Axiomyx Engine
The reusable decision architecture underneath Axiomyx
Axiomyx Engine
Arena and Decision Lab are different environments connected to one common tactical decision architecture. The environment describes the problem. The engine decides.
Core architecture
Three layers.
The top layer changes per problem. The middle layer does not. That is the whole design, and it is why a second environment does not mean a second intelligence.
Environment
Defined per problem. This is the part that changes.
- State representation
- Legal actions
- Rules and transitions
- Objectives and reward
Axiomyx Engine
Reusable across environments. This is the part that does not change.
- State encoding
- Neural policy
- Value estimation
- Monte Carlo Tree Search
- Simulation
- Decision selection
- Training
- Telemetry
Application
What the environment is used for.
- Arena
- Decision Lab
- Future custom environments
The environment adapter
A new environment should not require rebuilding the intelligence. It should require describing the problem.
A structured environment defines four things: what exists in the state, which actions are legal from a given state, how an action changes the environment, and how objectives and outcomes are measured. Given those four, the engine supplies the search, the learned evaluation, the decision selection and the training framework.
Where this actually stands: the architecture is being separated into a reusable core and environment-specific adapters. That separation is real work in progress rather than a finished interface. There is no adapter SDK, no documented integration surface and no public API for building a new environment today.
Axiomyx does not claim plug-and-play support for arbitrary environments, and it will not until a second environment has actually been built on the adapter boundary rather than described.
Why this matters
Arena is the first commercial proving environment. It is not the limit of the architecture.
A tabletop wargame was chosen because it is a genuinely hard decision problem that happens to be cheap to test: hidden intent, spatial reasoning, sequencing across turns, a large legal action space, and an opponent who is trying to win. Solve enough of that and the same machinery has something to say about other problems with the same shape.
The claim being made is deliberately narrow: one engine, several decision environments, demonstrated so far in one of them.
Components
What the engine actually contains.
State encoding
Implemented
- Turns the current environment into a structured numerical representation the models can read. Everything downstream depends on this being a faithful description of the situation rather than a summary of it.
Policy
Implemented
- A learned estimate of which candidate actions are worth exploring. With a large legal action space, this is what makes a modest search budget productive instead of arbitrary.
Value
Implemented
- A learned estimate of the expected quality of a state, so a line can be judged without playing it all the way to the end of the game.
Monte Carlo Tree Search
Implemented
- Guided search over candidate futures. Rather than scoring one immediate action, it expands promising lines, evaluates where they lead, and uses those results to decide what to expand next.
Self-play and simulation
Implemented
- Experience is generated by repeated interaction, including the system playing against itself. This is how the policy and value estimates improve without a human supplying labelled examples.
Telemetry
Implemented, with gaps
- Recorded today, where the build emits them: estimated state value, MCTS simulations, root visits, decision time, and the top candidate actions with their visit share, action value and prior.
- Not every decision in a recorded game carries telemetry. A field that is not recorded is shown as absent rather than filled in, and the replay viewer never invents a number.
Have an environment you want to point this at?
Register and pick Engine, or partnership and research collaboration. Integration conversations are early and genuinely open.