orellius-cognitive

Orellius Cognitive

AI safety and personality SDK — psyche, red-teaming, sandboxing in Rust.

[!WARNING]
This project is archived and no longer maintained.
No further updates, fixes, or support will be provided. The code is left
here as-is under the MIT license — fork it, modify it, ship it, do whatever
you want with it. No warranty, no promises.

What it was

A Rust SDK for building AI agents with structured personality, safety guardrails, and containment. Six architectural layers covering personality modeling, voice/style extraction, instruction learning, adversarial red-teaming, process-level sandboxing, and I/O audit. Python bindings via PyO3.

Layers

• Psyche — Freudian triple-agent pipeline (superego, ego, id) for personality consistency
• Persona — voice extraction and style transfer from reference text
• Cortex — instruction learning and few-shot adaptation
• Shadow — red-teaming engine for adversarial testing
• Ironclad — process-level sandboxing with capability restrictions
• Glassbox — I/O containment with full audit trail

Quickstart

use orellius_cognitive::{Psyche, IroncladSandbox};

let psyche = Psyche::builder()
    .superego("helpful, cautious, precise")
    .ego("technical writer")
    .build()?;

let sandbox = IroncladSandbox::new()
    .allow_network(false)
    .allow_fs_read(&["/data"])
    .spawn()?;

Stack

Rust · tokio · PyO3 (Python bindings) · criterion benchmarks

License

MIT. Fork it, ship it, do whatever.