Why does Tessary check every trace instead of sampling?

Because sampling is built to measure a rate, and by the time an agent is mature, nearly everything left to catch is rare, exactly what a sample misses. Tessary checks every trace because the checks are cheap enough to afford it.

Sampling is a rate instrument: accurate and cheap for telling you how often something happens, nearly useless for catching one instance of something uncommon. The failures a team already fixed are the frequent ones, so what’s left as an agent matures is low-probability by construction.

What matters is what one check has to cost before reading everything is affordable. Deep judgment on every trace isn’t that. The classifiers that run by default are rule checks and statistical tests over stored spans. Groundedness, which you switch on, runs a public model on a GPU you provide rather than calling an LLM per answer. It still reads only the answers it can check: those on call sites that answer from retrieved documents, summarize, or extract.