Datasets

I built a 300-case conformance suite for the code that gets JSON out of an LLM, and it immediately found 18 failures in my own parser

300 labelled malformed LLM outputs with ground truth and a scorer that exits 2 on a regression. json.loads 25/300; my own parser 282/300 with 5 invented values, named and left unfixed. Scorer + 30 cases CC0.