A research notebook for the labs the ELN market cannot reach: no server, no cloud account, no Docker, no IT department.
Unzip a folder, double-click, and it runs — on a laptop with the network cable pulled. Upload your own project documents and it builds a research-notebook system shaped around your project, then drafts each note from the lab logs and measurement files you feed it. Every sentence it writes carries the evidence ID it came from, and it will not seal a note containing a claim the evidence ledger does not support.
Korean · English · Japanese, at parity.
![AAA-RNS in thirty-four seconds: uploading a project plan, the extraction showing a source and a confidence per value, an experiment log and a measurement CSV attached by SHA-256, sections drafted with [E#] citations, four gates running, two people signing, and the note sealed](/hopejsh/aaa-rns/blob/main/.github/branding/demo.gif)
Unzip to sealed note. No server, no account, no build step.
| Platform | How to start |
|---|---|
| macOS | double-click start_mac.command |
| Windows | double-click start_windows.bat |
| Any | python3 server.py → open http://localhost:8777 |
First run only: the OS checks files extracted from a downloaded zip before running them — on macOS, right-click
start_mac.command→ Open (or System Settings → Privacy & Security → "Open Anyway"); on Windows, click More info → Run anyway on the SmartScreen prompt. After that, a double-click is enough.
No registration key. No account. No installer. Chrome or Edge recommended (shared-folder mode uses the File System Access API). Python 3 is the only prerequisite, and macOS and most Linux systems already have it.
First run walks you through four steps: upload your documents → deep analysis → review what
was extracted → generate the system. sample_docs/ has a sample plan and metrics table if you
want to try it before using real documents.
📘 Installation & user guide (PDF, with real screen captures) · User manual · Verification report · All documents, three languages
It drafts from your evidence, not from a template. Drop in an experiment log or a
measurement CSV; the file is attached with its SHA-256 hash, evidence is registered as [E#],
and the work-performed, result-data and interpretation sections are written from that evidence
alone. Sentences come only from the uploaded originals — the system does not compose facts.
It refuses. Four deterministic gates run before a note can be sealed:
| Gate | What it checks |
|---|---|
| G1 Evidence mapping | Sentences with no evidence ID, citations of evidence that does not exist, speculative wording |
| G2 Longitudinal consistency | Period overlap with sealed notes, contradictory values for one metric, status regression |
| G3 Numeric & unit audit | Recalculated sums and ratios, unit notation, direction against target, unsupported claims of achievement |
| G4 Compliance | Required fields, and originals attached for every cited evidence item |
A note with outstanding findings cannot be sealed. Nothing is discarded — the text and the findings are both kept, so a refusal is a record, not a deletion.
Sealed means sealed. On sealing, the note's hash covers the content and the signatures together and chains to the previous sealed note. Editing the text afterwards — or swapping an approver's name — is exposed by the chain check. Changes are made only by issuing a revision (-R1, -R2 …); the original and the revision history both remain.
Zero dependencies, and that is not a slogan. The PDF, DOCX, XLSX, HWPX and CSV parsers, the DOCX/XLSX generators, the charts and the hash chain are all implemented in this repository. There is no CDN, no npm install, no build step. The Content-Security-Policy blocks outbound requests entirely — except to the three AI endpoints (Anthropic, Google, OpenAI), and only if you choose to connect an AI engine, which is off by default and never required.
26 specialist agents. Seven layers with group leads, typed packets and a single-writer blackboard (agents/MAS_SPEC.md). The prompts are engine-neutral templates. The deterministic core — parsers, gates, hash chain — runs in the browser with no LLM at all.
Stated here rather than discovered later:
- This is not a LIMS. No sample or inventory management, no instrument integrations, no real-time co-editing, no mobile app.
- Identity is self-asserted. Signatures carry a per-user device key (ECDSA P-256, non-extractable) and optionally a passkey (WebAuthn — Touch ID and the like); both are locked under the seal hash, so swapping a signer after sealing breaks the chain. But no certificate authority vouches that the key belongs to that person. The PIN is stored with PBKDF2-SHA256 (210,000 iterations, salted) and is still not access control — real access control comes from the OS permissions on your shared folder.
- Timestamps default to the machine's clock. An admin can enable RFC-3161 timestamping in Settings: on sealing, the seal hash (32 hash bytes only — never content or files) is sent to a TSA and the signed token is attached. Off by default; when offline it silently falls back to the local clock. The default TSA (FreeTSA) is not an accredited authority.
- It still cannot meet any regime that requires a certificate-based electronic signature or an accredited timestamp — one technical fact that carries a different name in each jurisdiction: not 21 CFR Part 11-capable (US FDA); not a qualified signature or timestamp under eIDAS, and not compliant with EU GMP Annex 11 (EU); not compliant with the ER/ES guideline (Japan); and it does not satisfy 국가연구개발사업 연구노트 지침 전자서명인증 as required by 국가연구개발사업 연구노트 지침 (Korea). It is a research-integrity and reproducibility tool, not a regulated-record system. The clause-by-clause position, including what is and is not verifiable, is in the compliance matrix; the threat model is in SECURITY.md.
- The hash chain proves internal consistency. By default it is tamper-evident, not tamper-proof. With RFC-3161 enabled, each sealed note's hash and time are anchored outside the record-holder's control (under the TSA's signature) — but never retroactively.
If your lab is well funded, cloud-comfortable and needs deep sequence or chemistry tooling, Benchling or Revvity will serve you better. This is for the labs those products cannot reach.
Two campaigns, 21 cycles, 2,446,015 simulation runs — deliberately injecting missing citations, contradictory figures, tampered signatures, corrupted backups and forged licenses, then asserting the product blocks them. Five genuine defects were found and fixed; three findings turned out to be defects in the test code and were retracted rather than counted.
node simulation/run_simulation.mjs --campaign 2 --cycle 9 --iters 150000Runs are reproducible from the seed. Reports: simulation/reports/ —
including an honest note about nine campaign-1 reports that were
lost to an overwrite and are not reconstructed. Full write-up:
verification report.
The build in this repository is the Community edition: no registration key, no limit on the
number of projects. An Enterprise edition — the same code with one constant flipped in
js/core/edition.js — requires a signed license key and binds an
installation to a single project, for builds issued to a named company or institute.
Because the source is published, a key check cannot stop anyone able to edit the code. Its purpose is confirming who a build was issued to, not technical copy protection.
data/ project structure and audit log · notes/ research notes · notes_files/ your uploaded
originals · ledger/ evidence ledger · archive/ permanent copies of sealed notes. All of it
stays on your machine, or in the team folder you point it at. None of these directories is
tracked by git. Backups are a single ZIP with a per-file SHA-256 manifest, and a restore writes
nothing unless every file passes verification.
Issues and pull requests are welcome — see CONTRIBUTING.md. Two house rules: documentation changes may not introduce a claim the repository cannot evidence, and the three READMEs move together.
If this software contributed to work you publish, please cite it. The repository carries a
CITATION.cff, so GitHub's "Cite this repository" button on the sidebar
produces APA and BibTeX for you.
Jung, S. H. (2026). AAA-RNS: AI Agent-driven Autonomous Research Notebook System (Version 2.1.0) [Computer software]. https://doi.org/10.5281/zenodo.21937754
@software{jung_aaarns_2026,
author = {Jung, Seung Ho},
title = {{AAA-RNS: AI Agent-driven Autonomous Research Notebook System}},
version = {2.1.0},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.21937754},
url = {https://github.com/hopejsh/aaa-rns}
}Zenodo issues two DOIs. 10.5281/zenodo.21937754 is the concept DOI and always resolves to
the newest release — cite this one normally. 10.5281/zenodo.21955354 is the version DOI,
frozen to v2.1.0 (this release); cite it instead when the exact bytes matter, such as in a
reproducibility statement or a methods section describing a specific analysis. Every earlier
release keeps its own version DOI, listed under the concept DOI on Zenodo.
If your journal uses Research Resource Identifiers, cite AAA-RNS (RRID:SCR_028836) inline in the Methods section. That is the registry entry, and it is separate from the DOI: the DOI belongs in the reference list, the RRID in the running text.
Apache License 2.0 — LICENSE · NOTICE · Copyright © 2026 Seung Ho Jung
Seung Ho Jung — system design, core engines (parsers, analyzer, generator, gates, hash chain), the 26-agent MAS architecture, and the verification simulation harness.