An executable history of science.
English · Simplified Chinese
Learning paths · Paper catalog · Paper timeline · Case updates · Run an example
RunThePaper is building an executable history of science. Our paper reproduction agent, PRAgent, independently reconstructs the derivations, methods, and computations behind papers. RunThePaper brings the resulting code, notes, data, figures, and evidence together, organized by research field, learning path, and paper chronology. The current collection focuses on physics and quantum science, with English and Chinese notes.
The starting point is independent reconstruction from papers—not just summaries or wrappers around authors' code. PRAgent's reproduction workflow does not require authors to release runnable code. RunThePaper preserves the resulting implementations, evidence, and limitations so readers can inspect, rerun, and extend the work. Missing scientific inputs or experimental resources remain explicit case boundaries.
A paper compresses a research result. Getting started often requires working through intermediate derivations, parameter choices, and numerical decisions. We preserve that path so the next researcher can understand how a result was obtained, run the calculation, and build on it. Making these working foundations accessible to students, researchers, and scientific agents is our approach to research infrastructure for the AI era.
| What you want to do | Your starting point |
|---|---|
| Enter a research field | Prerequisites, a suggested paper order, and a first exercise for beginning graduate students and undergraduates with the relevant foundations. Choose a learning path. |
| Continue existing work | Derivations, code, results, and checks together for verification, teaching, new parameter studies, and extensions. Browse the cases. |
| Build AI for Science | A machine-readable case index, executable calculations, and evidence records as domain context and validation material for scientific agents. Explore the index. |
Start with one formula, run one calculation, explain its output, then change an assumption. Each case records the scope it reproduces and the work still open, so you can choose a starting point with its limitations in view.
How many quantum gates does Hamiltonian simulation require? The qDRIFT case reconstructs the resource estimates in A random compiler for fast Hamiltonian simulation, comparing qDRIFT with Trotter methods for three molecular examples.
Generated reproduction of Fig. 2. Follow the derivation, inspect the CSV data, or read the numerical checks. The case is awaiting independent review.
The qDRIFT example runs locally on a CPU with Python 3.11 or newer. From a macOS or Linux terminal:
git clone https://github.com/xi-zhao/runthepaper.git
cd runthepaper
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
cd cases/1811.08017/code
python scripts/run_reproduction.py --config config/paper_exact.jsonThis recomputes the numerical data behind Figs. 2 and 4 and writes:
cases/1811.08017/outputs/data/fig2_gate_counts.csvcases/1811.08017/outputs/data/fig4_phase_estimation_counts.csvcases/1811.08017/outputs/checks/target_checks.json
The check file reports "status": "passed" when the numerical checks pass.
The plotted figures are included in the repository; this command regenerates
the data and checks. See the case instructions
for its scope and remaining review work.
100 public paper cases, including partial and blocked reproductions. Open a collection in the full catalog for paper references, recorded status, bilingual notes, code, and evidence.
Learning paths · Paper timeline · Recent case updates
For another starting point, explore non-Hermitian edge states or disorder and Lyapunov band theory. Their case pages link the paper, derivations, results, and current limitations.
Three questions matter when using a reproduction:
| Question | Where to look |
|---|---|
| What was reproduced? | The case overview names the figures or claims covered, parameters used, and any missing inputs or compute limits. |
| Do the results agree? | Generated data and scientific checks show numerical agreement, discrepancies, and tolerances. Visual similarity is recorded separately. |
| What remains open? | The completion assessment records unresolved targets and independent review status. Passing a run does not close the whole case. |
The collection includes partial results and unsuccessful attempts with their evidence. A paper-error candidate is a finding to investigate, not a settled correction to the paper.
The frozen 100-paper audit maps 3,933 checks to 1,427 scientific claims. With each numerical claim weighted equally, its outcomes are:
| Outcome | Share |
|---|---|
| Successfully reproduced | 40.55% |
| Blocked by documented external limitations | 21.93% |
| Attempted but not reproduced | 37.52% |
These are claim-level outcomes, not a percentage of fully reproduced papers. The audit includes the full ledger and measurement method, including fidelity evidence and its limits.
Independent reproduction is the starting point; an executable history preserves what we learn; original research is the direction.
| Project | Role |
|---|---|
| PRAgent | Independently reconstruct and check: rebuild methods from papers, implement and run reproduction code, and organize validation and independent review. |
| RunThePaper | Preserve and share: accumulate derivations, code, data, figures, evidence, failure records, and review status for people and agents to inspect, run, and extend. |
| CyberEinstein | Build toward original research: connect field history, reproduction evidence, and new questions into a sustained research process. |
You do not need access to PRAgent to read or rerun the public cases here. The PRAgent execution system is developed separately and is not distributed in this repository. CyberEinstein builds on these existing assets; its sustained original-research workflow is the next stage, not a prerequisite for using this collection.
A paper is one entry into this history. Its case connects a research question to the claims, derivations, code, generated results, and checks behind it. You can follow that chain, rerun a calculation, inspect a discrepancy, or ask what happens when an assumption changes.
As these cases accumulate, they form a shared knowledge foundation for learning and further research. Tracing relationships between discoveries, retrieving across cases, and measuring their value to scientific agents are next-stage work.
Research collections organize the papers by field. Learning paths add prerequisites, reading order, and exercises. The paper timeline provides a chronological route into the collection. These views refer to the same cases and their recorded scientific state.
The update history is generated from actual commits. It separates new papers, updates, and removals, and shows whether code, derivations, data, figures, or validation evidence changed. Every entry links an exact revision. The catalog, learning paths, paper timeline, and update history follow one maintenance workflow.
- Request a paper: open an issue with its DOI or arXiv ID and the figure or claim you want to reproduce.
- Report a run or discrepancy: include the case, command, environment, and observed result so someone else can check it.
- Review or extend a case: contribute a derivation check, missing input, correction, or additional result through the contribution workflow.
- Improve a learning path: report a missing prerequisite, unclear derivation, or command that failed, or suggest a paper order and exercise. A specific learning experience can improve the shared research foundation.
When using a case in your work, cite the original paper and link the case at the commit you used so readers can inspect the same materials.
Code: MIT. Notes, generated data, and generated figures: CC BY 4.0, unless a case states otherwise. Third-party material, including attributed paper excerpts in comparison panels, retains its original terms; see NOTICE.md.