Last month a manufacturing customer lost its uplink to the outside world for nineteen hours. The cause was a damaged fiber conduit two kilometers from the site, and nothing the customer could have prevented. What mattered was that the shift supervisors kept using the maintenance assistant the whole time, because it never needed the link in the first place.
Pin the models to the edge
Every Gammaxon Workbench edge node carries a pinned copy of the model catalog it has been approved to use. Updates arrive as signed bundles and are staged beside the running version, so a node never depends on a download finishing while someone is waiting on an answer. When the uplink drops, routing continues against the local copy without a configuration change on any client.
Queue what can wait
Not every request needs an immediate answer. Batch summaries, report drafts and indexing jobs go into a durable local queue and replay in order when the link returns. Interactive requests are served first, and the queue reports its depth so operators can see the backlog instead of guessing at it.
Tell people what changed
During the outage the smaller local model answered most requests, and the median response time rose from 0.9 to 2.4 seconds. We show a plain banner whenever a node is running on its local fallback, because people adjust how much they trust an answer when they know where it came from. Nobody asked for the banner to be removed.
We now suggest every customer rehearse this once a quarter: unplug the uplink on a quiet afternoon, watch the node fail over, and write down what surprised you. The platform page describes the pieces, and our engineers are glad to run the rehearsal with you.