How the manager works
How your AI gets managed after launch.
The manager keeps the standards you set in view. It checks live work by task. It follows the response rules your team approved when the evidence changes.
The operating loop
A clear rule for each task.
The manager turns a concern in live work into a response your team can run.
This can sit beside an assistant, agent, ranking system, or classifier when trained people can judge its outputs against written standards. The public demo shows a conversational assistant only.
What it checks
Does this task still meet its standards?
The questions stay close to the work your team already knows.
Was the result right?
Compare it with the approved source or example for that task.
Did it use approved sources?
For a system that looks up information, check whether the result stayed within the material it found.
Did it follow policy?
Check the rules and the cases that need a person.
Is the evidence still current?
A result can be current, watch, or expired. An expired check earns more review before a hard response.
When a task needs attention
The response matches the evidence.
Your team sets the line and the response. The loop makes that decision visible and repeatable.
Keep serving
The task is meeting its standards. The manager logs the check.
Offer a person
The task can keep serving with a human option while the team checks the cause and gathers fresher evidence.
Pause this task
A serious, confirmed problem can take one task out of service while the rest of the system continues.
Resume with proof
The card names the source, follow-up, and clean evidence required before the task returns.
A narrow response preserves the service that is still working while the team gives attention to the part that needs it.
Bring the work your team already has
Start with one system that matters.
The first conversation tells both sides whether the records and standards can support the loop.
Useful records
A system that has been live long enough to leave work your team can review.
Written standards
The sources, policies, or examples that should guide the result.
Important work types
The tasks where a wrong result has a clear cost and a clear owner.
Decision owners
People who can choose the response for those work types and approve a repair.
See the workflow
One assistant, six failure stories, one page.
Wrenfield Health is fictional. Its assistant, traffic, knowledge base, human labels, and figures are synthetic. The demo shows the manager’s workflow and its failure cases. Production performance is still unproven.
Limits that belong with the method
Review windows over instant blocks
The manager works over batches of work. Existing request-time controls remain the right place for rules that act inside a single interaction.
Fresh human checks still matter
A score can look stable while the facts underneath it have changed. Fresh human checks renew the evidence when older checks no longer represent current work.
Safety needs enough evidence
When the evidence cannot support a useful number, the card says so and directs more review to the affected task.
The first pass needs time
The first set of checks in the synthetic model takes roughly two to four weeks of collection. That timing comes from simulation rather than production traffic.
Start with the system you will have to explain
Choose one system and one task where a bad result costs time, money, or trust. We will map the standards and response rules for it.