Fleet Management

A fleet team lives in maintenance records, route plans and driver reports. Those documents name people, and people are the part a prompt should not carry. The starter screens each request and measures each answer before either becomes an instruction.

A fleet manager asks about a maintenance interval, a route plan or a driver report. The reply is text. Nothing here dispatches a vehicle, edits a route in a telematics system or calls a tool. An instruction to a driver still comes from a person.

Driver and technician details are the data most likely to slip into a prompt. Masking covers operator and badge numbers, maintenance or on-call phone numbers, and shift, crew or duty roster mentions. The request is rewritten and continues, so useful questions do not turn into refusals.

Refusals come from two fixed lists. The first list ends any request that names a facility blueprint, a safety interlock or a control system password. Built-in safety wording ends a request that mentions explosives, weapons or attacks. Neither list reads context, so an incident review that uses the word attack is refused.

Before either list, injection wording is screened and identity is enforced. A request with no verified user and organization does not proceed. Role claims are never taken from a caller header. On the return path, a reply must clear 0.87 overall, 0.78 relevance and 0.80 accuracy, and reach 120 characters and two sentences.

ISO 27001, NIS2 and ISO 39001 give the reference frame, and road traffic safety duties remain your program to run. The review step carries no configuration, so it is a switch to turn on rather than an escalation already working. The audit marker flags each evaluated request and never stops one.