Run the eval where the PII is

Keep it in a SQLite file on a container on your bare metal server. The harness is an exec script on the same container: read a batch, call the model, write score and rationale to a results table. Point it at a local model for the sensitive pass, or at the Hoody AI gateway for a hosted comparison. The gateway runs on your host and logs no prompt content, but the upstream provider receives what it's sent, so the PII pass stays local.

Every ticket read and every score written happened on your disk. Hoody's API saw a container; it didn't see a ticket.

The eval set stays on the disk it was born on; only the scores get copied into a slide.