Supply Chain

Sourcing teams work with supplier records, shipment data and contingency plans. A model reads that faster than a person does, and the same prompt can carry operational detail out of the business. The starter sets the boundary for that traffic.

Sourcing and logistics teams ask about a supplier, a shortage or a contingency plan. The starter answers in text only. It cannot place an order, release a shipment or call a tool, so a commitment to a supplier stays a human act.

The screens are set for operational secrets rather than commercial ones. A prompt is refused when it names a facility blueprint, a safety interlock or a control system password. Supplier pricing, contract terms and volume data are not covered, so add wording of your own if that is your exposure.

The built-in safety list refuses prompts about explosives, weapons and attacks, matched literally. Sanctions research and disruption analysis often use the last of those words, and such requests are refused with it. Ahead of that screen, injection wording and identity are checked in that order, and masking hides badge numbers, hotline numbers and roster mentions.

Every reply is scored against 0.87 combined, 0.78 relevance and 0.80 accuracy. It is also measured against 120 characters and two sentences. A reply that misses one of those is refused. Scoring is skipped when the answer is streamed.

ISO 28000, ISO 27001 and NIS2 are the reference points. What the starter supplies is input to a security management case, not an assurance conclusion. Supplier due diligence and continuity testing stay with your team. Approval sits unset in the chain, so decide who signs off on an answer before you trust one. The audit step never blocks a request or keeps a copy of it.