Audit an AI workflow for evals, monitoring, and approval points
Decide where the machine stops and the operator steps in.
Source: docs/positioning.md
Input
Workflow description, tool permissions, model/provider choices, risk points, and expected operator handoff.
Decision
Name the automation boundary, approval point, evidence standard, rollback path, and escalation rule.
Evidence
Control gaps, eval/oracle plan, receipt expectations, least-privilege boundary, rollback path, and escalation rule.
Failure handling
If controls are undefined, keep the workflow in implementation substrate rather than client-facing packaging.
What to verify
- The input exists in live evidence.
- The decision has an oracle.
- The evidence path is captured before closeout.