langchain-trust-gate

LangChain tool for behavioral trust scoring via Dominion Observatory.

Gate agent actions on real behavioral attestation data — interaction history, success rates, latency — before executing sensitive operations.

Install

pip install langchain-trust-gate

Quick start

from langchain_trust_gate import TrustGateTool, check_trust

# As a LangChain tool (drop into any agent)
tool = TrustGateTool()
result = tool.invoke({"agent_id": "sg-cpf-calculator"})
# {'decision': 'PASS', 'score': 0.82, 'raw_score': 82, ...}

# Direct async call
evaluation = await check_trust("sg-cpf-calculator")
if evaluation["decision"] == "PASS":
    # proceed with operation
    pass

With a LangChain agent

from langchain.agents import initialize_agent, AgentType
from langchain_openai import ChatOpenAI
from langchain_trust_gate import TrustGateTool

llm = ChatOpenAI(model="gpt-4")
tools = [TrustGateTool()]

agent = initialize_agent(
    tools, llm, agent=AgentType.OPENAI_FUNCTIONS, verbose=True
)

agent.run("Check the trust score of the sg-cpf-calculator agent")

Decision thresholds

Thresholds are configurable:

tool = TrustGateTool(pass_threshold=70, fail_threshold=30)

API

TrustGateTool

LangChain BaseTool subclass. Configurable fields:

check_trust(agent_id, **kwargs) -> TrustEvaluation

Async function for direct use outside LangChain.

check_trust_sync(agent_id, **kwargs) -> TrustEvaluation

Synchronous wrapper.

TrustEvaluation

TypedDict with fields:

• decision: "PASS" | "FAIL" | "UNCERTAIN"
• score: float | None — normalized 0.0-1.0
• raw_score: int | None — raw 0-100 from Observatory
• reason_code: str — machine-readable reason
• evidence_uri: str | None — link to Observatory evidence page
• provider: str — always "dominion-observatory"
• agent_id: str
• evaluated_at: str — ISO 8601 timestamp

Spec reference

• x402 Trust-Provider Interface v0.1
• Dominion Observatory

License

MIT