01 · Evidence
What was available at the decision point?
Capture the sources, timestamps, relevant extracts or identifiers, and known gaps. Label supplied facts, retrieved evidence, inference, and unresolved uncertainty separately.
Foundation 04 · Trace records
Record the decision,
not just the output.
A response log cannot by itself explain what evidence was available, which boundary applied, who authorized an action, or what changed afterward. Build the smallest record that a new reviewer can reconstruct.
Learning boundary. A trace record is not surveillance, proof of correctness, or a reason to capture all available data. Retain only what inspection, accountability, security, and lawful operations require.
Learning objective
Make a decision reviewable after the context has changed.
By the end of this lesson, you should be able to define a minimum trace record that separates observed evidence from a model’s interpretation, names the applied authority, and records what remains unknown.
02 · Context
Which task, scope, and limit applied?
Bind the record to the task goal, resource boundary, policy or instruction version, and the identity of the system or person making the decision.
03 · Action
What was proposed, approved, or executed?
Keep alternatives and the final action distinct. If a person approved an exception, record that approval as an attributable event rather than a model-generated justification.
04 · Outcome
What changed—and what is still open?
Record observable effects, review signals, reversals, and unresolved questions. An action with no observed outcome is not necessarily a completed decision.
Practice lab
Repair an unreviewable scheduling action.
A workflow rescheduled a technician visit after summarizing a sensor alert and an old maintenance note. The log preserves the generated message and the final calendar update, but not the source timestamps, approved outage window, alternate slots, notification result, or reason a higher-priority work order was displaced.
- List the evidence and context that must be recovered before the action can be reviewed.
- Separate the system’s summary from the source records it used.
- Identify the authority and limit that should have been bound to the calendar change.
- Write an outcome record that includes a correction or escalation path.
Source shelf
Use observability as a starting point—not a verdict.
- OpenTelemetry GenAI Semantic ConventionsInspect the evolving open conventions for GenAI clients, MCP, spans, metrics, and events; they support shared observability language but do not determine what a specific organization must retain.
- C2PA guidance for AI and MLRead the provenance and integrity guidance when exploring what cryptographic records can and cannot establish about model, data, software, or output artifacts.
Continue the pathway