Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Principia

Principle-first scientific idea discovery

Grounded in evidence. Structured for scrutiny. Exported for validation.

PyPI Python versions Release v1.3.3 MIT license ICML 2026

GitHub · PyPI · Documentation · Examples · Release QA

Ideas from principles. Validated by evidence.

Principia is a local-first Python framework that turns public literature and private research materials into traceable Idea Cards, prior-art comparisons, and validation-ready research packs.

Release scope: this README documents the maintained Principia v1.3.3 framework only.

Why Principia · Workflow · Installation · Examples · Trustworthiness · Research foundations · Architecture · Contact


The thesis

Most LLM research assistants optimize for a convincing answer. Principia optimizes for an inspectable scientific object.

A serious research idea should make five things visible:

  1. Origin — which works and evidence records informed it;
  2. Principles — which mechanisms, constraints, and transferable abstractions it builds on;
  3. Construction — how those ingredients became the proposed method;
  4. Risk — which assumptions, failure modes, and prior-art overlaps may invalidate it;
  5. Testability — what experiment, baseline, metric, and falsification path should come next.

Principia therefore treats an idea as a typed research contract rather than a paragraph:

Idea Card
  = selected evidence
  + reusable principles
  + explicit mechanism
  + novelty contrast
  + assumptions and risks
  + validation contract

The goal is not to generate more ideas. The goal is to generate ideas whose provenance, scientific logic, uncertainty, and next experiment are visible.

Why this changes the workflow

Conventional workflow Principia v1.3.3
Ask an LLM to brainstorm from a prompt Build an exact evidence packet before generation
Retrieve whole documents or opaque chunks Extract typed ideas, principles, takeaways, comparators, evaluation contexts, and result facts
Accept a fluent one-shot proposal Run a strict candidate–critique–evolution–selection process
Trust citations emitted by the model Resolve every citation to a canonical (work_id, kind, record_id)
Hide provider failures behind generic output Surface failures, allow one grounded repair, then fail closed
Leave the result in chat history Export a portable, versioned research pack
Treat validation as future work Make baselines, metrics, risks, and falsification part of the output

From a research goal to a validation pack

Research goal
    │
    ├── Public literature
    │     arXiv · OpenAlex · Crossref · Semantic Scholar · Europe PMC
    │
    └── Private corpus (optional)
          PDF · Office · Markdown · LaTeX · code · structured text
                │
                ▼
      Cross-domain retrieval and identity reconciliation
                │
                ▼
      Structured scientific feature extraction
      ideas · principles · takeaways · baselines · benchmarks · result facts
                │
                ▼
      Canonical evidence packet
                │
                ▼
      SciDialect-Evo: 3 candidates → 2 evolved candidates → 1 final Idea Card
                │
                ▼
      Prior-idea comparison
                │
                ▼
      Deterministic validation plan + portable seven-file research pack

Principia is deliberately staged. Retrieval, extraction, evidence selection, generation, comparison, and export remain separately inspectable and resumable.


What ships in v1.3.3

Layer Capability Why it matters
Cross-domain retrieval Query planning, arXiv/OpenAlex/Crossref/Semantic Scholar access, automatic Europe PMC routing for biomedical topics, bounded retries, identity reconciliation, BM25 or embedding reranking A research idea should not be limited to one source, one naming convention, or one surface vocabulary
Private research context Recursive local ingestion with portable local:// identities; core support for PDF, HTML/XML, Markdown, RST, LaTeX, text/code, JSON, YAML, CSV/TSV; optional DOCX/PPTX/XLSX support Public literature can be combined with private notes, unpublished drafts, internal reports, and project-specific evidence
Structured extraction Typed prior ideas, principles, takeaways, comparators, evaluation contexts, and grounded result facts with provenance and content fingerprints Principia reasons over scientific objects, not only document chunks
Canonical evidence Exact record-level citations, source-text hydration, mixed public/private evidence checks, configurable global evidence budgets The generator cannot silently invent the evidence it claims to use
Strict SciDialect-Evo Three distinct proposals, explicit scoring, evolution of the strongest two, final selection, one evidence-grounded repair, fail-closed persistence Idea generation becomes a controlled search process rather than a single completion
Prior-art comparison Content-level shortlisting plus model-assisted comparison of mechanistic similarity, essential difference, potential advantage, and weakness Novelty is treated as a contrastive claim, not a self-awarded label
Validation hand-off Human-readable and JSON validation plans generated from the final Idea Card and canonical evidence without another LLM call The research hypothesis and the experiment contract stay synchronized
Controllable jobs Weighted progress, checkpoints, pause/resume/stop controls, notebook widgets, terminal display, event history Long research runs remain observable and controllable while the worker process is active
Portable state Shared project workspace, per-idea output folders, SQLite migrations, path-safe exports Research memory can compound without duplicating the evidence pool

Installation

Principia supports Python 3.10–3.13. The distribution name is principia-ai; the import package is principia.

python -m pip install principia-ai==1.3.3

Add Office-document and notebook support when needed:

python -m pip install "principia-ai[local,notebook]==1.3.3"

The examples below use SiliconFlow through an OpenAI-compatible API. Principia can also be configured with pc.LLMConfig(...) for other OpenAI-compatible endpoints.

export SILICONFLOW_API_KEY="your-key"

# Required to use OpenAlex under its current authentication policy.
export OPENALEX_API_KEY="your-openalex-key"

Never commit credentials to source files, notebooks, .env files, or exported artifacts.


Quick start

One-call research workflow

import os
import principia as pc

GOAL = (
    "Develop an evidence-grounded method for improving long-horizon "
    "reasoning efficiency in LLM agents under a fixed token budget."
)

ws = pc.Workspace.project(
    "principia_project",
    llm_config=pc.siliconflow_config(
        os.environ["SILICONFLOW_API_KEY"],
        max_calls=220,
    ),
)

job = ws.start(
    GOAL,
    pipeline_config=pc.PipelineConfig.research(),
)

result = job.result()
result.show()

PipelineConfig.research() is the opinionated v1.3.3 research preset:

  • exactly 50 public works requested;
  • embedding reranking requested and reported explicitly;
  • strict failure if the public target cannot be completed;
  • an exact 15-record evidence packet: 5 ideas, 5 principles, and 5 takeaways;
  • no more than 2 selected records from one work;
  • strict scidialect-evo generation.

The preset is intentionally demanding. Every field can be overridden through PipelineConfig, RetrievalConfig, or the staged API.

Add a private corpus

import os
import principia as pc

ws = pc.Workspace.project(
    "principia_project",
    llm_config=pc.siliconflow_config(
        os.environ["SILICONFLOW_API_KEY"],
        max_calls=220,
    ),
    allow_remote_private_content=True,
)

job = ws.start(
    "Your research objective",
    documents="private_sources",
    pipeline_config=pc.PipelineConfig.research(),
)

result = job.result()
result.show()

Private documents are supplemental: they never silently replace the requested public-literature target. Original files are not copied into the project. Absolute paths remain in hidden local state and are excluded from prompts and shareable exports.

Sending private text to a remote model requires explicit allow_remote_private_content=True. The provider receives document content and portable identifiers, not local absolute paths.

See Private corpus ingestion for parser limits, chunking, diagnostics, deduplication, custom parser registration, and cache cleanup.


Inspect every stage

The high-level workflow is concise, but the research process remains fully programmable:

works = ws.research.search(
    GOAL,
    target_count=50,
    rerank_mode="embedding_rerank",
    require_target=True,
    show_progress=True,
)

features = ws.research.extract(
    works,
    model="auto",
    show_progress=True,
)

evidence = pc.select_evidence(
    features,
    global_kind_limits={
        "ideas": 5,
        "principles": 5,
        "takeaways": 5,
    },
    max_per_work=2,
    require_exact=True,
    user_note=GOAL,
)

idea = ws.ideas.generate(
    evidence,
    user_note=GOAL,
    mode="scidialect-evo",
    model="auto",
    show_progress=True,
)

comparison = ws.ideas.compare(
    idea,
    features,
    model="auto",
    show_progress=True,
)

This staged interface is useful when you need to inspect retrieval diagnostics, curate an evidence packet, replace a model, modify a generation budget, or stop before an expensive stage.


Control long-running jobs

status = job.status()
print(status.status, status.stage, status.progress)

job.pause()   # pause at the next safe provider boundary
job.resume()
job.stop()    # schedule no further work

for event in job.events():
    print(event)

The same controls are available from the workspace or CLI while the original worker process remains active. Keep the persisted run ID so another notebook cell or terminal can address that worker:

run_id = job.run_id
ws.status(run_id)
ws.pause(run_id)
ws.resume(run_id)
ws.stop(run_id)

Pause and stop are cooperative. An in-flight bounded provider request may finish and checkpoint, but no subsequent paid call begins after the relevant control boundary.

Run status, events, and completed stage checkpoints persist in SQLite. A Python PipelineJob itself is thread-backed, however: reopening a workspace after its worker process has exited does not recreate that thread or restart the remaining pipeline automatically.


What Principia returns

Workspace.project(...) separates reusable research state from idea-specific outputs:

principia_project/
  workspace/
    works.json
    features.json
    manifest.json
    .principia/
  outputs/
    <idea_id>/
      idea.md
      idea.json
      evidence.json
      comparison.json
      result.json
      validation_plan.md
      validation_plan.json

The Idea Card

A generated Idea can include:

  • title and one-sentence thesis;
  • novelty claim and closest conceptual contrasts;
  • mechanism design and methodological details;
  • formulas and symbols when supported by evidence;
  • method variants;
  • validation protocol;
  • baselines, metrics, risks, and assumptions;
  • derived principles;
  • exact source-evidence references;
  • candidate, critique, evolution, selection, and repair metadata.

The validation pack

The validation plan is derived deterministically from the final Idea Card and selected evidence. It does not trigger another model call. Its Markdown and JSON forms preserve:

  • the research goal and thesis;
  • the proposed protocol;
  • comparators and baselines;
  • success and failure metrics;
  • risks and assumptions;
  • canonical evidence references;
  • model, mode, schema version, and timestamps.

The result is designed to move cleanly into experiment repositories, coding agents, review workflows, or reproducibility bundles.


Live v1.3.3 examples

The release contains three compact, output-bearing acceptance showcases across AI, computer vision, and physics.

Each originating run used 50 public works + 5 local documents, produced 55 feature bundles, selected an exact 15-record evidence packet, ran live non-degraded scidialect-evo, compared the final idea with extracted prior ideas, and exported seven portable artifacts.

These examples demonstrate the framework's retrieval, grounding, generation, comparison, mathematics, privacy, and export contracts. Their Idea Cards are generated research hypotheses, not experimentally confirmed discoveries.

The excerpts below are rendered from the accepted live Idea Cards. Only Markdown wrapping and deterministic LaTeX notation normalization are applied; no scientific claim is rewritten.

The following acceptance metrics are generated directly from the verified showcase manifests:

Task Online Local Features Evidence Mode Validation
Communication-efficient LLM multi-agent reasoning 50 5 55 15 scidialect-evo passed
Uncertainty-aware sparse-view dynamic 3D reconstruction 50 5 55 15 scidialect-evo passed
Broadband squeezed-state axion sensing 50 5 55 15 scidialect-evo passed
Research task Generated Idea Card Review
Communication-efficient LLM multi-agent reasoning Entropy-Constrained Discrete Codebook with Counterfactual Decoding and Diversity-Aware Calibration Notebook · Showcase
Uncertainty-aware sparse-view dynamic 3D reconstruction AnchorSplat-Dynamic: Sparse Anchor-Based Uncertainty for Uncalibrated Motion Notebook · Showcase
Broadband squeezed-state axion sensing Dynamic Heuristic Optimization of Squeezed-State Haloscopes with Continuum Noise Modeling Notebook · Showcase
Example 1 — Communication-efficient LLM multi-agent reasoning

Generated thesis

Imposing a per-task entropy floor on learned discrete messages prevents representational collapse while achieving token efficiency, provided that a counterfactual decoding protocol verifies causal interpretability and the entropy target is calibrated using real-time embedding diversity metrics.

Mechanism

Agents map internal states to indices in a shared codebook. The loss function includes a term penalizing codebook usage below a calculated entropy threshold ($H_{\mathrm{target}}$). $H_{\mathrm{target}}$ is adjusted if pre-training diagnostics show agent embedding cosine similarity exceeds 0.88, signaling potential collapse. A separate decoder module must reconstruct the original observation from the code index; if reconstruction fails under minimal input perturbations (counterfactuals), the dialect is rejected.

$$L_{\mathrm{total}} = L_{\mathrm{task}} + \lambda L_{\mathrm{tokens}} - \beta \max(0, H_{\mathrm{target}} - H(C))$$

Validation contract

Compare against standard CoT, uncompressed multi-agent baselines, and static codebook variants. Measure token reduction, accuracy, codebook entropy, success rate on counterfactual decoding tests, and embedding overlap. Specifically test scenarios where $S_{\mathrm{cos}}$ approaches 0.88 to verify the dynamic calibration of $H_{\mathrm{target}}$ prevents performance degradation.

The exported card used canonical evidence from five works and was compared against three extracted prior ideas.

Example 2 — Sparse-view dynamic 3D reconstruction

Generated thesis

By decoupling Gaussian primitives from the 2D pixel grid and anchoring them to a sparse set of 3D geometric proxies, we jointly optimize pose, motion, and heteroscedastic uncertainty for uncalibrated dynamic scenes, reducing redundancy in static regions while focusing capacity on moving objects.

Mechanism

A feed-forward encoder predicts sparse 3D anchors and associated Gaussian parameters (appearance, opacity, motion vectors, covariance). A differentiable renderer projects these into input views. A geometric prior module enforces epipolar and motion consistency constraints. An uncertainty head predicts per-Gaussian variance, gating gradient flow during test-time refinement to prevent overfitting to noise in uncalibrated views.

$$\mathcal{L}{\mathrm{total}} = \sum{i \in \mathcal{A}} (1 - \sigma_{u,i}) \cdot \mathcal{L}{\mathrm{render}}(i) + \lambda \mathcal{L}{\mathrm{geo}}(i)$$

Validation contract

Train on synthetic dynamic scenes (6–12 views). Evaluate on real-world uncalibrated sequences. Metrics: PSNR/SSIM, ECE, Risk-Coverage curves. Baselines: Dense pixel-aligned GS, Pose-dependent LRMs. Pass/Fail: $\mathrm{ECE} \le 0.10$; masking top 20% uncertain Gaussians must reduce median error.

The exported card used five canonical records across four works and was compared against five extracted prior ideas.

Example 3 — Broadband squeezed-state axion sensing

Generated thesis

Integrating real-time heuristic search algorithms with Josephson Parametric Amplifiers allows for dynamic stabilization of squeezed states against environmental drift, while modeling noise propagation as a continuum through lossy transmission lines maximizes broadband scan speed and maintains false-positive controls via configuration-independent rejection.

Mechanism

A feedback loop where a heuristic search algorithm continuously adjusts JPA pump frequency and power based on real-time noise spectral density measurements derived from a continuum noise model. The system employs a dual-readout chain to enforce configuration-independent signal rejection, ensuring candidates are physical resonator responses rather than readout artifacts.

$$\text{SNR}^{2} = T_{\mathrm{int}} \int \frac{S_{\mathrm{ax}}^{2}(f)}{S_{\mathrm{noise}}(f)} df$$

Validation contract

Validate squeezing advantage using reported metrics; verify candidate signals appear in both readout chains with predicted coherence width; test rejection of signals correlating with configuration changes; compare continuum noise model predictions against measured noise spectra.

The exported card used five canonical records across five works and was compared against three extracted prior ideas.


Trustworthiness is an architecture, not a disclaimer

Principia does not treat provenance and privacy as presentation features. They are enforced at the data-model and persistence layers.

Invariant v1.3.3 behavior
Record-level evidence Every citation must resolve to one canonical (work_id, kind, record_id) tuple. Quoted content is hydrated from the selected record rather than trusted from model output.
Evidence before generation The generator receives an explicit EvidencePacket; configuration, traces, warnings, and usage metadata cannot become scientific evidence.
Fail-closed generation Live output receives at most one evidence-grounded repair. If canonical references, grounding, required fields, or mathematical constraints remain invalid, the idea is not persisted.
No invisible live fallback A failed live model call is surfaced. Deterministic mock output exists only as an explicitly labelled mock_fixture for tests and cannot satisfy release showcase gates.
Private by explicit consent Local originals are not copied; absolute paths remain hidden; exported identifiers are portable; remote processing of private text requires explicit authorization.
Prompt-injection boundary Retrieved and local documents are delimited as untrusted research data; embedded instructions are not treated as executable directives.
Observable retrieval Diagnostics expose query plans, source health, retry outcomes, result counts, degraded-source warnings, ranking traces, and the rerank mode actually applied.
Strict mathematics Retained formulas are normalized to canonical LaTeX, structurally validated, and included in strict KaTeX release checks.
Deterministic hand-off ValidationPlan is generated from the accepted Idea Card and evidence registry without another probabilistic model call.

Principia's grounding checks are recoverability checks, not a truth certificate. They ask whether canonical evidence references and the task-critical anchors needed to audit a proposal remain recoverable. Scientific validity still requires external evaluation.


Release confidence

The v1.3.3 release tree records a fail-closed acceptance process. Highlights include:

  • the complete regression suite is rerun on Python 3.10, 3.12, and 3.13 against the final source and showcase artifacts before publication, with exact results recorded in the release QA report;
  • CI coverage configured for Python 3.10–3.13;
  • Ruff and mypy checks passed;
  • wheel and source distribution build and twine check passed;
  • clean-environment core and [local] installation smokes passed;
  • strict source, export, notebook, LaTeX, and KaTeX audits passed;
  • archive scans checked credentials, authorization headers, private paths, private sentinels, and stale artifacts;
  • all three live research showcases passed their retrieval, evidence, generation, comparison, privacy, mathematics, and export gates.

See the complete Principia 1.3.3 Release QA report for exact commands, thresholds, warnings, and artifact checksums.


Core research foundations

Principia is not merely a retrieval wrapper. It is a downstream research system built around a deeper question:

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

Two related research lines shape the framework.

1. Machine Dialectology

Machine Dialectology studies how heterogeneous LLM agents can create, inherit, exchange, and route compact machine-oriented languages.

Its concrete ICML 2026 precursor is Communicative Language Symbolism Routing (CLSR):

CLSR treats a Language Symbolism Framework (LSF) as a reusable protocol rather than a shorter prose prompt. An LSF contains compact symbols, usage rules, validity constraints, and a message-passing contract. The system:

  1. samples seed exemplars;
  2. lets LLM agents invent candidate LSFs;
  3. generates LSF-conditioned responses;
  4. selects correct and token-efficient traces;
  5. evolves the protocol pool through propose → evaluate → select → mutate cycles;
  6. routes, ensembles, or composes LSFs at inference time.

The ICML paper reports a 3–6× reduction in latency-oriented generated-token completion relative to standard Chain-of-Thought while maintaining accuracy across the studied benchmarks.

The broader Machine Dialectology agenda generalizes this idea from one model family to heterogeneous machine societies. Its central insight is that machine-to-machine reasoning does not have to inherit the full rhetorical overhead of human prose. Agents can develop task-conditioned protocols whose intermediate states are denser, reusable, and routable.

Principia's connection: scientific ideation also has an intermediate-language problem. Mechanisms, trade-offs, assumptions, analogies, and falsification rules must be represented before they can be composed. Machine Dialectology provides the conceptual substrate for compact symbolic handles and reusable operators that help an agent move from literature evidence to structured, testable hypotheses.

2. SciDialect

SciDialect: Symbolic Compression as an Intrinsic Reward for Scientific Discovery Agents (Zhengqi Pei, Qingming Huang, and Shuhui Wang; technical manuscript, 2026) studies a complementary problem: compact symbolic representations are useful only when their scientific meaning survives compression.

Scientific agents often receive sparse external rewards. A hypothesis may not be rewarded until an experiment finishes; a coding approach may not be rewarded until hidden tests run; a research trajectory may not be rewarded until review. SciDialect introduces a denser representation-level signal: grounded symbolic compression.

A symbolic state earns credit only when it becomes shorter while preserving the task-critical anchors needed for independent reconstruction and audit.

An evidence anchor has the conceptual form:

anchor = (type, normalized value, source or location, role in the task)

Anchors may include variables, units, regimes, time windows, effect directions, equations, function signatures, dataset paths, output contracts, visible tests, prior-work links, or experimental controls.

A scientific dialect primitive is not an arbitrary abbreviation. It carries:

symbol
+ natural-language definition
+ argument schema
+ evidence slots
+ validity conditions
+ anti-definition of nearby invalid meanings
+ decoder template

Independent decoders must reconstruct the proposed meaning without access to the original verbose proposition. A skeptical reviewer checks missing anchors, added assumptions, semantic drift, invalid arguments, unsupported evidence, and dictionary bloat.

The intrinsic objective can be summarized as:

$$R_{\mathrm{SD}}=w_{\mathrm{c}}G_{\mathrm{comp}}+w_{\mathrm{r}}S_{\mathrm{rec}}+w_{\mathrm{g}}S_{\mathrm{ground}}+w_{\mathrm{a}}S_{\mathrm{agree}}+w_{\mathrm{u}}U_{\mathrm{task}}-w_{\mathrm{b}}B_{\mathrm{dict}}.$$

The terms reward compression, reconstruction fidelity, evidence grounding, decoder agreement, and optional development utility while charging dictionary growth. Reconstruction, grounding, agreement, dictionary-cost, leakage, and task-validity checks act as hard feasibility gates: a highly compressed but ungrounded state is never allowed to guide the solver.

This creates three important principles:

  • No-free compression: a short private identifier has no scientific value when its definition cost and lost meaning are ignored;
  • Amortized abstraction: a primitive is valuable only when reuse savings exceed definition, redundancy, and checking costs;
  • Anchor-gated auditability: accepted states retain an explicit lower bound on recoverable task-critical anchors under the chosen grounding rule.

SciDialect does not certify that a claim is true. It certifies a stricter and more useful intermediate property: the compressed representation remains public enough to reconstruct, grounded enough to audit, and reusable enough to justify its symbolic cost.

How v1.3.3 operationalizes this line

Principia v1.3.3 implements a strict idea-evolution protocol inspired by this research:

3 mechanistically distinct candidate ideas
        ↓
score novelty · grounding · feasibility · discriminability
        ↓
evolve the strongest 2 with explicit critiques and recorded changes
        ↓
select 1 final evolved candidate with a rationale
        ↓
validate canonical evidence and scientific fields
        ↓
one grounded repair or fail closed

The broader SciDialect manuscript studies reward-conditioned symbolic scientific memory across several task adapters. Principia v1.3.3 implements the evidence-grounded idea-generation protocol relevant to this framework; it does not claim to reproduce every adapter or experiment in the broader manuscript.

How the two ideas fit together

Research question Research line Principia consequence
What machine-oriented representations can agents invent, evolve, and route? Machine Dialectology / CLSR Use reusable symbolic protocols and task-conditioned operators rather than relying only on verbose prose
When should a compressed scientific representation be trusted? SciDialect Motivate Principia's canonical-evidence and audit checks; decoder-agreement and dictionary-cost objectives remain outside v1.3.3
How can those representations become useful research outputs? Principia Convert an evidence packet into a traceable Idea Card, contrast it with prior ideas, and export a falsifiable validation contract

Together, they move scientific-agent design from “generate a plausible proposal” toward “evolve an auditable representation that can survive comparison and experimentation.”


Technical architecture

Principia v1.3.3 uses typed objects and explicit boundaries rather than a monolithic agent loop.

Component Responsibility
principia_retrieval Source-aware query planning, provider clients, identity reconciliation, ranking, embeddings, retries, and diagnostics
principia.research Public search, private ingestion, source acquisition, structured extraction, and checkpoint reuse
principia.features Scientific feature schemas, evidence selection, canonical registries, validation, and source-text hydration
principia.ideas Standard, calculus-compatible, and strict SciDialect-Evo generation; prior-idea comparison; fail-closed grounding checks
principia.validation Deterministic validation-plan construction and Markdown/JSON rendering
principia.pipeline Persisted end-to-end orchestration, weighted progress, pause/resume/stop, and result assembly
principia.storage SQLite state, migrations, manifests, content fingerprints, portable artifacts, and private-cache controls
principia.math LaTeX normalization, structural validation, and release-quality mathematical checks
principia.cli Search, extraction, generation, run inspection, and run control from the terminal

Public models, retrieval dataclasses, and job handles are importable directly from principia, including WorkItem, WorkFeatures, EvidencePacket, Idea, IdeaComparison, PipelineResult, ValidationPlan, RetrievalConfig, SciDialectConfig, SearchDiagnostics, and PipelineJob.


CLI

The package installs a principia command.

# Initialize and inspect a workspace
principia --workspace ./principia_project/workspace init
principia --workspace ./principia_project/workspace status

# Search public research metadata
principia --workspace ./principia_project/workspace \
  search "evidence-efficient scientific agents" \
  --target-count 20 \
  --rerank-mode embedding_rerank \
  --require-target

# Search, extract, and generate one strict SciDialect-Evo idea
principia --workspace ./principia_project/workspace \
  generate "evidence-efficient scientific agents" \
  --target-count 20 \
  --rerank-mode embedding_rerank \
  --require-target

# Inspect and control persisted runs
principia --workspace ./principia_project/workspace runs
principia --workspace ./principia_project/workspace status RUN_ID
principia --workspace ./principia_project/workspace pause RUN_ID
principia --workspace ./principia_project/workspace resume RUN_ID
principia --workspace ./principia_project/workspace stop RUN_ID

Use --mock-llm only for deterministic smoke tests. Mock output is explicitly labelled and is not scientific evidence.


Documentation


Development

git clone https://github.com/pzqpzq/Principia.git
cd Principia/Principia-v1.3

python -m pip install -e ".[dev,local,notebook]"

python -m ruff check src tests scripts
python -m mypy src
python -m pytest -q
python -m build --no-isolation
python -m twine check dist/*

Contributions are especially valuable in retrieval adapters, scientific schemas, evidence validation, domain-specific evaluation, private-document parsers, reproducibility tooling, and expert review of generated Idea Cards.


Responsible interpretation

Principia is a research framework, not an oracle.

  • Generated Idea Cards are hypotheses, not established results.
  • Canonical evidence improves provenance; it does not guarantee that a source is correct.
  • Prior-art comparison reduces obvious duplication risk; it does not replace a complete scholarly review or legal patent search.
  • Validation plans structure the next experiment; they do not predict its outcome.
  • Private documents should be sent to a remote model only under an appropriate data-governance policy.
  • Scientific claims should be accepted only after independent empirical, theoretical, or expert validation.

The intended standard is simple: every persuasive claim should be paired with an inspectable source, an explicit assumption, or a test that could prove it wrong.


Citation

To cite Principia as software:

@software{principia2026,
  title   = {Principia: Principle-First Scientific Idea Discovery},
  author  = {{Principia Contributors}},
  year    = {2026},
  version = {1.3.3},
  url     = {https://github.com/pzqpzq/Principia}
}

To cite the CLSR / Machine Dialectology research foundation:

@inproceedings{pei2026lsf,
  title     = {When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning},
  author    = {Pei, Zhengqi and Huang, Qingming and Wang, Shuhui},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}

Contact

Academic collaboration
In collaboration with the Institute of Computing Technology, Chinese Academy of Sciences.
Contact: peizhengqi22@mails.ucas.ac.cn

Business collaboration
In collaboration with Beijing Chipflow Technology Co., Ltd.
Contact: peizhengqi@chipflow.net


License

The principia-ai v1.3.3 framework is released under the MIT License.

The upstream multi-project repository contains separately scoped material. See UPSTREAM.md for the exact integration and license boundary.


Build ideas whose evidence, mechanism, and falsification path can be inspected.

If Principia is useful to your research, consider starring the repository and following the next releases in Principia, Machine Dialectology, and SciDialect.

About

Principia extracts reusable principles, composes those principles into traceable research ideas, and helps researchers inspect why an idea may be worth testing.

Topics

Resources

Stars

441 stars

Watchers

13 watching

Forks

Releases

Packages

Contributors

Languages