Summary

You can now measure a running system and reason about whether it is reliable enough to keep shipping - the two halves of one discipline.

In this module, you:

• Learned an AI-native telemetry model from the observability standard: OpenTelemetry-spec traces, metrics, and logs under one correlation, bound to a canonical model and queryable from MCP, where every change, access, and inference is a traceable record.
• Saw what AI/agent observability adds - the agent trace, token and cost, eval drift, and inference records kept as responsible-AI (42001) evidence.
• Defined an SLI, an SLO, and an error budget, and computed one: 99.9% over 30 days is ~43 minutes.
• Stated the error-budget policy from the reliability standard and ADR-0020 - a spent budget triggers a per-service feature freeze - plus the two non-deviable rules: freeze on an exhausted budget, and an SLA always looser than the SLO.

You have completed the path

This was the final module of Engineering standards for an AI-native org. Completing it earns the engineering-standards trophy. You now know the non-negotiable floor every repository works to - from the AGENTS rules through to how a service is measured and kept reliable - well enough to make changes that pass the org gates by default.

Where to go next

• Engineering standards for an AI-native org - return to the path to claim your trophy.
• The reliability standard, in the estate's standards repository, covers on-call, the toil cap, capacity headroom, and chaos/game-days - re-read it when you are ready to operate a service, not just build one.
• The observability standard is the source of truth for the SLI menu and the telemetry spine; link to it, never paraphrase it.