ledgerproof-langchain

LangChain and LangGraph adapter for the LedgerProof protocol. Emits cryptographically signed transparency receipts as a side channel, suitable for evidence of EU AI Act Article 50 disclosure obligations.

Install

pip install ledgerproof-langchain
# Optional LangGraph support:
pip install "ledgerproof-langchain[langgraph]"

Python 3.10+.

5-minute quickstart (callback handler)

from langchain_openai import ChatOpenAI
from langchain_ledgerproof import LedgerProofCallbackHandler, LogEmitter

handler = LedgerProofCallbackHandler(
    deployer_id="acme-corp-eu",
    schema="chatbot_session/v1",
    emitter=LogEmitter("./receipts.jsonl"),
)

llm = ChatOpenAI(model="gpt-4o-mini", callbacks=[handler])
llm.invoke("Hello, are you an AI?")
# Receipt appended to ./receipts.jsonl. LLM response is not modified.

See examples/01_chatbot_quickstart.py and examples/02_langgraph_editorial.py.

What this adapter does

• Captures a stream-aware SHA-256 transcript hash of the LLM interaction (commit-on-start, sign-on-end — never buffers the full body).
• Signs the receipt with Ed25519.
• Emits the receipt via a pluggable side channel (file, webhook, in-process queue).
• Validates receipt structure against one of three Pydantic schemas:
  • chatbot_session/v1 (Article 50(1) — AI interaction disclosure)
  • generated_content/v1 (Article 50(2) — AI-generated content marking)
  • human_review/v1 (Article 50(4) — editorial-control exemption)
• Supports both legacy LangChain callbacks and LangGraph node middleware.

What this adapter explicitly does NOT do

• Does not modify the LLM response payload. Receipts are side-channel only.
• Does not phone home during verification. Verification is offline against the Bitcoin chain and the published protocol public key.
• Does not provide Article 9 (risk management), Article 10 (data governance), Article 13 (transparency to users beyond Article 50), Article 15 (accuracy / robustness / cybersecurity), or Article 72 (post-market monitoring) coverage.
• Does not claim regulator endorsement.
• Does not claim Article 40 presumption of conformity (no harmonized standard has been published for Article 50 at the time of this release).
• Provides an Article 50 transparency receipt evidence layer only.

Schemas

All three schemas reject email-shaped strings in deployer_id, reviewer_role, and review_rationale fields as a GDPR safety check. Receipts should reference opaque pseudonymous identifiers, not personal data.

Phase 2+ roadmap

The MVP signer generates an ephemeral in-memory Ed25519 keypair. Production deployments will use HSM-backed signing (AWS KMS, GCP Cloud KMS, Azure Key Vault, YubiHSM). See signer.py for backend stubs.

License

Apache 2.0. See LICENSE.