How it works

The model only reports what it sees. Plain code checks it, sends errors back once, builds the SCL file, validates it against IEC's schema, and scores it. Each step links to a real saved run.

Next step: the checks as a reward signal planned, not built

The answer key and the checks above already turn a drawing into a number: does the JSON fit the schema, is the wiring consistent, does the SCL file pass IEC's XSD, how many items match. That is a verifiable reward. The plan is to wrap them as a NeMo Gym environment, generate many synthetic drawings with exact answer keys, and use the reward to compare prompting strategies and, if accuracy plateaus, to fine-tune a smaller vision-language model.

Nothing in this section has been built or measured yet.