Industrial Robotics

A robot cell is documented in programs, drawings and safety files. Engineers want a model to help them read that material, and none of it belongs in an unchecked prompt. The starter sets what may travel and what an answer must satisfy.

Controls engineers ask about a fault code, a cycle time or a program change on a cell. The starter governs that conversation only. No tool permissions are configured, so the model cannot touch a controller. A program change still moves through your change process, and a person still signs it.

One phrase in the block list is easy to misread. A request that names a safety interlock is refused, along with a facility blueprint or a control system password. That keeps the information out of a prompt. It does nothing for the interlock itself, and it is not a functional safety measure of any kind.

The rest of the request path is ordinary and ordered. Injection wording is screened first, so pasted text cannot switch off what follows. Identity is required, and roles are read from verified claims. Masking hides operator badge numbers, maintenance hotline numbers and roster references. Built-in safety wording then refuses explosives, weapons and attacks by exact term.

Replies must clear 0.87 overall, 0.78 relevance and 0.80 accuracy, with equal weighting between the two measures. They must also reach 120 characters and two sentences. A reply that misses any floor is refused, and scoring does not run on streamed output.

ISO 10218, ISO 27001 and NIST CSF 2.0 are the reference points. What the starter supplies is evidence for a risk assessment, not a safety rating. An approval path is available to build, and until you build it no reply waits for anyone. The closing audit step marks an evaluated request and stops none of them.