Remit and remit drift
A remitRemitA written list of an agent’s allowed actions, resources and destinations for information. The list states whether it is complete. records what an AI agent is allowed to do: which actions it may take, which resources it may use and where it may send information.
Example. A purchasing agent may prepare orders but must not submit them. An observed submission is outside that explicit prohibition. If a list merely mentions preparing orders and is not declared complete, an unlisted action needs review; absence from a partial list alone does not prove it was forbidden.
Read the result.
- Inside the recorded remit: the activity fits the parts of the remit that could be checked.
- Outside the recorded remit: the activity breaks an explicit exclusion or falls outside a list declared complete. The result names what failed.
- Authority not established / review required: no usable remit applies, or an incomplete list cannot settle the question. An incomplete inventory never proves prohibition.
- Cannot be stated: the records lack something needed for a check, such as the destination. This is not an all-clear.
What to do. Open the activity and its remit, check their dates and coverage, and take the result to the responsible person. They can review the declaration or arrange action in the runtimeRuntimeThe software that runs a model or agent. A model runtime answers model requests; an agent runtime carries out the customer's agent actions.. NovusLens records and compares evidence; it does not stop the agent.
For an operator investigating a refusal
A refusal and a blocked operation are different records. Refused and reported blocked means the resource reports stopping the attempt; it does not count as drift. Refused, execution unresolved remains unresolved and can count until blocking is reported. Refused yet recorded as completed identifies an enforcement gap in the runtime’s records. Unknown execution does not establish an outcome.
A source must map a native operation to a reviewed business activity before that mapping can support the comparison. The result names detector classifications and their rule versions where used. Workers sharing one credential are identified as a group, not attributed to a guessed individual. Corrections change the affected support while retaining history.