CoT Compliance Rate Achieves 9.91%→100%

As a sequel to Function Calling Harness, this covers a harness that verifies procedural compliance rather than outcome accuracy.

In the IAutoBeInterfaceEndpointReviewApplication case, GPT-5.4's first-attempt success rate was 9.91%, which it explains is caused by the procedural burden of having to classify dozens of endpoints without missing a single one. Free-form narrative CoT hides omissions, but typed submissions force slots like review, revises[], and reason, immediately exposing missing steps and duplicates.

The core message is that prompt requests the procedure while schema enforces it. This approach can extend to formats like investment memos, SOAP, IRAC, and code reviews, and it argues that even in areas where outcomes cannot be judged immediately, a minimum level of procedural quality can be guaranteed. It also suggests that the schema itself should be backtested against historical cases.

The latter part points out the one-shot limitation of traditional function calling, and proposes a method of validating streaming intermediate states using Typia's lenient parsing and incremental validation. It explains a structure where, even if output is cut off midway, the valid prefix remains as a checkpoint, pushing the procedural compliance rate up to 100%.