OutputProof

AI Agent Output Verification Platform — infrastructure for trusting AI work product.

OutputProof is a verification layer that sits between AI agents and downstream consumers. It asserts, scores, and logs output correctness before generated work is trusted or acted upon.

Packages

Package	License	Description
outputproof-sdk	Apache 2.0	Core SDK: assertion engine, LLM-as-judge, CLI, MCP/LangChain integrations
outputproof-server	BSL 1.1	Dashboard server: verification history, SQLite storage, agent/team reliability scoring

Install From Source

OutputProof is in pre-release development and is not published to PyPI yet. Install both packages from this repository:

git clone https://github.com/IntuitiveDesigns/outputproof.git
cd outputproof

python -m pip install -e .\outputproof-sdk
python -m pip install -e .\outputproof-server

Quick Start

Run a local verification:

cd outputproof-sdk

@'
assertions:
  - type: output_matches
    pattern: authenticated
'@ | Set-Content demo-assertions.yaml

python -m outputproof.cli.main verify `
  --agent-id demo-agent `
  --prompt "Create auth" `
  --output "authenticated" `
  -a demo-assertions.yaml

Populate the dashboard:

# Terminal 1
cd outputproof-server
python -m outputproof_server.cli --port 8080

# Terminal 2
cd outputproof-sdk

@'
assertions:
  - type: output_matches
    pattern: authenticated
'@ | Set-Content demo-assertions.yaml

$env:OUTPUTPROOF_SERVER_URL = "http://127.0.0.1:8080"
python -m outputproof.cli.main verify --agent-id demo-agent --prompt "Create auth" --output "authenticated" -a demo-assertions.yaml

Open http://127.0.0.1:8080. You should see demo-agent, a PASS verdict, and the dashboard counters update.

GitHub Actions Gate

OutputProof v1.1 adds a PR gate that exits non-zero when changed agent output fails verification. Add assertion rules such as .outputproof/github-gate.yaml, then run the gate from a workflow:

- name: OutputProof GitHub Actions gate
  env:
    OUTPUTPROOF_AGENT_ID: ${{ github.actor }}
    OUTPUTPROOF_DEVELOPER_ID: ${{ github.actor }}
    OUTPUTPROOF_TASK_TYPE: code_generation
    OUTPUTPROOF_BASE_REF: origin/${{ github.base_ref || 'main' }}
  run: python -m outputproof.cli.main github-gate --assertions .outputproof/github-gate.yaml

See .github/workflows/outputproof-gate.yml for a complete workflow. The gate writes a GitHub job summary, stores the verification locally, syncs to OUTPUTPROOF_SERVER_URL when configured, and blocks the PR unless the verdict is PASS.

Python SDK

from outputproof import verify, assertions as a


@verify(
    assertions=[
        a.output_matches("authenticated"),
        a.function_present("authenticate_user"),
    ],
    retry_on_fail=False,
)
async def generate_auth_module(prompt: str) -> str:
    return """
def authenticate_user(username: str, password: str) -> str:
    return "authenticated"
"""

Documentation

• SDK documentation
• Dashboard server documentation
• Commercial boundary

Search Keywords

AI agent verification, LLM output validation, LLM-as-judge, agent reliability, AI coding agent testing, LangChain verification, Claude Code MCP, Model Context Protocol, FastAPI dashboard, policy engine, local-first AI governance, output quality gates, Python SDK, CI guardrails.

For GitHub discovery, set repository topics in the GitHub About panel:

ai, llm, ai-agents, agent-verification, llm-evaluation, llm-as-judge,
output-validation, langchain, mcp, fastapi, python, pytest, ai-governance,
developer-tools, ci-cd

License

• Core SDK: Apache 2.0
• Dashboard server: Business Source License 1.1

---

A product of StreamKernel LLC · outputproof.io · steven.lopez@streamkernel.io