An answer becomes a starting point

Some questions require several connected steps. Before reaching an answer, a model may need to establish which information applies, derive a relationship, and use that relationship to continue.

This is the idea behind compounding in Paladin’s development: a result from one step becomes an input for the next.

What builds on what

The user supplies files and defines the reasoning the model should follow. Together, these establish the working field for a question.

Within that field, an intermediate result can make further reasoning possible. A relationship established in one step can support a conclusion in another. The model builds toward an answer through these connected results.

The question is whether each step contributes something useful—and whether the resulting answer remains supported by the supplied information.

A simple example

Suppose a user provides records from an experiment, the conditions required for a valid run, and instructions for interpreting the results.

The reasoning could proceed in three stages:

1. Determine which runs meet the stated conditions.
2. Apply the interpretation rules to those runs.
3. Use the interpreted results to address the original question.

The first result determines which records the second stage can use. The second produces information needed by the third.

This illustrates the intended structure of compounding. It is an example of the approach, rather than a reported evaluation result.

Accuracy across the steps

A mistake in an early step can affect everything that follows. If an invalid run is included, later reasoning may apply the interpretation rules correctly and still reach an unsupported conclusion.

Evaluation therefore needs to examine the intermediate results as well as the final answer. We need to understand where an error begins, how it moves through subsequent steps, and whether later reasoning detects it or builds on it.

What we want to measure

In developing Paladin, we want to establish:

• When connected reasoning steps improve an answer.
• Whether each intermediate result follows from its inputs.
• How errors and missing information affect later steps.
• How much computation the additional reasoning requires.

The aim is to understand when a result provides a sound foundation for further work, and when the reasoning needs to stop or be revised.