Skip to main content

Audit-grade execution kernel for AI workflows - immutable trails, artifact provenance, and replay capability

Project description

Stardive

Audit-grade execution truth layer for AI workflows

PyPI version License: AGPL v3 Python 3.10+ Test Coverage Status: Alpha

Quick Links: InstallationQuick StartKey FeaturesUse CasesDocumentationGitHub


What is Stardive?

Stardive makes AI execution provable — not smarter, faster, or more automated.

Stardive is an execution & observation kernel that provides audit-grade truth for AI workflows. It captures execution provenance, stores artifacts immutably, and enables replay — without replacing your existing orchestration or agents.

Perfect for regulated industries (finance, healthcare, legal, government) that need defensible AI systems with complete audit trails.

Core Principle

Observation and provenance, not control.

We capture truth, we don't orchestrate.


Why Stardive?

Add ~10-20 lines of Python to your existing AI workflow and get:

  • Immutable audit trails - Every step, artifact, and decision recorded with hash chains
  • Complete provenance - Know exactly what produced each output, when, and by whom
  • Lineage graphs - Visual DAG of Step → Artifact → Step relationships
  • Snapshot replay - Reproduce past executions from stored artifacts
  • Non-determinism transparency - Explicit marking of non-reproducible steps (LLM calls, etc.)
  • Zero replacement - Keep your existing orchestration (LangChain, CrewAI, custom code)

Installation

pip install stardive

Requirements: Python 3.10+

New to Stardive? Start with the 5-minute quick start below, then explore the documentation.


Quick Start

5-Minute Example

Track your AI workflow with just a few decorators:

from stardive import StardiveContext

# Create audit context
ctx = StardiveContext()

# Decorate your functions
@ctx.step_meta(step_id="fetch", produces=["raw_data"])
def fetch_data():
    return {"data": [1, 2, 3]}

@ctx.step_meta(step_id="analyze", produces=["result"], depends_on=["raw_data"])
def analyze(raw_data):
    return {"sum": sum(raw_data["data"])}

# Execute with full audit trail
record = ctx.execute()

# Every step, artifact, and decision is now auditable
print(f"Run ID: {record.run_id}")
print(f"Complete audit trail stored with hash chain integrity")

That's it! You now have:

  • Immutable execution records
  • Artifact provenance
  • Lineage graphs (API access)
  • Replay capability

Alternative: Zero-Replacement Instrumentation

Already have working code? Add audit trails without refactoring:

from stardive.instrumentation import emit_run_start, emit_step_start, emit_step_end

# Wrap your existing workflow
run_id = emit_run_start(initiator={"user": "bob"})

emit_step_start(run_id, step_id="process", inputs={"data": [1, 2, 3]})
result = your_existing_function(data)  # Your code stays UNCHANGED
emit_step_end(run_id, step_id="process", outputs={"result": result})

# Full audit trail generated with zero business logic changes

Perfect for:

  • LangChain/CrewAI workflows
  • Jupyter notebooks
  • Legacy batch jobs
  • Gradual migration to full SDK

Key Features

Feature Description Status
Python SDK Metadata decorators for audit capture ✅ Alpha
Instrumentation API Event ingestion for existing workflows ✅ Alpha
Artifact Storage Deterministic serialization + SHA256 hashing ✅ Alpha
Immutable Storage Append-only SQLite backend ✅ Alpha
Hash Chain Integrity Tamper-evident audit trails ✅ Alpha
Lineage Graphs Automatic DAG construction (API) ✅ Alpha
Snapshot Replay Reproduce executions from artifacts ✅ Alpha
Audit UI Web interface for trails & lineage ✅ Alpha
PostgreSQL Backend Enterprise storage 🔜 v0.2
Framework Integrations LangChain, CrewAI adapters 🔜 v0.2

What Stardive Does NOT Do

Stardive is observation-only. We don't replace your existing tools:

  • ❌ No workflow orchestration (use your existing orchestrator)
  • ❌ No agent planning (we observe, not control)
  • ❌ No dependency inference (you declare dependencies)
  • ❌ No AI models or business logic (bring your own)
  • ❌ No correctness checking (we capture truth, not judge it)

Architecture

┌──────────────────────────────────────┐
│   Your AI Workflow / Agent           │
│   (Existing Code + Stardive SDK)     │
└──────────────────────────────────────┘
                  │
                  │ metadata + events
                  ▼
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃   STARDIVE OBSERVATION KERNEL      ┃
┃   ─────────────────────────────────┃
┃   • Capture execution truth        ┃
┃   • Hash artifacts                 ┃
┃   • Build lineage                  ┃
┃   • Store immutably                ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
                  │
                  │ immutable trail
                  ▼
┌──────────────────────────────────────┐
│  Artifacts • Lineage • Replay        │
│  (Audit Evidence, Provenance)        │
└──────────────────────────────────────┘

Audit UI

Run the web UI for audits and lineage:

# Frontend
cd frontend && npm run dev

# Backend
uv run python -m backend

Audit UI Screenshots

Run list Run info Run lineage Run replay Run export


Use Cases

Stardive is designed for regulated environments where AI decisions must be defensible:

Industry Examples

Industry Use Case Why Stardive?
Financial Services Credit scoring, fraud detection, algorithmic trading Prove model decisions to regulators (FCRA, ECOA)
Healthcare Clinical decision support, diagnosis assistance Document AI's role in patient care (HIPAA, FDA)
Legal Contract analysis, case research, e-discovery Maintain chain of custody for AI evidence
Government Benefits determination, policy analysis Transparency & accountability (APA, FOIA)
Insurance Claims processing, underwriting Audit AI decisions for fairness & compliance

Key Requirements Stardive Solves

  • Regulatory compliance: Full audit trail for GDPR, AI Act, FDA, FCRA, etc.
  • Reproducibility: Prove AI decisions can be replayed from stored artifacts
  • Accountability: Answer "How did the AI reach this conclusion?"
  • Defensibility: Provide evidence in disputes, audits, or litigation
  • Trust: Demonstrate responsible AI deployment to stakeholders

How Stardive Compares

Tool Purpose Relationship to Stardive
LangChain / CrewAI Agent orchestration Use together - Stardive observes their execution
MLflow / Weights & Biases ML experiment tracking Different focus - Stardive is for production audit trails
Apache Airflow Workflow orchestration Use together - Stardive captures provenance
OpenTelemetry Observability/telemetry Complementary - Stardive adds artifact provenance
DVC / Pachyderm Data versioning Different - Stardive tracks execution, not just data

Stardive's unique value: Immutable audit trails with artifact provenance and replay capability, specifically designed for regulated AI.


Explicit Boundaries

Stardive Will NEVER

  • Infer dependencies (user must declare)
  • Plan workflows (user provides plan)
  • Optimize execution (no auto-optimization)
  • Judge correctness (only capture truth)
  • Replace orchestrators or agents (observation only)

Stardive ONLY Does

  • Capture execution truth
  • Store artifacts immutably
  • Build lineage graphs
  • Enable replay from snapshots
  • Provide audit evidence

Development

Prerequisites

  • Python 3.11+
  • uv (recommended) or pip

Setup

# Clone the repository
git clone https://github.com/stardive/stardive.git
cd stardive/stardive-core

# Create virtual environment with uv
uv venv
source .venv/bin/activate  # or `.venv\Scripts\activate` on Windows

# Install dependencies
uv pip install -e ".[dev]"

# Run tests
pytest

# Run linters
black src tests
ruff check src tests
mypy src

Project Structure

stardive-core/
├── src/stardive/
│   ├── models/          # Canonical IR (RunPlan/RunRecord)
│   ├── sdk/             # Python SDK (context, decorators)
│   ├── instrumentation/ # Event ingestion API
│   ├── storage/         # Storage backends (SQLite)
│   ├── lineage/         # Lineage graph construction
│   ├── replay/          # Snapshot replay engine
│   └── cli/             # Legacy CLI (not used)
├── tests/
│   ├── unit/            # Unit tests (99% coverage)
│   └── integration/     # Integration tests
├── docs/                # Documentation
└── pyproject.toml       # Project config

Project Status

Current Version: 0.1.0a1 (Alpha)

✅ v0.1 Complete (375 tests, 96% coverage)

  • Canonical IR (RunPlan/RunRecord)
  • Identity & Provenance tracking
  • Artifact management with deterministic hashing
  • SQLite append-only storage
  • Python SDK with context-scoped decorators
  • Instrumentation API for zero-replacement adoption
  • Lineage graph API
  • Snapshot replay engine
  • Audit UI (web interface)

🔜 v0.2 Roadmap

  • PostgreSQL backend for enterprise scale
  • Framework integrations (LangChain, CrewAI)
  • YAML workflow compiler
  • Enhanced lineage visualization
  • Multi-tenant isolation
  • Advanced replay modes (partial, conditional)

See PROJECT_STATUS.md for detailed roadmap.

Test Coverage

Phase 2 (Core Models):    127 tests, 98% coverage
Phase 3.1 (Storage):       63 tests, 91% coverage
Phase 3.2 (Artifacts):    165 tests, 96% coverage
─────────────────────────────────────────────────
Total:                    375 tests, 96% coverage

Contributing

We welcome contributions! Stardive is open source (AGPL-3.0) and community-driven.

How to contribute:

  1. Check open issues or start a discussion
  2. Fork the repository and create a feature branch
  3. Write tests (we maintain >90% coverage)
  4. Submit a PR with clear description

Priority areas for v0.2:

  • PostgreSQL backend implementation
  • LangChain/CrewAI integration adapters
  • Documentation improvements
  • Example workflows for regulated industries
  • Performance optimizations

See CONTRIBUTING.md for detailed guidelines.


License

This project is licensed under the GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later).

See LICENSE for details.

Why AGPL?

We chose AGPL to ensure that:

  • The code remains open source
  • Cloud providers offering Stardive-as-a-service must contribute improvements back
  • Enterprises modifying Stardive must either open source changes or obtain a commercial license

For commercial licensing options, contact: jiaye@stardive.xyz


Links & Resources

Support


Screenshots

Audit Trail Visualization

View complete execution history with detailed run information:

Run List - Browse all execution runs Browse all execution runs with status, timestamps, and metadata

Run Details - Comprehensive execution information View detailed run information including steps, artifacts, and provenance

Lineage & Provenance

Visualize artifact flows and dependencies:

Lineage Graph - Visual DAG of execution flow Interactive lineage graph showing step → artifact → step relationships

Replay & Export

Reproduce executions and export audit evidence:

Snapshot Replay - Reproduce past executions Replay past executions from stored artifacts with non-determinism transparency

Audit Export - Generate compliance reports Export complete audit trails for compliance and regulatory requirements


Making AI execution provable, not perfect.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

stardive-0.1.0a2.tar.gz (1.4 MB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

stardive-0.1.0a2-py3-none-any.whl (143.6 kB view details)

Uploaded Python 3

File details

Details for the file stardive-0.1.0a2.tar.gz.

File metadata

  • Download URL: stardive-0.1.0a2.tar.gz
  • Upload date:
  • Size: 1.4 MB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.11.13

File hashes

Hashes for stardive-0.1.0a2.tar.gz
Algorithm Hash digest
SHA256 f7e374e4e3fb14a0e7c268d86244a2a87ccc4f0920f40e418c7444a24115fccb
MD5 1243da1d11b6fcddec1139134834ae07
BLAKE2b-256 1881d0aae33b203e117039736a999053f7cd97e772b0337ce11cf107266da94a

See more details on using hashes here.

File details

Details for the file stardive-0.1.0a2-py3-none-any.whl.

File metadata

  • Download URL: stardive-0.1.0a2-py3-none-any.whl
  • Upload date:
  • Size: 143.6 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.11.13

File hashes

Hashes for stardive-0.1.0a2-py3-none-any.whl
Algorithm Hash digest
SHA256 af70a11f5eb551019fce62f5ace897bc09bf1a8366c875631109549a73a25f6c
MD5 00bfc1ada4d1a3749fb4a8c0c3f01796
BLAKE2b-256 cf500b016b6c80a4cecc3325fcaf92d7a39a60a9f1438a3237442f4c48419097

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page