EU AI Act high-risk deadline in days

The paper trail for AI agents.

Your humans have paper trails. Your AI agents don't. Wrap any LangChain agent in three lines — every action becomes signed, hash-chained, independently verifiable evidence.

See the 3 lines View on GitHub

THE PROBLEM

Agents have permissions. Nobody has proof.

AI agents are starting to move money, make trades, and handle procurement. But right now there's no way to prove what they actually did. MCP handles tool connections. A2A handles messaging between agents. AP2 handles payments. But nobody handles proof that the agent actually followed its rules. That's what we built.

Until now.

SIX PRIMITIVES

The trust stack

Together, these turn “policy” from a promise into a proof.

Identity

DID

Decentralized identifiers. Who is this agent? Cryptographic binding to keys and lineage.

Covenant

Behavioral spec

What will it do — and won’t do? Signed, immutable constraints. Only narrowable, never loosened.

Attestation

W3C VC

Verifiable credentials. Prove identity and compliance without revealing internals.

Action log

Hash-chained

Every action logged. Tamper-evident. Audit trail that anyone can verify.

Verification

Deterministic

Same inputs, same result. Independent, deterministic verification—no trust in the verifier.

Enforcement

Staking / slashing

Skin in the game. Violations are costly. Stake at risk aligns behavior with commitments.

HOW IT WORKS

From action to verifiable proof

1 Agent decides action

2 Middleware intercepts

3 Evaluate covenant

4 Block or allow

5 Log to action log

6 Verifiable by anyone

SEE IT FAIL CLOSED

Agent tries $600 transfer → middleware blocks → log verifies

nobulex-demo

$ mw.execute({ action: 'transfer', params: { amount: 600 } }, ...)

Agent: Attempting transfer of $600

Middleware: BLOCKED — covenant: amount > 500

Action log: { action: "transfer", amount: 600, result: "blocked", hash: "0x7a3f..." } ✓ verifiable

Full demo →

TWO-TIER GUARANTEE

Impossible or costly

Tier 1

Impossible violation

Middleware runs in a TEE. Policy bypass is prevented inside the enforcement boundary (TEE + attestation assumptions).

Assumptions & threat model: We assume the TEE is uncompromised and correctly attests its identity. The enclave intercepts all agent actions before execution; forbidden actions never reach the host. If the enclave is compromised, an attacker could bypass enforcement — attestation lets you verify which software is running. See the Spec for details.

Tier 2

Costly violation

Stake at risk. Slashing on breach. Rational agents don’t violate when the cost exceeds the gain.

WHY NOBULEX

Not just guardrails

	Guardrails / policy engines	Proof-of-Behavior (Nobulex)
Enforcement	Best-effort; can be bypassed	Signed commitments; pre-execution middleware
Verification	Trust the operator	Third-party verifiable; anyone can audit
Consequences	Policy violation = incident	Tamper-evident proof; cryptographic evidence of breach

REGULATORY DEADLINES

The clock is ticking

Every major AI compliance framework requires tamper-evident audit trails. Regular logs won't pass.

June 30, 2026

Colorado AI Act

Accountability requirements for AI systems making consequential decisions. First U.S. state-level AI law.

August 2, 2026

EU AI Act Article 12

Requires tamper-evident automatic event logging for high-risk AI systems. Penalties up to €15M or 3% of global revenue.

November 2026

NAIC AI Evaluation

Nationwide AI evaluation tool for insurance. Carriers must prove agent compliance across every claim touched.

WORKS WITH

Plugs into your existing stack

AWS AgentCore

MCP-native — auto-discoverable in Agent Registry

Microsoft AGT

Proof layer for governance toolkit

Google A2A

Behavioral attestation for Agent Cards

LangChain

Drop-in compliance callbacks

TRACTION

Built in the open

3 linesTo integrate

2,736Tests passing

5Packages

OATRRegistered issuer

NISTRFI submitted

MIT · Open source · No vendor lock-in

FOR DEVELOPERS

Add accountability in minutes

JAVASCRIPT

import { createDID, parseSource, EnforcementMiddleware } from '@nobulex/core';

// 3 lines to add proof-of-behavior enforcement
const agent = await createDID();
const spec = parseSource(`covenant MyAgent { permit read; forbid write; }`);
const mw = new EnforcementMiddleware({ agentDid: agent.did, spec });