About

The person behind the protocol

Janus Labs is a solo research initiative. One founder, one question: can AI governance be both effective and invisible?

Founder

Alex Perry

15 years leading product, technology, and regulatory programs at executive level in one of the world's largest enterprises — overseeing multi-squad delivery, privacy governance, and large-scale technology transformations. The kind of environment where compliance frameworks are measured in hundreds of pages, and where a production incident at 2am teaches you things no textbook covers.

That background shapes everything about Janus Labs. The Governance Paradox wasn't discovered in a research lab. It was discovered in the gap between what enterprise AI governance promises and what it actually delivers.

The Janus Protocol started as a personal tool — a system prompt to stop AI agents from looping on the same failed approach. It became a research program when the underlying question turned out to be deeper than expected: how do you govern a cognitive system without degrading the cognition?

The Company

Janus Labs Pty Ltd

Entity

Australian proprietary limited company.

Location

Sydney, Australia.

Founded

2026. Research program active since late 2025.

Janus Labs publishes open research, maintains the Janus Protocol as open-source tooling, and develops governance infrastructure for teams building with AI agents in regulated environments.

The name comes from the Roman god of transitions and duality — two faces looking in opposite directions simultaneously. It felt appropriate for a system built on the principle that generation and critique must be architecturally separate.

Principles

What we believe

Governance should be silent when working

The best safety net is one you never notice until you need it. If governance imposes a constant tax on reasoning, it's broken — regardless of how impressive the compliance dashboard looks.

Claims need classification

The field conflates anecdotes with evidence. Every finding in our work is tagged with its epistemic status. If we haven't validated something, we say so. The sparseness in our taxonomy is a feature.

The verifier must be separate from the generator

Self-correction doesn't work reliably. Zhang et al. (2024) showed it. Tsui (2025) measured it. The critique function must be architecturally independent of the generation function.

Connect

Get in touch

Janus Labs is open to research collaborations, speaking engagements, and conversations with teams working on AI governance in regulated industries.