Archibald312 builds the switchboard before the switches

GordonOSS shelved local-model support to first lay down the wiring that decides which AI answers which question.

The original next step was hooking up self-hosted inference engines so firms could run the system on their own hardware. Archibald312 punted that and instead built the routing layer those connectors would have needed - the piece that decides, for a given chat, whether the answer comes from Claude, Gemini, OpenAI, or eventually something running in-house.

The logic walks a hierarchy: a preference set on a specific document wins over one set on a project, which wins over what the user asked for. Unknown models get rejected and logged rather than silently used. Today nothing actually sets those preferences, so every chat still goes where the user pointed it - the plumbing is in, the taps are off. Every routing decision is written to the audit trail that shipped earlier.