> Third, aggressive compression can remove action-critical evidence: on SWE-bench-derived Go tasks, compression reduced successful patch application from 27/40 to 15/40 by corrupting verbatim edit anchors.
The only way to avoid this is to hand-craft the translation layer (harness). You cannot rely on some general purpose technique to satisfy domain-specific concerns.
The most extreme version of this I've seen so far is in browser automation. "Compressing" the raw DOM down to a plaintext description (which was parsed deterministically by human-authored code) is a completely different universe of performance than trying to fight the raw DOM directly. Purpose-built compressors can massively outperform generic ones. I've got a case where ~500kb of DOM converts into maybe 1kb of plaintext. There is absolutely no "loss" in terms of information the business cares about in this context. These kinds of decisions cannot be made for you if they are to be made efficiently.
> Third, aggressive compression can remove action-critical evidence: on SWE-bench-derived Go tasks, compression reduced successful patch application from 27/40 to 15/40 by corrupting verbatim edit anchors.
The only way to avoid this is to hand-craft the translation layer (harness). You cannot rely on some general purpose technique to satisfy domain-specific concerns.
The most extreme version of this I've seen so far is in browser automation. "Compressing" the raw DOM down to a plaintext description (which was parsed deterministically by human-authored code) is a completely different universe of performance than trying to fight the raw DOM directly. Purpose-built compressors can massively outperform generic ones. I've got a case where ~500kb of DOM converts into maybe 1kb of plaintext. There is absolutely no "loss" in terms of information the business cares about in this context. These kinds of decisions cannot be made for you if they are to be made efficiently.