Mark the controls you can evidence today. Critical items must have an evidence description to clear.
Technology chosen before the problem is understood wastes money and hides the real issue.
Selecting “Verified” only takes effect with a full evidence set below — a selection alone is never treated as verified.
Without accountable ownership no one can accept or stop the work.
Operational changes need an owner who can act on the system.
An unmeasurable goal cannot be evaluated or stopped.
The allowed decision or action scope must be explicit.
Unknown data sources cannot be governed or traced.
Unauthorised data use is a legal and ethical failure.
If essential data does not exist the use case cannot function.
Material decisions must be traceable to their source.
Data kept without rules creates privacy and security exposure.
Sensitive data with no access model is a critical exposure.
Without an action boundary the system's authority is unbounded.
Unrestricted production credentials allow unbounded damage.
Secrets in client code or prompts are immediately compromised.
Material actions that are not logged cannot be reconstructed.
Without containment a failure cannot be stopped.
External models may receive sensitive operational data.
Oversight without a named person is not oversight.
'Human in the loop' is meaningless without timing and authority.
Exceptions need a defined path to a person who can act.
Reviewers must be able to override an incorrect result.
Reviewers cannot challenge a result they cannot see the evidence for.
Without a baseline there is nothing to compare performance against.
A target defines what success means.
Output volume is not a measure of quality.
Without a stop threshold a failing system keeps running.
Performance must be comparable to the process it replaces.
Metrics without an owner and cadence are not maintained.
Happy-path-only testing hides real failures.
Untested permissions allow unauthorised actions in production.
Instruction-injection and misuse are common failure modes.
Containment that has never been tested may not work.
Material actions that cannot be reconstructed cannot be governed.
Without versions, incidents cannot be attributed or reproduced.
A system that cannot be paused cannot be made safe.
Incidents with no owner are not resolved.
Failures that are not detected in time cause compounding harm.
Deploying beyond the tested scope ships untested risk.
A pilot with no expiry becomes an ungoverned production system.
A deployment with no approver has no accountability.
A pilot needs pre-agreed success and stop thresholds.