Consulting Firm

Consulting work moves fast across many accounts, and the same laptop often touches two competitors in one week. A model makes that speed cheaper and the confidentiality boundary thinner. This starter is a first control layer for that pattern, not a finished policy.

The chain runs the prompt screen, then the identity check, then masking, then the block list. Order matters, because an injected instruction cannot switch off a check that already ran. A call with no verified identity stops at the second stage, since authentication is on by default.

The shipped block list is small. It stops a secret and a private key, and nothing else. Client names, pricing and deal terms are not covered until you add them. Masking handles national identifier strings inside a request.

Response scoring uses a 0.78 aggregate floor, with relevance at 0.78 and accuracy at 0.80 carrying equal weight. Short output fails as well, since the gate wants at least 120 characters and two sentences. A failing draft is blocked instead of returned, which keeps thin text out of a client deliverable. Scoring does not apply to streamed responses.

The strategy starter in this family carries the same controls as this one, down to that aggregate floor. Neither one checks sources, so a sourced claim needs a named reviewer either way. Pick this starter for broad delivery work across many clients. The review step and the block list are the two parts you must finish before rollout.