DSL wins CAD bracket test 10/15 where code generation hits 4/15

This experiment follows an earlier comparison where the same DSL and code-generation approaches were tested on eight simpler tasks, such as boxes, cylinders and holes. In that round, code generation used about 26% fewer tokens than the DSL, because loops and functions let it compress several operations into one code block. The new bracket task was chosen to be closer to real work: the AI had to avoid surrounding parts and repeatedly measure and correct its shape, not just produce a basic form.

The bracket conditions interact with each other. Widening a plate to increase contact area with the sensor can eat into the required 5 mm clearance from an obstacle, and rerouting around that obstacle can split the plate or make it impossible to fold from a single 5 mm sheet. In the logs, the DSL method, which had checks built into each operation, tended to halt when a condition was violated and never reach a final shape. Code generation often kept going and left a shape that looked finished but had internal defects, such as three separate solids instead of one. One run even reported a single solid while the verifier found three.

The article's own suggested setup is to use the checked DSL as the standard path, and only hand tasks the DSL cannot express to code generation in a temporary copy of the data, verifying results before they touch production. The key open question, and the limit of this result, is that the test involved one bracket design, one model and one CAD tool, so whether the DSL's advantage holds on other parts or with a stronger model remains to be seen.