Suppose you have a directory full of text documents. Most are repetitive, padded with boilerplate, or otherwise low-signal. A few contain the useful material. You could read every file manually, feed them all into an embedding pipeline, or ask an LLM to rank them. Or you could ask gzip . The basic idea is simple: Repetitive text compresses well. Varied text usually does not. That makes compression ratio a crude but surprisingly useful proxy for redundancy. It will not tell...