langchain-alex

Verify signed ALEX Evidence Packages inside LangChain and LangGraph workflows. The verifier checks the package against a public key obtained separately from the evidence bundle and fails closed on unsupported schemas, signature failures, incomplete evidence, or inconsistent declared outcomes.

valid=True means that the evidence package passed verification and carries a verified outcome. Authentic bundles documenting failed or inconclusive runs remain rejected; inspect reason and outcome separately.

Requirements

• Python 3.10 or newer
• OpenSSL available as openssl on PATH

Install

pip install langchain-alex

Verify an Evidence Package

from langchain_alex import AlexEvidenceVerifier

verifier = AlexEvidenceVerifier.from_key_files("alex-public-key.pem")
result = verifier.verify_file("evidence-bundle.json")

print(result.valid, result.reason, result.outcome)

The trust anchor is intentionally supplied separately. A public key embedded only in the bundle is not trusted.

LangChain tool

tool = verifier.as_tool()
result = tool.invoke({"bundle": evidence_bundle})

The tool returns a JSON-serializable object with valid, reason, bundle_id, schema_version, outcome, and claim_ladder.

LangGraph node

from typing_extensions import TypedDict
from langgraph.graph import END, START, StateGraph


class State(TypedDict):
    evidence_bundle: dict
    alex_verification: dict


graph = StateGraph(State)
graph.add_node("alex_verify", verifier.as_langgraph_node())
graph.add_edge(START, "alex_verify")
graph.add_edge("alex_verify", END)
app = graph.compile()

result = app.invoke({"evidence_bundle": evidence_bundle})
assert result["alex_verification"]["valid"] is True

The node returns a state update and does not mutate its input state. Route on both valid and outcome when the graph controls a consequential action.

Development

python -m pip install -e ".[dev]"
python -m pytest
ruff check .
python -m build

The verifier implementation is also used by the repository's standalone Python CLI at tools/verify-bundle/verify.py; the integration does not maintain a second verdict algorithm.