Autonomous Scientific Discovery, grounded in Principles and tested against data.

Explore scientific knowledge. Discover interpretable Rules. Inspect the evidence.

Quick start · Included projects · Discovery workflow · ASD benchmark · v1.4.1 · Research

Bring a research goal and a local dataset. Principia connects scientific context with executable analysis to look for compact, interpretable relationships—and keeps the evidence behind each result available for inspection.

Version 1.4.2 brings dataset-native Autonomous Scientific Discovery (ASD) into the Principles workspace. Literature Principles provide scientific context; observations describe computed evidence; extracted Rules carry equations, calibration results, validation decisions, and a stated scope. A study map connects these objects so you can move from a promising relationship to the records that support it.

The release includes five demo public-data projects, analyzed with DeepSeek-V4-Pro. Their saved maps, results, equations, and packaged evidence open locally without the original datasets or an API key. Connect your own data and models when you are ready to begin a new discovery.

_{A shared scientific workspace: explore Principles by area, enter a research goal, or return to a saved project.}

The result is a scientific object you can examine. A strong fit alone does not establish a mechanism. Known identities, empirical relationships, and tentative interpretations retain their own evidential scope; unsuccessful candidates remain part of the study record. A dataset may produce observations without a Rule that passes the available checks.

_{The Rydberg project: move between the study map, discovery results, and individual evidence records.}

The workspace also supports Discover again with a new project by default, configurable reasoning and vision models, visible task activity and stop controls, resizable information panels, and project deletion. Data and workspace state remain separate, so repeated discovery does not require copying the source dataset into each project.

These examples span atomic sensing, seismology, particle physics, thin-film metrology, and human movement. Three were selected from the initial twenty-scenario discovery campaign; two were added during subsequent user testing. They illustrate different kinds of useful scientific output, rather than a claim that every recovered relationship is a new law of nature.

Open Dataset & evidence for each project's provenance and selection rationale. Open a Rule to inspect its equation, calibration, comparisons, and limitations. Rounded display values are backed by full-precision records. The demo guide explains portability, reruns, and deletion.

Inspect a Principle. Read the scientific statement, its scope, and supporting context without leaving the project.

Interrogate a Rule. Follow a typeset equation through interpretation, quantitative checks, and evidence. Click either image for full resolution.

Use Python 3.11 or 3.12 in a virtual environment. This is the GitHub source release; the commands below do not depend on a matching version being published to PyPI.

Download the application without the separate public test corpus.

git clone --depth 1 --filter=blob:none --sparse https://github.com/pzqpzq/Principia.git
cd Principia
git sparse-checkout set Principia-v1.4.2

For Linux or macOS

python3 -m venv .venv
source .venv/bin/activate
python -m pip install "./Principia-v1.4.2/core-v1.4.2[asd,local]"
principia open --working-directory ./principia-workspace --port 8142

For Windows

python -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install "./Principia-v1.4.2/core-v1.4.2[asd,local]"
principia open --working-directory ./principia-workspace --port 8142

# for Windows users after frontend changes
# If the frontend code has been modified, rebuild the frontend and reinstall the local package using the following commands:
cd "Principia\Principia-v1.4.2\core-v1.4.2\frontend"
pnpm build

cd ..
pip install -e .

The application opens at http://127.0.0.1:8142/. Its frontend is already built; Node.js is needed only for frontend development. Dependency installation requires internet access. Once installed, the five demos and their packaged evidence can be browsed offline; external publisher links and remote model calls require a connection.

Demos initialize once in an empty workspace. Existing projects are preserved, and deleted demos do not reappear on restart. To try the included projects separately from existing work, choose a new working directory.

1. Choose New research, use Add data to connect a local folder, and describe your research goal.
2. Open Sources & model to configure your provider credentials and select reasoning and vision models. The included examples used DeepSeek-V4-Pro via SiliconFlow; new runs use your own configuration.
3. Start discovery and review the provider permission request. View activity shows task progress, models, source information, and stop controls.
4. Inspect observations and Rules, follow the evidence, or use Discover again to change the goal, data, or models. Keep a new attempt as a separate project or explicitly overwrite the current project.

Raw files remain in the selected folders. Remote analysis can send bounded profiles, excerpts, or previews after per-run permission; it is not a promise of zero network egress. No author credentials or original demo datasets are packaged with the application. A new computation on a demo requires reconnecting the relevant public data. See ASD, privacy, and storage.

Principia-100 brings 100 source-verifiable scientific scenarios to autonomous discovery: native data, practical research questions, and a protocol for testing the resulting equations. It is a core part of Principia's research ecosystem and is also usable with other discovery systems.

The collection spans semiconductor devices, materials, physics, biology, medicine, environmental science, society, computing and mathematics. It retains heterogeneous source formats: spreadsheets, instrument logs, scientific arrays, images, audio, video, spatial records and computational trajectories. Every case includes source attribution, reuse terms, checksums, a neutral brief and a scientific task card with validation units and limitations.

Start with membrane permeation, industrial screw driving, CHO bioreactor cultivations, or Arctic ocean profiles. Download a selected case, connect it through New research → Add data, and use its USER_BRIEF.txt as a starting research goal. For a formal benchmark study, follow the input allowlist and freeze the case-specific validation design first.

Explore all 100 → · Download selected cases · Read the protocol

The full benchmark is about 4.80 GB and uses Git LFS for large assets; it is optional and separate from the application installation above. This is an open, source-aware corpus: published analyses are disclosed, the original 20 scenarios were used during Principia development, and native-reader support is distinguished from application ingestion. Reproduction, validated extension, novelty candidate and justified abstention are separate outcomes. The release supplies research inputs and evaluation standards, without claiming 100 unknown laws or an aggregate discovery score.

Dataset reuse follows the source-specific terms. The earlier scenario/ directory remains a historical collection; use the versioned ASD-benchmarks/ corpus for new benchmark studies.

The release separates source data, runtime state, and portable demonstration results. Its demo bundle is compressed and checksummed; it carries saved study records, executable Rules, calibrations, and selected derived artifacts without shipping an old runtime database. A clean-release size gate excludes environments, caches, private folders, and credentials. Backend and frontend checks, schema validation, archive hygiene, and a fresh-workspace demo check are included in v1.4.2 CI.

Explore the source, read the demo guide, or browse the recorded release checks. Computational reproduction requires the original inputs and the recorded analysis conditions; a packaged result remains evidence to assess, not independent replication.

Help improve Principia. Share a reproducible issue, a difficult public dataset, or a scientifically meaningful failure case in GitHub Issues. Star or watch the repository to follow development of the Principles Cloud and data-driven scientific discovery.

Principia is shaped by a broader question:

What representations should AI agents create, exchange, preserve, and evolve when the objective is scientific discovery rather than fluent conversation?

Machine Dialectology studies how heterogeneous LLM agents can invent, exchange, route, and evolve compact machine-oriented languages. Its ICML 2026 precursor, When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning, introduces Communicative Language Symbolism Routing (CLSR).

• ICML 2026
• OpenReview
• arXiv:2606.29354
• Code

Its connection to Principia is representational: mechanisms, constraints, analogies, trade-offs, and falsification rules must become reusable intermediate objects before they can be composed efficiently.

SciDialect studies grounded symbolic compression as an intrinsic reward for scientific discovery agents. Compact states are useful only when task-critical meaning, evidence anchors, and reconstruction remain recoverable.

Principia operationalizes this philosophy through typed scientific objects, explicit evidence references, bounded generation trace, strict validation, and fail-closed persistence.

Principia/
  README.md                       # v1.4.1 project overview
  assets/                         # product screenshots
  global-cloud/                   # canonical, paper-free Cloud data and schemas
  Principia-v1.4.1/
    core/                         # regular-user package, API, UI, tests, and docs
  Principia-v1.3/                 # maintained v1.3.3 framework and examples
  legacy/                         # historical releases

Principia is a research framework, not an oracle.

• A fluent claim is not automatically a Principle.
• A reviewed record is not automatically universal outside its stated scope.
• A foundation link is not proof of truth.
• A relation score is not a probability of correctness.
• A virtual Principle is a hypothesis, not a confirmed contribution.
• A generated connection suggests a reasoning path; it does not replace empirical validation.

The intended standard is simple:

Every persuasive claim should be paired with an inspectable source, an explicit scope or assumption, and a test that could prove it wrong.

@inproceedings{
pei2026when,
title={When {LLM}s Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning},
author={Zhengqi Pei and Qingming Huang and Shuhui Wang},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=ovpL0ujD6j}
}

The repository root is distributed under the Apache License 2.0. The v1.4.1 regular-user core is separately released under the MIT License.

Academic collaboration
Institute of Computing Technology, Chinese Academy of Sciences
peizhengqi22@mails.ucas.ac.cn

Business collaboration
Beijing Chipflow Technology Co., Ltd.
peizhengqi@chipflow.net

Let scientific knowledge compound: from works, to Principles, to solutions.

If Principia is useful to your research, consider starring the repository and following the evolution of the Principles Cloud.