In July a defense contractor asked us to stand up a complete Gammaxon Workbench environment inside a facility that has no connection to the outside world, not even a monitored one. Two of our engineers spent nine days on site. This is what we carried in, what we did each day, and what we would change.
Everything arrives on media
The first two days went to inventory. Every model, container image, package and certificate we needed was assembled beforehand into one signed bundle, checked against a manifest, and then re-checked by the customer's own security staff before it crossed the boundary. Anything missing meant a day lost, so the manifest was longer than the installation guide.
Install, then prove it
Hardware was racked and burned in by day four, the platform was installed by day six, and the last three days were acceptance. The customer's team wrote the tests: sixty scenarios drawn from their own work, run with their own data, judged by their own staff. We changed two defaults as a result, both about how long a long-running job may hold a GPU.
What we would do differently
- Mirror the package index inside the boundary before the visit, not during it.
- Rehearse the update procedure on day one, because the second bundle will arrive months later and be applied by someone who was not there.
- Leave a one-page runbook for the operators, written with them rather than for them.
Offline updates are routine work for the team now, and the security and trust page describes how bundles are signed and verified. If you are planning something similar, we would like to hear about it.