ReproWeave

Decide what can be rebuilt before committing the lab.

Local-first · Python 3.11+ · zero runtime dependencies · MIT

ReproWeave is a file-native research tool for engineers and AI researchers. It connects each paper to the claims you care about, the exact figure or appendix that supports them, the experiments and resources they depend on, a transparent reconstructability assessment, and a dependency-aware plan for testing the work yourself. Its rule-based triage joins those records into a practical queue: run now, prepare, collect evidence first, or finish the plan.

It deliberately does not scrape PDFs, call a language model, assign a hidden “paper quality” score, or require a cloud account. You supply the evidence; ReproWeave keeps the reasoning reviewable.

See it before installing

• Project website
• PyPI package
• Interactive synthetic evidence report
• Synthetic evidence matrix
• Synthetic replication triage
• Synthetic replication plan

Every record in the demonstration is fictional. It exists to show the workflow without misrepresenting a real paper, author, repository, score, or experimental result.

The problem it solves

A citation library answers “what did I save?” A reproducibility checklist asks “what did the authors disclose?” A project tracker asks “what should I do next?” ReproWeave joins these views without pretending they are the same:

_{The workflow keeps source evidence, missing information, and human decisions separate.}

The result is useful when deciding which paper to reproduce, preparing a lab meeting, scoping a benchmark, handing work to a collaborator, or explaining why an experiment is blocked.

Quick start

ReproWeave needs Python 3.11 or newer and has no runtime dependencies. Install from PyPI:

python -m pip install reproweave
reproweave demo my-review
reproweave audit --workspace my-review
reproweave triage --workspace my-review --format markdown --output my-review/reports/triage.md
reproweave report --workspace my-review

For an isolated one-off demonstration with pipx:

pipx run reproweave demo my-review

PyPI distributions are published from the tagged GitHub workflow with a short-lived OIDC credential and a public provenance attestation. GitHub Release assets include the same wheel, source archive, and SHA-256 checksums.

For development directly from the repository:

git clone https://github.com/CAOShurong/reproweave.git
cd reproweave
python -m pip install -e .
reproweave demo scratch/demo

Open scratch/demo/reports/evidence-report.html in any modern browser. The report is one self-contained HTML file: no server, JavaScript package, font request, or analytics endpoint.

A real workflow

1. Create the review boundary

reproweave init reviews/edge-ai \
  --title "Edge AI replication candidates" \
  --question "Which latency claims can be reconstructed on our hardware?" \
  --owner "Your Lab"

Edit reviews/edge-ai/reproweave.json to record inclusion criteria before screening.

2. Import references offline

reproweave import bibtex library.bib --workspace reviews/edge-ai
reproweave import csl-json selected.json --workspace reviews/edge-ai

The importers cover the portable subset needed for review records. Keep the original export in version control because ReproWeave is not a full BibTeX macro engine or citation manager.

3. Add evidence artifacts

Artifacts are plain JSON. Add them by writing files in their documented directories or through the validated CLI:

reproweave add claim claim.json --workspace reviews/edge-ai
reproweave add assessment assessment.json --workspace reviews/edge-ai
reproweave add task task.json --workspace reviews/edge-ai

Each claim requires an evidence_locator such as Figure 4, Appendix B.2, or repository/results.csv. A rating requires a written evidence note. Missing facts remain unknown; they are never filled by inference.

4. Triage candidates before committing resources

reproweave assess --workspace reviews/edge-ai
reproweave matrix --workspace reviews/edge-ai --format csv --output matrix.csv
reproweave backlog --workspace reviews/edge-ai
reproweave triage --workspace reviews/edge-ai --format markdown --output triage.md
reproweave plan --workspace reviews/edge-ai --format markdown --output plan.md

Triage combines the recorded assessment, experiment resources, task states, dependencies, and remaining effort. Hard blockers are evaluated before effort; there is no hidden composite score. Test a resource-access scenario without editing the workspace:

reproweave triage --workspace reviews/edge-ai \
  --resource rf-spectra-v2=available \
  --format markdown --output triage-with-data.md

The override changes only the generated scenario. Task dependencies then become execution waves. Tasks in one wave can run in parallel; later waves wait for declared prerequisites. Availability and effort remain reviewer-supplied inputs.

The next figure shows why that distinction matters. Two papers can look equally interesting, yet one can start now while the other first needs missing data or hardware details.

_{The labels “run now” and “collect evidence first” are explained by visible facts, not a secret score.}

5. Audit and seal the review

reproweave audit --workspace reviews/edge-ai
reproweave seal --workspace reviews/edge-ai
reproweave verify --workspace reviews/edge-ai

The audit checks artifact schemas, cross references, task cycles, and assessment coverage. The seal hashes the manifest and every source artifact with SHA-256. It proves that a specific workspace snapshot has not changed; it does not prove that a scientific claim is true.

The eight-dimension rubric

Ratings are yes, partial, no, unknown, or na. Scores summarize documented reconstructability only. They do not measure correctness, novelty, importance, statistical validity, fairness, or research integrity. Read the scoring methodology before comparing projects.

Workspace anatomy

review/
├── reproweave.json          # scope and screening criteria
├── papers/                  # bibliographic identity
├── claims/                  # bounded statements + evidence locators
├── experiments/             # protocol summaries and resource links
├── resources/               # code, data, environments, models, hardware
├── assessments/             # explicit rubric ratings and next actions
├── tasks/                   # replication DAG, states, estimates, acceptance
├── screening/               # inclusion/exclusion decision trail
├── reports/                 # generated HTML, CSV, Markdown, and JSON
└── reproweave-seal.json     # content-addressed source snapshot

Seven published JSON Schemas document the durable artifacts. Runtime validation uses the same conservative vocabulary without adding a JSON Schema dependency.

Commands

Run reproweave COMMAND --help for exact options.

Why local-first?

Literature assessments often contain unpublished research questions, early replication choices, reviewer notes, or links to restricted datasets. ReproWeave keeps those records in ordinary files that can stay on your device, live in a private repository, or enter your lab's existing backup system. No data is uploaded by the application.

“Local-first” is not a security guarantee. Repository permissions, disk encryption, backups, and the sensitivity of quoted material remain your responsibility. See SECURITY.md.

What ReproWeave is not

• It is not Zotero, a PDF reader, a systematic-review search engine, or a reference formatter.
• It does not claim formal PRISMA compliance or automate database search.
• It does not execute third-party repositories or sandbox untrusted research code.
• It does not reproduce an experiment merely because the audit passes.
• It does not use LLMs to infer evidence, judge papers, or fill missing fields.
• It is not an institutional record of research integrity.

These boundaries are features: the tool makes a narrow, inspectable record instead of hiding uncertainty behind automation.

Where it fits

Reference managers organize literature, systematic-review platforms support search and screening, and experiment trackers record runs. ReproWeave starts after a candidate set has been chosen. It preserves the claim-to-artifact chain and turns missing evidence and resources into an inspectable execution decision.

That positioning is intentionally narrower than “a complete systematic-review platform.” The project does not claim that evidence maps, reproducibility checklists, or replication planners are new ideas. Its contribution is a dependency-free, file-native bridge between them, including what-if resource triage and content-addressed verification. See the competitive landscape for tested boundaries and alternatives.

Documentation

• Methodology and score interpretation
• Data model
• Replication planning
• Audit and evidence seals
• Import behavior
• Competitive landscape
• Validation evidence
• Roadmap

Development

python -m unittest discover -s tests -v
ruff check src tests scripts
ruff format --check src tests scripts
python -m compileall -q src tests scripts
python scripts/check_repository.py
python -m build

The test suite uses only the Python standard library. CI runs on Windows and Ubuntu with Python 3.11 and 3.13, then installs the built wheel in a clean environment and exercises the CLI.

Contributing

Bug reports, schema discussions, and focused pull requests are welcome. Please read CONTRIBUTING.md and the Code of Conduct. Scientific disagreement is welcome; personal evaluation of authors is not.

License and citation

ReproWeave is released under the MIT License. Cite the versioned software release using CITATION.cff.

Questions and real replication-planning workflows belong in Discussions. Reproducible defects and methodology proposals belong in Issues.

Created and maintained by Shurong Cao.