The question
Can an AI assistant recognize when an explicit update is possible under one interface and impossible under another? The current prototype focuses on exact state changes and a declared single-request contract.
The evidence to collect
A useful study would separate response validity, executability, exact success, unintended changes, and appropriate abstention. Paired cases help expose strategies that always act or always refuse.
Scope and status
The prototype uses synthetic development fixtures. No model-performance finding is reported here, and these fixtures are not a representative sample of real-world requests. Standards-defined behavior is background knowledge, not a newly discovered contribution.
Start here
Read: Valid JSON is only the beginning
Inspect the synthetic worked example