Verification, Human-in-the-Loop & Control Gates

8 min

The output looks fine. Should the run just continue?

In Module 3 you made a run able to continue after an interruption. A run that continues reliably will also continue reliably with a wrong result. The core question of this module is: which results may continue automatically, and which must be confirmed first?

• Good format does not mean correct content
• Some checks a program can do; some need a person
• Not every action needs the same level of checking
• Where the checks sit is a design decision made in the harness

Key takeaway: Reliable continuation is not enough. Decide which results can move on automatically and which must be checked or approved first.

9 min

The gate model: pass, retry, human, fail

A gate is a check placed between steps. Verification asks whether the output is right. Validation asks whether it meets the need it was made for. A gate applies these checks and decides what happens next. Gates come in kinds: a quality gate checks whether the result is good enough, a safety gate checks whether it could cause harm, and a policy gate checks whether it follows the rules.

Text equivalent: the output enters a gate. The gate has four possible outcomes. Pass means the run continues. Retry means the step is attempted again. Human means a person reviews before anything continues. Fail means the run stops or moves to recovery. The gate never leaves the result in an undefined state.

Key takeaway: A gate checks the output and decides one of four outcomes: pass, retry, human review or fail.

9 min

Where does verification come from?

A gate is only as good as the check behind it. Different sources of verification suit different questions, and they can be combined. Prefer the most reliable source that fits the question, and do not treat a model checking itself as strong evidence on its own.

Key takeaway: Match the verification source to the question, combine sources, and do not rely on a model vouching for itself.

9 min

Human-in-the-Loop: review, approval, escalation, override

Human-in-the-Loop means designing the points where a person takes part. It is not a vague promise that someone is watching. Each pattern has a clear trigger and a clear action for the person, and the run keeps an approval state so it is known what was decided.

Approval state: a run waiting for a person should be visibly paused, with a record of who approved or rejected, when, and on what evidence. Combined with Module 3, a paused run is a run that has been checkpointed and can resume once the person decides.

Key takeaway: Design human control as clear patterns, with an approval state and the evidence a reviewer needs.

9 min

Evidence, thresholds and safe stop

To decide which outcome a gate gives, the design needs rules. Three ideas help: what evidence is required, what confidence is high enough, and what risk is too high to continue without a person.

Text equivalent: the gate first checks that evidence is present. High risk sends the result to a person for approval. Low confidence sends it to a person for review. Only a result that clears all of these passes automatically. A confidence score alone is not enough, since it can be wrong.

Key takeaway: Set evidence requirements, confidence and risk thresholds, and safe stop conditions so gate decisions are explicit.

11 min

Practice: Automatic or Human Gate?

Activity 4 is called Automatic or Human Gate? First work the example, then do it for your own workflow. Please do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise.

1. 1List the actions
   Write the actions in your workflow that produce output or change something.
2. 2Choose a gate for each
   Choose one: auto-continue, auto-verify, require human review, require human approval, or prohibit. Give a reason.
3. 3Name the verification source
   For each auto-verify action, name the source (rule, schema, trusted source, second model or policy check) and the threshold.
4. 4Define stop and escalation
   For each human gate, say what the reviewer sees, what the approval state is, and the safe stop condition if nobody responds.

• Go to the Practice page: Automatic or Human Gate?
• Next: Retry, Failure Classification & Recovery (Module 5)

Key takeaway: Human-AI-Human: a person decides which actions need a gate, automatic checks catch clear errors, and a person approves high-impact actions and stays responsible.