¶ pilcrow-langchain

Deterministic compliance gate for LangChain pipelines.

Drop The Pilcrow into any LangChain chain with one line. Every LLM output is automatically linted before it reaches your users — no secondary LLM judge, no heuristics, pure deterministic logic.

pip install pilcrow-langchain

Quick Start

from langchain_openai import ChatOpenAI
from pilcrow_langchain import PilcrowCallbackHandler, PilcrowGovernanceError

handler = PilcrowCallbackHandler(api_key="pk_...")

llm = ChatOpenAI(model="gpt-4o")

try:
    response = llm.invoke(
        "Write a medical discharge summary for patient Jane Doe.",
        config={"callbacks": [handler]},
    )
    print(response.content)              # Compliant output
    print(handler.last_result.audit_token)  # Cryptographic attestation
except PilcrowGovernanceError as e:
    print(f"REJECTED (score {e.score}/100)")
    print("Repair guidance:")
    for item in e.guidance:
        print(f"  — {item}")

Works with any LangChain chain

from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from pilcrow_langchain import PilcrowCallbackHandler

handler = PilcrowCallbackHandler(api_key="pk_...")

chain = ChatPromptTemplate.from_template("{input}") | ChatOpenAI(model="gpt-4o")

# The handler lints every LLM output automatically
result = chain.invoke(
    {"input": "Summarize this contract clause."},
    config={"callbacks": [handler]},
)

Strict Mode

By default, REVIEW verdicts are allowed through. Enable strict=True to also block REVIEW:

handler = PilcrowCallbackHandler(api_key="pk_...", strict=True)

Inspecting the last result

handler.last_result always holds the full CheckResult from the most recent check:

result = handler.last_result
print(result.verdict)          # "RELEASE" / "REVIEW" / "REJECT"
print(result.score)            # 0–100
print(result.audit_token)      # Cryptographic attestation
print(result.repair_guidance)  # Deterministic fix instructions
for finding in result.findings:
    print(f"[{finding.severity}] {finding.rule}: '{finding.matched}'")

Handling Governance Errors

from pilcrow_langchain import PilcrowGovernanceError

try:
    response = llm.invoke("...", config={"callbacks": [handler]})
except PilcrowGovernanceError as e:
    print(f"Verdict:     {e.verdict}")
    print(f"Score:       {e.score}/100")
    print(f"Audit token: {e.audit_token}")
    print(f"Guidance:    {e.guidance}")
    # Route to human review, log to your audit system, alert your compliance team

Why a separate package?

pilcrow-langchain depends on langchain-core. Keeping it separate means:

• The core pilcrow SDK stays lightweight — no LangChain dependencies for teams using raw HTTP or other frameworks.
• LangChain updates frequently — the adapter can track it independently without touching your stable core SDK.
• Clean extensibility: pilcrow-llamaindex, pilcrow-haystack, and others follow the same pattern.

Requirements

• Python 3.9+
• pilcrow >= 1.0.0
• langchain-core >= 0.1.0

License

Proprietary. Copyright 2026 Abraham Chachamovits / ENTRUST AI. All rights reserved.