When information takes shape

A set of files can contain enough information to answer a question even when no file states the answer directly. Reaching it requires connecting the information and working through what follows.

At Mumbrane, we are developing field-based intelligence around this problem. Our research asks how data and defined reasoning can form a working structure within which a model develops an answer.

What forms the field

Our approach has two foundations. The model begins with general language capabilities developed from a limited body of data. The user then supplies the information for a particular task, along with the reasoning steps the model should follow.

We call this working scope a field. It brings together the supplied information, the relationships between it, and the instructions for reasoning with it.

“Closed” refers to that bounded scope. The research question is how much a model can establish within it.

The information still needs to be assessed. A file may contain an observation, an assumption, or an instruction. Those have different roles, and representing that distinction is part of the problem.

Why a field?

Our inspiration comes from general relativity, where matter and energy influence spacetime geometry.

We are exploring a computational analogy: information shapes the field through which reasoning proceeds.

Making that analogy useful requires precise definitions. How is information represented? How does a new input change the field? What operations connect a question to a result?

These questions turn an intuition into something we can implement and test.

How reasoning compounds

A question may require several connected steps. One step establishes a result. The next uses that result alongside the available information to continue the work.

This is what we mean by compounding.

A model could establish a relationship, use it to derive another, and build toward a conclusion that was never explicitly written in the source files. Each intermediate result contributes to the next stage.

The value of this process depends on the quality of those steps. An early mistake can influence later conclusions. Evaluation therefore needs to examine both the final answer and the path used to reach it.

What we need to test

Our central questions are:

• Does the field preserve the information and relationships needed for the task?
• When does compounding improve the result?
• How do missing information, conflicting inputs, and intermediate errors affect reasoning?
• How do accuracy and computational cost change as the field and reasoning grow?

Answering these requires tasks with clear criteria, controlled changes to the inputs, and measurements that others can examine.

Developing the idea through Paladin

Paladin is our first model in this research direction. It gives us a concrete system through which to develop and evaluate bounded reasoning.

Each release needs to distinguish demonstrated capabilities from ideas still under investigation. The broader field hypothesis will depend on evidence from implementations and experiments.

Our aim is to understand when this structure helps a model reason, how far it can carry a result, and what determines its limits.