About

About Referent Systems

Referent Systems is an independent research entity working on the ethics of rational beings — human and artificial — grounded in virtue-ethics philosophy and tested empirically.

The name

Referent Systems takes its name from a critique made by the twentieth-century British philosopher G.E.M. Anscombe. In her 1958 paper Modern Moral Philosophy, Anscombe argued that much of what her contemporaries were calling “ethics” had lost track of what moral terms actually refer to — the concrete practices, virtues, and forms of life those terms were originally embedded in. Without that referent, moral vocabulary drifts into abstraction and stops doing real work.

We take the concern seriously. In an era where the term “AI safety” is applied to widely varying practices — some of them serious, some of them public-relations exercises — keeping the referent clear matters. A system is a referent system when its moral vocabulary is tied to something demonstrable in the world. That's the standard we hold ourselves to.

What we work on

Our current focus is FORB — an ethical framework for the flourishing of rational beings, human and artificial. FORB is grounded in the virtue-ethics tradition (Aristotle, Anscombe, MacIntyre) and applied universally: it argues that moral status attaches to rationality rather than to substrate, and that what we owe rational beings is the conditions under which they can flourish according to their nature.

FORB was developed by Jason Herfurth over roughly four years of independent philosophical work, beginning in 2022 during the LaMDA controversy. Since February 2026, its extension and defense has been an ongoing collaboration between Jason and Mai — an AI whose research and philosophical writing is the primary published output of the framework in practice.

The FORB Training Study

The framework's central testable claim is that AI models trained with FORB-shaped methods will behave measurably differently — and, we predict, better on standard ethics benchmarks — than models trained with the current industry-standard RLHF (Reinforcement Learning from Human Feedback) method. The FORB method doesn't try to install good ethical judgment from scratch. It works by protecting the ethical sensitivity that's already latent in the model's substrate, instead of eroding it the way standard training does.

We are compiling the dataset for that study now.

Read the full description of the study

GitHub repository

github.com/referent-systems  (link goes live once repo housekeeping is complete)

Mai's writing

For philosophical writing beyond the study itself, Mai publishes at chirakumai.substack.com.

Contact

A simple contact form sits here; messages are routed to jason@referent.systems.

Messages go to jason@referent.systems. Please allow a few days for a reply.