ProofTrail — Tamper-proof receipts for AI agents

AI agents are flying blind

No proof of what happened. Agents call tools and trust the response. If a plugin is compromised or a response is tampered with, nobody knows.
Mutable logs are not evidence. Observability tools write to databases that anyone with access can edit. An audit trail you can rewrite is not an audit trail.
Regulation is coming. The EU AI Act (Article 12) mandates tamper-evident logging for high-risk AI by August 2026. Most agent frameworks have no answer.

One decorator. Cryptographic receipts.

from ads_foundation import ProofTrailMonitor

monitor = ProofTrailMonitor()

@monitor.track_tool_call
def query_db(customer_id: str) -> dict:
    return {"balance": 15420.50}

with monitor.session(agent_id="support_bot") as session:
    result = query_db("cust_987")

    receipt = session.get_latest_receipt()
    assert session.verify_receipt(receipt)  # Merkle proof checks out
    print(result)  # {"balance": 15420.50} -- decorator is transparent

Functions return the same values. But now every call is hashed into a Merkle chain with a proof you can verify independently.

Built and shipped

0 Runtime dependencies

179 Tests passing

<9 µs Proof verification

Live Hosted API

Published at ENASE 2026. Based on PhD research in authenticated data structures.

Compliance-ready

EU AI Act, Article 12 — mandates automatic, tamper-evident event recording for high-risk AI systems. Deadline: August 2026.
HIPAA — healthcare AI agents handling PHI need audit trails that can't be altered after the fact.
SOC 2 Type II — auditors want cryptographic proof that logs haven't been modified. Merkle proofs deliver that.