Chemical Industry

A plant runs on procedures that are precise for a reason. When staff ask a model to explain one, the prompt can carry interlock detail and system credentials with it. This starter sets a fixed screen for that traffic.

A process engineer or a control room supervisor asks about a procedure, an alarm write-up or a batch record. Nothing in this starter reaches the plant. It governs prompts and answers, and it carries no permission for a tool call or a write action. Advice from the model must travel the same approval path as any other advice.

Two lists decide what never reaches the model. A term list rejects any request that names a facility blueprint, a safety interlock or a control system password. A second screen carries built-in wording and refuses talk of explosives, weapons and attacks. Both compare words rather than purpose. A genuine process hazard question that uses one of those words is refused as well.

Order matters on the way in. Injection screening runs before the access check, so a pasted instruction cannot cancel it. The access stage refuses a request with no verified user and organization behind it. Redaction then masks operator and badge numbers, on-call and maintenance phone numbers, and duty roster references. Redaction rewrites the prompt instead of rejecting it.

Returning text meets a quality gate set at 0.87 overall, 0.78 for relevance and 0.80 for accuracy. Short replies fail too, because the gate wants at least 120 characters and two sentences. Text that misses a floor is withheld. Streamed replies skip this gate, so leave streaming off when the floor must hold.

Responsible Care, ISO 27001 and NIST CSF 2.0 are the references behind the starter. The human review step is empty, which means no answer waits for a signature until your team turns it on. The final audit step flags the evaluated request without blocking, and it neither keeps nor seals the entry. Process hazard analysis, alarm management and emergency response stay where they are today.