Skip to the content.

Observability & the operations report (design — lands with tracing)

Principle: instrument once, export anywhere

Tracing is built on the open standard — OpenTelemetry / OpenInference — not a single vendor. assay emits standard spans; the backend is a swappable exporter. No lock-in, and the same instrumentation lands in whichever tool a team already runs.

Backend Role Why
Arize Phoenix default open-source, runs locally, OpenInference-native — zero setup
W&B Weave supported export instrument once, view in Weave too
Self-hosted (Phoenix / Langfuse) regulated / on-prem data residency — traces never leave your environment
Others (Galileo, Braintrust, raw OTLP) swap the exporter standard OTLP out

Tracing

SLIs

Operations report — clean vs. issues

A rollup across runs for the operations team:

Each issue links to its run directory + audit log for triage, so the report is a queue, not just a dashboard. Feeds ESCALATION.md.

Data residency: with a self-hosted backend, no trace data leaves the deployment — the same property the audit log guarantees for evidence.