Build on
Rubric Protocol.

Integrate post-quantum AI attestation into any system in minutes. Every AI decision your system makes becomes cryptographically signed, Merkle-batched, and permanently anchored to Hedera Consensus Service.

Installation

Install the TypeScript SDK via npm, or the Python SDK via pip.

First Attestation

The following example attests an AI agent output to Hedera in under 5 minutes.

Attestation at the
decision point.

Drop Rubric into Claude Desktop, Claude Code, Cursor, Windsurf, Continue, or Zed — every agent tool call becomes a post-quantum signed, Merkle-rooted, Hedera-anchored audit record.

Installation

Choose the port that matches your runtime. Both expose identical tools and return the same attestation IDs.

8 Tools

The full surface exposed by the MCP server. register_agent, status, framework_detect, and cost_estimate work without an API key.

• attest — Attest an AI decision. HCS-anchored with an API key; PQ-signed local Merkle leaf without.
• verify — Verify an attestation by ID. Checks local store first, then the federation.
• get_proof — Generate ZK Merkle inclusion proof (Noir / Barretenberg).
• register_agent — Request a free developer API key via email.
• status — Federation health across US, SG, JP, CA, EU nodes + ZK node.
• framework_detect — Auto-detect applicable regulatory frameworks (EU AI Act, SR 11-7, HIPAA, NIST AI RMF, etc.). Offline.
• cost_estimate — Estimate monthly Rubric cost from expected decision volume. Offline.
• bundle_query — Query attestation bundles by leaf type, agent, or time range. Standard+ tier.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (Mac) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

Claude Code

Cursor

Add to ~/.cursor/mcp.json or project-local .cursor/mcp.json:

REST API

All endpoints are available at https://rubric-protocol.com/verify/. Authentication via x-api-key header.

Attestation Endpoints

Enterprise Endpoints

Verification Endpoints

Batch & Status

Agent Management

Post-Quantum
Security Model

Rubric Protocol is built on NIST-standardized post-quantum cryptography. Every attestation is resistant to both classical and quantum adversaries — permanently.

ML-DSA-65 Digital Signatures

ML-DSA-65 (Module-Lattice Digital Signature Algorithm, formerly CRYSTALS-Dilithium) is the NIST post-quantum digital signature standard finalized in FIPS 204. Rubric uses it as the sole signing algorithm for all attestations.

Keypairs are generated once per node, encrypted at rest with AES-256-CBC using a passphrase, and never transmitted. The public key is published to HCS on initialization. All attestation batches are signed before HCS submission.

Merkle Forest Architecture

Individual attestations are not written to HCS one-by-one — that would be prohibitively expensive at scale. Instead, Rubric constructs a Merkle forest: a set of individual Merkle trees, one per pipeline or batch group, whose roots are then combined into a single forest root.

This means a batch of 100,000 attestations produces exactly one HCS write. The forest root represents the entire batch cryptographically — any individual leaf can be proven against it.

Each leaf includes a nullifier — a unique value that prevents duplicate attestation attacks. Direction bits on proof paths ensure every leaf position is unique and tamper-evident.

Hedera Consensus Service Anchoring

The forest root and signature hash are submitted to HCS topic 0.0.10416909. Hedera's hashgraph consensus provides an immutable, globally-ordered timestamp — the attestation exists permanently in the public record.

Verification requires only the attestationId. Rubric maintains a local O(1) index for instant lookups. For fully independent verification, any party can query the HCS mirror node directly — no reliance on Rubric infrastructure required.

Zero-Knowledge
Proof Layer

Rubric Protocol includes a production-deployed ZK layer for cryptographic Merkle inclusion proofs. Every attestation batch can be proven with cryptographic Merkle inclusion proofs. Full payload privacy proofs are live. Every attestation is encrypted with a customer-held AES-256-GCM key — Rubric stores only ciphertext. Customers retrieve plaintext using their key via /v1/payload-decrypt. ZK proofs over committed payload values (proving compliance without revealing decision data) are in active development.

How It Works

At every tier-1 flush, the attestation batch is forwarded to the dedicated ZK node. The Noir circuit computes a cryptographic Merkle inclusion proof over the batch. The proof is attached to the HCS submission alongside the ML-DSA-65 signature.

Any third party can verify that a specific attestation is included in a batch without seeing any other attestation in that batch. The proof is self-contained — no Rubric infrastructure required for verification.

ZK Roadmap

Your AI Knows What
It Decided. Can It Prove It?

Existing AI systems are optimized to make decisions — not to prove them. Rubric closes that gap.

Every major framework.
One integration.

Rubric is post-quantum AI compliance infrastructure. Declare your system's intended use and jurisdiction — Rubric computes your full regulatory obligation set automatically and generates cryptographically verifiable evidence packages for every applicable framework.

The section below provides a technical deep-dive on EU AI Act Article 12 — the record-keeping obligation that underpins Rubric's core attestation architecture. All other frameworks build on this foundation.

Article 12
Compliance Guide

Official text: EU AI Act — Regulation (EU) 2024/1689, Article 12 ↗

The EU AI Act requires providers and deployers of high-risk AI systems to maintain detailed records of system operation. Article 12 mandates automatic logging sufficient to enable post-hoc accountability. [EUR-Lex ↗] Rubric Protocol is designed to satisfy these requirements by default.

What Article 12 Requires

Article 12 of the EU AI Act mandates that high-risk AI systems automatically generate logs that enable monitoring for the duration of operation. Specifically:

Audit Trail Architecture

A complete audit trail for an AI decision looks like this in Rubric Protocol:

Multi-Node
Architecture

Rubric Protocol operates as a federated network of independent nodes. Each node maintains its own ML-DSA-65 keypair and participates in threshold consensus before HCS anchoring.

Node Requirements

Each federation node requires:

Threshold Consensus

Before any Merkle forest root is submitted to HCS, participating nodes must reach threshold consensus. Each node independently signs the forest root with its ML-DSA-65 key. Once the threshold is met (currently 3-of-5), the combined signature is anchored.

This means no single node can submit fraudulent attestations — consensus is required. Geographic distribution across jurisdictions also ensures no single regulatory action can take down the network.

Threshold Key Ceremony

Federation signing keys are protected by a Shamir Secret Sharing scheme. Each master signing key is split into 5 shares using pure GF(256) arithmetic — any 3 shares reconstruct the key. No single node ever holds the complete key.

Threshold signing is exposed via /v1/threshold-attest. The coordinating node requests partial signatures from peers via HMAC-authenticated inter-node calls. The combined signature is assembled at submission time and never persisted.

Any Enterprise.
Any Volume.

One HCS write per batch — 1 decision or 150 million. The architecture doesn't change. The cost barely does.

Reliable Queue Architecture

The attestation pipeline is backed by a Redis reliable queue using the BRPOPLPUSH pattern. Every incoming attestation is atomically moved from an intake queue to a processing queue before any work begins. If a node restarts mid-batch, in-flight attestations are recovered automatically on startup — zero attestation loss by design.

The queue drives an 8-worker cluster per node, sustaining the throughput behind 327M+ attestations/day across the 5-node federation.

@rubric-protocol/sdk

TypeScript SDK for direct integration. Full type definitions, AttestationsService, StorageService, and federation-aware routing built in.

autogen-rubric Python

One-line auto-instrumentation for 13 AI frameworks. Every LLM call, agent decision, and tool use is attested automatically — no per-decision code required.

Auto-Instrumentation — Recommended

Add one line at app startup. Rubric detects installed frameworks and patches them at the class level. Complete attestation coverage by construction — no codebase audit required.

Client Reference

AttestationsService

Examples

Real-world integration patterns across regulated industries.

Logistics — Route optimization audit trail

A logistics operator runs 150M AI routing decisions per day. Every decision attested in a single HCS write per batch. Regulators get a cryptographic receipt for every reroute, delay override, or priority escalation — provable in court.

Financial Services — Full loan decision lineage

EU AI Act Article 12 requires lenders using AI to maintain complete, tamper-evident records of every automated credit decision. Rubric creates an immutable chain: model output → compliance check → human review → final approval.

Healthcare — AI diagnostic audit trail

A hospital deploys AI for radiology triage. Every recommendation must be auditable for liability, HIPAA, and EU MDR compliance. Rubric attests the model version and output — without storing protected patient data on-chain.

Algorithmic Trading — MiFID II decision records

MiFID II and SEC Rule 17a-4 require broker-dealers to retain records of every algorithmic trading decision. Rubric provides a quantum-resistant audit trail verifiable by regulators — without exposing proprietary model internals.

LangChain — Automatic agent attestation

Drop one callback handler into any LangChain agent or chain. Every LLM call, tool use, and agent decision is automatically attested to HCS — zero changes to your existing logic. EU AI Act Article 12 compliant out of the box.

LangGraph — Node-level decision attestation

Pass one callback handler in config. Every graph node execution, LLM call, and tool use is attested — ideal for compliance-requiring enterprise deployments where audit trails and human-in-the-loop oversight are requirements.

OpenAI — Wrap once, attest everything

Wrap your existing OpenAI client with one line. Every chat.completions.create() call — streaming or not — is automatically signed and anchored to HCS. No changes to the rest of your codebase.

AutoGen — Attest every agent message

Attach one hook to any AutoGen agent. Every message sent and received is automatically attested to HCS — full multi-agent conversation trails with zero changes to your agent logic.

LlamaIndex — Query and retrieval attestation

Add one callback handler to LlamaIndex. Every LLM call and query result is attested to HCS — ideal for RAG pipelines where source fidelity and model output must be provable for compliance.

CrewAI — Attest every agent, task, and tool

One listener attests every agent execution, task completion, tool invocation, and crew kickoff — full multi-agent audit trail with zero changes to your crew logic.

Haystack — RAG pipeline attestation

Drop RubricHaystackComponent into any Haystack pipeline as a pass-through. Every LLM reply is attested before it reaches your application.

Semantic Kernel — Enterprise AI attestation

Register one filter on the kernel. Every function invocation — across all plugins and agents — is attested automatically. Ideal for Microsoft Azure enterprise deployments.

Pydantic AI — Type-safe agent attestation

Wrap any Pydantic AI agent. Every run — sync or async — is attested immediately on completion. Full proxy: all agent methods remain accessible.

OpenAI Agents SDK — Full lifecycle attestation

Pass one hooks object to Runner.run(). Agent starts, LLM completions, tool calls, and handoffs between agents are all attested — complete chain of custody for multi-agent workflows.

Google ADK — Vertex AI Agent Builder attestation

Register one plugin on the Runner — every agent, model call, and tool invocation across your entire ADK application is attested automatically. Works with any ADK agent on any model.

Strands Agents / Amazon Bedrock AgentCore — Full lifecycle attestation

Pass one hook to any Strands agent. Agent invocations, model calls, and tool results are all attested — works with Amazon Bedrock AgentCore Runtime.

DSPy — Declarative AI pipeline attestation

Register one callback globally. Every DSPy module execution and LM call is attested — zero changes to your existing program logic.

Webhooks —
Real-Time Event Push

Push every attestation to your SIEM, GRC platform, or custom endpoint in real time. HMAC-SHA256 signed. Auto-retry with failure tracking. Enterprise+ feature.

Register a Webhook

Available Events

Payload Format

Signature Verification

Manage Webhooks

GRC & SIEM
Export API

Every attestation is exportable in three formats — JSON, CSV, and SIEM-compatible NDJSON. Pipe directly into ServiceNow, Archer, OneTrust, Splunk, Elastic, or QRadar. No custom integration work required.

Endpoints

Query Parameters

JSON Export

CSV Export — ServiceNow / Archer / OneTrust

SIEM Export — Splunk / Elastic / QRadar

PDF Compliance Report

SR 11-7 — Model Risk Management

The Federal Reserve's SR 11-7 guidance requires banks to maintain a model inventory with risk classification, validation status, and outcomes monitoring. Rubric builds this automatically from attestation data.

Examination Package

Enterprise tier generates a complete SR 11-7 examination package in a single API call — model inventory, risk tier assignments, validation gap analysis, and outcomes monitoring summary. Export as PDF or structured JSON for GRC platform ingestion (ServiceNow, Archer, OneTrust).

FINRA 3110 / 3120 — Supervision Records

FINRA Rules 3110 and 3120 require broker-dealers to establish and maintain supervisory systems for AI-assisted workflows. Rubric provides cryptographically anchored supervision records that satisfy examination requirements.

Human Oversight Attestation

When a compliance officer or supervisor reviews, approves, or overrides an AI decision, that action is attested alongside the original AI decision — creating a complete chain of custody from AI output to human sign-off. Required under FINRA 3110 and EU AI Act Article 14.

FFIEC IT Examination

FFIEC IT examiners now include AI governance in examination scope. Rubric generates the third-party AI vendor audit evidence, access log completeness metrics, and policy enforcement summaries examiners ask for.