Where the ladder ends

Compounding uses the result of one reasoning step as an input for the next. For that process to remain useful, each step needs a sound basis.

Sometimes the available information supports the next step. Sometimes a fact is missing, the inputs conflict, or the model cannot complete the required operation.

Understanding these limits is part of evaluating Paladin. We need to examine where reasoning stops and what prevents it from continuing.

One missing link

Consider the experiment example from our previous article. The user supplies records and instructions for deciding which runs are valid before interpreting their results.

Suppose a run is valid only when the instrument was calibrated and the temperature stayed within a specified range. Its record includes the temperature but leaves out the calibration status.

The temperature condition is satisfied. Calibration remains unknown from the supplied information, so the run’s validity cannot yet be established.

That gap matters to every later step that depends on treating the run as valid. Continuing would require the missing information or an explicit change to the conditions.

This is an illustrative example of a reasoning boundary, rather than a reported model response.

Different limits need different explanations

Evaluation should distinguish several situations:

• Missing information. A required fact has not been supplied or established.
• Conflicting information. The inputs give incompatible accounts of something the reasoning depends on.
• Unsupported reasoning. The task requires an operation the current model cannot perform.
• Incomplete computation. Processing ends before the reasoning reaches a result.

Each calls for a different response. More data may resolve a missing fact. A conflict needs clarification. A computation limit says something about execution, while an unsupported operation points to a capability limit.

Changing the field changes the task

Adding a calibration record supplies new evidence. Removing calibration from the validity conditions changes the question being evaluated.

Both could allow reasoning to continue, but for different reasons.

Tests should record which input changed, which step it affected, and how that change influenced the result. This helps distinguish progress in reasoning from a change in the task itself.

What we want to establish

In developing Paladin, we want to evaluate whether the model can identify these boundaries accurately and explain which step remains unresolved.

A useful outcome may be a completed answer. It may also be a precise account of the information or operation needed to continue.

The goal is to make the stopping point understandable: what has been established, what remains unresolved, and why.