Enter the Centaur

Bruce Lee entered our lexicon with his iconic movie, Enter the Dragon. Can artificial intelligence enter our daily corporate decisions with Enter the Centaur? But then what on earth is this Centaur?

Before explaining the mechanics of this animal, it is worth asking a simple question. If AI systems are to assist in serious decision making, how should their cognitive roles be structured?

Most current uses of large language models collapse description, evaluation and strategy into a single stream of text. The same system describes the world, judges options, anticipates adversaries and recommends action. That is perfectly adequate for drafting emails. It is less obvious that it is ideal for structured strategic reasoning.

Centaur is an experiment in separating these functions and creating an artifact that has a novel approach to complex problem solving. It is a human–AI hybrid software construct designed to act as an adjudicator in complex strategic simulations. The current alpha version, Alpha Centaur, is available as a Google Colab notebook in public GitHub repository Centaur.

Centaur began life as a teaching tool.

I was designing an MBA elective course on Geopolitics of Technology at Praxis Business School. The course explored how technologies such as Rare Earth mining, green nuclear power and GPU infrastructure shape national and international politics. Since the students were management students rather than engineers, the course is not very technical but includes a simulation game. Crises would be constructed. Students would respond. Responses would need to be evaluated.

The difficulty lay in evaluation.

No single faculty member can realistically claim deep expertise across mining supply chains, nuclear regulation, semiconductor geopolitics, energy markets, trade regimes and military signaling. Yet credible adjudication was essential. That requirement led to Centaur.

At its simplest, Centaur is a Python program that connects to an LLM through an API. In principle it could work with OpenAI, Gemini, Grok, Llama, Claude or comparable systems. The model is given two inputs: a description of the world and a proposed action. It generates an adjudication in text, which a human can review or modify before proceeding.

The real challenge is not coding but describing the world.

Geopolitical reality is messy, narrative and multi-layered. A rambling story does not work well as input. A purely numerical representation strips away too much context. The compromise is what I call a faceted description. The world is broken into a set of aspects, each described through focused text fragments. The result is structured text rather than free narrative.

To construct this world-state, I built a second module called ZeitWorld. It converts large amounts of geopolitical text into the faceted representation that Centaur uses for adjudication. Human refinement remains possible at every stage.

A third and optional construct, Chanakya, represents a strategic actor seeking to advance its own interests. Unlike Centaur, it is not neutral. It generates incentive driven responses within the same simulated environment.

The three components play distinct roles. ZeitWorld structures the world. Centaur evaluates actions. Chanakya advances strategic interests. For each construct, the role is sharply defined and supplied to the LLM as a dedicated text file. This ensures that the role remains stable across successive turns and that each construct operates under explicitly different instructions. Human oversight can intervene between steps. The intention is not automation, but disciplined simulation.

The project is currently in alpha phase. Alpha Centaur runs as a Google Colab notebook. The roles of ZeitWorld, Centaur and Chanakya are defined through editable text instructions. Access to LLMs requires an API key. The prototype currently uses the inexpensive gpt-4o-mini model.

The MBA course may or may not run. That is no longer the central issue. Centaur has become an experiment in integrating large language models into structured cognitive activity that goes beyond drafting copy. It is also a small exploration of how clearly defined roles and human oversight can shape AI mediated reasoning.

The notebook along with code and rather sparse documentation are available. Feedback from thoughtful readers is welcome.