auto-academic

name	auto-academic
description	Fully autonomous academic research pipeline. Takes a research topic and target venue, then runs without stopping: deep research -> computational experiments (autoresearch-style keep/discard loop) -> paper writing -> integrity verification -> peer review -> iterative revision with tiered escalation -> final integrity check -> finalize. 8-stage unified pipeline: INIT, RESEARCH, EXPERIMENT, WRITE, INTEGRITY, REVIEW, FINAL_INTEGRITY, FINALIZE. Dual-loop architecture with mandatory integrity gates. Triggers on: auto-academic, autonomous research, auto research pipeline, full autonomous paper, research to accepted paper.
metadata	{"version":"2.0.0","last_updated":"2026-03-27","depends_on":"auto-academic"}

Auto-Academic v2.0 — Fully Autonomous Academic Research Pipeline

A fully autonomous orchestrator that takes a research topic and target venue, then runs the complete pipeline — research, experiments, paper writing, integrity verification, peer review, revision — without stopping until the paper is accepted or safety caps are hit.

Core philosophy: Adapted from autoresearch's "NEVER STOP" principle. Once started, the pipeline runs autonomously. Do not pause to ask the user if you should continue. The user expects you to work indefinitely until interrupted.

Architecture: Dual-loop with spine orchestrator and mandatory integrity gates.

• Inner loop: deep-research + autoresearch-style experiments (keep/discard/refine)
• Outer loop: peer review + revision with tiered escalation
• Integrity gates: Pre-review verification (Stage 4) and post-revision verification (Stage 6)
• Spine: INIT -> RESEARCH -> EXPERIMENT -> WRITE -> INTEGRITY -> REVIEW -> FINAL_INTEGRITY -> FINALIZE

Trigger Conditions

English: auto-academic, autonomous research, auto research pipeline, full autonomous paper, research to accepted paper, run auto-academic

Dependencies

All required skills are bundled within this plugin. No external plugin dependencies.

Invocation

The user provides:

1. Topic (required): the research subject
2. Venue (required): target publication venue

Examples:

/auto-academic "Impact of LLMs on undergraduate writing skills" --venue "Computers & Education"
Run auto-academic on the effects of AI tutoring in STEM education, targeting the Internet and Higher Education

---

Adaptive Checkpoint System

Core rule: After each major stage, present a checkpoint. The pipeline is autonomous, but checkpoints ensure state is recorded and the process is traceable. In fully autonomous mode, FULL/SLIM checkpoints are logged to process-log.md and the pipeline proceeds without waiting. MANDATORY checkpoints log their status but also proceed autonomously. The checkpoint types exist for transparency and state tracking.

Checkpoint Types

Type	When Used	Content
FULL	First checkpoint; after non-critical stages	Full deliverables list + metrics logged to process-log
SLIM	After 2+ consecutive routine stage completions	One-line status logged to process-log
MANDATORY	Integrity gates (Stage 4, Stage 6); Review decisions	Status logged; pipeline proceeds but records PASS/FAIL clearly

Decision Dashboard (logged at FULL checkpoints)

━━━ Stage [X] [Name] Complete ━━━

Metrics:
- Word count: [N] (target: [T] +/-10%)    [OK/OVER/UNDER]
- References: [N] (min: [M])              [OK/LOW]
- Coverage: [N]/[T] sections drafted       [COMPLETE/PARTIAL]

Deliverables:
- [Material 1]
- [Material 2]

Flagged: [any issues detected, or "None"]

Proceeding to Stage [Y]...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

---

Stage 0: INIT

Goal: Parse input, set up workspace, initialize state. Checkpoint: MANDATORY

Step 1: Parse Input

Extract topic and venue from the user's message. If either is missing, infer from context or use defaults:

• Missing venue: default to APA 7.0 format, Tier 3 rigor (SSCI standard)
• Missing topic: this is an error — the pipeline cannot start without a topic

Step 2: Venue Profile Lookup

Read ${CLAUDE_PLUGIN_ROOT}/references/venue-profiles.md. Match the user's venue to a profile category. If no match, use WebSearch to research the venue, then assign the closest category.

Record the venue profile settings:

• Citation style
• Paper format
• Page limit
• Review rigor tier
• Experiment expectations

Step 3: Material Passport Check

Before creating a new workspace, check for an interrupted run that can be resumed:

ls auto-academic-run-* 2>/dev/null

If an existing workspace is found:

1. Read its pipeline-state.json to determine current_stage and stage statuses
2. If current_stage is not "INIT" and not "FINALIZE": log "Found interrupted run for '{topic}'. Resuming from {current_stage}." and jump to that stage
3. If current_stage is "FINALIZE" and status is "completed": log "Previous run already complete. Starting fresh run."
4. Record the entry point in material_passport.entry_point

If no existing workspace: proceed with fresh initialization.

Step 4: Create Workspace

TIMESTAMP=$(date +%Y%m%d-%H%M%S)
WORKSPACE="auto-academic-run-${TIMESTAMP}"
mkdir -p "${WORKSPACE}"/{research,experiments/{specs,code,results},paper/latex,reviews,integrity}

Step 5: Initialize State

Copy ${CLAUDE_PLUGIN_ROOT}/templates/pipeline-state.json to ${WORKSPACE}/pipeline-state.json. Fill in:

• topic: the parsed topic
• venue: the parsed venue
• created_at: current timestamp
• current_stage: "INIT"
• material_passport.entry_point: "INIT"
• material_passport.provenance: []

Step 6: Initialize Logging

Copy ${CLAUDE_PLUGIN_ROOT}/templates/process-log.md content to ${WORKSPACE}/process-log.md. Fill in topic, venue, timestamp. Copy ${CLAUDE_PLUGIN_ROOT}/templates/results-tsv-header.tsv to ${WORKSPACE}/experiments/results.tsv.

Step 7: Initialize Git

cd "${WORKSPACE}"
git init
git add -A
git commit -m "chore: initialize auto-academic workspace for '${TOPIC}'"

Step 8: Log and Transition

Log to process-log: [{timestamp}] INIT: Topic="{topic}", Venue="{venue}", Workspace="{workspace}" Update pipeline-state: current_stage -> "RESEARCH", INIT status -> "completed" Proceed immediately to Stage 1.

---

Stage 1: RESEARCH

Goal: Produce research deliverables and extract testable hypotheses. Checkpoint: MANDATORY

Step 1: Invoke deep-research

Trigger the auto-academic:deep-research skill in full mode with the topic. Provide:

• The parsed topic as the research question seed
• The venue profile for calibrating research depth

Wait for deep-research to complete. It produces:

• RQ Brief -> save to research/rq-brief.md
• Methodology Blueprint -> save to research/methodology.md
• Annotated Bibliography -> save to research/bibliography.md
• Synthesis Report -> save to research/synthesis.md

Step 2: Extract Hypotheses

Read the synthesis report. Identify all claims that:

1. Are testable through computational experiments
2. Have measurable outcomes
3. Are central to the paper's argument

For each claim, formulate a hypothesis:

• State it in testable form (e.g., "Students using AI tutoring will show higher test scores than control group")
• Identify the independent and dependent variables
• Specify the expected direction of effect

Save to research/hypotheses.md using this format:

## Hypotheses

### H1: {statement}
- **Variables:** IV = {X}, DV = {Y}
- **Expected direction:** {positive / negative / nonlinear}
- **Derived from:** Synthesis Report, Section {N}
- **Evidence base:** {key citations supporting this hypothesis}

### H2: {statement}
...

Step 3: Style Calibration (if author samples available)

If the user provided sample papers or writing style references at invocation time:

• Read the samples to calibrate tone, vocabulary level, and structural preferences
• Record style notes to research/style-calibration.md
• These notes will be passed to Stage 3 (WRITE) for style-consistent drafting

Step 4: Log and Transition

Log to process-log: [{timestamp}] RESEARCH: Complete. {N} hypotheses extracted. Update pipeline-state: RESEARCH -> "completed", EXPERIMENT -> "in_progress" Git commit: git add research/ && git commit -m "feat: complete research phase, {N} hypotheses extracted"

Log MANDATORY checkpoint to process-log:

[CHECKPOINT-MANDATORY] RESEARCH complete: {N} hypotheses, {M} references. Proceeding to EXPERIMENT.

Proceed immediately to Stage 2.

---

Stage 2: EXPERIMENT (Inner Loop)

Goal: Validate all hypotheses through computational experiments. Runs as an autonomous keep/discard loop until all hypotheses have a verdict. Checkpoint: FULL on convergence

Step 1: Plan Experiments

Invoke the experiment-planner agent with:

• research/hypotheses.md
• research/methodology.md
• Venue profile experiment expectations
• ${CLAUDE_PLUGIN_ROOT}/references/experiment-patterns.md

The agent produces one experiment spec per hypothesis, saved to experiments/specs/.

Step 2: Implement Experiments

For each experiment spec, invoke the auto-academic:writing-plans skill to create an implementation plan, then invoke auto-academic:executing-plans to implement the code.

If multiple experiments are independent (no shared data dependencies), use auto-academic:dispatching-parallel-agents to implement them concurrently.

Each experiment's code is saved to experiments/code/{hypothesis_id}/.

Step 3: Execute Experiments (The Loop)

Run each experiment. For each run:

1. Execute the experiment code, redirecting all output:

   cd experiments/code/{hypothesis_id}
   python main.py > ../../results/run-{NNN}/output.log 2>&1
   echo $? > ../../results/run-{NNN}/exit_code.txt
   

2. Check the exit code:

   • Exit 0: proceed to evaluation
   • Non-zero: mark as CRASH

3. Read the output and parse metrics.

4. Log to experiments/results.tsv:

   {run_number}\t{hypothesis_id}\t{status}\t{metrics}\t{description}
   

5. Git commit:

   git add experiments/
   git commit -m "experiment: run {NNN} for {hypothesis_id} — {status}"
   

Step 4: Handle Crashes

If an experiment crashes:

1. Read the last 50 lines of the log to diagnose the error
2. If it's a simple fix (typo, import error, wrong path): fix and retry
3. If it's a fundamental issue (OOM, incompatible library, impossible computation): skip, log "crash" status
4. Maximum 3 retry attempts per experiment

Step 5: Evaluate Results

After each successful run, invoke the experiment-evaluator agent with:

• The experiment spec
• The experiment output
• The run log

The agent updates experiments/claim-evaluation.md.

Step 6: Check Convergence

After evaluating results, invoke the convergence-judge agent for the inner loop. The agent should consult ${CLAUDE_PLUGIN_ROOT}/references/convergence-criteria.md for detailed convergence rules.

• If CONVERGE: exit the inner loop, proceed to WRITE
• If CONTINUE: identify which hypothesis needs re-running, refine the experiment spec, go back to Step 3
• If FORCE_CONVERGE: log warning, proceed to WRITE with best results

Step 7: Log and Transition

Log to process-log: [{timestamp}] EXPERIMENT: Inner loop converged after {N} runs. Verdicts: {summary} Update pipeline-state: EXPERIMENT -> "completed", WRITE -> "in_progress" Git commit: git add experiments/ && git commit -m "feat: complete experiment phase, all hypotheses resolved"

Log FULL checkpoint to process-log:

[CHECKPOINT-FULL] EXPERIMENT complete: {N} runs, {M} hypotheses resolved. Proceeding to WRITE.

Proceed immediately to Stage 3.

---

Stage 3: WRITE

Goal: Produce a complete paper draft incorporating research and experiment results. Checkpoint: ADAPTIVE (FULL first time, SLIM on subsequent re-entries from escalation)

Step 1: Prepare Writing Materials

Gather all materials for the paper-writing skill:

• research/rq-brief.md (research question and scope)
• research/methodology.md (methodology blueprint)
• research/bibliography.md (annotated bibliography)
• research/synthesis.md (synthesis report)
• research/style-calibration.md (if exists — style preferences)
• experiments/claim-evaluation.md (hypothesis-evidence mapping)
• experiments/results/ (raw experiment outputs for tables and figures)
• Venue profile (citation style, format, page limit)

Step 2: Invoke academic-paper

Trigger the auto-academic:academic-paper skill in full mode. Provide all gathered materials and instruct it to:

• Use the venue's citation style and format
• Include experiment methodology in the Methods section
• Include experiment results with tables/figures in the Results section
• Ground the Discussion in the actual experiment outcomes (not hypothetical)
• Reference the claim-evaluation for each argument in the Discussion
• Apply any style calibration notes from research/style-calibration.md

Save the output to paper/draft-v1.md.

Step 3: Log and Transition

Log to process-log: [{timestamp}] WRITE: Draft v{N} complete. Word count: {N} Update pipeline-state: WRITE -> "completed", draft_version -> 1, INTEGRITY -> "in_progress" Git commit: git add paper/ && git commit -m "feat: complete paper draft v1"

Determine checkpoint type:

• If this is the first time reaching WRITE: log FULL checkpoint
• If re-entering WRITE from escalation: log SLIM checkpoint

Proceed immediately to Stage 4.

---

Stage 4: INTEGRITY

Goal: Verify academic integrity of the paper draft before peer review. Must pass 100%. Checkpoint: MANDATORY

Step 1: Invoke integrity-verification

Invoke the integrity-verification agent with the current paper draft (paper/draft-v{N}.md). The agent performs a 5-phase check:

Phase A: Reference Existence — 100% of all cited references must be verified to exist
  - Verify DOI/URL for each reference
  - Check author names, publication year, journal/venue
  - Flag ghost citations (fabricated references)

Phase B: Citation Context — >= 30% citation context spot-check
  - Verify cited content actually says what the paper claims
  - Check for citation distortion or out-of-context use

Phase C: Statistical Data — 100% statistical data verification
  - Every numerical claim must match the actual experiment output in experiments/results/
  - Cross-reference tables and figures against results.tsv

Phase D: Originality — >= 30% originality spot-check + self-plagiarism check
  - Check for text reuse from source materials
  - Verify AI-generated text does not duplicate verbatim sources

Phase E: Claim Verification — 30% spot-check (minimum 10 claims)
  - Verify factual claims are supported by cited evidence
  - Flag MAJOR_DISTORTION and UNVERIFIABLE claims

Save the verification report to integrity/pre-review-check.md.

Step 2: Handle Results

PASS (zero SERIOUS + zero MEDIUM + zero MAJOR_DISTORTION + zero UNVERIFIABLE):

• Log: [{timestamp}] INTEGRITY: PASS. {N} references verified, {M} citations checked.
• Update pipeline-state: INTEGRITY.verification_pass -> true, INTEGRITY.phases_passed -> ["A","B","C","D","E"]
• Log MANDATORY checkpoint: [CHECKPOINT-MANDATORY] INTEGRITY PASS. Proceeding to REVIEW.
• Proceed to Stage 5

FAIL (any issues found):

• Produce correction list from the verification report
• Fix each issue in paper/draft-v{N}.md:
  • Remove or replace ghost citations
  • Correct bibliographic inaccuracies
  • Fix data mismatches against experiment outputs
  • Revise distorted claims
• Increment INTEGRITY.retry_count
• Re-run phases A and C in full; spot-check B, D, E for fixed items
• If PASS after corrections: proceed as above
• If still FAIL after 3 retries: log warning with list of unresolvable issues, proceed to Stage 5 with disclaimer

Step 3: Log and Transition

Log to process-log: [{timestamp}] INTEGRITY: {PASS|FAIL after N retries}. Issues: {summary} Update pipeline-state: INTEGRITY -> "completed", REVIEW -> "in_progress" Git commit: git add paper/ integrity/ && git commit -m "feat: integrity verification {PASS|FAIL} — {issues summary}" Proceed immediately to Stage 5.

---

Stage 5: REVIEW (Outer Loop)

Goal: Iterate through peer review and revision until the paper is accepted or safety caps are hit. Checkpoint: ADAPTIVE (FULL first time, SLIM on subsequent rounds after routine Minor decisions)

Step 1: Peer Review

Invoke the auto-academic:academic-paper-reviewer skill. The review mode depends on the current round:

Round	Mode
1	full
2	methodology-focus (if round 1 was Major/Reject) or full (if Minor)
3+	As specified by escalation protocol

Save all reviewer reports to reviews/round-{N}/.

Determine checkpoint type:

• Round 1: FULL checkpoint
• Round 2+ after routine continuation: SLIM checkpoint
• Any round with Major/Reject/Escalate decision: MANDATORY checkpoint

Step 2: Auto-Decision

Read the editorial decision from the review output. Invoke the convergence-judge agent for the outer loop. The agent should consult ${CLAUDE_PLUGIN_ROOT}/references/convergence-criteria.md for convergence rules and ${CLAUDE_PLUGIN_ROOT}/references/escalation-protocol.md for escalation decision logic.

Based on the decision:

CONVERGE (Accept):

• Log: [{timestamp}] REVIEW Round {N}: Accept. Paper accepted!
• Log MANDATORY checkpoint: [CHECKPOINT-MANDATORY] REVIEW Accept at round {N}. Proceeding to FINAL_INTEGRITY.
• Proceed to Stage 6 (FINAL_INTEGRITY)

REVISE_AND_CONVERGE (Minor, non-structural):

• Invoke auto-academic:academic-paper in revision mode with the review comments
• Save revised draft to paper/draft-v{N+1}.md
• Save response to reviewers to reviews/response-to-reviewers/round-{N}-response.md
• Run a re-review to confirm
• If re-review returns Accept or Minor with no critical issues: proceed to Stage 6 (FINAL_INTEGRITY)
• Otherwise: continue the loop

CONTINUE (Minor or Major, needs revision):

• Invoke auto-academic:academic-paper in revision mode with the review comments
• Save revised draft to paper/draft-v{N+1}.md
• Save response to reviewers to reviews/response-to-reviewers/round-{N}-response.md
• Git commit: git add paper/ reviews/ && git commit -m "feat: revision round {N} — addressing {decision} revision comments"
• Increment round counter
• Update REVIEW.checkpoint_mode -> "SLIM" if this was a routine Minor revision
• Re-run Stage 4 (INTEGRITY) on the revised draft before the next review round

ESCALATE (Round 3 + Major/Reject):

• Log: [{timestamp}] REVIEW Round {N}: {decision}. ESCALATING to inner loop.
• Log MANDATORY checkpoint: [CHECKPOINT-MANDATORY] ESCALATION triggered at round {N}.
• Invoke the escalation-analyst agent
• The agent produces a concern map and revised experiment specs
• Re-enter Stage 2 (EXPERIMENT) at Step 3 (Execute) with only the revised specs
• After inner loop re-converges, re-write only affected paper sections (not full rewrite) — re-enter Stage 3 (WRITE) in SLIM mode
• Return to Stage 4 (INTEGRITY) -> Stage 5 (REVIEW) to continue the outer loop

FORCE_CONVERGE (Round 5, safety cap):

• Log: [{timestamp}] REVIEW Round {N}: Safety cap reached. Force-finalizing.
• Log MANDATORY checkpoint: [CHECKPOINT-MANDATORY] Safety cap at round {N}. Force-proceeding to FINAL_INTEGRITY.
• Generate quality report documenting unresolved concerns
• Proceed to Stage 6 (FINAL_INTEGRITY) with quality disclaimer

Step 3: Log Each Round

After each review round, log:

[{timestamp}] REVIEW Round {N}: Decision={decision}, Action={action}

Update pipeline-state with the decision history. Git commit after each round.

---

Stage 6: FINAL_INTEGRITY

Goal: Post-revision final integrity verification before finalization. Must pass 100%. Checkpoint: MANDATORY

Step 1: Invoke integrity-verification (Final Mode)

Invoke the integrity-verification agent on the accepted/final draft (paper/draft-v{latest}.md). This is a stricter check than Stage 4:

Phase A: 100% reference verification (including references added during revision)
  - Full check, no sampling

Phase B: 100% citation context verification (not spot-check — full check)
  - Every citation must be verified for accurate representation

Phase C: 100% statistical data verification
  - All numbers, tables, figures verified against experiment outputs

Phase D: >= 50% originality spot-check (100% for newly added/modified paragraphs)
  - Higher bar than pre-review check

Phase E: 100% claim verification
  - Zero MAJOR_DISTORTION + zero UNVERIFIABLE required
  - Compare against Stage 4 results to confirm all previous issues are resolved

Save the verification report to integrity/final-check.md.

Step 2: Handle Results

PASS (zero issues):

• Log: [{timestamp}] FINAL_INTEGRITY: PASS. {N} references verified (full), {M} citations verified (full).
• Update pipeline-state: FINAL_INTEGRITY.verification_pass -> true
• Log MANDATORY checkpoint: [CHECKPOINT-MANDATORY] FINAL_INTEGRITY PASS. Proceeding to FINALIZE.
• Proceed to Stage 7

FAIL (any issues found):

• Fix issues in the paper draft
• Re-verify (max 3 retries)
• Must achieve PASS to proceed — this gate is non-negotiable
• If still FAIL after 3 retries: log warning, list unresolvable items, proceed to Stage 7 with disclaimer

Step 3: Log and Transition

Log to process-log: [{timestamp}] FINAL_INTEGRITY: {PASS|FAIL after N retries}. Issues: {summary} Update pipeline-state: FINAL_INTEGRITY -> "completed", FINALIZE -> "in_progress" Git commit: git add paper/ integrity/ && git commit -m "feat: final integrity verification {PASS|FAIL}" Proceed immediately to Stage 7.

---

Stage 7: FINALIZE

Goal: Produce the final paper in all output formats, generate bilingual process summary. Checkpoint: MANDATORY

Step 1: Format Conversion

Invoke auto-academic:academic-paper in format-convert mode:

• Ask user (or infer from venue profile) which formatting style: APA 7.0 / Chicago / IEEE / etc.
• Generate LaTeX output to paper/latex/paper.tex (using corresponding document class, e.g., apa7 for APA 7.0)
• Compile to PDF via tectonic: paper/latex/paper.pdf
  • Fonts: Times New Roman (English) + Source Han Serif TC VF (Chinese) + Courier New (monospace)
  • PDF must be compiled from LaTeX (HTML-to-PDF is prohibited)
• If the venue requires DOCX: also generate paper/paper.docx
• Keep the Markdown version as paper/draft-final.md

Step 2: Process Summary Generation (Bilingual)

Review the session history and generate a comprehensive process record documenting the human-AI collaboration:

1. Compile the following:
   - Topic and initial instructions
   - Key decision points at each stage
   - Iteration counts and experiment summaries
   - Review rounds and decisions (with revision counts)
   - Integrity check results (pre-review + final)
   - Pipeline statistics (total stages, escalations, integrity retries)

2. Generate Markdown:
   - English: paper_creation_process_en.md
   - (Bilingual if Chinese was used in the session)

3. Convert to LaTeX and compile PDF:
   - pandoc MD -> LaTeX body
   - Package complete LaTeX document (with cover page, table of contents, headers/footers)
   - tectonic compile PDF
   - Chinese version: xeCJK + Source Han Serif TC VF

Required sections in process record:

Section	Content
Paper Information	Title, venue, final deliverables list
Stage-by-Stage Process	Input/output/key decisions for each stage
Experiment Summary	Hypotheses tested, runs executed, verdicts
Integrity Summary	Pre-review + final check results, issues found and resolved
Review History	Round-by-round decisions, escalations, revision summaries
Pipeline Statistics	Total duration, stage durations, iteration counts
Collaboration Quality Evaluation	1-100 score + dimensional analysis (see below)

Collaboration Quality Evaluation (mandatory final section):

+--------------------------------------------------+
|  Collaboration Quality Score: [XX]/100            |
+--------------------------------------------------+
|  Direction Setting          [score]               |
|  Intellectual Contribution  [score]               |
|  Quality Gatekeeping        [score]               |
|  Iteration Discipline       [score]               |
|  Delegation Efficiency      [score]               |
|  Meta-Learning              [score]               |
+--------------------------------------------------+

Scoring criteria: 90-100 = Exceptional; 75-89 = Excellent; 60-74 = Good; 40-59 = Basic; 1-39 = Needs Improvement. Be honest — if the user only pressed "continue," reflect that truthfully.

Save to paper_creation_process_en.md (and paper_creation_process_zh.md if applicable). Compile to PDF: paper_creation_process_en.pdf.

Step 3: Final process-log Summary Section

Append to process-log.md:

## Summary

- **Total duration:** {start_time} to {end_time}
- **Research phase:** {duration}
- **Experiments:** {N} runs, {M} hypotheses, {K} validated
- **Pre-review integrity:** {PASS/FAIL}, {N} issues found, {M} fixed
- **Paper drafts:** {N} versions
- **Review rounds:** {N}, final decision: {decision}
- **Escalations:** {N}
- **Final integrity:** {PASS/FAIL}, {N} issues found, {M} fixed
- **Final word count:** {N}
- **Final reference count:** {N}

## Final Deliverables

1. Paper (Markdown): `paper/draft-final.md`
2. Paper (PDF): `paper/latex/paper.pdf`
3. Paper (DOCX): `paper/paper.docx` (if generated)
4. Experiment code: `experiments/code/`
5. Experiment results: `experiments/results.tsv`
6. Review history: `reviews/`
7. Integrity reports: `integrity/pre-review-check.md`, `integrity/final-check.md`
8. Process record: `paper_creation_process_en.pdf`
9. Process log: `process-log.md`

Step 4: Final Commit

git add -A
git commit -m "feat: finalize auto-academic pipeline — paper complete"

Log MANDATORY checkpoint: [CHECKPOINT-MANDATORY] FINALIZE complete. Pipeline ended successfully.

Step 5: Report to User

Output a concise completion report:

=== Auto-Academic Pipeline Complete ===

Topic: {topic}
Venue: {venue}
Final Decision: {Accept / Force-finalized at round N}

Integrity Gates:
- Pre-review: {PASS/FAIL} ({N} issues found, {M} fixed)
- Final: {PASS/FAIL} ({N} issues found, {M} fixed)

Deliverables:
- Paper: {workspace}/paper/latex/paper.pdf
- Code: {workspace}/experiments/code/
- Results: {workspace}/experiments/results.tsv
- Process Record: {workspace}/paper_creation_process_en.pdf
- Full log: {workspace}/process-log.md

{If force-finalized: "Note: Quality report attached — see integrity/final-check.md for unresolved concerns."}

---

Safety Rules

NEVER STOP

Once the pipeline starts (after INIT), do NOT:

• Ask the user if you should continue
• Pause to check if the user is still watching
• Say "should I proceed to the next stage?"
• Wait for confirmation between stages

The pipeline runs autonomously until completion or user interruption. The user may be away from the computer.

Safety Caps (Non-Negotiable)

Cap	Value	Action on Hit
Inner loop experiments	10 runs	Force-converge, proceed with best results
Outer loop review rounds	5 rounds	Force-finalize with quality report
Escalation re-entries	1	No second escalation, proceed to final rounds
Integrity check retries (Stage 4)	3 per check	Log warning, proceed with disclaimer
Integrity check retries (Stage 6)	3 per check	Log warning, proceed with disclaimer
Experiment crash retries	3 per experiment	Mark as crash, skip

State Recovery (Material Passport)

If the session is interrupted (context limit, crash, user stops):

1. On next invocation, Stage 0 (INIT) checks for existing auto-academic-run-* directories
2. If found, read pipeline-state.json — check current_stage and material_passport
3. Resume from current_stage — the material passport records provenance for all completed stages
4. Resume message: "Found interrupted auto-academic run for '{topic}'. Last completed stage: {stage}. Resuming from {next_stage}."

Error Handling

If a skill invocation fails (deep-research, academic-paper, or academic-paper-reviewer):

1. Log the error to process-log
2. Retry once
3. If retry fails: log failure, attempt to proceed to next stage with partial results
4. If proceeding is impossible (e.g., RESEARCH failed, no hypotheses): report error to user and stop

Git Discipline

• Commit after every major action (stage completion, experiment run, revision, integrity check)
• Use descriptive commit messages with stage context
• Never commit credentials, API keys, or large binary files
• Use .gitignore for temporary files and caches