Frequently asked questions
Do I need to implement all 61 ORF telemetry items?
Not always. The requirements that apply depend on your system scope and risk profile.
Review the full set, identify what applies, and record the decision and rationale in your requirements register.
How does ORF relate to IM8?
ORF is separate from IM8 and does not replace existing IM8 requirements.
ORF focuses specifically on the observability and evidence needed to operate, investigate, and govern generative and agentic AI systems.
Is there a central platform or tool I can use?
Yes. Harness AI provides a central Langfuse-based observability capability.
Use the Observability section for tracing and related observability capabilities.
Does ORF require a specific observability technology?
No. ORF defines what evidence and telemetry should be available, not which product you must use.
You can use your existing logging, tracing, monitoring, SIEM, or observability stack.
Do we need to store prompts and model responses?
Not always. The goal is to retain enough evidence to understand and investigate what happened.
If full prompts or responses are too sensitive or unnecessary to retain, you can use metadata such as request and trace IDs, model and prompt versions, tool calls, retrieval references, timestamps, classifications, hashes, or redacted excerpts.
How should an existing production AI system start adopting ORF?
Start by mapping the system boundary and key user journeys, then compare your existing telemetry against the ORF requirements that apply.
Focus first on whether you can reconstruct important requests and incidents. Add more advanced analytics later.
When should ORF scope be reassessed?
Reassess it when the system changes in a meaningful way, such as a new model, prompt flow, tool, data source, user group, permission, or level of autonomy.
These changes may affect the system risk profile and the evidence you need to retain.
Contact
For questions about ORF or implementation guidance, contact the AI Practice team.