Glossary / LLM Observability

Definition

What is LLM Observability?

LLM observability is instrumenting AI applications to see latency, errors, prompts, outputs, and drift — and explain them later. Trinitite adds OpenTelemetry tracing across ops, security, and audit streams plus a signed daily compliance number, turning observability into reproducible evidence rather than a dashboard.

A “98% pass rate” rendered from your own database rows is not evidence an auditor will accept, and an LLM-as-judge on a shared GPU pool gives different verdicts at high utilization than at low. Signed observability fixes both.

Every event returns a signed verdict; every day rolls into one Merkle-rooted, KMS-signed, externally anchored compliance number with Population Stability Index drift detection that fires a signed webhook when the model retrains or the system prompt changes.