 內建期刊庫（麻醉學 Top 20）
     → 有匹配 → 讀取 YAML → 複製到專案 → 用戶確認
     → 無匹配 → 從 template 建立 + 補全欄位
  c) 無明確期刊 → 使用 paper_type 預設值
     → 產出 YAML 後提醒用戶日後補充

  💡 提醒用戶：可直接請 Copilot 讀取內建期刊設定（templates/journal-profiles/），
     例如「幫我套用 BJA 的設定」即可自動產生 journal-profile.yaml。
     用 read_file 閱讀 _index.yaml 可查看所有 20 個可用期刊。

── Step 2: 產生 journal-profile.yaml ──
  1. 從 templates/journal-profile.template.yaml 複製模板
  2. 填入期刊資訊、字數限制、圖表上限
  3. 設定 pipeline 行為（loop 上限、review 輪數）
  4. 設定 reporting guidelines checklist
  5. 🗣️ 呈現 YAML 摘要給用戶確認
     → 特別強調：字數上限、圖表上限、必要文件、loop 設定

── Step 3: 解析 submission guide（如有）──
  IF 用戶提供 submission guide 文件：
    a) 讀取文件內容（PDF/URL）
    b) 提取結構化資訊：
       - 字數限制（total + per section）
       - 圖表限制（數量 + 格式 + DPI）
       - 引用格式 + 上限
       - 必要文件清單
       - 報告指引要求
    c) 填入 YAML 對應欄位
    d) 標註 ⚠️ 不確定的欄位 → 請用戶確認

journal-profile.yaml constraints on the entire Pipeline

YAML field	Affected Phase / Hook
`paper.type`	Phase 1 project setup, Phase 4 writing order
`paper.sections`	Phase 4 outline, Phase 5 writing order
`word_limits.*`	Hook A1 word count, Hook C6 total word count
`assets.figures_max/tables_max`	Phase 4 Asset Plan, Phase 5 Asset generation
`references.max_references`	Phase 2 literature count, Phase 8 citation limit
`references.reference_limits`	Citation limit by paper type (overrides max_references)
`references.style`	Phase 8 citation format
`reporting_guidelines.checklist`	Hook B5 methodology, Hook C2 submission checklist
`required_documents.*`	Phase 9 export, Hook C2 submission checklist

source-materials.yaml constraints on the entire Pipeline

YAML field	Affected Phase / Hook
`summary.total_candidates`	Whether Phase 0 has actually scanned the user's materials
`materials[].evidence_priority`	Phase 3 concept、Phase 4 plan、Phase 5 results/methods
`materials[].ingestion.status`	Whether Phase 2.1/Phase 5 need to first parse DOCX/XLSX with asset-aware
`agent_next_steps.asset_aware_file_paths`	The specific file list for which the Agent must call asset-aware
`agent_next_steps.asset_aware_required`	Prevent reading only summaries and missing formal tables/data files

Data Anchor Provenance Rule：data-artifacts.yaml of data_anchors must not use concept.md, agent summary, or inferred/estimated values as sources. Every numeric/statistical anchor must cite ready/ingested source_material_id、asset_aware_doc_id、data_artifact_id, or a trusted data file; if the source material is still pending_asset_aware, Hook F4 must CRITICAL fail. | pipeline.hook_*_max_rounds | Hook A/B/C cascading limit | | pipeline.review_max_rounds | Phase 7 Autonomous Review rounds | | pipeline.writing.anti_ai_* | Hook A3 Anti-AI strictness | | pipeline.writing.citation_density | Hook A2 citation density standard | | pipeline.assets.* | Phase 5 Asset Sub-Pipeline behavior | | pipeline.autopilot | Fully automatic mode (default true) |

Autopilot mode (enabled by default)

pipeline:
  autopilot: true # 預設全自動，用戶可設 false 啟用逐步審閱
  autopilot_pause_at: [] # 用戶可指定暫停點，例如 ["phase4", "phase5_section", "phase7_rewrite"]

autopilot setting	Behavior
`true`(default)	The Agent makes autonomous decisions, self-reviews, and self-corrects throughout, and only asks the user when a safety valve is triggered
`false`	Ask the user at every Phase confirmation point (traditional mode)
`autopilot_pause_at`	Specify particular pause points (fine-grained control), e.g. Phase 4 confirm outline, Phase 5 section by section

Safety valves (the user is asked even if autopilot=true):

concept score < 75 and still below standard after 2 automatic corrections
Phase 6 still has CRITICAL after 3 rounds of cascading fix
Phase 7 reaches max_rounds and quality is still below standard
regression_count > 2 (to prevent infinite loops)

Gate: journal-profile.yaml exists + the user has confirmed key fields (word count, figure/table limits)

Phase 1: PROJECT SETUP

Skill: project-management

project_action(action="current") → switch if a project exists, otherwise create one
project_action(action="create", name, paper_type) or project_action(action="switch", slug)
project_action(action="setup")
Load journal-profile.yaml → verify paper_type consistency
project_action(action="update", target_journal=journal.name)
workspace_state_action(action="sync", doing="phase 1 setup", next_action="phase 2 literature")

Gate: project exists + paper_type is consistent with journal-profile

Phase 2: LITERATURE SEARCH

Skill: literature-review, parallel-search External MCP: pubmed-search, zotero-keeper（optional）

generate_search_queries(topic, strategy="comprehensive")
Parallel unified_search() × 3-5 sets
merge_search_results()
get_citation_metrics(sort_by="relative_citation_ratio")
Select the first 15-20 papers → save_reference_mcp(pmid, agent_notes)
[Optional] Zotero: search_items(query) → retrieve PMIDs → save_reference_mcp()

Gate: Paper-type-aware minimum literature count (Code-Enforced)

Paper Type	Minimum literature count
original-research	20
review-article	30
systematic-review	40
meta-analysis	40
case-report	8
letter	5
fallback	15

Parsing priority order:journal-profile.yaml references.minimum_reference_limits → DEFAULT_MINIMUM_REFERENCES → DEFAULT_MIN_REFERENCES (15)

Automatic decision: < minimum → expand_search_queries; >500 → add MeSH restriction; if the user has Zotero → proactively ask about importing.

Phase 2.1: FULLTEXT INGESTION & ASSET EXTRACTION

Skill: literature-review External MCP: asset-aware-mcp, pubmed-search Input: Literature saved in Phase 2 (must meet the paper-type-specific minimum count) Output: Full-text parsing result for each paper + analysis.json + references/fulltext-ingestion-status.md

Core principle: Before citing a paper, you must read the full text and complete a structured analysis. Abstract/metadata alone is insufficient to support citations in Methods/Results/Discussion.

Context Isolation: Each paper is read and analyzed by an independent subagent, avoiding 20+ full texts polluting the main agent context.

Process:

list_saved_references() → retrieve the list of all saved literature
For each paper: a. Look up the PDF source:get_fulltext(pmcid) (pubmed-search MCP) or analyze_fulltext_access(pmids) confirm Open Access status b. IF the full text is available:
- Retrieve full text: pubmed-search get_fulltext(pmcid, sections="all") or download the PDF
- Parse full text: asset-aware ingest_documents(file_paths=[pdf_path], use_marker=true)
- Wait for processing:get_job_status(job_id) until complete
- Build index:inspect_document_manifest(doc_id) retrieve the complete asset list
- Extract key passages:get_section_content(doc_id, section_path) → Methods, Results, Discussion
- Update metadata.json: add fulltext_ingested: true, asset_aware_doc_id, fulltext_sections c. IF the full text is unavailable:
- Record the reason: add to metadata.json fulltext_ingested: false, fulltext_unavailable_reason: "..."
- ⚠️ When writing, this paper's abstract may only be cited for conclusive information; specific methods or data must not be cited
Subagent reading and analysis (CRITICAL — one subagent per paper): a. Main Agent calls get_reference_for_analysis(pmid) → retrieve the structured full-text package (≤30K characters) b. Use runSubagent to launch an independent subagent, with a prompt containing:
- Full-text content (result of step a)
- Analysis instructions: summary, methodological assessment, key findings, limitations, which sections of the article they can be used in
- Require the subagent to call save_reference_analysis(pmid, summary, methodology, key_findings, limitations, usage_sections, relevance_score) c. After the Subagent completes,analysis.json has been written to the reference directory,metadata.json mark analysis_completed: true d. For subsequent writing, the Main Agent uses only analysis_summary, and does not re-read the full text → zero context pollution
Produce references/fulltext-ingestion-status.md：

PMID	Title	Full-text status	Asset-Aware	Analysis status	Usable sections	Notes
...	...	✅	doc_abc	✅	I/M/R/D
...	...	❌	—	⚠️	I only	Non-OA

Gate: fulltext-ingestion-status.md has been created + each paper's full-text status is marked CRITICAL Gate: All papers must have analysis_completed: true(enforced check by Pipeline Gate Phase 2.1) WARNING Gate: If >50% of papers lack full text → remind the user to consider supplementing with OA literature

Phase 2.5: CRITICAL APPRAISAL

Input: Literature completed in Phase 2 + 2.1 (including full-text parsing results) Output: references/appraisal-matrix.md

Purpose: Before entering the conceptual stage, assess the quality of core literature to avoid building research on low-quality evidence. requires: Phase 2.1 complete (quality assessment is more precise when full-text parsing results are available)

For each core reference (≥5 papers), assess:
- Study design quality (RCT > cohort > case series > case report > expert opinion)
- Risk of bias (selection, performance, detection, attrition, reporting)
- Adequacy of sample size
- Appropriateness of statistical methods
- Whether conclusions are supported by the data
- [NEW] Full-text verification status: Has the full text been read via asset-aware reading? References whose full text has not been read are marked ⚠️
Produce a quality matrix (appraisal-matrix.md):

PMID	1st Author	Design	Quality	Bias Risk	Key Limitation
...	...	RCT	High	Low	...

Mark low-quality references (Quality = Low); they should not be relied on excessively in the concept phase

Gate: appraisal-matrix.md has been created, with at least 5 papers assessed

Phase 3: CONCEPT DEVELOPMENT

Skill: concept-development External MCP: cgu（when novelty needs boost）

Analyze saved references → identify Gap
Write concept.md (including 🔒 NOVELTY + 🔒 SELLING POINTS)
write_draft(filename="concept.md", content=..., skip_validation=True)
validate_concept(project=...)
IF score < 75:
- Auto-revise once → re-verify
- Still < 75 → CGU: deep_think / spark_collision / generate_ideas → revise → re-verify
- Still < 75 → STOP, report back to the user (with CGU suggestions)

Gate: concept score ≥ 75 OR the user explicitly says "continue"

CGU tool mapping:

Problem	CGU tool	Output
Find weaknesses	`deep_think`	Reviewer attack angles
Find arguments	`spark_collision`	Collide novel angles
Broad ideation	`generate_ideas`	3-5 novelty candidates
Multiple perspectives	`multi_agent_brainstorm`	Multi-role debate

Phase 4: MANUSCRIPT PLANNING

Deliverables:manuscript-plan.yaml(exists under projects/{slug}/ )

Read concept.md + saved references + journal-profile.yaml
Produce manuscript-plan.yaml(paragraph-level Section Brief):
- writing_order: based on journal-profile.paper.sections, fallback to paper type default
- sections: each section includes word_budget + paragraphs[]
  - Each paragraph:id, function, topic, key_claims, must_cite, word_target
  - 🔒 Protected paragraph marker protected: true（Novelty Statement / Selling Points）
  - Methods paragraphs may include checklist_items(corresponds to B5 methodology)
- asset_plan: figures/tables/flowcharts (including tool, tool_args, caption, caption_requirements）
  - ⚠️ total verified count ≤ assets.figures_max + assets.tables_max
  - ⚠️ verified ≤ assets.total_display_items_max
- submission_checklist: list items to prepare based on required_documents.* List required items to prepare
- metadata: generated_at, based_on, changelog[]
Plan modification rules:
- Agent may add paragraphs / modify claims / adjust word count → must be written to metadata.changelog
- Agent ProhibitedDelete protected: true paragraphs
- User edits freely; changelog is tracked automatically
Autopilot mode（pipeline.autopilot: true, default):
- Agent verifies the plan's completeness on its own (number of figures/tables, word count budget, section coverage)
- Verification passes → save directly to projects/{slug}/manuscript-plan.yaml → continue to Phase 5
- Verification fails → Agent revises on its own → re-verifies (max 2 rounds) → only ask the user if it still fails
- Plan summary is recorded in .audit/pipeline-run-{ts}.md, which the user can review later
Manual mode（pipeline.autopilot: false or autopilot_pause_at includes "phase4"）：
- 🗣️ Present manuscript-plan.yaml summary to user for confirmation
- User adjusts → confirms → saves

Gate: manuscript-plan.yaml exists + number of figures/tables does not exceed limit

Writing order (according to journal-profile.paper.sections, fallback to paper type default):

Paper Type	Suggested order
original-research	Methods → Results → Introduction → Discussion → Abstract
systematic-review	Methods → Results → Discussion → Introduction → Abstract
case-report	Case Presentation → Discussion → Introduction → Abstract

Phase 4.5: SUPPLEMENTARY PLANNING

Input: manuscript-plan.yaml(Phase 4 output) Output: manuscript-plan.yaml updated (add supplementary section)

Purpose: Plan Supplementary Materials before writing, to avoid having to go back and supplement after the main manuscript is completed.

Determine whether supplementary materials are needed based on the journal profile
Plan supplementary content:
- Supplementary Tables (large data tables, subgroup analyses)
- Supplementary Figures (additional result figures)
- Supplementary Methods (detailed methodology, code)
- Sensitivity analysis results
- Full EQUATOR checklist (e.g., CONSORT flow diagram)
In manuscript-plan.yaml add supplementary section:

supplementary:
  - id: S1
    type: table
    title: "Baseline characteristics by subgroup"
    linked_section: Results
  - id: S2
    type: figure
    title: "Sensitivity analysis forest plot"
    linked_section: Results

Create supplementary/ directory structure

Gate: manuscript-plan.yaml contains a supplementary section (can be empty = no supplementary materials needed)

Phase 4.9: SUB-SECTION WRITING TEMPLATES (paragraph-level writing guidance)

Purpose: Supplement the paragraph-level prior knowledge of Phase 4 manuscript-plan.yaml . When the Agent writes each paragraph in Phase 5, itmust cross-checkthe following templates to ensure complete structure and coherent logic.

Introduction paragraph template (Funnel Structure)

Paragraph	Function	Content requirements	Citation requirements	Hook check
P1	Broad Context	Importance of the research field, epidemiological data, clinical significance	≥2 citations (epidemiology/guidelines)	B12（context markers）
P2	Evidence Review	Main findings, methods, conclusions of existing literature	≥3 citations (core literature)	A2（citation density）
P3	Knowledge Gap	Deficiencies, contradictions, unresolved issues in existing literature	≥1 citation (identify the source of the gap)	B12（gap markers）
P4	Study Objective	Purpose of this study, hypothesis (if applicable)	0-1 citation	B12（objective markers）

Anti-patterns(automatic FLAG):

P1 jumps directly to the study purpose (missing context/gap)
P2 only lists names without describing content (name-dropping)
P3 lacks a specific gap statement (e.g., "however, ... remains unclear")
P4 exceeds 2 sentences (overly lengthy purpose statement)

Methods paragraph template (Reproducibility Structure)

Paragraph	Function	Content requirements	Citation requirements	Hook check
M1	Study Design	Study type, period, institution, IRB approval	≥1 (design basis)	B14（ethical statements）
M2	Participants	Inclusion/exclusion criteria, sample size calculation	≥1 (criteria basis)	B9（past tense）
M3	Intervention/Exposure	Detailed description of intervention measures, control group definition	≥1（intervention protocol）	B5 (methodology checklist)
M4	Outcome Measures	Primary/secondary outcome measures, measurement methods, definitions	≥2（validated instruments）	C11 (Methods citations)
M5	Statistical Analysis	Statistical methods, software version, significance level	≥1 (statistical method citations)	B8（data-claim alignment）

Anti-patterns：

M1 missing IRB/ethics approval → B14 CRITICAL
M2 no sample size justification → WARNING
M5 statistical methods inconsistent with Results → B8 CRITICAL

Results paragraph template (Data-First Structure)

Paragraph	Function	Content requirements	Citation requirements	Hook check
R1	Study Flow / Demographics	Screening process, baseline characteristics (→ Table 1)	0 (pure data)	C13 (figure/table citations)
R2	Primary Outcome	Main results, statistical tests, effect size + CI	0 (pure data)	B11 (objectivity), B16 (effect size)
R3	Secondary Outcomes	Secondary results, subgroup analyses	0-1 (method citations)	B8 (statistical alignment)
R4	Subgroup / Sensitivity	Subgroup analysis, sensitivity analysis (if applicable)	0 (pure data)	C13 (figure/table citations)

Anti-patterns：

R1-R4 contain interpretive language ("suggesting", "indicating") → B11 WARNING
Results contain no figure/table citations → C13 WARNING
Effect size missing CI → B16 WARNING

Discussion paragraph template (Interpretation Structure)

Paragraph	Function	Content requirements	Citation requirements	Hook check
D1	Main Findings	Summarize main findings (do not repeat Results numbers)	0-1	B13（main finding markers）
D2-D3	Literature Comparison	Consistency/contradiction with existing literature	≥3 (comparative literature)	C11 (Discussion citation distribution)
D4	Mechanisms / Implications	Possible mechanistic explanations, clinical/practical implications	≥1 (mechanistic literature)	B13（implications markers）
D5	Limitations	Study limitations (design, bias, generalizability)	0-2	B13（CRITICAL if missing）
D6	Conclusion / Future	Summary + future research directions	0-1	B15 (appropriate amount of hedging)

Anti-patterns：

D1 repeats statistical numbers from Results → WARNING
D2-D3 no citations (not compared with literature) → C11 CRITICAL
D5 absent → B13 CRITICAL
Excessive hedging throughout the text (> 6/1000 words) → B15 CRITICAL

Abstract paragraph template (Structured Abstract)

Paragraph	Function	Word count target	Content requirements
A1	Background	40-60 words	Background + gap (1-2 sentences)
A2	Methods	60-80 words	Design, sample, main measurements
A3	Results	80-120 words	Main results + statistical numbers
A4	Conclusions	30-50 words	Clinical significance (1-2 sentences, avoid overgeneralization)

Note: Abstract word count is strictly limited by journal-profile.yaml.

Figure/table insertion specifications (integrated with C13 Hook)

Rule	Description	Hook
Sequential citation	Figure 1 must first appear before Figure 2	C13
Complete caption	≥10 words, including axes/variables/abbreviations	C13
Results must have	Results with data claims → must cite at least 1 figure/table	C13
Orphan detection	Present in Manifest but not cited in text → WARNING	C7d
Phantom detection	Cited in text but not in Manifest → CRITICAL	C7d
Quantity limit	Does not exceed the max set in journal-profile	C7a

Citation Decision Record specification (integrated with C12 Hook)

Each cited reference must be recorded in citation_decisions.json with a decision record:

{
  "smith2023_12345678": {
    "justification": "Largest RCT comparing X vs Y in same population, supports our primary outcome finding",
    "cited_sections": ["Introduction", "Discussion"],
    "relevance_score": 9,
    "decision_date": "2025-01-15",
    "decided_by": "agent"
  },
  "jones2022_87654321": {
    "justification": "Validated the measurement instrument used in our study (M4 outcome measures)",
    "cited_sections": ["Methods"],
    "relevance_score": 8,
    "decision_date": "2025-01-15",
    "decided_by": "agent"
  }
}

Workflow：

Phase 2 → save_reference_analysis() Create usage_sections
Phase 5 → every time insert_citation() , simultaneously update citation_decisions.json
Phase 6 → C12 Hook validation: every citation has justification + cited_sections matching the actual text

Phase 5: SECTION WRITING (core + Cascading Audit Loop)

Skill: draft-writing External MCP: drawio（diagrams）, cgu（Discussion）, data-tools(figures/tables) Input: manuscript-plan.yaml(Phase 4 output)

plan = load("manuscript-plan.yaml")

FOR section IN plan.writing_order:

  brief = plan.sections[section]

  ── Step 1: 準備 ──
  1a. 讀取 brief.paragraphs + 已完成 sections + get_available_citations()
  1b. 讀取 plan.asset_plan[section] → 確認需要哪些圖表

  ── Step 2: Asset 生成（先於寫作，見下方 Asset Sub-Pipeline）──
  2a. IF plan.asset_plan 有該 section 的 asset:
      → 執行 Asset Generation Sub-Pipeline
      → 產出 asset manifest entries（圖表路徑 + caption）

  ── Step 3: 撰寫（段落級 Brief 引導） ──
  3a. IF Methods + asset_plan 有 diagram:
      → drawio.create_diagram → XML
      → drawio.export_diagram → PNG/SVG
      → save_diagram(project, content, rendered_content, rendered_format)
      → insert_figure(rendered_filename or source filename, caption, draft_filename, after_section)
  3b. FOR paragraph IN brief.paragraphs:
      → 依據 paragraph.topic + key_claims + must_cite 撰寫
      → 尊重 paragraph.word_target
      → 若 paragraph.protected = true → 確保 🔒 內容完整
      → 若 paragraph.checklist_items → 逐條涵蓋
      → ✍️ 每段開始前呼叫 `workspace_state_action(action="checkpoint", section=..., plan=..., notes=..., references_in_use=...)`
        → 儲存寫作大綱與推理狀態，防止 compaction 遺失
  3c. draft_section() 或 write_draft()
      → 撰寫時整合 Step 2 產出的 asset references
      → ✅ 成功後自動 checkpoint 至 .mdpaper-state.json（無需手動）
  3d. IF Discussion + 論點弱:
      → cgu.deep_think → 強化邏輯鏈

  ── Step 4: Cascading Audit Loop（最多 3 rounds） ──
  4a. 🔔 HOOK A: post-write audit
      → Round 1: run A1-A4 → collect issues
        A1: 字數 vs brief.word_budget
        IF issues → patch_draft 修正 → re-run A1-A4
      → Round 2: IF still issues → patch_draft（不同策略）→ re-run
      → Round 3: IF still issues → LOG to .audit/ + FLAG for Phase 6

  4b. 🔔 HOOK B: post-section audit
      → run B1-B7 → collect issues
      → B7: Section Brief 合規（段落級比對 + caption_requirements，見 Hook B7 定義）
      → IF critical (B1 concept 不一致 / B2 🔒 缺失 / B5 <5):
        → targeted patch_draft → re-run Hook A on patched areas (1 round)
        → re-run Hook B → IF still critical → FLAG for Phase 6
      → IF advisory only → LOG + continue

  ── Step 5: Agent 自主審閱 + 即時修正（Autopilot 預設行為）──

  **核心原則**：Agent 在寫完 section 後立即自我評估，發現問題就地修正，
  不需等到 Phase 7 Review 才回頭改。這才是真正的 autopilot。

  5a. Agent 自行評估 section 品質：
      - Hook A/B 結果 → 有無未解 CRITICAL/WARNING
      - 與 concept.md 的一致性
      - 與已完成 sections 的邏輯連貫性
      - 段落是否完整涵蓋 manuscript-plan.yaml 的 key_claims
  5b. IF 發現問題 → Agent 立即修正：
      - 針對性 patch_draft() → re-run 相關 Hook
      - 修正直到 0 CRITICAL，或達 cascading 上限後 FLAG for Phase 6
      - 決策和修正記入 `.audit/pipeline-run-{ts}.md`
  5c. 自我評估通過 → Agent 呼叫 `pipeline_action(action="approve_section", section=section, decision="approve")`
      → 自動標記 approved → 繼續下一個 section
  5d. **用戶可隨時介入**：
      - 用戶說「我要看一下」→ Agent 暫停，呈現 section 摘要
      - 用戶說「每個 section 都讓我看」→ Agent 對後續 section 切換為手動審閱模式

  **手動審閱模式**（`pipeline.autopilot: false` 或 `autopilot_pause_at` 含 `"phase5_section"`）：
  5e. 呈現 section 內容摘要（字數、主要論點、Hook A/B 結果）
  5f. 🗣️ 詢問用戶：「approve / 請修改（附具體意見）」
  5g. Agent 呼叫 `pipeline_action(action="approve_section", section=section, decision=action, feedback=feedback)`
  5h. IF action == "revise":
      → 根據 feedback 執行 patch_draft / write_draft 修正
      → re-run Hook A/B on modified content
      → 回到 5e 再次呈現（最多 3 輪修改）
  5i. IF action == "approve":
      → 繼續到下一個 section

    ⏸️ 暫停點：用戶可隨時說「暫停」→ Agent 呼叫 `pipeline_action(action="pause")`
     → 用戶自由編輯草稿
      → 用戶說「繼續」→ Agent 呼叫 `pipeline_action(action="resume")`
     → 根據修改偵測結果重新跑相關 Hook

  ── Step 6: 記錄 ──
  6a. Log section audit results 到 .audit/pipeline-run-{ts}.md
  6b. Log 到 .memory/progress.md
  6c. 更新 checkpoint.json: { last_section: section, audit_status }
  6d. IF plan 需修改 → 寫入 plan.metadata.changelog + 存檔

Asset Generation Sub-Pipeline（Phase 5 Step 2）

Before writing each section, generate the required assets according to asset_plan:

FOR asset IN asset_plan[section]:
  SWITCH asset.type:

    CASE "table_one":
      1. list_data_files() → 確認資料存在
      2. detect_variable_types(file) → 自動分類
      3. generate_table_one(file, groups, variables) → 產出表格
      4. insert_table(filename, table_data, caption) → 插入草稿
      → ❌ 無資料檔 → 提示用戶提供，SKIP asset

    CASE "statistical_test":
      1. analyze_dataset(file) → 了解資料分佈
      2. run_statistical_test(file, test_type, ...) → 執行檢定
      3. 結果整合到 section 文字中
      → ❌ 不確定檢定方法 → 詢問用戶

    CASE "plot":
      1. create_plot(file, plot_type, ...) → 產生圖片
      2. insert_figure(filename, image_path, caption) → 插入草稿
      → ❌ 圖表不合理 → 提供替代建議

    CASE "flow_diagram":  (CONSORT, PRISMA, study flow)
      1. IF drawio MCP 可用:
         → drawio.create_diagram(type, data) → XML
        → drawio.export_diagram(format="png" or "svg") → rendered asset
        → save_diagram(project, content, name, rendered_content, rendered_format)
        → insert_figure(filename=name or rendered filename, caption, draft_filename, after_section)
      2. ELSE:
         → 產生 Mermaid 文字描述 → 存入 results/diagrams/
         → FLAG: 建議用戶手動轉換
      3. reference 在草稿中，確保導出的 Word/PDF 可見圖像而不只是 `.drawio` 原始檔

    CASE "custom_figure":
      1. Agent 依 asset_plan 說明 → create_plot 或手動描述
      2. insert_figure() 或 FLAG 需外部工具

Asset Type	Required MCP Tool	External MCP	Fallback
table_one	`generate_table_one`	—	Manually provide table
statistical_test	`run_statistical_test`	—	Describe expected analysis
plot	`create_plot`	—	Describe figure/table requirements
flow_diagram	`save_diagram` + `insert_figure`	`drawio` 🔸	Mermaid text
forest_plot	❌ Missing	`meta-analysis` 🔸	R/Python script
funnel_plot	❌ Missing	`meta-analysis` 🔸	R/Python script
PRISMA_diagram	`save_diagram` + `insert_figure`	`drawio` 🔸	Mermaid text

Agent-Initiated Asset Generation (autonomously add figures/tables during writing)

The asset_plan in Phase 4 cannot foresee all needs. During writing, the Agent may discover the need for literature comparison table、method comparison table , etc.

觸發條件（Phase 5 Step 3 寫作中）：
  - 引用 ≥3 篇文獻做比較 → 考慮 literature_summary_table
  - 描述多種方法/技術差異 → 考慮 comparison_table
  - 概念或架構複雜 → 考慮 concept_diagram

流程：
  1. 檢查 journal-profile.yaml → pipeline.assets.agent_initiated.enabled
  2. 檢查類型是否在 allowed_types 中
  3. 檢查目前圖/表數量是否已達 assets.figures_max / tables_max
  4. IF 可新增:
     a. 用 Markdown 表格撰寫（文獻比較表、方法對照表等）
     b. 或 create_plot / save_diagram（如需圖形）
     c. 附 caption + 標記來源為 "agent-initiated"
     d. insert_table() 或 insert_figure() → 插入草稿
     e. 更新 asset_plan（追加到 plan.metadata.changelog）
     f. 記錄到 .audit/: 為何新增、依據哪些文獻
  5. IF 已達上限:
     → 評估是否替換低優先級 asset
     → 或以文字描述替代，不生成實際圖表

常見 Agent-Initiated Assets：
  - 「Table X. Comparison of [topic] across studies」 → 文獻整理表
  - 「Table X. Characteristics of included studies」 → SR/Review 常見
  - 「Figure X. Conceptual framework of [approach]」 → 架構圖

Phase 6: CROSS-SECTION CASCADING AUDIT

Three-stage audit: first full-manuscript check → then retrospective correction → final validation.

── Stage 1: 全稿掃描 ──
1. 🔔 HOOK C: post-manuscript (C1-C8, C9-C14)
2. 收集所有 issues → 分類為 CRITICAL / WARNING / INFO
3. 收集 Phase 5 FLAG（未解決的 Hook A/B issues）

── Stage 2: 分層回溯修正（Cascading Fix，最多 3 rounds）──
Round 1:
  FOR each CRITICAL issue:
    a. 定位到具體 section + 段落
    b. patch_draft() 修正
    c. re-run 該 section 的 Hook A（確認 patch 沒破壞原有品質）
  → re-run Hook C (C1-C8, C14 where strong claims changed)

Round 2 (IF still CRITICAL):
  FOR each remaining CRITICAL:
    a. 分析 Round 1 修正為什麼失敗
    b. 嘗試不同修正策略（重寫段落 vs 微調）
    c. patch_draft()
  → re-run Hook C

Round 3 (IF still CRITICAL):
  → LOG all remaining issues to .audit/quality-scorecard.md
  → 呈現給用戶，附具體修改建議
  → 等待用戶決定（修改 / 接受 / 回到 Phase 5 重寫）

── Stage 3: 最終驗證 ──
4. re-run Hook C → 確認 0 CRITICAL
5. 生成 quality-scorecard.md（量化分數）
6. 更新 checkpoint.json

Gate: 0 critical issues (warnings recorded but acceptable)

Cascading retrospective rules

Hook C Issue	Retrospect to	Triggered Hook
C1 Manuscript inconsistency	Weaker section	Hook B4 → Hook A
C3 N value inconsistent across sections	All sections containing N	Hook A → patch
C4 Abbreviation undefined	Section where it first appears	Hook A4 → patch
C5 Wikilinks unresolvable	Corresponding section	Hook A2 → A4
C6 Total word count over limit	Longest section	Hook A1 → patch
C7a Figures/tables over limit	—	Merge or move to supplementary
C7b Citations over limit	—	Mark low-citation refs → user decides
C7c Exact word count comparison	Longest section	Hook A1 → patch
C7d Phantom citations	Corresponding section	Insert missing figure/table or remove citation
C7e Wikilink inconsistency	Corresponding section	Hook A4 → patch
C11 Uneven citation distribution	Sections lacking citations	Hook A2 → add citations
C12 Citations missing decision records	—	Generate citation_decisions.json
C13 Figure/table order/caption issues	Corresponding section	Hook C7d → patch

Phase 6.5: EVOLUTION GATE (mandatory entry into Review) 🆕

Purpose: establish a revision baseline, ensuring Phase 7 Review always executes(not skipped just because Hooks A-C all pass). Trigger: After Phase 6 completes MANDATORY(no skip conditions).

── Step 1: Snapshot Baseline ──
1. DraftSnapshotManager.snapshot_all(reason="pre-review-baseline")
   → 快照所有 section 的當前版本
2. 記錄 quality-scorecard Round 0 分數
3. 記錄到 .audit/evolution-log.jsonl:
   {"event": "baseline", "round": 0, "timestamp": "...",
    "scorecard": {6 維度分數}, "word_count": N,
    "instruction_version": git_short_hash()}

── Step 2: Force Review Entry ──
4. 無論 Hook C 結果如何 → 設定 review_required = true
5. 載入 journal-profile.yaml → reviewer_perspectives + quality_threshold
6. IF journal-profile.yaml 不存在:
   → 從 templates/ 生成預設值 → 存入專案
   → LOG: "Auto-generated journal-profile.yaml with defaults"

Gate: baseline snapshot completed → pipeline_action(action="validate_phase", phase=65) must PASS → enter Phase 7

Phase 7: AUTONOMOUS REVIEW (Structured Review Loop — MANDATORY) 🆕

Purpose: Simulate peer review, produce a structured Review Report + Author Response, and ensure every issue is responded to. Trigger：ALWAYS(Phase 6.5 mandatory entry, cannot be skipped). Even if Hook A-C all pass and quality has met the standard, at least 2 rounds must still be executed (code-enforced min_rounds=2）。 upper limit：pipeline.review_max_rounds(default 3). minimum rounds：min_rounds=2（Code-Enforced：AutonomousAuditLoop._determine_verdict() When round < min_rounds, forcibly return CONTINUE; early termination is not allowed). Hard Gate: Each round MUST call pipeline_action(action="start_review", min_rounds=2) to start + pipeline_action(action="submit_review", scores=...) to end. After the loop ends, pipeline_action(action="validate_phase", phase=7) must PASS (verify rounds_completed >= min_rounds).

Multi-Reviewer design (simulating 2-3 independent reviewers) 🆕

Human peer review is usually conducted by 2-3 independent reviewers, each with different knowledge backgrounds and thinking patterns. To simulate this process:

Each Review round launches 2-3 subagents, each subagent usinga different top-tier model(to avoid identical thinking patterns):

Reviewer	Model	Role
R1	Claude Opus 4.6	Methodology Expert + Statistician
R2	OpenAI Codex 5.3	Domain Specialist + Editor
R3	Gemini Pro 3.1	Independent Critical Reviewer

Each Reviewer must：

Search literature for verification: use unified_search / find_related_articles to search relevant literature, compare whether the manuscript's citations are sufficient and whether the conclusions are consistent with existing evidence
Independently read the full manuscript: do not refer to other Reviewers' opinions
Raise 5-10 structured questions: each question must include section, paragraph, category, issue, suggestion
Literature comparison: cite at least 2-3 searched references to support or challenge the manuscript's viewpoints
Give an independent score: score each of the 6 quality dimensions

Merge process：

After all Reviewers independently complete, merge into a unified review-report-{round}.md
Deduplicate repeated issues, keeping the strictest severity
Author Response must respond to all issues from all Reviewers

載入 journal-profile.yaml → 取得 reviewer_perspectives + quality_threshold

── Review Loop（最少 min_rounds 輪，最多 max_rounds 輪）──

FOR round = 1 TO N:

  ── Stage A: Multi-Reviewer Review（獨立 subagent） ──
  LAUNCH 2-3 subagents in parallel (each using different model):

  FOR each reviewer subagent:
    1. 搜尋文獻：unified_search() 搜尋與稿件主題相關的最新文獻（至少 5 篇）
    2. 比對文獻：find_related_articles() / find_citing_articles() 驗證稿件引用
    3. 獨立審查全稿，提出 5-10 個結構化問題
    4. 以搜尋到的文獻為依據，標記稿件中的不足或錯誤
    5. 獨立評分（6 維度）

    Reviewer 角色覆蓋：
    "methodology_expert":
      - 研究設計是否嚴謹？統計方法是否恰當？
      - 方法是否可再現？偏差控制是否充分？
      - 搜尋同類方法學文獻 → 比對是否有遺漏的 best practice

    "domain_specialist":
      - 文獻引用是否全面且最新？
      - 對領域 gap 的理解是否準確？臨床意義是否明確？
      - 搜尋最新系統性回顧 → 驗證稿件是否與當前共識一致

    "statistician":
      - 統計假設是否合理？結果呈現是否清晰？
      - 圖表是否有效傳達數據？
      - 搜尋統計方法指引 → 驗證分析方法的恰當性

    "editor":
      - 寫作品質（清晰度、邏輯流、語法）
      - 是否符合期刊風格？圖表品質與必要性

  MERGE all reviewer outputs → 統一 review-report-{round}.md

  產出 .audit/review-report-{round}.md：
  ┌─────────────────────────────────────────────────┐
  │ ---                                             │
  │ round: 1                                        │
  │ date: "YYYY-MM-DD"                              │
  │ reviewers:                                      │
  │   - role: "Methodology Expert"                  │
  │     issues_major: 2                             │
  │     issues_minor: 1                             │
  │   - role: "Domain Specialist"                   │
  │     issues_major: 1                             │
  │     issues_minor: 3                             │
  │ total:                                          │
  │   major: 3                                      │
  │   minor: 5                                      │
  │   optional: 2                                   │
  │ ---                                             │
  │                                                 │
  │ # Review Report — Round {round}                 │
  │                                                 │
  │ ## Reviewer 1: Methodology Expert               │
  │ ### MAJOR                                       │
  │ - id: R1-M1                                     │
  │   section: Methods                              │
  │   paragraph: methods-p3                         │
  │   category: methodology                         │
  │   issue: 缺少 sample size justification         │
  │   suggestion: 加入 power analysis               │
  │ ### MINOR                                       │
  │ - id: R1-m1                                     │
  │   section: Results                              │
  │   ...                                           │
  └─────────────────────────────────────────────────┘

  ── Stage B: Author Response（全 issue 須回應） ──
  產出 .audit/author-response-{round}.md：
  ┌─────────────────────────────────────────────────┐
  │ ---                                             │
  │ round: 1                                        │
  │ response_to: "review-report-1.md"               │
  │ actions:                                        │
  │   accepted: 6                                   │
  │   accepted_modified: 2                          │
  │   declined: 2                                   │
  │ ---                                             │
  │                                                 │
  │ # Author Response — Round {round}               │
  │                                                 │
  │ ### R1-M1: Sample size justification            │
  │ - **Action**: ACCEPT                            │
  │ - **Response**: Added power analysis...         │
  │ - **Change**: methods-p3 (+45 words)            │
  │ - **Verified**: Hook A re-run ✅               │
  │                                                 │
  │ ## Completeness Check                           │
  │ | Issue ID | Addressed | Action      |          │
  │ | R1-M1    | ✅       | ACCEPT      |          │
  │ | R2-o1    | ✅       | DECLINE     |          │
  │ ALL issues addressed: ✅                        │
  └─────────────────────────────────────────────────┘

  ── Completeness Check（每輪必過）──
  FOR issue IN review_report.all_issues:
    IF issue.id NOT IN author_response → FAIL（禁止忽略 issue）
  → 未回應的 issue 必須標記 DECLINE + 理由

  ── Stage C: 執行修正（MANDATORY — 稿件必須被修改） ──
  ⚠️ HARD ENFORCEMENT: `pipeline_action(action="submit_review", ...)` 會比對稿件 hash。
  若稿件未修改 → 提交會被 REJECT。

  FOR each ACCEPTED issue:
    1. 定位 paragraph ID（from manuscript-plan.yaml）
    2. patch_draft() 修正
    3. re-run Hook A on patched section
    4. 記錄到 author-response

  FOR each MINOR ACCEPTED issue:
    1. batch patch_draft 修正
    2. 快速 Hook A 驗證

  OPTIONAL + DECLINED issues → LOG only（不自動修正）

  ── Stage C2: 敘事強化（MANDATORY — 即使無結構問題） ──
  人類 reviewer 即使不要求結構修改，也會提升敘事品質。
  AI reviewer 必須做到至少同等標準：
    1. 選擇至少 3 段落進行「敘事強化」（tighten prose, improve transitions,
       strengthen claims, eliminate redundancy, improve word economy）
    2. 每段 patch_draft() 修正，附帶改善理由
    3. 記錄到 author-response 的 "Narrative Enhancements" section

  IF 沒有結構性 issue（全部 DECLINE 或 OPTIONAL only）：
    → Stage C2 仍然 MANDATORY（最低改 3 段落）
    → 目標：人類讀者應能看到「review 前」vs「review 後」有明顯差異

  ── Stage C3: Anti-AI 自然度審查（MANDATORY — Code-Enforced Gate） ──
  AI 生成文字常被偵測器（Gemini 3.1 等）秒辨。Stage C3 強制 Agent 自我檢測並修正。
  更關鍵的是：有經驗的人類審稿人會注意到段落間的「語體斷裂」——前段 ESL 風格突然切換到
  流暢的 corporate-academic prose，即使通過自動偵測也會引起懷疑。

  Step 1: Code-Enforced Hook Gate（三層掃描）
    → run_quality_checks(action="writing_hooks", hooks="A3,A3B,A3C")
    → A3: 禁止詞掃描（~150 個 AI 特徵用語）
    → A3b: 結構信號偵測（句長均勻度、轉折詞密度、句首多樣性、三連列舉、段落均勻度）
    → A3c: 語體一致性偵測（段落級可讀性指標 z-score 異常偵測 + 詞彙精緻度落差）
    → 任何 CRITICAL → 必須 patch_draft 修正後重跑，直到 PASS

  Step 2: Agent 自我審閱（Anti-AI Naturalness Check）
    Agent 以下列視角重新閱讀全文：
    a) 是否有「四平八穩」的語調？（每段都同一節奏 → 改）
    b) 是否過度使用被動語態或名詞化？（nominalization → 改為動詞）
    c) 是否有 AI 特有的「禮貌性冗餘」？（"It is important to note that" → 刪）
    d) 段落是否缺乏個人學術觀點？（全是 hedging → 加 assertive claim）
    e) 是否有「假平衡」？（on the one hand... on the other hand → 取立場）
    f) **語體斷裂**：A3c 標記的 outlier 段落是否「太流暢」而脫離作者自然語體？
       → 降級過度精煉的句子、加回作者的「笨拙但真實」語法特徵
    → 每項至少修正 2 處 → patch_draft()

  Step 3: 雙 Subagent 交叉審計（RECOMMENDED — 至少最終輪必做）
    使用兩個不同模型的 subagent 分別審計，交叉比對結果：

    Subagent 1 — Surface Scanner（concept-challenger 或主 Agent）：
      → 審計目標：黑名單詞、結構信號、自動偵測器可偵測的 pattern
      → 問題：「能否通過 GPTZero？哪些具體詞/句需要替換？」

    Subagent 2 — Voice Analyst（domain-reviewer）：
      → 審計目標：語體一致性、ESL baseline vs polished prose 對比
      → 問題：「人類審稿人會不會因為語體突變而起疑？逐段評分 1-5」
      → 重點：段落級 AI 機率評分，標出 TOP 5 最可疑句子

    交叉比對規則：
      → 兩者都標 🔴 → 必須重寫該段落
      → 僅一方標 ⚠️ → Agent 判斷是否需要調整
      → 兩者都 ✅ → 安全

  Step 4: 修正後驗證
    → 重跑 run_quality_checks(action="writing_hooks", hooks="A3,A3B,A3C")
    → 必須全部 PASS 或僅 WARNING（無 CRITICAL）才能進入 Stage D
    → IF 仍有 CRITICAL → 回到 Step 2 再修，最多 3 輪

  ── Stage D: 品質重評 ──
  更新 quality-scorecard.md：
    | 維度 | Round N-1 分數 | Round N 分數 | 變化 |

  計算總分 → 比對 quality_threshold

  IF 總分 ≥ quality_threshold:
    → ✅ PASS: 結束 review loop
    → LOG: "Review completed at round {round}, score={score}"

  IF 總分 < quality_threshold AND round < N:
    → 繼續下一輪（重新審稿）
    → 分析哪些維度分數最低 → 下一輪重點關注

  IF round = N AND 總分 < quality_threshold:
    → ⚠️ PARTIAL: 呈現剩餘問題 + 分數趨勢
    → 詢問用戶：
      a) 接受當前品質（記錄風險）
      b) 繼續 N 輪（用戶延長 loop）
      c) 手動修改後重新 review
      d) 🔁 回退到 Phase 5 重寫特定 section → call `pipeline_action(action="rewrite_section", ...)`

  ── Stage D2: 回退決策（Agent 自主判斷 + 自主執行） ──
  IF Agent 判斷某些 section 的問題無法透過 patch_draft 修正，需要整個 section 重寫：
    1. 識別需要重寫的 section（通常是 persistent critical issues 集中的 section）
    2. **Autopilot 模式**（預設）：
      → Agent 直接呼叫 `pipeline_action(action="rewrite_section", sections="Methods,Results", reason="...")`
       → 決策和理由記入 `.audit/pipeline-run-{ts}.md`
       → Pipeline 回退到 Phase 5 → 僅重寫指定 sections
    3. **手動模式**（`autopilot: false` 或 `autopilot_pause_at` 含 `"phase7_rewrite"`）：
       → 🗣️ 詢問用戶確認：「建議重寫 [section]，是否同意？」
      → 用戶確認後執行 `pipeline_action(action="rewrite_section", ...)`
    4. 重寫後走 Phase 5 → 6 → 7 完整流程
    5. ⚠️ regression_count > 2 → **安全閥：強制停下詢問用戶**（防止無限循環）

  ── Stage E: Evolution Tracing（每輪結束後） ──
  追加到 .audit/evolution-log.jsonl：
  {"event": "review_round", "round": N,
   "timestamp": "...",
   "scorecard": {6 維度分數},
   "scorecard_delta": {與 Round N-1 的差異},
   "review_issues": {"major": M, "minor": m, "accepted": A, "declined": D},
   "draft_snapshots": ["snapshot_path_1", ...],
   "patches_applied": [{"section": "...", "issue_id": "R1-M1", "words_changed": N}],
   "instruction_version": git_short_hash()}

Hook E: EQUATOR Reporting Guidelines Compliance (Phase 7 Stage A addition) 🆕

In each Review round of Phase 7, additionally execute EQUATOR reporting guideline compliance check. This is one of the system's core selling points: AI canitem by item, without omissionexecute checklist items that human reviewers often overlook.

Trigger: Stage A of each Review round in Phase 7 (in parallel with the 4-role review)

E1: Automatically detect applicable guidelines

1. 讀取 journal-profile.yaml → reporting_guidelines.checklist
2. IF checklist 已指定 → 使用指定指引
3. IF checklist 為空 → 依據 paper.type + Methods 內容自動偵測：

   paper.type × 內容特徵 → 適用指引：
   ┌─────────────────────────┬───────────────────────────────────┐
   │ Paper Type / 特徵       │ 主要指引          │ AI 擴展指引   │
   ├─────────────────────────┼───────────────────────────────────┤
   │ RCT                     │ CONSORT 2010      │ CONSORT-AI    │
   │ RCT + AI intervention   │ CONSORT-AI        │ SPIRIT-AI     │
   │ Observational cohort    │ STROBE            │ RECORD        │
   │ Observational + routine │ RECORD            │ STROBE        │
   │ Systematic review       │ PRISMA 2020       │ PRISMA-S      │
   │ Meta-analysis           │ PRISMA 2020       │ MOOSE         │
   │ MA of observational     │ MOOSE             │ PRISMA 2020   │
   │ Diagnostic accuracy     │ STARD 2015        │ STARD-AI      │
   │ Diagnostic + AI         │ STARD-AI          │ CLAIM         │
   │ Prediction model        │ TRIPOD 2015       │ PROBAST       │
   │ Prediction + AI/ML      │ TRIPOD+AI         │ MI-CLAIM      │
   │ AI clinical decision    │ DECIDE-AI         │ TRIPOD+AI     │
   │ AI clinical trial       │ SPIRIT-AI         │ CONSORT-AI    │
   │ Medical imaging AI      │ CLAIM             │ STARD-AI      │
   │ Case report             │ CARE 2013         │ —             │
   │ Qualitative research    │ SRQR / COREQ      │ —             │
   │ Quality improvement     │ SQUIRE 2.0        │ —             │
   │ Economic evaluation     │ CHEERS 2022       │ —             │
   │ Animal research         │ ARRIVE 2.0        │ —             │
   │ Protocol (trial)        │ SPIRIT 2013       │ SPIRIT-AI     │
   │ Protocol (SR)           │ PRISMA-P 2015     │ —             │
   │ Software/Methods paper  │ —                 │ see Note      │
   └─────────────────────────┴───────────────────────────────────┘

   Note: Software/Methods papers 無專屬 EQUATOR 指引，
         但若涉及 AI/ML → 適用 TRIPOD+AI 相關項目（選擇性）

E2: Checklist item-by-item verification

FOR guideline IN detected_guidelines:
  checklist = load_checklist(guideline)  # 內建 checklist 資料庫

  FOR item IN checklist.items:
    # 定位：該項目應出現在哪個 section
    target_section = item.expected_section  # e.g. "Methods", "Results"
    content = read_draft(target_section)

    # 三級判定
    IF item clearly addressed in content:
      → ✅ REPORTED (記錄: section + paragraph + 摘要)
    ELIF item partially addressed:
      → ⚠️ PARTIAL (記錄: 缺什麼、建議補充)
    ELSE:
      → ❌ NOT REPORTED (記錄: 建議加入的位置和內容)

  # 合規率計算
  compliance_rate = REPORTED / total_applicable_items
  partial_rate = PARTIAL / total_applicable_items

  # 閾值判定（依 item 重要性分級）
  IF any ESSENTIAL item is NOT REPORTED → MAJOR issue
  IF compliance_rate < 80% → WARNING
  IF compliance_rate ≥ 80% AND all ESSENTIAL reported → PASS

E3: EQUATOR guideline built-in database

Guideline	Full name	Scope	Number of items	Source
CONSORT 2010	Consolidated Standards of Reporting Trials	RCT	25 items	equator-network.org
CONSORT-AI	CONSORT Extension for AI	RCTs with AI intervention	14 extension items	Lancet Digital Health 2020
STROBE	Strengthening Reporting of Observational Studies	Observational studies	22 items	equator-network.org
PRISMA 2020	Preferred Reporting Items for SR and MA	Systematic reviews	27 items	BMJ 2021
PRISMA-S	PRISMA Search Extension	Search Strategy Report	16 items	Systematic Reviews 2021
PRISMA-P	PRISMA for Protocols	SR protocol	17 items	BMJ 2015
MOOSE	Meta-analysis of Observational Studies	Observational study MA	35 items	JAMA 2000
STARD 2015	Standards for Reporting Diagnostic Accuracy	Diagnostic accuracy	30 items	BMJ 2015
STARD-AI	STARD for AI Diagnostic	AI diagnostic studies	extension items	Nature Medicine 2021
TRIPOD 2015	Transparent Reporting of Prediction Models	Prediction models	22 items	BMJ 2015
TRIPOD+AI	TRIPOD for AI/ML Prediction	AI/ML prediction models	27+extensions	BMJ 2024
PROBAST	Prediction Model Risk of Bias Assessment	Prediction model bias	20 items	Annals Int Med 2019
DECIDE-AI	AI Decision Support Early Evaluation	AI decision support	17 items	Nature Medicine 2022
SPIRIT 2013	Standard Protocol Items for Trials	Trial protocol	33 items	equator-network.org
SPIRIT-AI	SPIRIT Extension for AI	AI trial protocol	15 extension items	Nature Medicine 2020
CLAIM	Checklist for AI in Medical Imaging	Medical imaging AI	42 items	Radiology 2020
MI-CLAIM	Minimum Information about Clinical AI Modeling	Minimum information for AI modeling	15 items	Nature Medicine 2020
CARE 2013	Case Report Guidelines	Case reports	13 items	equator-network.org
ARRIVE 2.0	Animal Research Reporting	Animal studies	21 items	PLOS Biology 2020
SQUIRE 2.0	Standards for Quality Improvement	Quality improvement	18 items	BMJ Quality Safety 2015
SRQR	Standards for Reporting Qualitative Research	Qualitative research	21 items	Academic Medicine 2014
COREQ	Consolidated Criteria for Qualitative Research	Qualitative (interviews/focus groups)	32 items	Int J Qual Health Care 2007
CHEERS 2022	Consolidated Health Economic Evaluation	Health economic evaluation	28 items	BMJ 2022
RECORD	Reporting of Studies Using Routine Data	Routine data studies	13+extensions	PLOS Medicine 2015
AGREE II	Appraisal of Guidelines Research and Evaluation	Clinical guidelines	23 items	CMAJ 2010

E4: Compliance Report output

Output per Review round .audit/equator-compliance-{round}.md：

# EQUATOR Compliance Report — Round {round}

## Detected Guidelines: TRIPOD+AI (primary), MI-CLAIM (secondary)

## Detection Basis: paper.type=prediction-model, Methods mentions "deep learning"

### TRIPOD+AI Compliance: 85% (23/27 items)

| #   | Item                                 | Section      | Status          | Notes            |
| --- | ------------------------------------ | ------------ | --------------- | ---------------- |
| 1   | Title identifies as prediction model | Title        | ✅ REPORTED     |                  |
| 2   | Abstract: structured summary         | Abstract     | ✅ REPORTED     |                  |
| 3a  | Background and objectives            | Introduction | ✅ REPORTED     |                  |
| 4a  | Source of data                       | Methods      | ✅ REPORTED     |                  |
| 4b  | Data collection dates                | Methods      | ⚠️ PARTIAL      | Missing end date |
| ... |                                      |              |                 |                  |
| 10d | Handling of missing data             | Methods      | ❌ NOT REPORTED | → Add to Methods |
| 15a | Model performance metrics            | Results      | ✅ REPORTED     |                  |

### MI-CLAIM Compliance: 73% (11/15 items)

...

### Summary

| Guideline | Compliance | ESSENTIAL items | Status                 |
| --------- | ---------- | --------------- | ---------------------- |
| TRIPOD+AI | 85%        | 20/22 ✅        | ⚠️ 2 ESSENTIAL missing |
| MI-CLAIM  | 73%        | 9/10 ✅         | ⚠️ 1 ESSENTIAL missing |

### Action Items (for Author Response)

- [E-M1] TRIPOD+AI Item 10d: Add missing data handling to Methods
- [E-M2] TRIPOD+AI Item 4b: Add data collection end date
- [E-m1] MI-CLAIM Item 12: Add model interpretability discussion

E5: Integration with Phase 7 Review Loop

Phase 7, Stage A（每輪）:
  1. 執行 4 角色 Reviewer 審查 → review-report-{round}.md
  2. 執行 Hook E → equator-compliance-{round}.md
  3. 合併 issues: Review issues + EQUATOR issues → 統一編號
     - Review issues: R1-M1, R2-m1, ...
     - EQUATOR issues: E-M1, E-m1, ... (E = EQUATOR)
  4. Author Response 必須回應兩種 issue

Phase 7, Stage D（品質重評）:
  quality-scorecard 新增維度：
  | 維度 | 評分標準 | 權重 |
  | EQUATOR 合規 | checklist compliance rate + ESSENTIAL 完整度 | 15% |
  → 原有 6 維度權重等比調降，總和仍 = 100%

Review quality dimensions (quality-scorecard)

Dimension	Scoring criteria (0-10)	Weight
Citation quality	Citations sufficient, up-to-date, high-impact, correctly formatted	10%
Methodological reproducibility	Study design, statistics, reproducible	18%
Text quality	Clarity, logical flow, no AI traces, grammar	16%
Conceptual consistency	NOVELTY embodied, SELLING POINTS, full-manuscript consistency	16%
Format compliance	Word count, figures and tables, number of citations, journal requirements	7%
Figure and table quality	Necessity of figures and tables, clarity, captions, data presentation	8%
EQUATOR compliance	checklist compliance + ESSENTIAL items complete	13%
Reproducibility/data availability	Data sharing statement, code availability, protocol registration	12%

Total score = Σ(dimension score × weight)

Division of labor between Review and Hook

Aspect	Hook A-C（Phase 5-6）	Autonomous Review（Phase 7）
Purpose	Technical compliance	Academic quality + reporting guideline compliance
Trigger timing	During writing / after full manuscript completion	Phase 6.5 mandatory entry (MANDATORY)
Focus	Format, citations, word count, consistency	Content quality, logic, academic persuasiveness + EQUATOR
Correction method	patch_draft (local correction)	May rewrite paragraphs or restructure arguments
Role	Automated linter	Simulated Reviewer + EQUATOR compliance
Stop condition	0 CRITICAL	quality_threshold met + ESSENTIAL complete
Output	audit log	review-report + equator-compliance
Granularity	pass/fail	MAJOR/MINOR/OPTIONAL + checklist item by item

Phase 8: REFERENCE SYNC

sync_references(filename=manuscript) → generate References section
Confirm all [[wikilinks]] resolved
format_references(style=journal-profile.references.style)
Validate citation count ≤ references.reference_limits[paper.type]（fallback max_references）
IF over the limit → flag the least-cited refs → suggest deletion

Gate: 0 broken links + citation count compliant

Phase 9: EXPORT（Code-Enforced — CRITICAL Gate）

Skill: word-export

⚠️ Code-Enforced: pipeline_action(action="validate_phase", phase=9) The checks for docx and pdf are CRITICAL severity. Both docx and pdf must be produced, otherwise the Gate will not pass.

export_document(action="docx") → produce DOCX
export_document(action="pdf") → produce PDF
inspect_export(action="list_templates") → select template (prefer matching journal-profile.journal.name)
export_document(action="session_start") → export_document(action="session_insert") × N
Validate required document checklist (from required_documents.*）：
- cover_letter → if not yet generated, prompt the user or auto-draft
- author_contributions → prompt the user to fill in
- ethics_statement → prompt the user to provide
inspect_export(action="verify_document") → export_document(action="session_save")
inspect_export(action="docx_smoke") or inspect_export(action="xml_smoke") → validate DOCX zip/XML structure,word/document.xml, paragraphs and visible text

Gate: pipeline_action(action="validate_phase", phase=9) — docx + pdf must exist (CRITICAL) + required document checklist complete; DOCX XML smoke should PASS before entering final delivery

Phase 9.5: SUBMISSION READINESS

Input: documents exported by Phase 9 External MCP: —

Purpose: final systematic check before submission, ensuring all submission requirements are complete.

File checklist verification(according to journal profile):
- Title page (including author info, ORCID, corresponding author)
- Manuscript（blinded if required）
- Abstract（structured / unstructured）
- Cover letter
- Conflict of interest declarations
- Funding statement
- Ethics approval statement
- Data availability statement
- Supplementary files（if any）
- Figure files（high-res, correct format）
- Table files（if separate）
- EQUATOR checklist（if required）
- Author contributions (CRediT if applicable)
Final format confirmation：
- Word count limit (abstract + main text)
- Figure/table count limit
- Reference format (journal CSL style)
- File naming conventions
Content consistency：
- Title page authors == manuscript authors
- Abstract numbers == main text numbers
- Cover letter mentions the correct journal name
Produce exports/submission-checklist.md(✅/❌ checklist)

Gate: submission-checklist.md all ✅ or user confirms acceptance of exceptions

Phase 10: RETROSPECTIVE (closing-loop core)

🔔 HOOK D: meta-learning (see definition below, including D7 Review Retrospective, D8 EQUATOR Retrospective)

Review execution records + Hook trigger statistics + Review round statistics
Update SKILL.md Lessons Learned
Update AGENTS.md (if applicable)
Update complete .memory/ records
Analyze whether journal-profile settings are reasonable → suggest fine-tuning
🆕 D7: analyze review-report + author-response → evolve Reviewer instructions
🆕 D8: analyze equator-compliance → evolve EQUATOR detection and classification logic

Phase 11: FINAL DELIVERY（Code-Enforced — CRITICAL Gate）🆕

Purpose: confirm paper deliverables are complete and auditable. Git serves only as optional provenance and must not block the paper-only workflow. Trigger: automatically entered after Phase 10 completes.

⚠️ Code-Enforced: pipeline_action(action="validate_phase", phase=11) Validate the following CRITICAL items:

Phase 9 exported .docx + .pdf exists
Phase 10 retrospective / meta-learning artifacts complete
Phase 7 review loop completed

The following items are optional provenance / WARNING or INFO, and must not block final delivery:

Whether the Git repository exists
Whether the working directory is clean
Whether the latest commit includes the project files
Whether there is a remote/upstream / whether to push

Flow：

Confirm that Phase 9 .docx + .pdf has been produced
Confirm that Phase 10 .audit/pipeline-run-*.md、.audit/meta-learning-audit.yaml、.audit/meta-learning-feedback.yaml is complete
pipeline_action(action="validate_phase", phase=11, response_format="json", compact=true) → Verification PASS
If the project uses Git, commit/push may optionally be used as provenance; if the user is only writing a paper, no remote is needed

Gate: pipeline_action(action="validate_phase", phase=11) — final paper delivery artifacts complete; Git remote/push is not a hard requirement

Audit Trail generation (automatically executed when each Phase completes)

The Agent must maintain the following files in projects/{slug}/.audit/ :

Incremental updates (at the end of each Phase)

Phase 完成 → 更新以下檔案：

1. pipeline-run-{YYYYMMDD-HHmm}.md（增量 append）
   格式：
   ## Phase {N}: {名稱}
   - 開始時間 / 結束時間
   - 輸入：{files read}
   - 輸出：{files written}
   - Hook 結果：
     | Hook | 觸發 | 通過 | 修正次數 | 最終結果 |
   - 決策紀錄：{任何自動決策 + 理由}
   - Issues flagged for later: {list}

2. checkpoint.json（覆寫更新）
   {
     "last_completed_phase": N,
     "current_section": "Methods",
     "phase_outputs": { "P1": "project_created", "P2": "15_refs_saved", ... },
     "flagged_issues": [...],
     "audit_stats": { "hook_a_triggers": 5, "hook_a_passes": 3, ... },
     "timestamp": "ISO-8601"
   }

Generated after Phase 6 completes

3. quality-scorecard.md
   | 維度 | 分數 (0-10) | 說明 |
   | 引用品質 | | |
   | 方法學再現性 | | |
   | 文字品質 | | |
   | 概念一致性 | | |
   | 格式合規 | | |
   | 圖表品質 | | |
   | 總分 | | |

4. hook-effectiveness.md
   | Hook | 觸發次數 | 通過次數 | 修正次數 | 誤報次數 | 有效率 |
   → 供 Hook D 分析用

Generated for each round of Phase 7

5. review-round-{N}.md
   ## Round {N} Summary
   | 角色 | MAJOR issues | MINOR issues | OPTIONAL |
   | Issue | 角色 | 嚴重度 | 修正策略 | 修正後狀態 |

   ## Quality Score Trend
   | 維度 | Round 0 | Round 1 | Round 2 | ... | 變化趨勢 |
   | 總分 | 6.2 | 7.1 | 7.8 | ... | ↑ |

6. equator-compliance-{N}.md  🆕
   ## EQUATOR Compliance Report — Round {N}
   ## Detected Guidelines: {auto-detected or specified}
   | # | Item | Section | Status | Notes |
   → 見 Hook E4 格式定義

Evolution Tracing（Phase 6.5 → 7 → 10）🆕

7. evolution-log.jsonl（append-only，每事件一行 JSON）

   ## 事件類型：
   - baseline: Phase 6.5 建立的基線快照
   - review_round: Phase 7 每輪 review 結果
   - equator_check: Hook E 每輪 compliance 結果
   - hook_correction: Hook A-C 修正事件
   - meta_learning: Phase 10 自我改進事件

   ## Schema:
   {"event": "baseline", "round": 0, "timestamp": "ISO-8601",
    "scorecard": {"citation": 7, "methodology": 8, "text": 6, ...},
    "word_count": 3200,
    "instruction_version": "abc1234"}

   {"event": "review_round", "round": 1, "timestamp": "...",
    "scorecard": {"citation": 7.5, "methodology": 8, "text": 7, ...},
    "scorecard_delta": {"text": +1, "total": +0.8},
    "review_issues": {"major": 3, "minor": 5, "accepted": 7, "declined": 1},
    "patches_applied": [{"section": "Methods", "issue_id": "R1-M1", "words_changed": 45}],
    "instruction_version": "abc1234"}

   {"event": "equator_check", "round": 1, "timestamp": "...",
    "guideline": "TRIPOD+AI", "compliance_rate": 0.85,
    "essential_complete": false, "missing_essential": ["Item 10d", "Item 4b"],
    "action_items": 3}

   {"event": "meta_learning", "phase": 10, "timestamp": "...",
    "skill_updates": ["Lessons Learned +1"],
    "hook_adjustments": [{"hook": "B5", "param": "threshold", "old": 5, "new": 6}],
    "total_rounds": 2, "final_score": 7.8}

   ## 用途：
   - Phase 10 D1-D7 分析 → 產出 evolution summary
   - 跨 run 比較（如有多次執行）
   - 論文自身的 Fig 1 / Table 2（框架效果量化證據）

Generated after Phase 2 completes

5. search-strategy.md
   - 搜尋策略 + MeSH terms
   - 每組搜尋的結果數量
   - 篩選標準 + 排除原因

6. reference-selection.md
   - 最終選擇的文獻列表 + 選擇理由
   - 排除的重要文獻 + 排除原因

Generated after Phase 3 completes

7. concept-validation.md
   - 驗證分數 (每輪)
   - 修正歷史
   - CGU 使用紀錄（如有）

Checkpoint Recovery Mechanism

Implementation：CheckpointManager（infrastructure/persistence/checkpoint_manager.py）

Pipeline 啟動時：
1. 檢查 projects/{slug}/.audit/checkpoint.json
2. IF exists:
   → ckpt = CheckpointManager(audit_dir)
   → summary = ckpt.get_recovery_summary()  # 顯示上次進度
   → 詢問：
     a) 從 Phase {N+1} 繼續
     b) 從當前 section 繼續
     c) 重新開始（保留文獻和 concept）
     d) 完全重來 → ckpt.clear()
3. IF 選擇繼續：
   → state = ckpt.load()
   → 載入 phase_outputs
   → 載入 flagged_issues（帶入 Phase 6）
   → 跳到對應 Phase/Section

Phase 轉換時：
  ckpt.save_phase_start(phase, name)
  ... 執行 ...
  ckpt.save_phase_completion(phase, name, outputs, flagged_issues)

Copilot Hooks definition (triggered during writing)

Hook A: post-write (immediately after each write, at most N rounds, N = `pipeline.hook_a_max_rounds`）

#	Check item	MCP Tool	Failure behavior	Threshold source
A1	Word count within target ±20%	`count_words`	`patch_draft` condense/expand	`paper.sections[].word_limit`
A2	Citation density meets standard	`get_available_citations`	`suggest_citations` + `patch_draft`	`pipeline.writing.citation_density.*`
A3	No Anti-AI patterns	`run_quality_checks(action="writing_hooks", hooks="A3")`	`patch_draft` rewrite	`pipeline.writing.anti_ai_strictness`
A3b	AI structural signal detection 🆕	`run_quality_checks(action="writing_hooks", hooks="A3B")`	`patch_draft` rewrite	Code-Enforced (5 structural analyses)
A3c	Voice consistency detection 🆕	`run_quality_checks(action="writing_hooks", hooks="A3C")`	`patch_draft` rewrite	Code-Enforced (z-score outliers)
A4	Wikilink format correct	`validate_wikilinks`	Auto-fix	—
A5	Language consistency (BrE/AmE)	`run_quality_checks(action="writing_hooks", hooks="A5")`	`patch_draft` unify spelling	`pipeline.writing.prefer_language`
A6	Paragraph duplication detection	`run_quality_checks(action="writing_hooks", hooks="A6")`	`patch_draft` rewrite duplicated paragraphs	`pipeline.writing.overlap_threshold`
A7	Sufficiency of literature quantity 🆕	`run_quality_checks(action="writing_hooks", hooks="A7")`	BLOCK writing, prompt to expand search	`DEFAULT_MINIMUM_REFERENCES` per type

Hook A Cascading Protocol

Round 1: 執行 A1-A4 → 收集所有 issues
  → IF issues:
    → batch patch_draft（一次修正多個問題）
    → 記錄：哪些問題、修正策略

Round 2: re-run A1-A4
  → IF same issues persist:
    → 改變策略（例：A1 字數超標 → Round 1 精簡句子 → Round 2 刪除次要段落）
    → patch_draft
    → 記錄：策略變更原因

Round 3: re-run A1-A4
  → IF still issues:
    → LOG to audit trail（不阻擋）
    → FLAG issue type + section → 帶入 Phase 6 Cascading Audit
    → 繼續下一步（Hook B）

每 round 的修正策略升級：
  A1 字數: trim sentences → remove paragraphs → restructure section
  A2 引用: suggest_citations → search more refs → flag for user
  A3 Anti-AI: rephrase → rewrite paragraph → flag specific phrases
  A3b AI結構信號: 句長CV/轉折詞密度/句首多樣性/三連列舉/段落CV → patch_draft
  A3c 語音一致性: 段落風格離群值 → patch_draft 改寫風格斷裂段落
  A4 Wikilink: auto-fix → manual check → flag broken refs

A2 Citation density standard: Introduction ≥1/100w, Methods ≥0, Results ≥0, Discussion ≥1/150w

A3 Anti-AI prohibited words (~75): covering six major categories: Classic AI filler, Grandiose/buzzword, Metaphorical cliché, Vague academic padding, Formulaic conclusion/transition, and AI-specific phrasing. Common ones such as In recent years, plays a crucial role, has garnered significant attention, a comprehensive understanding, This groundbreaking → replace with specific content.

A3b AI structural signal detection (Code-Enforced):

Sentence length uniformity: CV < 0.25 = WARNING (AI tends to produce sentences of equal length)
Transition word density: sentence-initial transition words >20% WARNING, >35% CRITICAL
Sentence-initial diversity: unique ratio < 0.50 = WARNING
Tricolon: X, Y, and Z appears >3 times = WARNING
Paragraph length uniformity: CV < 0.20 = WARNING

A3c Voice consistency detection (Code-Enforced):

Paragraph-level features: avg_sent_len, avg_word_len, type_token_ratio, punct_complexity
Document baseline: calculate the mean + std of each metric across the full text
Outlier detection: any metric z-score > 1.8 → WARNING (style break)
Vocabulary refinement discrepancy: max − min avg_word_len > 1.2 → WARNING (ESL vs polished paragraphs)
Typical scenario: an ESL grammar paragraph suddenly shifts to a polished corporate academic style → a human reviewer can notice

Hook B: post-section (after the section is completed, including retrospective revision)

#	Check item	MCP Tool	Failure behavior
B1	Consistent with concept.md	`read_draft("concept.md")` + Agent comparison	Rewrite inconsistent paragraphs
B2	🔒 NOVELTY reflected in Intro	`read_draft` check	`patch_draft` add
B3	🔒 SELLING POINTS in Discussion	Item-by-item comparison	`patch_draft` supplement
B4	Not contradictory with written sections	`read_draft` cross-check	Fix contradictions
B5	Methodological reproducibility	Agent evaluates according to paper_type	`patch_draft` add details
B6	Writing order verification	`check_writing_order`	⚠️ Advisory (does not block)
B7	Section Brief compliance 🆕	Agent compares `manuscript-plan.yaml`	`patch_draft` fill in omissions

Hook B Cascading Protocol

執行 B1-B7 → 分類 issues:
  CRITICAL: B1 concept 不一致, B2 🔒 NOVELTY 缺失, B5 方法學 <5 分
  WARNING:  B3 selling points 不完整, B4 sections 矛盾, B7a/b/d/e Brief 遺漏
  ADVISORY: B6 寫作順序, B7c 段落順序偏離

IF CRITICAL issues:
  → 針對性 patch_draft 修正
  → re-run Hook A on patched areas（1 round，確認 patch 沒破壞品質）
  → re-run 失敗的 Hook B checks
  → IF still CRITICAL after 2 attempts:
    → FLAG for Phase 6 Cascading Audit
    → 記錄到 .audit/pipeline-run-{ts}.md

IF WARNING issues:
  → patch_draft 修正（1 round）
  → IF still WARNING → LOG + continue（Phase 6 會再檢查）

IF ADVISORY only:
  → LOG to audit trail → continue

B5 Methodology Checklist

Basic methodology check(all paper types):

Check item	Original	Case	Systematic	AI/ML Prediction	AI Clinical
Study design description	✅	✅	✅	✅	✅
Primary/secondary outcomes	✅	⬜	✅	✅	✅
Sample size/power	✅	⬜	⬜	✅	✅
Inclusion/exclusion criteria	✅	⬜	✅	✅	✅
Statistical method match	✅	⬜	✅	✅	✅
Variable definitions	✅	✅	⬜	✅	✅
Ethics statement	✅	✅	⬜	✅	✅
Collection period	✅	✅	✅	✅	✅
EQUATOR	CONSORT/STROBE	CARE	PRISMA	TRIPOD+AI	DECIDE-AI

AI/ML-specific check items(AI/ML paper types only):

Check item	Prediction	Diagnostic	Imaging	Decision
Data split strategy	✅	✅	✅	✅
Model architecture/hyperparameters	✅	✅	✅	✅
Training/validation/test set ratio	✅	✅	✅	⬜
External validation	✅	✅	✅	✅
Missing value handling	✅	✅	⬜	✅
Calibration	✅	⬜	⬜	⬜
Bias/Fairness analysis	✅	✅	✅	✅
Interpretability/explainability	⬜	⬜	✅	✅
Human vs AI comparison	⬜	✅	✅	✅
Applicable guidelines	TRIPOD+AI	STARD-AI	CLAIM	DECIDE-AI

Division of labor between B5 ↔ Hook E: B5 performs a "rapid methodology scan" (10 items or fewer) in Phases 5-6, while Hook E performs a "full EQUATOR checklist item-by-item verification" (20-42 items) in Phase 7. The two are complementary and non-overlapping.

If any required item < 5 points → patch_draft → still < 5 after 2 rounds → human intervention.

B6 Prerequisites

Target	Prerequisite	Reason
Results	Methods	Results depend on the Methods definition
Discussion	Results + Intro	Discussion addresses Results in response to Intro
Conclusion	Discussion	Summarizes the Discussion
Abstract	All main sections	Extract key points

Advisory only (§22 may be reorganized); skipped items are recorded in the audit trail.

B7 Section Brief compliance 🆕

Based on manuscript-plan.yaml compare paragraph by paragraph to ensure the Section Brief is correctly implemented.

Sub-item	Check content	Failure behavior
B7a	Each paragraph `key_claims` whether it is covered in the text	`patch_draft` add argument
B7b	Each paragraph `must_cite` whether it appears	`patch_draft` Add citation
B7c	Paragraph order/count vs plan	⚠️ Advisory (non-blocking)
B7d	Whether the `asset_plan` items of that section have been inserted	WARNING + prompt to insert figures/tables
B7e	Whether the `caption_requirements` of inserted figures/tables satisfies	`patch_draft` Supplement figure captions

FOR paragraph IN plan.sections[section].paragraphs:
  content = extract_paragraph(draft, paragraph.id)
  FOR claim IN paragraph.key_claims:
    IF claim NOT conveyed in content → ISSUE (B7a)
  FOR ref IN paragraph.must_cite:
    IF ref NOT in content → ISSUE (B7b)

FOR asset IN plan.asset_plan WHERE asset.section == section:
  IF asset NOT inserted in draft → ISSUE (B7d)
  IF asset inserted AND caption_requirements NOT met → ISSUE (B7e)

B7a/B7b omissions are WARNING (1 round patch_draft fix), B7d is WARNING, B7e is WARNING. B7c is ADVISORY (order deviation acceptable).

Hook B8: Data-Claim Alignment

Triggered after Results completion. Verify statistical claims are consistent with Methods:

Statistical methods: tests used in Results must be declared in Methods
P-value threshold: p < X in Results must match the α declared in Methods
CI width: the 95% CI in Results must match Methods
Software: statistical software mentioned in Results must be declared in Methods

#	Check item	MCP Tool	Failure behavior	Threshold source
B8	Statistical claim ↔ method alignment	`run_quality_checks(action="writing_hooks", hooks="B8")`	`patch_draft` Methods/Results fill-in	`pipeline.writing.stat_alignment`

Hook C: post-manuscript (after full manuscript completion, including layered backtracking, at most N rounds, N = `pipeline.hook_c_max_rounds`）

#	Check item	MCP Tool	Failure behavior	Backtracking layer	Threshold source
C1	Manuscript consistency	`check_formatting("consistency")`	`patch_draft`	→ B4	—
C2	Submission checklist	`check_formatting("submission")`	Fixed-point correction	—	`required_documents.*`
C3	N value consistent across sections	`read_draft` × N + number comparison	`patch_draft` unify	→ A	—
C4	Abbreviation first definition	`read_draft` + full-text scan	`patch_draft` add definition	→ A	—
C5	Wikilinks resolvable	`scan_draft_citations`	`save_reference_mcp` save	→ A4	—
C6	Total word count compliant	`count_words`	Condense overly long sections	→ A1	`word_limits.total_manuscript`
C7	Count and cross-reference compliance 🆕	See C7 subitems below	Handle by subitem	By subitem	`assets.`, `word_limits.`, `references.max_references`
C8	Temporal consistency	`read_draft` × N + Agent scan	`patch_draft` update outdated descriptions	→ B	—
C9	Supplementary material cross-references	`run_quality_checks(action="writing_hooks", hooks="C9")`	`patch_draft` add citation	—	—
C10	Reference fulltext + analysis verification	`run_quality_checks(action="writing_hooks", hooks="C10")`	Add fulltext/analysis	—	—
C11	Balanced citation distribution 🆕	`run_quality_checks(action="writing_hooks", hooks="C11")`	Redistribute citations to sections lacking citations	→ A2	—
C12	Citation appropriateness audit 🆕	`run_quality_checks(action="writing_hooks", hooks="C12")`	Add decision records to citation_decisions.json	—	—
C13	Figure/table quality and ordering 🆕	`run_quality_checks(action="writing_hooks", hooks="C13")`	Fix order/add captions	→ C7d	—
C14	Claim-Evidence alignment 🆕	`run_quality_checks(action="writing_hooks", hooks="C14")`	Add citations/source material/figure-table evidence for strong claims, or weaken the wording	→ Phase 5/8	Strong claims are not classified as A3 style issues

Hook C Cascading Protocol

Stage 1: Full Scan
  → 執行 C1-C7 → 收集 ALL issues
  → 分類: CRITICAL (C1不一致, C3數字錯, C5斷鏈) / WARNING (C2, C4, C6, C7)

Stage 2: Cascading Fix (最多 3 rounds)
  Round N:
    FOR each CRITICAL issue:
      1. 定位到具體 section + 段落
      2. patch_draft() 修正
      3. 觸發 回溯層 的 Hook（見上表）確認 patch 品質
    → re-run Hook C

Stage 3: C8 Temporal Consistency Pass
  → C1-C7 全過後，執行 C8 逆向掃描
  → IF C8 發現過時引用 → patch_draft 更新 → 重跑 C1 確認一致性

Stage 4: 人工介入（3 rounds 後仍 CRITICAL）
  → 生成 quality-scorecard.md（量化分數）
  → 呈現具體問題 + 建議修改方案
  → 用戶選擇：修改 / 接受風險 / 回到 Phase 5

Hook C 修正策略：
  C1 不一致: 統一術語 → 統一語氣 → 重寫弱 section
  C3 N 值: 以 Methods 定義為準 → 更新所有 sections
  C4 縮寫: 找首次出現 → 加全稱 → 後續只用縮寫
  C5 斷鏈: validate_wikilinks → save_reference_mcp → manual
  C6 字數: 刪冗餘 → 合併段落 → 詢問用戶刪哪段
  C7 數量: 依子項分別處理（見下方）
  C8 過時: 逆向掃描 → patch_draft 更新 → 重跑 C1
  C11 引用分布: 找缺引用 section → 補引用 → re-run A2
  C12 引用審計: 生成 citation_decisions.json → 每篇引用留 justification
  C13 圖表品質: 修正順序 → 補 caption → 確保 Results 有圖表引用

C7 Count and cross-reference compliance (D5 extension) 🆕

The original C7 checked only figure/table counts; it is extended into a comprehensive count/citation compliance check with five subitems.

Subitem	Check content	MCP Tool	Failure behavior	Backtracking layer	Threshold source
C7a	Total figure/table count ≤ limit	`list_assets`	Merge or move to supplementary	—	`assets.figures_max/tables_max`
C7b	Reasonable range for total citation count	`scan_draft_citations`	Flag low-citation refs → user decides	—	`references.reference_limits[paper.type]` fallback `max_references`
C7c	Total word count vs journal-profile	`count_words`	Trim overly long sections	→ A1	`word_limits.total_manuscript`
C7d	Figure/table cross-references (orphan/phantom)	`list_assets` + `read_draft`	orphan=WARNING, phantom=CRITICAL	—	—
C7e	Wikilink citation consistency	`validate_wikilinks`	`save_reference_mcp` Save	→ A4	—

orphan = manifest 中有但 draft 沒引用 → WARNING（有圖沒用）
phantom = draft 引用但 manifest 沒有 → CRITICAL（有引用沒圖）

NOTE: C6 (total word count) overlaps functionally with C7c. C6 performs a quick word count check, C7c performs a journal-profile-driven precise comparison. At implementation time, choose: C6 is retained as a Phase 6 quick pre-check (only looks at the total), C7c performs precise section-level comparison.

C8 Temporal Consistency Pass

Root cause: writing order (e.g., Methods → Results → Introduction) causes sections written earlier to potentially reference the state of sections “not yet written.” Once later sections are completed, earlier descriptions become outdated.

Trigger timing: Phase 6, after Hooks C1-C7

Check flow：

Scan in reverse writing order (the earliest-written section is checked last)
Search each section for the following patterns:
- "not yet written" / "not yet written"
- "Deferred" / "deferred"
- "will be" + section name (e.g., "will be discussed in Discussion")
- Hook status references (e.g., "B2: Deferred")
- Assertions about the content of other sections (e.g., "the Introduction contains...")
For each match, verify: does the referenced section already exist? Is the description still correct?
If incorrect → patch_draft update to the actual state

Implementation：

FOR section IN reverse(writing_order):
  content = read_draft(section)
  FOR pattern IN temporal_patterns:
    matches = scan(content, pattern)
    FOR match IN matches:
      referenced_section = extract_section(match)
      IF referenced_section EXISTS AND match.claim != actual_state:
        patch_draft(section, old=match, new=actual_state_description)
        log_to_audit("C8: Updated stale reference in {section}")

Failure behavior：patch_draft Update outdated descriptions. Maximum 2 rounds.

Hook C9: Supplementary Cross-Reference

Triggered after the full manuscript is complete. Verify bidirectional references between the main manuscript and supplementary materials:

Every "Supplementary Table/Figure N" reference in the main manuscript has a corresponding entry in the supplementary materials
Every item in the supplementary materials is referenced at least once by the main manuscript (orphan detection)
e-prefix references (eTable, eFigure) are also tracked

#	Check item	MCP Tool	Failure behavior	Threshold source
C9	Supplementary material cross-reference verification	`run_quality_checks(action="writing_hooks", hooks="C9")`	`patch_draft` Add citation / delete orphan citation	—

Hook D: meta-learning (Phase 10, core of the closed loop)

Hook D does not only improve the SKILL — it improves the Hook itself (CONSTITUTION §23).

Infrastructure（infrastructure/persistence/）：

Component	File	Purpose
`HookEffectivenessTracker`	`hook_effectiveness_tracker.py`	Record hook events, calculate trigger rate/correction rate/false-positive rate, produce recommendations
`QualityScorecard`	`quality_scorecard.py`	6-dimensional quality score (0-10), threshold checks, weakness detection
`MetaLearningEngine`	`meta_learning_engine.py`	D1-D8 orchestrator,`ThresholdAdjustment` (±20%)、`LessonLearned`

Usage：

from med_paper_assistant.infrastructure.persistence.hook_effectiveness_tracker import HookEffectivenessTracker
from med_paper_assistant.infrastructure.persistence.quality_scorecard import QualityScorecard
from med_paper_assistant.infrastructure.persistence.meta_learning_engine import MetaLearningEngine

tracker = HookEffectivenessTracker(audit_dir)
scorecard = QualityScorecard(audit_dir)
engine = MetaLearningEngine(audit_dir, tracker, scorecard)
result = engine.analyze()  # → {adjustments, lessons, suggestions, audit_trail, summary}

D1: Performance statistics

HookEffectivenessTracker Record each hook evaluation event (trigger/pass/fix/false_positive), persist to .audit/hook-effectiveness.yaml。

Performance judgment (§23):

Trigger rate > 80% → Hook is too strict, recommend relaxing the threshold
Trigger rate < 5% (over more than 5 runs) → Hook is too loose/outdated, consider removal
False-positive rate > 30% → Judgment criteria need revision

QualityScorecard Track quality scores across 7 standard dimensions, persist to .audit/quality-scorecard.yaml。

D2: Quality dimension analysis

MetaLearningEngine._d2_analyze_quality() Perform in-depth analysis of the 7-dimensional quality scores of the QualityScorecard:

Weakness detection: dimensions with score < 6.0 → produce quality_gap lesson
Missing item detection: unevaluated dimensions → produce process_gap lesson
Trend judgment: average score ≥ 8 → achievement, < 6 → critical review needed
Dimension → Hook mapping: methodology → B5, text_quality → A3, equator_compliance → E1-E5

D2's lessons feed into D3 (adjust thresholds) and D4-D5 (modify SKILL), forming an analysis chain.

D3: Hook self-improvement

MetaLearningEngine._d1_d3_analyze_hooks() Generate based on tracker recommendations ThresholdAdjustment：

Automatic adjustment（auto_apply=True, change magnitude ≤ ±20%):

Trigger rate too high → loosen threshold +15%
Trigger rate too low → tighten threshold -15%

Requires user confirmation（auto_apply=False): logic fixes, add/remove Hook

Prohibited modifications: CONSTITUTION principles, 🔒 rules, save_reference_mcp priority, Hook D's own logic

Hook propagation procedure (automatically executed after user confirmation)

When D3 proposes a new Hook and the user confirms, automatically sync all files according to the following spec:

Hook Spec format(D3 output):

hook_id: C7 # 類型字母 + 編號
category: C # A/B/C/D
name: 時間一致性 # 簡短中文名
description: 逆向掃描修正因寫作順序造成的過時引用
check_tool: "`read_draft` × N + Agent 掃描"
fix_action: "`patch_draft` 更新過時描述"
detailed_definition: |
  #### Hook C7: 時間一致性
  C1-C6 完成後，逆向掃描每個 section...

Propagation list(5 files, updated in order):

#	File	Update content	Mode
1	`.claude/skills/auto-paper/SKILL.md`	Hook table add row + detailed definition + Phase flow	Insert row
2	`AGENTS.md`	`### Hook 架構（N checks）` N+1 + table description row	Replace
3	`.github/copilot-instructions.md`	`## Hook 架構（N checks）` N+1 + table row `Copilot X1-M` → `X1-(M+1)`	Replace
4	`vscode-extension/copilot-instructions.md`	Same as above	Replace
5	`vscode-extension/skills/auto-paper/SKILL.md`	Same as #1 (VSX mirror)	Insert row

Automatic calculation：

new_count = grep all Hook 架構（(\d+) checks） get old value + 1
new_range = parse Copilot {cat}1-{M} → {cat}1-{M+1}
Use multi_replace_string_in_file to complete all replacements in one pass

Verification: after propagation completes grep -rn "Hook 架構" AGENTS.md .github/ vscode-extension/ confirm the numbers are consistent

D4-D5: SKILL + Instruction improvement

MetaLearningEngine._d4_d5_skill_suggestions() detection:

A certain Hook triggers >2 times and correction rate <50% → add pre-check
Weak quality dimensions → strengthen corresponding Hooks (methodology → B5, text_quality → A3)

D6: Record audit trail

MetaLearningEngine._d6_build_audit_trail() Append-write .audit/meta-learning-audit.yaml(append-only array), and by run_quality_checks(action="meta_learning") generate .audit/meta-learning-feedback.yaml, feeding pending evolutions back into the project as an agent-readable summary.

Other updates:.audit/hook-effectiveness.md, .audit/quality-scorecard.md, .memory/progress.md, .memory/activeContext.md, memory-bank/decisionLog.md, Lessons Learned in this file

D7: Review Retrospective 🆕

Analyze Phase 7 Review Loop output, evolve Reviewer instructions (CONSTITUTION §23 L2 level).

Trigger: Phase 10, after D1-D6

Flow：

1. 讀取 review-report-*.md + author-response-*.md
2. 統計分析：
   - 哪些 reviewer 角色的 MAJOR issues 最有價值（ACCEPT 率高）？
   - 哪些 suggestions 被 DECLINE 最多？→ 可能過度嚴格
   - 哪些 issues 跨 round 反覆出現？→ 修正策略不佳
   - 各 reviewer 角色的有效問題數量分布
3. 產出建議：
   - 調整 reviewer_perspectives 的檢查重點描述（±20%）
   - 記錄到 SKILL.md Lessons Learned
   - 更新 .audit/reviewer-effectiveness.json
4. 禁止：修改 CONSTITUTION 原則、修改 Hook D 自身邏輯

D8: EQUATOR Compliance Retrospective 🆕

Analyze Hook E's execution effect in Phase 7, continuously improving checklist accuracy.

Trigger: Phase 10, after D7

Flow：

1. 讀取 equator-compliance-*.md
2. 統計分析：
   - 哪些 checklist items 被標為 N/A 最多？→ 可能不適用該 paper type
   - 哪些 items 反覆 PARTIAL？→ 可能定義不清楚
   - compliance rate 趨勢（Round 0 → N）
   - ESSENTIAL items 的修補成功率
3. 產出建議：
   - 調整 E1 偵測邏輯的 paper_type 映射
   - 記錄到 SKILL.md Lessons Learned
   - 建議新的 ESSENTIAL 分類（若某非 ESSENTIAL 項反覆 miss）
4. 更新 evolution-log.jsonl with meta_learning event

Automatic decision logic

Scenario	Automatic behavior	Stop condition
Phase 0 no journal information	Use paper_type default value	Overridden after user provides it
Phase 0 has PDF/URL	Parse + auto-fill YAML	⚠️ Field needs confirmation
Insufficient search	Expand search	After 3 rounds still <10 papers
Concept 65-74	Auto-correct once	Still <75
Hook A word count over limit	Round 1-N stepwise correction	FLAG after N rounds
Hook A insufficient citations	suggest + patch, N rds	No available citations
Hook B 🔒 missing	patch added	Research direction needs change
Hook B5 <5 points	patch fills in details, 2 rounds	After 2 rounds still <5
Hook C CRITICAL	cascading fix, N rds	After N rounds ask user
Hook C WARNING	patch 1 round	LOG + continue
Hook C7 charts/tables over limit	Move to supplementary	User decides which to delete
Hook C7d phantom	Insert missing chart or remove citation	User decides
C11 citation distribution uneven	Add citations to sections lacking citations	LOG + continue
C12 citation decision record missing	Generate citation_decisions.json	LOG + continue
C13 chart/table quality issue	Fix ordering/add caption	LOG + continue
Hook B7 Brief omission	patch_draft fills omissions	LOG after 1 round
Phase 6 FLAG	Trace back Hook B → A	After 2 cascades ask user
Review MAJOR issue	patch/rewrite	quality ≥ threshold
Review score stagnant	Change strategy or ask user	No improvement for 2 consecutive rounds
Asset missing data	Prompt user to provide	Skip that asset
Asset tool unavailable	Fallback (see Sub-Pipeline)	LOG + alternative
Citations exceed limit	Mark low-citation refs	User decides which to delete
Hook D threshold fine-tuning	±20%	Out of range
Hook D add/remove	Propose suggestions	Always requires confirmation

Must stop: Concept < 60 (still low after two times), Phase 4 outline approve, research direction change, Phase 6 still CRITICAL after N rounds of cascading, Review no score improvement for 2 consecutive rounds, modify AGENTS.md core principles.

Auto-continue: Hook A/B WARNING → LOG → next step. Hook C WARNING → LOG → Phase 7. Review MINOR → batch fix → next round. Asset fallback succeeds → continue.

Cross-Tool Orchestration Map

Core principle: Pipeline defines "when" → Skill defines "how" → Hook defines "quality" → Review defines "completeness".

Phase × Tool Matrix

Phase	Internal MCP Tools	External MCP	journal-profile field
0 Pre-Planning	—	`fetch_webpage` 🔸	Generate all fields
1 Project Setup	`create_project`, `update_settings`	—	`paper.type`, `journal.*`
2 Literature Search	`save_reference_mcp`	`pubmed-search`, `zotero`	`references.max_references`
2.1 Fulltext Ingest	`list_saved_references`	`asset-aware-mcp` 🔸, `pubmed-search`	—
3 Concept Dev	`write_draft`, `validate_concept`	`cgu` 🔸	—
4 Planning	`read_draft`	—	`paper.sections`, `assets.*`
5 Writing	`draft_section`, `patch_draft`, etc.	`drawio` 🔸, `cgu` 🔸	`word_limits.`, `assets.`
6 Audit	`check_formatting`, `count_words`	—	All `pipeline.*` thresholds
7 Review	`read_draft`, `patch_draft`	`cgu` 🔸	`pipeline.autonomous_review.*`
8 Ref Sync	`sync_references`, `format_references`	—	`references.*`
9 Export	`save_document`, `verify_document`	—	`required_documents.*`
10 Retrospective	—	—	Analyze reasonableness of all fields

Cross-MCP passing rules

Source	Target	Passed item	Rule
pubmed-search	mdpaper	PMID	`save_reference_mcp(pmid)` passes only PMID
pubmed-search	asset-aware	Full text / PMC ID	`get_fulltext(pmcid)` → PDF/text
asset-aware	mdpaper	doc_id + paragraph	Update metadata.json after parsing
zotero-keeper	mdpaper	PMID/DOI	Fetch PMID → `save_reference_mcp()`
cgu	concept.md	Text suggestions	Agent integrates into `write_draft()`
drawio	mdpaper	XML	`save_diagram(project, content)`
data tools	drafts	Tables/figures	Agent integrates into draft text

When to skip external MCP

Scenario	Skip
No Zotero	zotero-keeper
Concept ≥ 75 first pass	CGU
No dataset	table_one / plot
Pure review (no Methods flow)	drawio
Full text unavailable (non-OA)	asset-aware (but must record ⚠️ in metadata)

Skill dependencies

auto-paper → Phase 0(pre-plan) → project-management(P1) → literature-review + parallel-search(P2) → fulltext-ingestion(P2.1) → concept-development(P3) → draft-writing(P4,5) → evolution-gate(P6.5) → autonomous-review+equator(P7) → reference-management(P8) → word-export(P9) → submission-preparation(P9)

Closed-loop checklist

Phase 0: source-materials.yaml generated; all pending_asset_aware files handed off to asset-aware or have explicit fallback reason
Phase 0: journal-profile.yaml generated + user confirmed
Phase 2.1: fulltext-ingestion-status.md created, each reference marked with full-text status + analysis_completed
All sections pass Hook A (cascading)
All sections pass Hook B (including retrospective fixes)
All Phase 5 FLAGs handled in Phase 6
Full manuscript passes Hook C (cascading fix)
Phase 6.5: Evolution Gate baseline snapshot created
Phase 7: Autonomous Review reaches quality_threshold (MANDATORY, at least 2 rounds, unless pipeline config explicitly adjusted)
Hook E: EQUATOR compliance rate ≥ 80% + all ESSENTIAL items reported
quality-scorecard.md generated (all dimensions ≥ 6 points, including EQUATOR dimension)
review-round-*.md generated (complete record per round)
equator-compliance-*.md generated (checklist report per round)
evolution-log.jsonl contains baseline + all round events
hook-effectiveness.md generated
.audit/pipeline-run-{YYYYMMDD-HHmm}.md covers all Phases and includes ## D7 retrospective: + ## D8 retrospective:
meta-learning-feedback.yaml generated, pending evolutions fed back into workspace state
checkpoint.json marked complete
All Asset Plan items generated or with fallback record
Reference count ≤ journal-profile.references.max_references
Figure/table count ≤ journal-profile.assets limits
Required documents list (required_documents) completed
.memory/ updated
Hook D meta-learning executed (including D7 Review + D8 EQUATOR Retrospective)
SKILL.md Lessons Learned updated
Word exported

Lessons Learned (Hook D auto-update section)

Run 2026-02-20: Self-Referential Paper (Software/Methods)

Pipeline: 9 phases, fully autonomous, 0 human interventions

Proactive Hook Effect: All 6 sections passed Hook A on first write (0 corrections needed). The hook criteria documented in this SKILL.md act as proactive constraints — the LLM avoids prohibited patterns because it knows they will be checked. This means hooks serve dual purpose: reactive verification AND proactive generation shaping.
Hook A4 False Positive: Example text [[author_year_pmid]] in the System Architecture section was flagged as invalid wikilink. Action: Detection criteria should exclude placeholder/example patterns. Recommended regex exclusion: patterns matching author_year_pmid or similar template-style tokens. FP rate 100% (1/1).
CGU Engine Dependency: CGU in "simple" mode returns empty results for deep_think and spark_collision. Pipeline must handle graceful degradation — log the limitation and proceed. Full meta-learning assessment requires CGU engine upgrade.
Template Path Bug: start_document_session resolves template path via __file__ traversal to src/templates/ but workspace templates live in root templates/. Workaround: direct python-docx export. Fix needed in template_reader.py initialization.
Concept Gate Override: Score 70 (< 75 threshold) was overridden after 2 correction attempts + CGU failure. The 60 hard-stop threshold proved appropriate — the paper was completed successfully at quality score 91.4%. Consider formally documenting the override decision tree: score ≥ 75 → auto-proceed; 60-74 + consistency 100 + user authority → override with audit; < 60 → hard stop.
Non-PubMed References: CS/AI system papers frequently cite arXiv preprints lacking PMIDs. The tiered trust architecture only covers PubMed-indexed literature. Future: extend save_reference_mcp to accept DOIs with CrossRef verification as a secondary verified channel.
Writing Order Validated: Methods → Results → Introduction → Discussion → Abstract order for Software/Methods papers produced coherent flow. The Introduction was contextually richer because System Architecture, Methods, and Results were already written.
Self-Referential Circularity: The system writing about itself creates a bootstrapping challenge — Results data (hook statistics) are generated during writing, but the Results section must describe them. Resolution: write Results with partial data, then verify final numbers match in Hook C. Acceptable for n=1 demonstration.
Temporal Inconsistency from Writing Order: When Results is written before Introduction, deferred hook statuses (e.g., "B2: Deferred — Introduction not yet written") become stale after Introduction is completed. Action: Added Hook C7 (Temporal Consistency Pass) to systematically scan for and correct such stale references in Phase 6. Root cause: writing order creates forward-references that need backward-patching.

About this skill

Closed-loop architecture

Execution audit trail (Audit Trail)

Dual Hook system

🚧 Hard Gate Enforcement (Code-Level, cannot be skipped)

Required MCP Tool call

Enforce rules

🛡️ Anti-Compaction Recovery Protocol

11-Phase Pipeline (Phase 0-11, including Phase 6.5)

Phase 0: PRE-PLANNING（Source Materials + Journal Profile + Pipeline Config）🆕

journal-profile.yaml constraints on the entire Pipeline

source-materials.yaml constraints on the entire Pipeline

Autopilot mode (enabled by default)

Phase 1: PROJECT SETUP

Phase 2: LITERATURE SEARCH

Phase 2.1: FULLTEXT INGESTION & ASSET EXTRACTION

Phase 2.5: CRITICAL APPRAISAL

Phase 3: CONCEPT DEVELOPMENT

Phase 4: MANUSCRIPT PLANNING

Phase 4.5: SUPPLEMENTARY PLANNING

Phase 4.9: SUB-SECTION WRITING TEMPLATES (paragraph-level writing guidance)

Introduction paragraph template (Funnel Structure)

Methods paragraph template (Reproducibility Structure)

Results paragraph template (Data-First Structure)

Discussion paragraph template (Interpretation Structure)

Abstract paragraph template (Structured Abstract)

Figure/table insertion specifications (integrated with C13 Hook)

Citation Decision Record specification (integrated with C12 Hook)

Phase 5: SECTION WRITING (core + Cascading Audit Loop)

Asset Generation Sub-Pipeline（Phase 5 Step 2）

Agent-Initiated Asset Generation (autonomously add figures/tables during writing)

Phase 6: CROSS-SECTION CASCADING AUDIT

Cascading retrospective rules

Phase 6.5: EVOLUTION GATE (mandatory entry into Review) 🆕

Phase 7: AUTONOMOUS REVIEW (Structured Review Loop — MANDATORY) 🆕

Multi-Reviewer design (simulating 2-3 independent reviewers) 🆕

Hook E: EQUATOR Reporting Guidelines Compliance (Phase 7 Stage A addition) 🆕

E1: Automatically detect applicable guidelines

E2: Checklist item-by-item verification

E3: EQUATOR guideline built-in database

E4: Compliance Report output

E5: Integration with Phase 7 Review Loop

Review quality dimensions (quality-scorecard)

Division of labor between Review and Hook

Phase 8: REFERENCE SYNC

Phase 9: EXPORT（Code-Enforced — CRITICAL Gate）

Phase 9.5: SUBMISSION READINESS

Phase 10: RETROSPECTIVE (closing-loop core)

Phase 11: FINAL DELIVERY（Code-Enforced — CRITICAL Gate）🆕

Audit Trail generation (automatically executed when each Phase completes)

Incremental updates (at the end of each Phase)

Generated after Phase 6 completes

Generated for each round of Phase 7

Evolution Tracing（Phase 6.5 → 7 → 10）🆕

Generated after Phase 2 completes

Generated after Phase 3 completes

Checkpoint Recovery Mechanism

Copilot Hooks definition (triggered during writing)

Hook A: post-write (immediately after each write, at most N rounds, N = pipeline.hook_a_max_rounds）

Hook A Cascading Protocol

Hook B: post-section (after the section is completed, including retrospective revision)

Hook B Cascading Protocol

B5 Methodology Checklist

B6 Prerequisites

B7 Section Brief compliance 🆕

Hook B8: Data-Claim Alignment

Hook C: post-manuscript (after full manuscript completion, including layered backtracking, at most N rounds, N = pipeline.hook_c_max_rounds）

Hook C Cascading Protocol

C7 Count and cross-reference compliance (D5 extension) 🆕

C8 Temporal Consistency Pass

Hook C9: Supplementary Cross-Reference

Hook D: meta-learning (Phase 10, core of the closed loop)

D1: Performance statistics

D2: Quality dimension analysis

D3: Hook self-improvement

Hook propagation procedure (automatically executed after user confirmation)

D4-D5: SKILL + Instruction improvement

D6: Record audit trail

D7: Review Retrospective 🆕

Hook A: post-write (immediately after each write, at most N rounds, N = `pipeline.hook_a_max_rounds`）

Hook C: post-manuscript (after full manuscript completion, including layered backtracking, at most N rounds, N = `pipeline.hook_c_max_rounds`）