Making good agents better — safe, auditable, and verifiable — without making them slow.

---

Lightweight, deterministic, policy-bound, auditable, and verifiable execution runtime for AI agents operating in regulated environments.

Reference domain: Healthcare

The Idea

Agent runtimes like ZeroClaw and OpenClaw proved that AI agents can be fast, tiny, and deployable anywhere. But they were not built for environments where every action must be traceable, policy-constrained, and verifiable.

VERITAS does not replace them. It wraps them with trust.

Linux Kernel        →  ZeroClaw / OpenClaw    (fast, minimal, runs anywhere)
Red Hat Enterprise  →  VERITAS                (trusted, governed, auditable)

Execution Model

State → Policy → Capability → Audit → Verify → Next State

Architecture

┌─────────────────────────────────────────────────────┐
│              Application / Agent Code               │
│         (LangGraph, CrewAI, OpenClaw, etc.)         │
├─────────────────────────────────────────────────────┤
│                    VERITAS                           │
│   Policy Engine │ Audit Trail │ Verifier │ Caps     │
├─────────────────────────────────────────────────────┤
│              Agent Runtime Kernel                    │
│            (ZeroClaw or equivalent)                  │
└─────────────────────────────────────────────────────┘

Core Components

Component	Purpose
veritas-core/	Deterministic runtime (ZeroClaw lineage)
veritas-policy/	Deny-by-default permission & risk engine
veritas-audit/	Immutable, append-only execution trace
veritas-verify/	Output validation before delivery
veritas-contracts/	Capability / policy / audit schemas

Trust Boundary

Trusted	Untrusted
Runtime core	LLM
Policy engine	Tools
Audit engine	Input data
Verifier	External environment

Design Principles

1. Control over autonomy
2. Evidence over intelligence
3. Determinism over emergence
4. Deny by default
5. Capability-based security
6. Minimal trusted computing base
7. Auditability by design
8. Verifiable execution
9. Human override always possible
10. Data-model independence

Lightweight by conviction. Governance must not be the reason agents become slow, heavy, or hard to build.

Documentation

See docs/ for full documentation.

Whitepaper v0.3

Language	Link
English	WHITEPAPER.en.md
简体中文	WHITEPAPER.zh.md
日本語	WHITEPAPER.ja.md
Français	WHITEPAPER.fr.md

Quick Start

Prerequisites

• Rust 1.70+ (install)

Build & Test

git clone https://github.com/veritas-rt/veritas.git
cd veritas
cargo test --workspace       # 45 tests, all passing

Run the Healthcare Demo

cargo run -p demo -- run-all

Interactive TUI Demo

cargo run -p veritas-tui

The TUI lets you select scenarios, toggle patient consent and agent capabilities, and watch VERITAS enforce policy in real time.

Contributing

VERITAS is open source. Community contributions — including new translations — are welcome. See CONTRIBUTING.md for details.

License

Licensed under Apache License 2.0. See LICENSE for details.