Principia
Distills scientific literature into reusable principles and composes them into traceable, testable research ideas—every discovery step stays inspectable.
Evidence: connected folders unselected by default, per-run permission for remote analysis, OS-level credential storage excluded from frontend/logs/artifacts, prior Cloud snapshot retained for rollback, hash verification before activation. Deductions: dependency security shows only 'requires internet access' with no lockfile or audit evidence, scored 1. No malicious behavior evidence, but all claims are static assertions.
Tests cover control tokens, interruptible retry waits, transport close, strict validation errors; Cloud clients degrade visibly to SQLite FTS. Deductions: the README contradicts itself — badges and CI claim MIT while the LICENSE file and metadata are Apache-2.0; the README is truncated mid-tag. Self-consistency scored 1.
Twenty cross-domain public scenarios and five demo projects span many formats; capability-boundary statements are unusually thorough ('does not predict event times', 'not evidence of a new mechanism', 'not a promise of zero network egress'), so boundaries scored 3. Deductions: stated environment is Python 3.11/3.12 only while v1.3 CI tests 3.10–3.13, and the stable PyPI release is two minor lines behind, so environment fit scored 2.
Install notes cover venv, sparse checkout, Windows PowerShell differences, port and working directory; five demos plus twenty scenarios, a demo guide, and recorded release checks justify 3 for examples and limitations. Deductions: severe license inconsistency (README badge/CI say MIT; LICENSE and task metadata say Apache-2.0), license scored 1; no visible CHANGELOG and PyPI lags at 1.3.3, versioning scored 1; maintainer identity and response commitments unclear, only a GitHub Issues link, maintenance scored 1.
Outputs are inspectable scientific objects (equations, calibrations, linked evidence, stated scope) with full-precision backing records, stop/overwrite/delete controls, and offline-browsable demos requiring no API key. Deductions: real utility depends on user-supplied models and data, and discovery quality cannot be verified statically; remote-LLM and embedding costs are disclosed but not quantified, so cost_benefit scored 2.
Strongest dimension: Work→Principle provenance links, checksummed demo bundles, SHA256SUMS verified by redownload, and explicit separation of 'packaged result as evidence to assess, not independent replication' justify 3 for traceability. Deductions: cross-corroboration is internal-only (qa JSON, CI); the arXiv/ICML badge cannot be verified from the given files (and the 2606.29354 identifier looks implausible); demo result correctness is unverifiable statically. Both scored 2.
- License contradiction: README badges and v1.3 CI claim MIT while the LICENSE file and task metadata say Apache-2.0; confirm actual terms with the author before use.
- Remote analysis sends 'bounded profiles, excerpts, or previews' to external model providers; the docs explicitly state this is not a promise of zero network egress. Review the privacy doc and grant permissions per run before handling sensitive data.
- Dependencies require internet install with no lockfile or vulnerability-audit evidence in the reviewed files; audit them yourself before installing.
- The stable PyPI release is stuck at 1.3.3 while 1.4.x is source-only; the README itself is truncated at an unclosed details tag, raising documentation-completeness concerns.
- Publisher identity is unverified; the arXiv/ICML badge cannot be confirmed from the repository (the 2606.29354 identifier looks implausible), so verify independently before citing.
- Packaged demo results are 'evidence to assess, not independent replication'; do not treat them as verified scientific conclusions.
What does this agent do, and when should you use it?
Principia (GitHub: pzqpzq/Principia) is a local-first autonomous scientific discovery workbench that represents mechanisms, constraints, boundary conditions, and falsifiers from literature as versioned Principle objects. Version 1.4.2 adds dataset-native Autonomous Scientific Discovery (ASD): it ingests a local dataset, profiles variables, fits candidate expressions, runs held-out checks, and produces Rules carrying equations, calibration results, and a stated scope—all traceable through a saved study map. The repository ships five demo projects analyzed with DeepSeek-V4-Pro (Rydberg field calibration, earthquake magnitude scaling, ATLAS particle physics, thin-film wafer metrology, and walking dynamics) plus twenty independent public test scenarios. The system runs locally as a Python backend with a pre-built React frontend launched via `principia open`, while the Global Principles Cloud is distributed as reviewed JSON through GitHub. The repository root is Apache-2.0 licensed; the v1.4.2 core is MIT licensed.
Principia reads local data files and a user-defined research goal, profiles measurements, and connects variables with scientific context from literature Principles. It proposes candidate expressions, fits them on development data, compares alternatives with validation evidence, and applies recorded challenge checks (split definitions, errors, controls, failure reasons) before promoting a Rule. Each Rule carries an executable expression, calibration metrics, validation decisions, and scope, stored alongside observations and Principles in a study map. Version 1.4.1 provides a Discover–Inspect–Derive flow: global retrieval locates Works and Principles, a WebGL Principles Map exposes arguments, conditions, boundaries, and revisions, and Derive Principles performs multi-level reasoning over up to 20 selected records. Version 1.3.3 (PyPI package principia-ai==1.3.3) runs cross-domain retrieval (arXiv, OpenAlex, Crossref, Semantic Scholar, Europe PMC), structured evidence extraction, SciDialect-Evo candidate generation with prior-art comparison, and exports a seven-file research pack including a validation plan. Core commands include pip install "./Principia-v1.4.2/core-v1.4.2[asd,local]" and principia open --working-directory ./principia-workspace --port 8142.
- A researcher with a local experimental dataset who wants a systematic search for compact, interpretable relationships with a complete evidence chain
- A scientist converting mechanisms and constraints from papers into a reusable, versioned, transferable principle library rather than one-off reading
- A data scientist exploring public data (USGS earthquakes, NOAA storms, NIST wafer metrology) using the 20 ready-made scenarios
- An industrial R&D team needing local-first durable research memory that turns technical reports and private materials into engineering constraints without uploading local documents
- A graduate student or lab generating hypothesis Idea Cards with prior-art comparison and validation plans grounded in exact evidence records
- An evaluator wanting to browse the five packaged demo projects and their full-precision evidence offline before adopting the tool
What are this agent's strengths and limitations?
- Inspection is the core design: every Rule, Principle, and hypothesis carries evidence links, explicit scope, and falsifiability—scientific objects, not black-box output
- Local-first architecture: raw files stay in user folders, credentials use the OS credential mechanism, and Global search never uploads local content
- Five demo projects ship with packaged evidence and full-precision records, openable offline without the original datasets or an API key
- Twenty independent public scenarios across physics, biology, medical imaging, economics, and climate support reproducibility testing
- The Principles Cloud is published through deterministic, checksummed builds; clients validate hashes and schemas before atomic activation with rollback support
- v1.4.2 is published from GitHub source only; the stable PyPI release remains v1.3.3, and the two workflows differ substantially, so adopters must pick deliberately
- Autonomous discovery depends on an external LLM provider (e.g., DeepSeek-V4-Pro via SiliconFlow), with credential setup and ongoing inference costs
- Remote analysis sends bounded profiles and excerpts after per-run permission; teams with strict egress requirements must review the privacy documentation
- Some demo fits have limited scope (e.g., only two held-out power settings per readout sharing one RF chain in the Rydberg project), so results do not extrapolate as general laws
- Computational reproduction requires the original inputs and recorded conditions; packaged results are evidence to assess, not independent replication
How do you install or deploy this agent?
Requires Python 3.11 or 3.12. Use a sparse checkout to skip the 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
python3 -m venv .venv
source .venv/bin/activate
python -m pip install "./Principia-v1.4.2/core-v1.4.2[asd,local]"On Windows, activate with .venv\Scripts\Activate.ps1 in PowerShell. Dependency installation requires internet access. Alternatively, install the stable v1.3.3 workflow with pip install principia-ai==1.3.3.
How do you use this agent?
- Run
principia open --working-directory ./principia-workspace --port 8142and open http://127.0.0.1:8142/ in the browser; the five demos initialize once in an empty workspace and can be browsed offline. - For your own data: choose New research, use Add data to connect a local folder, and describe your research goal.
- Open Sources & model to configure provider credentials and select reasoning and vision models (demos used DeepSeek-V4-Pro via SiliconFlow; new runs use your own configuration).
- Start discovery and review the provider permission request; View activity shows task progress and stop controls.
- Inspect observations and Rules, follow the evidence, or use Discover again to change goal, data, or models. Note: remote analysis may send bounded profiles, excerpts, or previews after per-run permission—this is not a zero-egress guarantee.
How does this agent compare with similar options?
The README explicitly contrasts Principia with conventional document-centric RAG: Principia retrieves and composes scientific objects and relations (Principles, Meta-Principles, derived hypotheses), not just passages. It also positions itself as neither a paper database nor a generic chat interface.