EYDII

The trustless verification layer for autonomous systems. Content-blind. Mathematical.

Trustless infrastructure for autonomous systems.
Content-blind, mathematical behavioral intelligence — from AI agents to robots to power plants.

---

The Problem

Autonomous systems are not software. They are neural, adaptive, and unpredictable. An AI agent writing code, a robot picking orders in a warehouse, a controller managing coolant flow in a nuclear facility — they all share the same failure mode: silent drift. They don't crash. They don't throw errors. They gradually deviate from what they were supposed to do, and nobody sees it until something breaks.

Traditional security answers one question: "Who has access?" It checks credentials, rotates keys, scans payloads. It was built for deterministic software — programs that execute the same instruction the same way every time.

Autonomous systems don't work that way. An AI agent silently expands its scope. A warehouse robot subtly changes its path optimization. An industrial controller drifts its operating parameters by fractions of a degree per hour. No credentials were stolen. No vulnerability was exploited. The system simply drifted.

You will not see it in your access controls. You will not see it in your firewall logs. In healthcare, defense, pharmaceuticals, and industrial infrastructure — the environments where autonomous systems carry the highest stakes — you cannot even inspect the content to look for problems. Privacy law, classification, and regulatory compliance make content inspection illegal.

EYDII is the trust layer built for exactly this.

It monitors what autonomous systems do — action types, timing, frequency, behavioral patterns — without ever seeing what they say, produce, or access. Content-blind by architecture, not by configuration. Trustless by design, not by promise. Mathematical proof that behavior matches intent, or proof that it doesn't.

Where It Matters

• AI Agents — Coding agents, research agents, multi-agent organizations across all major agent frameworks
• Physical AI / Robotics — Warehouse robots, surgical systems, autonomous vehicles, manufacturing floors
• Industrial Autonomous Systems — Power plants, railway control, water treatment, pharmaceutical manufacturing
• Regulated Environments — HIPAA, ITAR, defense, financial services — where content inspection is forbidden by law

How It Works

EYDII operates as three independent layers. Compromising one cannot compromise another.

                         YOUR APPLICATION
                               |
              +----------------+----------------+
              |                                 |
         [ SIDECAR ]                     [ SIDECAR ]
         Runtime interception            Runtime interception
         Action classification           Action classification
         Content never leaves            Content never leaves
              |                                 |
              +----------------+----------------+
                               |
                          [ ENGINE ]
                          Behavioral analysis
                          Drift detection
                          Trust Score computation
                          Pattern recognition
                               |
                          [ PROOF ]
                          Mathematical attestation
                          Denial proofs
                          Immutable audit chain
                          Third-party verifiable

Layer	What it does	What it never sees
Sidecar	Intercepts actions at runtime, classifies action type and metadata	Message content, tool outputs, model responses
Engine	Analyzes behavioral patterns, detects drift, computes Trust Score	Raw data, payloads, user inputs, generated content
Proof	Generates mathematical attestations and denial proofs	Anything about the content -- proofs are over behavior, not data

Content-Blind by Architecture

This is not a filter. EYDII never receives, processes, stores, or transmits the content of agent actions. The architecture makes inspection impossible, not just disabled.

What EYDII sees:

• Action type (file_write, api_call, tool_use)
• Timestamp and sequence
• Agent identity
• Action frequency and patterns
• Behavioral deviation from established baselines

What EYDII never sees:

• What the agent wrote to the file
• What the API returned
• What the user said
• What the model generated
• What the tool produced

This is the difference between a security camera that watches where people walk and one that reads their mail. EYDII is the former. By construction, not configuration.

Quick Start

JavaScript

npm install @veritera.ai/eydii

import { EydiiClient } from "@veritera.ai/eydii";

const client = new EydiiClient({
  apiKey: process.env.EYDII_API_KEY,
});

const result = await client.verify({
  action: "file_write",
  agentId: "researcher-01",
  metadata: { target: "/reports/q2.pdf" },
});

if (!result.verified) {
  // Action violates established behavioral rules
  console.error(`Blocked: ${result.reason}`);
}

Python

pip install veritera

import os
from veritera import EydiiClient

client = EydiiClient(api_key=os.environ["EYDII_API_KEY"])

result = client.verify(
    action="file_write",
    agent_id="researcher-01",
    metadata={"target": "/reports/q2.pdf"},
)

if not result.verified:
    # Action violates established behavioral rules
    print(f"Blocked: {result.reason}")

Get your API key at id.veritera.ai.

Framework Adapters

Drop EYDII into any major agent framework with a single install.

Framework	Install	Language
LangChain / LangGraph	pip install langchain-eydii	Python
CrewAI	pip install crewai-eydii	Python
OpenAI Agents SDK	pip install eydii-openai	Python
LlamaIndex	pip install llama-index-tools-eydii	Python
Pydantic AI	pip install pydantic-ai-eydii	Python
Agno	pip install agno-eydii	Python
Google ADK	pip install google-adk-eydii	Python
Mastra	npm install @veritera.ai/eydii-mastra	TypeScript
n8n	Community node	n8n
OpenClaw	npm install @veritera.ai/eydii-openclaw	TypeScript

Each adapter wraps the native framework's tool/callback interface so EYDII operates transparently. No changes to your agent logic.

MCP Support

EYDII ships as a Model Context Protocol (MCP) server for local development tools:

• Claude Code -- Verify actions from Claude's CLI agent
• Claude Cowork -- Verify actions across collaborative Claude sessions
• Cursor -- Verify actions from Cursor's AI assistant

See documentation at veritera.ai for configuration.

Repository Structure

eydii/
  sdks/
    javascript/          # @veritera.ai/eydii — install from npm
    python/              # veritera — install from PyPI

SDKs are distributed via npm and PyPI. Framework adapters are distributed via pip and npm (see table above). Source is maintained in private repositories.

Architecture

EYDII was built for environments where content inspection is forbidden — healthcare, defense, pharmaceuticals, industrial control, multi-tenant SaaS, financial services. Where HIPAA, ITAR, classification rules, or regulatory compliance makes reading system outputs a legal violation, EYDII operates without compromise because there is nothing to compromise. The same architecture that monitors an AI agent writing code monitors a robot picking orders in a warehouse and a controller managing coolant in a power plant. Only the sidecar connector changes.

Seven patent applications cover the trust layer architecture, behavioral drift detection, mathematical attestation, and content-blind trust methods.

Links

EYDII Platform	id.veritera.ai
ROAM (Agentic OS)	github.com/veritera-ai/roam
Veritera Corporation	veritera.ai
Documentation	veritera.ai

License

This repository and its contents are proprietary software of Veritera Corporation. See LICENSE for terms.

---

_{Built by Veritera Engineering}