Architecture and Engineering IP

Design practices hold drawings, specifications and client material that a contract keeps closed. A model that helps write a specification can also carry that material into a prompt. The controls here aim at that single risk, and they leave design judgment untouched.

Access comes first here. After the prompt screen runs, the call must carry a verified identity, because authentication stays on even when no role list is present. The gateway takes roles from verified claims and never from a header that a caller supplies. A shared script with no identity therefore stops at the gate.

The block list holds two terms only. A prompt that mentions a secret or a private key is rejected before the model reads it. Drawing numbers, site addresses and client names are not on that list. Add the terms that your contracts treat as restricted, because the shipped list is a starting point.

Redaction masks national identifier strings inside a request, which covers staff and contractor records more than design data. Scoring then reads the returned brief and requires 0.78 in aggregate, 0.78 for relevance and 0.80 for accuracy. Text under 120 characters or a single sentence also fails. A design brief that misses a floor is refused rather than returned.

This starter checks no sources, so a cited standard or load figure needs a person to verify it. The review step ships with an empty setting and stays inactive until your team configures it. Audit marking notes an allowed call and blocks nothing. Reference material points at ISO 27001, at GDPR and at client confidentiality terms, which you still map to each contract.