"Nothing throws" is exactly why this goes unnoticed for three weeks. The advertised-versus-usable gap deserves the treatment we already give disk: nobody runs a filesystem at 100% and expects normal latency, yet people plan for a 200K window as if all of it is equally usable.
Two additions to the telemetry schema. Track the position of retrieved evidence within the assembled context, not just total tokens, because material in the middle of a long window gets used less reliably and a growing history quietly pushes evidence into that zone. And log the compaction event with what was dropped, otherwise a quality regression right after compaction is indistinguishable from a retrieval regression and you will debug the wrong stage.
"Nothing throws" is exactly why this goes unnoticed for three weeks. The advertised-versus-usable gap deserves the treatment we already give disk: nobody runs a filesystem at 100% and expects normal latency, yet people plan for a 200K window as if all of it is equally usable.
Two additions to the telemetry schema. Track the position of retrieved evidence within the assembled context, not just total tokens, because material in the middle of a long window gets used less reliably and a growing history quietly pushes evidence into that zone. And log the compaction event with what was dropped, otherwise a quality regression right after compaction is indistinguishable from a retrieval regression and you will debug the wrong stage.