Human Review and Stop Conditions

Unit ID: PAW-M11-U04 Estimated active time: 20-25 minutes

Decide in advance where you stop

A stop condition is a rule, written before you start, that says: if this is true, do not proceed with AI assistance.

Written in advance it is easy to follow. Decided in the moment, under deadline, it is not.

Stop conditions worth having

STOP CONDITIONS

Stop and do it another way if:

- the output would go to a person about a decision affecting them,
  and nobody named will read it first
- you cannot check whether the output is correct
- the task carries accountability that is personally yours
- the content cannot pass the privacy gate
- an error would be hard to reverse or hard to detect
- you are using it because you are tired rather than because it helps
- three follow-ups have not converged

If any is true: write it yourself, or ask a person.

The tiredness one is not a joke. It is when review quality drops and acceptance thresholds fall.

Human review is a step with a name

"Someone should check this" means nobody will.

The reviewer is a named person, with a stated thing to check and a time budget. Module 10 mapped it; the stop conditions decide when that review must happen before anything leaves.

Escalation needs a route

A gate that stops work without saying what happens next gets ignored.

For each stop condition, write the alternative: who to ask, what to do instead, how long that takes. A gate with a route is followed; a gate without one is an obstacle.

Review the conditions periodically

Stop conditions age. New tools, new approved categories, new kinds of work.

Look at them when you review your pattern library — quarterly is enough — and check that each still describes a real risk rather than an old one.

Mini practice

Write your own stop conditions, no more than seven.

For each, write the alternative route in one line.

Then check your three most frequent tasks against them.

Mark complete