OpenTelemetry

Every invocation emits one span pecca.call with attributes:

Attribute	Meaning
pecca.path	workspace/project/call
pecca.mode	record, shadow or live
pecca.version	model version used
pecca.served_by	llm, model or fallback
pecca.confidence	calibrated confidence
pecca.agreement	model agreed with the LLM (shadow)
pecca.latency_ms	wall time of the call

Your wrapped LLM span nests under it, so GenAI semantic-convention (gen_ai.*) spans stay intact. Without a configured exporter the tracer is a no-op.

workspace:
  telemetry: {exporter: otlp, endpoint: ${OTEL_EXPORTER_OTLP_ENDPOINT}}

Filter pecca.served_by = fallback to see where the LLM is still used. Pecca has no anonymous usage telemetry. Traces can be read back as training data with the otel-traces datasource.