Maritime

A vessel and a port both run on procedures, crew lists and system detail. A model can summarize any of it, and the prompt travels ashore to reach the provider. The starter defines what is allowed to travel with it.

Ship managers, port planners and superintendents ask about a procedure, a defect or a port call. The answer is advice in text. No tool permission is configured, so nothing here touches a bridge system, a cargo plan or a crane. The master and the terminal keep their own decisions.

Crew data is the pattern most likely to fire on this traffic. Masking hides mentions of a shift, crew or duty roster, along with operator or badge numbers and maintenance hotline numbers. The prompt continues in masked form, which keeps a routine crewing question usable.

Refusals are narrow. A prompt that names a facility blueprint, a safety interlock or a control system password is stopped. Built-in safety wording stops explosives, weapons and attacks. Piracy and armed robbery reporting uses that last word, so expect refusals until you set a list of your own.

The order in front of those checks is deliberate. Injection screening runs first and identity second, so a pasted instruction cannot disable the access rule. Every call needs a verified user and organization behind it. Replies must then clear 0.87 combined, 0.78 relevance and 0.80 accuracy, at 120 characters and two sentences.

IMO cyber risk guidance, NIS2 and ISO 27001 are the references. The rail starter runs the same chain against rail safety material, so pick by duty rather than by feature. No superintendent is called when a reply lands, because the oversight point has no action set yet. The audit step marks a request without blocking or keeping it.