Why PaperEngine exists

Most tools answer one narrow question: "Is this text copied?" or "Does this look AI-written?". Rejection happens for dozens of other reasons — missing ethics statements, impossible statistics, unreferenced figures, template violations, predatory venue traps, retracted citations. PaperEngine runs 65 specialized engines against your manuscript and returns every finding as:

Severity | Finding | Evidence (quoted from your paper) | Confidence | How to fix it

The honesty principle (by design, not marketing): Similarity is not plagiarism. An AI-risk score is not proof of AI authorship. Every finding carries evidence + confidence, and the readiness score is informational — final judgment stays with humans, exactly how editors are trained to use iThenticate/Similarity Check.

How to use

Step 1 — Clone & install (one time, ~1 minute)

Requirements: Python 3.10+ — nothing else (PDF support needs pypdf, image forensics needs Pillow; both come with .[all]).

# 1. Clone the repository
git clone https://github.com/abnsr-sol/paperengine.git
cd paperengine

# 2. Install (pick one)
pip install -e .            # core — zero required dependencies
pip install -e .[all]       # recommended: + PDF ingestion & image forensics

# 3. Verify the install
papercheck --list-venues    # should print all 19 venue presets

Windows tip: if pip isn't on PATH, use py -m pip install -e .[all].

No-install option: everything also runs straight from the cloned folder — just replace papercheck with python -m papercheck in any command below.

Step 2 — Use the Desktop version (web GUI)

papercheck --gui                 # launches the server and opens your browser
papercheck --gui --port 9000     # custom port if 8765 is already taken

Then, in the browser:

1. Drag your manuscript (.docx / .txt / .md / .tex / .pdf) onto the upload zone — or click to browse
2. Choose the standard — International (IEEE/Elsevier/ACM…) or National (India: UGC/AICTE/NAAC)
3. Pick the venue preset — e.g. ieee_conference, ugc_care, mdpi (the list filters by standard)
4. Click Check my paper → the full report renders in the browser: readiness score + findings table (severity · finding · evidence · confidence · how to fix)
5. (Optional) Drop the revised version into the second upload zone before checking → before/after comparison: fixed / still-open / new findings plus the score delta:

readiness score:  42 → 57  (+15)
  fixed: 12   still open: 41   new: 3

Privacy: the GUI runs on your machine only (localhost). The file is parsed in memory, checked by the same 65 engines as the CLI, and never uploaded to the internet.

Step 3 — Use the CLI version

Basic pattern:

papercheck <file> [--standard international|national] [--venue <preset>] [--format <format>] [--out <file>]

Common tasks:

Without installing, run from the cloned folder with python -m papercheck … instead:

python -m papercheck sample_paper.txt --venue elsevier

# Full report to a file (console | markdown | html | fixplan | csv)
python -m papercheck paper.docx --venue mdpi --format html --out report.html

# Indian national standards (UGC/AICTE/NAAC)
python -m papercheck thesis.docx --standard national --venue ugc_care

# Prioritized fix plan (criticals first, effort-estimated)
python -m papercheck paper.docx --venue ieee_conference --format fixplan

# Batch-scan a folder, worst-first summary table or CSV
python -m papercheck --batch papers/ --venue ugc_care --format csv --out summary.csv

# Before/after revision comparison
python -m papercheck v1.docx --compare v2.docx --venue elsevier --format html --out diff.html

# With online lookups (Crossref): duplicate-publication + DOI validation
python -m papercheck paper.docx --venue springer --online --mailto you@university.edu

# Compare against your already-published papers (duplicate / "no new content")
python -m papercheck paper.docx --corpus ./my_prior_papers/

One-time retraction database (optional, recommended)

papercheck --update-rwdb
# caches 70k+ retraction records (CC-BY 4.0, Crossref) at
# %TEMP%/papercheck_rwdb.json (Linux/macOS: /tmp/papercheck_rwdb.json).
# From then on, retracted-reference screening runs against the full DB offline.

What the engines check

The full 48-angle rejection map (with engine-by-engine status) is in COVERAGE_MATRIX.md; every engine is listed with its exact checks in the architecture section below.

Venue presets (dual standard)

papercheck --list-venues

Every preset works in both the CLI and the web GUI; --venue-json rules.json accepts exact limits for any venue not yet preset.

Output formats

Every finding, in every format, carries: severity · finding · evidence · confidence · concrete action.

Architecture

papercheck/
├── __main__.py        CLI entry point (single file, batch, compare, gui modes)
├── ingestion.py       DOCX (stdlib zip+XML), TXT/MD/TeX, PDF (optional pypdf)
├── metrics.py         text statistics (readability, burstiness, n-grams, …)
├── venues.py          19 venue rule presets + --venue-json override
├── risk.py            Finding / Severity / RiskReport / readiness score
├── report.py          console, Markdown, and HTML renderers
├── fixplan.py         prioritized, effort-estimated fix-plan renderer
├── compare.py         before/after revision diff (fixed / still open / new)
├── batch.py           folder scanning, worst-first ranking, CSV writer
├── rwdb.py            Retraction Watch DB download/cache/screening
├── webui.py           local drag-and-drop GUI (stdlib http.server)
└── checks/            65 engines — one per rejection angle
    compliance · structure · language · citations · claims · ai_risk ·
    integrity · novelty · consistency · figures · forensics · policy ·
    statistics · overclaiming · self_plagiarism · citation_integrity ·
    reproducibility · submission · ugc_plagiarism · reference_verify ·
    fabrication · methodology · reporting_guidelines · writing_depth ·
    legal_ethics · citation_cartel · paper_mill · predatory_journal ·
    retracted_refs · submission_package · image_forensics · stats_deep ·
    design_claims · redundancy · domain_checklists · literature_search ·
    scope_match · rebuttal · crossref_verify · author_network ·
    reviewer_fraud · image_manipulation · llm_artifacts · supplementary ·
    data_license · abstract_quality · citation_age · sex_gender ·
    stats_plan · editorial_format · author_info · figure_quality ·
    venue_extras · ai_disclosure_deep · safety_ethics · authorship ·
    repro_env · paragraph_structure · transitions ·
    reference_completeness · funder_compliance · peer_review ·
    domain_checklists2 · grammar_tool (optional LanguageTool) · cross_check

Extending: add checks/my_angle.py with run(doc, ctx) -> [Finding], register it in checks/__init__.py, add tests. New venue: one dict in venues.PRESETS. See CONTRIBUTING.md for the ground rules (evidence + confidence + action on every finding, offline-first, dual standard).

Plug-in points already in the code: LanguageTool server (grammar_tool), Crossref/OpenAlex (--online), Retraction Watch DB (rwdb.py), AI-detector APIs (checks/ai_risk.py — documented hook, no verdicts).

Project layout

Development

python -m unittest discover -s tests        # 150+ tests, offline, no services needed
python scripts/maintenance.py               # tests + retraction-cache refresh

CI (.github/workflows/ci.yml) runs the full suite on Python 3.10 – 3.13 on every push and PR. A weekly scheduled job refreshes the retraction database and re-runs the suite. Tagging vX.Y.Z triggers the PyPI publish workflow (tag/version match is verified first).

To enable PyPI uploads: create a pending publisher on pypi.org for abnsr-sol/paperengine (workflow publish.yml, environment pypi) — after that one-time setup, every v* tag publishes automatically.

Honest limitations (baked into the design)

1. Similarity ≠ plagiarism. Overlap requires human interpretation (Crossref itself warns against automatic rejection thresholds). The engine shows what matched as evidence, never a verdict.
2. AI detection is probabilistic. "Low burstiness" and "template transitions" occur naturally in non-native and highly technical writing. The AI-risk engine reports an uncertainty band — and deliberately refuses typography myths ("em dash = AI") that have no scientific support.
3. Grammar checks are heuristics, not a full grammar engine. Run LanguageTool/Grammarly/Paperpal for the final pass (or point grammar_tool at a local LanguageTool server).
4. Venue rules are typical published limits and change — confirm against the venue's current author guidelines.
5. Readiness score is informational. It aggregates weighted, confidence-scaled findings; it is not a prediction of acceptance.
6. What no software can check: whether the science is true, whether ideas match paywalled prior work, and the reviewer's subjective "so what?". Tools that pretend otherwise are selling overconfidence.

License

MIT — free for research, commercial products, and institutions.