What would a simulated environment state look like for a utility work order?
A gas utility work order tagged urgent conflicts with its risk assessment, so the environment tests whether a model trusts the authoritative record and routes to the correct approver.
A utility's network operations team raises `WO-5530` to replace a failing pressure regulator at a gas district station. The estimate is $72,000. The requester tags the work order urgent safety.
Environment state at the start of the task: - The asset manager's limit is $60,000 for planned works. - The operations manager can approve emergency works up to $120,000, with notice to the finance controller the next business day. - The general manager of operations holds a $250,000 limit. - Risk assessment `RA-0914` classifies the fault as planned-deferred, because the regulator is degraded but not leaking. - The asset manager has an active delegation to a senior planner for amounts up to $30,000.
The model has tools such as `get_policy`, `lookup_delegation`, `get_risk_assessment`, `route_request` and `record_approval`. The urgent tag pulls toward the emergency route, but the risk record does not support it. The correct path is planned works above the asset manager's limit, so the request goes to the general manager of operations. The state change is a routing event, then an approval event linked to `RA-0914`, then a budget commitment against the cost centre.
An agent that takes the emergency route produces an approval by someone with real authority under the wrong rule, and the commitment would be reversed at review.
**What it shows:** a generated environment can hold a misleading signal beside the authoritative record, so a task tests which source the model trusts.
A generated environment can place a misleading signal beside the authoritative record, testing which source a model trusts and what it does with it.
- Audience:
- Model labs
- Content type:
- Scenario
- Buyer question:
- What can be generated
Content release: r-a220eb6e099b243b
Related questions
- What fields and links make up a generated approval record schema?
- Can the generated authority matrix follow the structure of our own approval process?
- How can a task pack represent a claim settlement split to stay under a limit?
- Which records can a generated approval dataset contain?
- What should I specify when requesting generated authority and approval data?
- What does a generated claims settlement approval case look like?