# Observability acceptance record

Use with [Observability for LLM apps](/articles/observability-llm-apps).

| Signal | Diagnostic purpose | Source/version | Access boundary | Retention/deletion | Response owner |
| --- | --- | --- | --- | --- | --- |
| Request outcome | Distinguish held from delivered | | | | |
| Model metadata | Attribute configuration, usage and latency | | | | |
| Retrieval metadata | Identify eligible source/version decisions | | | | |
| Validation result | Separate schema from evidence support | | | | |
| Missing telemetry | Detect exporter/buffer failures | | | | |

Raw payload collection needs a separate justified purpose, approved data scope and controls. Pseudonymous IDs remain sensitive where re-identification is possible. Moving data to cold storage is not deletion.

## Acceptance probes

- Unsupported dispatch-to-arrival claim: HTTP success and valid schema, but held before delivery.
- Same-tenant authorised investigation succeeds; other-tenant trace search is denied.
- Queue handoff preserves trace relationship.
- Exporter outage produces bounded buffering/drop behaviour and a missing-telemetry signal.
- Deletion reaches primary storage, indexes and approved exports under the recorded policy.
- Replay cannot invoke live external side effects.

Record expected versus observed results, configuration, date, remaining gaps and owner. The article's incident is invented; replace it with your own measured evidence before claiming implementation acceptance.
