In our condition-tree work, a legal-rule example requires an amount to be strictly greater than a threshold. A baseline answer accepts the value exactly at the threshold. The structured version asks the strict condition explicitly and combines the answers through fixed logic.
The example is useful because the error is so small. The model does not need a grand theory of law to distinguish greater than from greater than or equal to. Yet stating or recognizing a rule does not ensure that the final answer consistently applies it.
The broader computational question is how much of a decision should be left to free-form generation once its logical structure is already known.
Move the composition outside the model
A condition tree decomposes a rule into questions. The model evaluates individual conditions; deterministic logic combines them. This can prevent the final generation step from casually changing an AND into an OR or relaxing a strict inequality.
It also gives the evaluator intermediate failures to inspect. If the final answer is wrong, was a fact misread, a condition misclassified, or the rule structure wrong? A single end-to-end answer makes these sources harder to separate.
That decomposition is useful only if the conditions are well specified. Some legal concepts are comparative, contextual, or dependent on interpretation. Breaking them into independent binary questions can discard the relationship that makes the rule meaningful.
Structure can carry extra information
Our paper identifies an important comparison problem: expert-authored condition prompts can contain guidance absent from the baseline. If a condition explicitly reminds the model that equality is insufficient, a gain cannot be attributed entirely to tree structure.
A fairer test separates the effects. Give the unstructured baseline the same substantive guidance. Compare intermediate judgments, aggregation, and total model cost. Then vary rule structure and model capability instead of reporting one average improvement as a universal effect.
The reported results include regimes where decomposition helps and regimes where it regresses. That is more useful than a blanket recommendation to structure every task. The structure can reduce one kind of error while introducing another.
A correct judgment still needs an execution boundary
The connection to agent control is direct but limited. A model can correctly judge that an action is disallowed and still emit a tool call that performs it. A structured answer can make the rule application more legible without creating an enforcement mechanism.
An executing service needs to bind the relevant judgment to the actual action and current state. If the state changes between evaluation and execution, an earlier correct answer may no longer authorize the effect. This adds time, identity, and consistency to the logical problem.
Conversely, an enforcement layer can faithfully enforce the wrong predicate. Deterministic composition is only as useful as the conditions and data it receives. Separating interpretation, composition, and execution helps locate the error rather than making the whole system automatically trustworthy.
Which parts deserve a model call?
I would use a model where interpretation is needed, an explicit computation where the relation is already formal, and an execution check where an external effect becomes real. The empirical question is whether that division improves correctness at an acceptable cost for a particular task family.
A strong evaluation should include boundary cases, missing facts, contradictory facts, and rules whose conditions interact. It should also include a capable unstructured model with the same information. Otherwise the experiment can mistake a helpful prompt for a necessary architecture.
The small threshold failure is a good starting point because the intended relation is inspectable. The larger ambition is a system whose intermediate commitments remain visible and whose eventual action actually respects them. Neither a fluent explanation nor a tree diagram establishes that final property on its own.