1. Understand your system

The objective is to define the smallest complete boundary within which a material outcome can be reconstructed. Start with user and service outcomes, then follow the components that influence them.

1. Describe the service

Record:

• intended and explicitly prohibited uses;
• users, affected parties, and service owners;
• outputs, recommendations, decisions, and external actions;
• the consequences of incorrect output, delay, disclosure, or unintended action;
• expected human review and fallback paths; and
• production environments and operating hours.

Use the system profile as the working record.

2. Draw the execution boundary

Show the main path and control points:

User / channel
  → application and identity
  → prompt / policy / orchestration
  → model
  → retrieval / memory
  → tools / APIs / other agents
  → guardrails and human approval
  → response or external action
  → telemetry pipeline and evidence stores

Include asynchronous work, queues, scheduled jobs, retries, callbacks, and human handovers. These are common places for trace context and accountability to be lost.

3. Inventory changeable artefacts

Artefact	Minimum information
Model	provider, model ID/version, endpoint, parameters, deployment environment
Prompt/instruction	stable ID, version/hash, owner, approval status
Policy/guardrail	policy ID/version, decision categories, enforcement point
Retrieval source	collection/index ID, content owner, classification, refresh process
Memory	type, owner, scope, write/read controls, retention
Tool/API	tool ID, operation, allowed scope, credentials, side effects
Agent	agent ID/version, delegated capabilities, parent/child relationships

This inventory supports ORF-AGT-12, ORF-AGT-13, and ORF-GOV-03 later in the lifecycle.

4. Map representative journeys

Choose at least:

1. a normal successful journey;
2. an expected refusal or safe failure;
3. a downstream timeout or partial failure;
4. a malicious or out-of-scope request; and
5. for action-taking systems, an approval, tool execution, and recovery journey.

For each journey, identify the events needed to answer:

• What did the user ask?
• What system version handled it?
• What context or knowledge was used?
• What decisions, refusals, or approvals occurred?
• Which tools ran, under which identity, and with what result?
• What did the user receive, and did any external state change?

5. Identify material failure scenarios

Failure	Potential impact	Evidence needed
Incorrect or ungrounded answer	user acts on false information	prompt/output reference, sources, model/prompt version, evaluation label
Runaway tool loop	cost, denial of service, repeated external action	task tree, call count, retries, circuit-breaker decision
Lost tool response after successful action	duplicated or unknown external state	idempotency key, tool result, callback, retry lineage
Sensitive data in output or log	privacy/confidentiality breach	policy decision, detection event, trace reference, access audit
Silent regression after change	degraded quality or safety	change record, before/after metrics, rollout cohort

6. Establish ownership

Assign owners for the service, application, model/prompt, retrieval data, tools, telemetry platform, security/privacy response, operations, and risk acceptance.

Deliverables

• completed system profile;
• architecture and data-flow diagram;
• component and artefact inventory;
• representative journey maps;
• initial failure-scenario list; and
• named owners.

Common mistakes

• treating only the model API as the system boundary;
• omitting human review, asynchronous callbacks, or external side effects;
• using tool names without recording the operations and privileges they expose;
• failing to distinguish model output from the final content shown to the user; and
• assuming existing application logs contain enough context to reconstruct an AI interaction.

Done when: another engineer can use the profile and diagram to follow a user request through every component that materially affects its outcome.

Next: Determine risk and scope