Developer's compression format fails readability test for LLMs

A developer built ctxfold, a lossless compression format for LLM prompts, and tested its CSV encoder by asking GPT-4o-mini to look up records. Raw CSV scored 24/24, but folded CSV scored 0/24, as the model couldn't reconstruct values from the compressed format. Even GPT-4o only managed 6-9/24, revealing that models struggle with indirection through headers.