Does reading every trace cost the same as grading every trace?

No, reading a trace and grading it are different operations with very different costs. Reading is deterministic work: parsing the trace, extracting fields, running a rule or a small classifier over it. That runs on ordinary CPU (or a tiny model) at a cost close to zero per event, so you can read every trace you ingest without the bill moving much.

Grading with a language model is different. It’s an inference call that reasons about the trace and produces a verdict, and every one of those calls costs real money that scales with volume. Treating “read everything” as if it meant “a language model judges everything” is the mistake. The two decisions are independent: you can read 100% of traffic cheaply and still reserve the expensive judge calls for whatever reading flags as worth a closer look.