What happens when a Tessary classifier flags a trace?

It depends on the classifier. Some file a finding on the flag itself, while the rate-tested ones count the flag and file a finding only when a call site’s rate of flags rises above its own normal. Either way, nothing pages anyone yet.

secret_leak is the first kind: one high-confidence leak is a finding. malformed_output, frustration, and groundedness are the second, so one failed schema check, frustrated message, or unsupported answer is counted and nothing more.

LLM judgment enters at the finding, reading it instead of the raw trace. Triage separates a real deviation from legitimate change, the harder half of the job. A tool your team changed on purpose looks identical to drift until something checks it against what actually happened. A high-confidence leak and a frustration finding are filed already ruled and skip triage. A groundedness finding goes through it.

What survives triage becomes a case, and a case is what gets attributed and, if it clears the alerting rules, what reaches a human. A flag that doesn’t survive triage never does.